{ "meta": { "built": "2026-08-29T09:52:07.225Z", "as_of": "2026-08-29", "method": "progress = (trl/9) × spec_fraction ; TI = 1000 × Σ(w·progress)/Σ(w)", "baseline": "0 = state of the art in 1984 · 1000 = a canonical T-800 could be built today", "components": 44, "total_weight": 313, "ti": 172.1 }, "index": { "ti": 172.1, "by_category": { "chassis-and-materials": 61.59, "power": 45.98, "locomotion": 238.45, "manipulation": 60.95, "perception": 387.55, "cognition": 89.25, "language-and-deception": 589.77, "autonomy-and-command": 241.42, "networking-and-c2": 151.83, "weapons": 129.01, "biology-and-camouflage": 27.5, "self-repair-and-durability": 18.71, "manufacture": 116.83, "time-displacement": 1.04 } }, "components": [ { "id": "hyperalloy-combat-chassis", "name": "Hyperalloy combat chassis", "category": "chassis-and-materials", "weight": 8, "one_liner": "A skeleton that survives what kills the flesh.", "canon": { "requirement": "A bipedal, fully armoured, human-scale structural skeleton machined from an unspecified 'hyperalloy', carrying every actuator, the power cell and the processor, and supporting a living-tissue envelope. It must remain fully operational - walking, aiming, gripping - after total loss of that envelope to fire. Canon defines the material entirely by what fails to destroy it; its composition is never given.", "quantified": [ { "metric": "material named", "value": "hyperalloy; composition never specified anywhere in T1 or T2", "source_ref": "t1-reese-briefing", "tier": "PRIMARY" }, { "metric": "operational after full-body hydrocarbon fireball", "value": "yes - walks out of the burning tanker as a stripped endoskeleton", "source_ref": "t1-tanker", "tier": "PRIMARY" }, { "metric": "confirmed destruction mechanisms", "value": "industrial hydraulic press (T1); immersion in molten steel, approx 1370-1600 C (T2)", "source_ref": "t1-press", "tier": "PRIMARY" }, { "metric": "actuation type", "value": "hydraulic - 'complex hydraulics and cables'", "source_ref": "t2-script-47j", "tier": "TERTIARY" } ], "sources": [ { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-reese-briefing", "evidence": "Reese describes the endoskeleton to Sarah Connor in the moving car.", "quote": "Underneath, it's a hyperalloy combat chassis... microprocessor controlled, fully armored... very tough.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-john-meets-t800", "evidence": "The Terminator identifies its own construction to John in the parking lot.", "quote": "I'm a cybernetic organism. Living tissue over metal endoskeleton.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day - revised final shooting script", "year": 1991, "medium": "shooting-script", "ref": "t2-script-opening", "evidence": "Stage direction, 2029 battlefield. NOT spoken dialogue.", "quote": "It looks like a CHROME SKELETON... a high-tech Death figure. It is the endoskeleton of a Series 800 terminator.", "verified": true, "tier": "TERTIARY" } ], "canon_confidence": "explicit", "canonical_target": { "metric": "survival envelope of the structural chassis", "value": "Fully functional after engulfment in a hydrocarbon fireball; destruction threshold above immersion in molten carbon steel (~1370-1600 C) and below the platen force of an industrial hydraulic press", "reasoning": "Canon never states a yield strength, density or composition, so a target in MPa would be invented rather than derived. What canon states twice is the FAILURE ENVELOPE: everything short of a foundry ladle or a press leaves it walking. The two destruction events (T1 press, T2 molten steel) bracket it from above; the tanker explosion brackets it from below. Real chassis are scored on how much of that envelope they cover." }, "assumed_by_sota_agent": "A load-bearing bipedal skeleton that continues to walk and fight after full-body incineration (the tanker fire and the steel-mill crucible in T2), after point-blank explosions, and after vehicle impact. Kyle Reese: 'Hyperalloy combat chassis, microprocessor-controlled, fully armored.' Assumed survival temperature of order 1,500 degrees C, structure intact and functional." }, "real": { "status": "in_progress", "trl": 7, "spec_fraction": 0.12, "spec_fraction_rationale": "On structural specific strength the real world already matches canon: Ti-6Al-4V gives 880-920 MPa yield at 4.43-4.51 g/cm3 and is what production humanoids are built from, and refractory high-entropy alloys hold 571 MPa yield at 1600 C, above the ~1,700 C of a blast furnace. On rated operating envelope the best production humanoid is certified only from -20 C to +40 C, i.e. 40/1500 = 0.027 of the canonical thermal requirement, and no humanoid has ever been demonstrated functioning after incineration (0). Weighting strength, thermal envelope and demonstrated survival roughly equally gives (1.0 + 0.03 + 0) / 3 = 0.34, discounted to 0.12 because no integrated combat-survivable humanoid frame has been built at all and the structural credit is therefore hypothetical.", "gap": "The alloys exist; the machine does not. Nobody has built a humanoid frame designed to survive fire, blast or impact, because there is no market for one. More fundamentally, a frame that survives 1,500 C is useless when the neodymium-iron-boron magnets, bearing lubricants, wire insulation, elastomer seals and silicon it carries all fail between roughly 150 C and 400 C.", "why_hard": "The binding constraint is not the structural material but everything mounted inside it. Refractory alloys also pest-oxidise in air far below their melting points, and the best silicide protective coatings survive on the order of ten hours at 1800 C, so even the frame is a consumable. Sealing the machine for environmental survival (IP67 on the production Atlas) simultaneously removes the convective path that its own waste heat needs.", "movers": [ { "name": "Boston Dynamics", "kind": "company", "country": "US", "what": "Ships the only production humanoid with a published environmental envelope: 90 kg, 56 DOF, 3D-printed titanium and aluminium structure, IP67, -20 C to +40 C." }, { "name": "Beijing Institute of Technology / Xi'an Jiaotong (HfMoNbTaW RHEA authors)", "kind": "university", "country": "CN", "what": "Demonstrated 571 MPa yield strength at 1600 C in an as-homogenised refractory high-entropy alloy, the current elevated-temperature strength benchmark." }, { "name": "Oak Ridge National Laboratory", "kind": "lab", "country": "US", "what": "Runs the US structural-materials programmes for extreme environments, including irradiation-tolerant refractory high-entropy alloys." }, { "name": "AgiBot", "kind": "company", "country": "CN", "what": "Demonstrated 106 km of continuous walking without power-down, the longest structural-durability field test on a humanoid frame to date." } ], "evidence": [ { "date": "2026-01-05", "claim": "Boston Dynamics unveiled the production Atlas: 56 degrees of freedom with fully rotational joints, 2.3 m reach, 50 kg lift, rated -20 C to +40 C, extremely water-resistant, entering production immediately.", "source": "Boston Dynamics", "title": "Boston Dynamics Unveils New Atlas Robot to Revolutionize Industry", "url": "https://bostondynamics.com/blog/boston-dynamics-unveils-new-atlas-robot-to-revolutionize-industry/", "kind": "product", "delta": "+" }, { "date": "2026-01-24", "claim": "Production Atlas specifications reported as 1.9 m, 90 kg, IP67, 50 kg instantaneous and 30 kg sustained payload, approximately 4 hours on dual self-swappable batteries.", "source": "Humanoids Daily", "title": "The Alien in the Factory: Boston Dynamics Launches Production-Ready Atlas at CES 2026", "url": "https://www.humanoidsdaily.com/news/the-alien-in-the-factory-boston-dynamics-launches-production-ready-atlas-at-ces-2026", "kind": "product", "delta": "0" }, { "date": "2022-10-17", "claim": "HfMoNbTaW refractory high-entropy alloy achieves 571 MPa yield strength at 1600 C, against approximately 400 MPa for earlier MoNbTaW and MoNbTaVW compositions.", "source": "Science and Technology of Advanced Materials 23:642-654", "title": "Edge-dislocation-induced ultrahigh elevated-temperature strength of HfMoNbTaW refractory high-entropy alloys", "url": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9586648/", "kind": "paper", "delta": "+" }, { "date": "2025-11-24", "claim": "AgiBot's A2 humanoid walked 106 km from Suzhou to Shanghai over 56 hours without powering down, using 15 hot battery swaps, taking the Guinness record for longest journey walked by a humanoid robot.", "source": "UPI", "title": "Watch: Robot breaks world record with 66-mile walk in China", "url": "https://www.upi.com/Odd_News/2025/11/24/AgiBot-A2-Guinness-World-Records-robot-walk/3681764010764", "kind": "demo", "delta": "+" } ], "indicators": [ { "metric": "Rated maximum ambient operating temperature, best production humanoid", "unit": "degC", "value": 40, "as_of": "2026-01-05", "direction": "up_is_progress", "canon_target": 1500, "source": "https://bostondynamics.com/blog/boston-dynamics-unveils-new-atlas-robot-to-revolutionize-industry/" }, { "metric": "Yield strength of best refractory high-entropy alloy at 1600 degC", "unit": "MPa", "value": 571, "as_of": "2022-10-17", "direction": "up_is_progress", "canon_target": 571, "source": "https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9586648/" }, { "metric": "Longest continuous walking distance by a humanoid without power-down", "unit": "km", "value": 106, "as_of": "2025-11-24", "direction": "up_is_progress", "canon_target": 100000, "source": "https://www.upi.com/Odd_News/2025/11/24/AgiBot-A2-Guinness-World-Records-robot-walk/3681764010764" } ] }, "commentary": "The franchise sold the alloy as the hard part. It is not. Refractory high-entropy alloys have held 571 megapascals at 1,600 degrees since 2022, and titanium at 900 megapascals and 4.4 grams per cubic centimetre has been routine since the 1950s. What fails in a steel mill is everything the frame is carrying: neodymium magnets, bearing grease, wire insulation and silicon, all of which surrender between 150 and 400 degrees. The most capable humanoid in production is rated from minus twenty to plus forty and sealed to IP67, which keeps water out and heat in. The metallurgy arrived decades ago. The machine did not.", "last_reviewed": "2026-08-29", "review_cadence_days": 30, "progress": 0.0933 }, { "id": "armor-and-ballistic-survivability", "name": "Armor and ballistic survivability", "category": "chassis-and-materials", "weight": 7, "one_liner": "Taking sustained gunfire and continuing to walk.", "canon": { "requirement": "Defeat sustained, multi-shooter small-arms fire from contact range to ~25 m with no loss of mobility, aim or mission tempo. Rounds deform against the chassis rather than penetrate it. The living tissue is consumable; the chassis is not. Canon is careful to show the one infantry weapon that does work: a 40 mm grenade, and only temporarily.", "quantified": [ { "metric": "rounds absorbed in one engagement", "value": "dozens - Sarah extracts slugs until the glass is 'nearly full of flattened bullets'", "source_ref": "t2-script-garage", "tier": "TERTIARY" }, { "metric": "calibres defeated on screen", "value": "9 mm, .45 ACP, 12-gauge, 5.56 mm, .38", "source_ref": "t1-technoir", "tier": "PRIMARY" }, { "metric": "police officers killed in one engagement", "value": "17", "source_ref": "t2-silberman-photos", "tier": "PRIMARY" }, { "metric": "CS/tear gas effect", "value": "none - 'Terminator emerges from the smoke. Not even misty-eyed.'", "source_ref": "t2-script-173e", "tier": "TERTIARY" }, { "metric": "effective disabling weapon", "value": "40 mm grenade - temporary disablement only", "source_ref": "t2-steel-mill", "tier": "PRIMARY" } ], "sources": [ { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-body-armor", "evidence": "The police explain away the Terminator's survivability to Sarah Connor.", "quote": "Sarah, this is what they call body armor. Our TAC guys wear these. It'll stop a 12-gauge round.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-silberman-photos", "evidence": "Detectives show Sarah the 1984 police-station surveillance stills.", "quote": "He killed 17 police officers that night.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day - revised final shooting script", "year": 1991, "medium": "shooting-script", "ref": "t2-script-garage", "evidence": "Stage direction during the field repair in the garage.", "quote": "Sarah nods, pulling out another slug. CLINK. The glass nearly full of flattened bullets.", "verified": true, "tier": "TERTIARY" } ], "canon_confidence": "implied", "canonical_target": { "metric": "small-arms impacts defeated per engagement with zero mission degradation", "value": ">=50 impacts (9 mm through 12-gauge slug, plus 5.56 mm) at contact to 25 m, zero degradation of gait, grip or aim; retains mobility after a 40 mm HE hit", "reasoning": "'Dozens of flattened bullets in a glass' is the only quantity canon offers; 50 is a conservative floor on 'nearly full'. The 40 mm clause is included deliberately so the target is falsifiable rather than 'invulnerable' - canon itself shows the one weapon that works." }, "assumed_by_sota_agent": "Whole-body immunity to sustained automatic rifle fire, shotgun at contact range, and explosive blast, with no loss of function and no mobility penalty. Assumed of the order of several hundred rifle hits distributed anywhere over the body, absorbed without degradation." }, "real": { "status": "in_progress", "trl": 9, "spec_fraction": 0.06, "spec_fraction_rationale": "The fielded reference is the ESAPI plate: 2.50 kg medium at 241 x 318 mm, rated to the NIJ RF3 threat (.30-06 M2 armour-piercing at 878 m/s); the SAPI generation it replaced was rated for three hits of the marked round. That is an areal density of 32.6 kg/m2. Covering an adult's ~1.8 m2 body surface at that density costs approximately 59 kg of ceramic, against a production Atlas total mass of 90 kg, and yields roughly 70 idealised hits over the whole body if perfectly distributed. Against a canon requirement of several hundred hits anywhere with no degradation, the hit ratio is roughly 70/500 = 0.14, discounted heavily for the mobility penalty (59 kg of armour on a 90 kg machine that must also run) and for the fact that real armour preserves a body that then stops fighting. Net 0.06.", "gap": "Coverage, mass and multi-hit capacity, in that order. Ceramic armour defeats a rifle round by shattering, so protection is consumed where it is used; the canon requirement is protection that is not consumed. Nothing in the fielded inventory protects joints, and joints are where a humanoid is stopped.", "why_hard": "Defeating a rifle projectile requires dissipating roughly 3-4 kJ in a few hundred microseconds, and the only materials that do it are brittle ceramics that fracture in the process. Areal density has been stuck near 30-40 kg/m2 for two decades because the mechanism is energy-limited, not manufacturing-limited. Full-body coverage at that density is incompatible with bipedal agility.", "movers": [ { "name": "National Institute of Justice", "kind": "agency", "country": "US", "what": "Publishes NIJ Standard 0123.00 and 0101.07, the ballistic protection level and test-threat definitions the whole industry certifies against." }, { "name": "US Army PEO Soldier", "kind": "program", "country": "US", "what": "Owns the SAPI/ESAPI/XSAPI plate family, the fielded reference for rifle-threat areal density and multi-hit rating." }, { "name": "DSM / Avient (Dyneema)", "kind": "company", "country": "NL", "what": "UHMWPE fibre supply for the backing architectures behind hard-plate systems; the incremental route to lower areal density." } ], "evidence": [ { "date": "2023-11-01", "claim": "NIJ Standard 0123.00 defines the current ballistic protection levels HG1, HG2, RF1, RF2 and RF3, with RF3 tested against .30-06 M2 armour-piercing at 2,880 ft/s (878 m/s). Addenda issued 2025-07-08.", "source": "National Institute of Justice", "title": "Specification for NIJ Ballistic Protection Levels and Associated Test Threats, NIJ Standard 0123.00", "url": "https://nij.ojp.gov/topics/equipment-and-technology/specification-nij-ballistic-protection-levels-and-associated-test-threats-nij-standard-012300", "kind": "regulation", "delta": "0" }, { "date": "2026-08-29", "claim": "Fielded ESAPI plates mass 2.50 kg for a medium at 241 x 318 mm and are rated against .30-06 M2 armour-piercing; the preceding SAPI generation was rated for survivability of three hits of the marked round.", "source": "Wikipedia (compiled from US Army specifications)", "title": "Small Arms Protective Insert", "url": "https://en.wikipedia.org/wiki/Small_Arms_Protective_Insert", "kind": "product", "delta": "0" }, { "date": "2026-08-29", "claim": "NIJ 0101.07 rifle levels map to specific threats: RF1 to 7.62x51 M80, 7.62x39 MSC and 5.56 M193; RF2 adds 5.56 M855; RF3 to .30-06 M2 AP. No areal density or shot-count requirement is specified by the standard itself.", "source": "Wikipedia", "title": "List of body armor performance standards", "url": "https://en.wikipedia.org/wiki/List_of_body_armor_performance_standards", "kind": "regulation", "delta": "0" } ], "indicators": [ { "metric": "Areal density of fielded rifle-threat armour (ESAPI medium)", "unit": "kg/m2", "value": 32.6, "as_of": "2026-08-29", "direction": "down_is_progress", "canon_target": 5, "source": "https://en.wikipedia.org/wiki/Small_Arms_Protective_Insert" }, { "metric": "Rated multi-hit capacity per rifle-threat plate", "unit": "hits", "value": 3, "as_of": "2026-08-29", "direction": "up_is_progress", "canon_target": 100, "source": "https://en.wikipedia.org/wiki/Small_Arms_Protective_Insert" }, { "metric": "Mass of whole-body RF3-equivalent coverage over 1.8 m2", "unit": "kg", "value": 59, "as_of": "2026-08-29", "direction": "down_is_progress", "canon_target": 10, "source": "https://en.wikipedia.org/wiki/Small_Arms_Protective_Insert" } ] }, "commentary": "Fielded, mature, two decades in service, and still nowhere near. An ESAPI plate stops a .30-06 armour-piercing round at 32.6 kilograms per square metre, and the generation it replaced was rated for three hits. Cover an adult's 1.8 square metres at that density and you have carried 59 kilograms of ceramic, two-thirds of the entire mass of a production Atlas, to buy perhaps seventy idealised hits distributed perfectly, which incoming fire never is. Canon asks for several hundred anywhere on the body with no loss of function. Readiness level nine, specification fraction six per cent. The two numbers are not in tension.", "last_reviewed": "2026-08-29", "review_cadence_days": 60, "progress": 0.06 }, { "id": "mimetic-polyalloy", "name": "Mimetic polyalloy", "category": "chassis-and-materials", "weight": 4, "one_liner": "T-1000 liquid metal.", "canon": { "requirement": "A single homogeneous metallic material that holds arbitrary human-scale form at sub-millimetre fidelity (face, clothing texture, a police badge); reversibly transitions between fluid and rigid load-bearing solid on command; reintegrates severed mass; conserves total mass; and CANNOT produce chemically heterogeneous or multi-part mechanisms. The exclusions are stated in dialogue and are as important as the capabilities.", "quantified": [ { "metric": "copy acquisition", "value": "physical contact only", "source_ref": "t2-pescadero-drive", "tier": "PRIMARY" }, { "metric": "mass conservation", "value": "'only an object of equal size'", "source_ref": "t2-pescadero-drive", "tier": "PRIMARY" }, { "metric": "hard exclusion", "value": "cannot form complex machines - chemicals or moving parts", "source_ref": "t2-pescadero-drive", "tier": "PRIMARY" }, { "metric": "permitted rigid forms", "value": "knives and stabbing weapons; solid metal shapes", "source_ref": "t2-pescadero-drive", "tier": "PRIMARY" }, { "metric": "thermal tolerance", "value": "walks out of a burning tanker; script: 'Unruffled by thousand-degree heat'", "source_ref": "t2-script-47j", "tier": "TERTIARY" }, { "metric": "wound closure", "value": "instantaneous - 'Shiny liquid metal visible in the hole, which then closes'", "source_ref": "t2-script-80a", "tier": "TERTIARY" }, { "metric": "known degradation mode", "value": "morphing malfunctions after cryogenic shattering and re-formation", "source_ref": "t2-se-bugs", "tier": "PRIMARY-SE" } ], "sources": [ { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-mimetic-polyalloy", "evidence": "The T-800 names the T-1000's construction for John.", "quote": "Yes. A mimetic poly-alloy.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-pescadero-drive", "evidence": "Sarah asks the T-800 what the T-1000 can and cannot do.", "quote": "It can't form complex machines. Guns and explosives have chemicals, moving parts. It doesn't work that way.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-pescadero-drive", "evidence": "The T-800 states the mass-conservation limit.", "quote": "No, only an object of equal size.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-pescadero-drive", "evidence": "The T-800 states the sampling mechanism.", "quote": "Anything it samples by physical contact.", "verified": true, "tier": "PRIMARY" } ], "canon_confidence": "explicit", "canonical_target": { "metric": "programmable metallic material - morph, fidelity, self-heal, mass conservation", "value": "Ambient-temperature morphing between fluid and rigid load-bearing solid in <1 s; arbitrary human-scale geometry to <=0.5 mm; full-penetration self-healing in <1 s; reintegration of detached mass; exact mass conservation; ~1000 C tolerance; cannot form chemically heterogeneous or multi-part mechanisms", "reasoning": "The 0.5 mm fidelity figure is extrapolated from what the disguise must defeat: close-range inspection of a face by people who know that face (Janelle's husband, John's foster parents, Sarah at the steel mill). Sub-millimetre is where a human observer at conversational distance stops seeing the difference. The exclusions are taken verbatim from dialogue, because a real material that could form a working firearm would EXCEED canon and should score above 1.0 on that axis." }, "assumed_by_sota_agent": "A self-powered humanoid of roughly 70-90 kg composed of a liquid-metal alloy that flows, reassembles from a dispersed or liquid state, mimics the appearance and texture of people and objects it touches, forms rigid load-bearing blades and edged weapons, and recovers from being shattered or cryogenically frozen, with onboard sensing, computation and power." }, "real": { "status": "on_horizon", "trl": 4, "spec_fraction": 0.002, "spec_fraction_rationale": "The best load-bearing phase-transitional material demonstrated is magnetoactive phase transitional matter: neodymium-iron-boron microparticles in gallium (melting point 29.8 C), 21.2 MPa strength and 1.98 GPa stiffness in the solid phase, bearing 30 times its own weight, flowing at up to 15 cm/s liquid. The demonstrators are millimetre-scale and of order one gram, and are heated and driven entirely by an external alternating magnetic field. Against a self-powered 70 kg humanoid the mass ratio alone is approximately 1 g / 70 kg = 1.4e-5, before crediting nothing at all for onboard power, sensing, cognition or mimicry. Crediting the demonstrated sub-capabilities (reversible field-driven solid-liquid transition, programmable shape memory over 100+ cycles, load-bearing self-assembling lattices at 3,445 N) gives approximately 0.002.", "gap": "Scale, autonomy and every function that is not shape change. No demonstrated system carries its own power, its own actuation, its own sensing or any computation; all of them are puppets of an external field or an external robot arm. None mimics appearance. None is load-bearing at human scale in a reconfigurable state.", "why_hard": "Gallium-based phase-transitional matter works because gallium melts at 29.8 C, which is also why it cannot be structural at body temperature. Scaling the concept requires an internal power source of the kind ruled out under energy-density-and-endurance, and delivering actuation energy without an external field has no known mechanism. Programmable matter faces an independent hard limit in inter-module connection strength and self-reconfiguration planning complexity.", "movers": [ { "name": "Carnegie Mellon University (Majidi group)", "kind": "university", "country": "US", "what": "Co-authored the magnetoactive phase transitional matter work, the load-bearing benchmark for solid-liquid phase-transitional robots." }, { "name": "Sun Yat-sen University (Jiang group)", "kind": "university", "country": "CN", "what": "Lead institution on the 2023 MPTM paper; millimetre-scale demonstrators that liquefy, escape confinement and re-solidify." }, { "name": "Wuhan University (Fu Lei group)", "kind": "university", "country": "CN", "what": "Published the 2026 liquid metal shape memory robot: 90 percent volumetric compression, over 100 shape-memory cycles, deformation above 220 C, recovery within 50 s." }, { "name": "MIT Center for Bits and Atoms (Gershenfeld)", "kind": "lab", "country": "US", "what": "August 2026 work on self-aligning interlocking lattice modules assembled by mobile robots into architectural-scale structures at 4,556 N/mm stiffness." }, { "name": "University of Pennsylvania / EPFL / University of Macau", "kind": "university", "country": "US", "what": "Authors of the February 2026 Science Robotics review setting out what modular reconfigurable robots can and cannot yet do in real applications." } ], "evidence": [ { "date": "2023-01-25", "claim": "Magnetoactive phase transitional matter (neodymium-iron-boron microparticles in gallium, melting point 29.8 C) achieves 21.2 MPa strength and 1.98 GPa stiffness in the solid phase, bears 30 times its own weight, and flows at up to 15 cm/s in the liquid phase; demonstrators are millimetre-scale and externally actuated by magnetic field.", "source": "Matter 6(3):855-872, DOI 10.1016/j.matt.2022.12.003", "title": "Magnetoactive liquid-solid phase transitional matter", "url": "https://doi.org/10.1016/j.matt.2022.12.003", "kind": "paper", "delta": "+" }, { "date": "2026-06-01", "claim": "A liquid metal shape memory robot achieves programmable deformation with automatic reversible shape recovery: 90 percent volumetric compression ratio, over 100 one-way shape-memory cycles, deformation at temperatures above 220 C, and recovery within 50 seconds.", "source": "Matter 9:102759, DOI 10.1016/j.matt.2026.102759", "title": "Liquid metal shape memory robot", "url": "https://doi.org/10.1016/j.matt.2026.102759", "kind": "paper", "delta": "+" }, { "date": "2026-02-25", "claim": "A Science Robotics review of modular reconfigurable robots concludes the field has progressed from laboratory settings toward real-world applications, but that general hardware challenges, general software challenges and application-specific challenges all remain open.", "source": "Science Robotics 11, DOI 10.1126/scirobotics.adz1999", "title": "Modular reconfigurable robots: Toward on-demand multifunctional applications", "url": "https://pubmed.ncbi.nlm.nih.gov/41739909/", "kind": "paper", "delta": "0" }, { "date": "2026-08-04", "claim": "Self-aligning compound nested lattice modules interlocking on three axes were assembled by robot arms and mobile assemblers into furniture- and architectural-scale structures, reaching 4,556 N/mm stiffness and 3,445 N maximum load; assembly is performed by external robots, not by the modules themselves.", "source": "arXiv (ACADIA 2026)", "title": "Reconfigurable Structural Robotic Assembly: Interlocking 3D Aggregations with Self-Aligning Compound Nested Lattice Modules", "url": "https://arxiv.org/abs/2608.07576", "kind": "paper", "delta": "+" } ], "indicators": [ { "metric": "Mass of largest self-actuating phase-transitional robot", "unit": "g", "value": 1, "as_of": "2023-01-25", "direction": "up_is_progress", "canon_target": 70000, "source": "https://doi.org/10.1016/j.matt.2022.12.003" }, { "metric": "Solid-phase strength of phase-transitional matter", "unit": "MPa", "value": 21.2, "as_of": "2023-01-25", "direction": "up_is_progress", "canon_target": 500, "source": "https://doi.org/10.1016/j.matt.2022.12.003" }, { "metric": "Reversible shape-memory cycles demonstrated in a liquid-metal robot", "unit": "cycles", "value": 100, "as_of": "2026-06-01", "direction": "up_is_progress", "canon_target": 100000, "source": "https://doi.org/10.1016/j.matt.2026.102759" } ] }, "commentary": "The T-1000 remains the index's cleanest reminder of what fiction costs. The best load-bearing phase-transitional matter in the literature is neodymium-iron-boron suspended in gallium: 21 megapascals, thirty times its own weight, fifteen centimetres per second liquid, and about one gram, driven entirely by an external magnetic field. June's Matter paper adds reversible shape memory over a hundred cycles at a hundredth the required scale. Against a self-powered seventy-kilogram humanoid that reassembles from a puddle, the mass ratio alone is one in a hundred thousand, before onboard power, sensing or cognition. Half the popular coverage reports the load figure as thirty kilograms. It is not.", "last_reviewed": "2026-08-29", "review_cadence_days": 45, "progress": 0.0009 }, { "id": "nuclear-power-cell", "name": "Nuclear power cell", "category": "power", "weight": 9, "one_liner": "A sealed cell that runs a combat humanoid for decades.", "canon": { "requirement": "A single sealed on-board power source giving 120 years of continuous operation to a combat android, with no refuelling, no recharge, no external cooling and no maintenance - sited inside a torso otherwise full of hydraulics, and thermally invisible through human skin. Its failure mode is a very large explosion.", "quantified": [ { "metric": "runtime on the existing cell", "value": "120 years", "source_ref": "t2-garage-repair", "tier": "PRIMARY" }, { "metric": "power source type", "value": "'Nuclear'", "source_ref": "salvation-fuel-cells", "tier": "SECONDARY" }, { "metric": "failure mode", "value": "'Enough to level this place'", "source_ref": "salvation-fuel-cells", "tier": "SECONDARY" }, { "metric": "full-power runtime (novelization)", "value": "1,095 days (3 years) at full power, 24 h/day", "source_ref": "t1-novelization", "tier": "TERTIARY-NOVELIZATION-UNVERIFIED" }, { "metric": "economy mode (novelization)", "value": "power cut to 40% of nominal; optics drop to IR only", "source_ref": "t1-novelization", "tier": "TERTIARY-NOVELIZATION-UNVERIFIED" }, { "metric": "total output (novelization)", "value": "'enough power to run the lights of a small city for a day' - NOVELIZATION TEXT, NOT A FILM LINE. 1984 Frakes & Wisher novelization; wiki-transcribed and NOT verified against the printed book.", "source_ref": "t1-novelization", "tier": "TERTIARY-NOVELIZATION-UNVERIFIED" }, { "metric": "cell count, T-850 (T3) - CONTRADICTS the 120-year figure", "value": "two hydrogen fuel cells", "source_ref": "t3-fuel-cells", "tier": "SECONDARY" }, { "metric": "'iridium' - NOT CANON", "value": "Appears in no film, script or novelization. Earliest attributable source is an item card in The Terminator CCG (Precedence, 2000), where it is an optional Skynet enhancement, not the standard power plant.", "source_ref": "ccg-iridium-card", "tier": "NOT-CANON" } ], "sources": [ { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-garage-repair", "evidence": "John asks the Terminator how long it lasts while Sarah digs bullets out of its back.", "quote": "A hundred and twenty years with my existing power cell.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator Salvation", "year": 2009, "medium": "film", "ref": "salvation-fuel-cells", "evidence": "Resistance fighters identify T-800 power cells inside a Skynet facility.", "quote": "Fuel cells. Life source for the T-800. Nuclear. Enough to level this place.", "verified": true, "tier": "SECONDARY" }, { "work": "Terminator 3: Rise of the Machines", "year": 2003, "medium": "film", "ref": "t3-fuel-cells", "evidence": "The T-850 explains its own power architecture to John and Kate.", "quote": "I am powered by two hydrogen fuel cells. My primary cell was damaged by a plasma attack.", "verified": true, "tier": "SECONDARY", "note": "Cannot be reconciled with T2's 120-year runtime: even at a 100 W average draw that requires ~3,155 kg of hydrogen. See contradiction 'power-source-chemistry'." }, { "work": "The Terminator (novelization, Frakes & Wisher)", "year": 1985, "medium": "novelization", "ref": "t1-novelization", "evidence": "Describes a nuclear-energy cell sited where a man's heart would be; 1,095 days at full power; 40% economy mode; output likened to a small city's lights for a day.", "quote": "nuclear-energy cell", "verified": false, "tier": "TERTIARY", "note": "Wiki transcription; NOT verified against the printed novelization. Quote fragment only." }, { "work": "The Terminator Collectible Card Game (Precedence)", "year": 2000, "medium": "collectible-card-game", "ref": "ccg-iridium-card", "evidence": "Item card 'Iridium Power Cell' (Item - SkyNet, Uncommon). Confirmed on two independent third-party checklists; printed rules text NOT obtained.", "quote": "An Iridium Power Cell is the power source of a Series 8xx Terminator. The cell itself is inexhaustible, though it is unstable.", "verified": false, "tier": "NOT-CANON", "note": "This is a fan-wiki article's summary of a card, not the card's printed text. Recorded so the 'iridium power cell' claim can be deliberately rejected." } ], "canon_confidence": "explicit", "canonical_target": { "metric": "usable stored electrical energy in a torso-sized sealed package", "value": "~3.5e11 to 1.1e12 J (~100-315 MWh) over 120 years of continuous duty", "reasoning": "120 years = 3.786e9 s. Canon states no power draw, so the target is bracketed by plausible continuous averages for a machine that outperforms a human athlete indefinitely: 100 W -> 3.79e11 J (105 MWh); 300 W (headline mid case) -> 1.14e12 J (315 MWh); 1 kW -> 3.79e12 J (1.05 GWh). CONSISTENCY CHECK 1: at the mid case, 1.14e12 J / 4.184e9 J per tonne TNT = 272 tonnes TNT equivalent - which genuinely would 'level this place' (Salvation). CONSISTENCY CHECK 2: the 1984 novelization's 1,095-day full-power runtime is exactly 1/40 of 120 years, so both hold if average duty is 2.5% of full power, fixing full power at ~12 kW - which sits beside the ~7 kW single-arm peak derived independently in actuation-and-strength. The same energy over one day is ~13 MW, a good match for the novelization's 'lights of a small city for a day'. Three figures from two media thirty years apart collapse into one model." }, "assumed_by_sota_agent": "A sealed, torso-mounted nuclear-energy cell (1984 novelization) of at most a few kilograms, supplying a combat humanoid's full electrical load of approximately 300 W continuous with kilowatt-class peaks, for approximately 120 years and a total delivered energy of approximately 1.14e12 J, with no external radiator and no servicing. Implied specific power 60-150 We/kg. Canon confidence: extrapolated. NOTE: the widely repeated 'iridium' descriptor is NOT canon (it originates in a 2000 collectible card game) and is deliberately not used here." }, "real": { "status": "in_progress", "trl": 9, "spec_fraction": 0.05, "spec_fraction_rationale": "One flight-proven GPHS-RTG delivers exactly the canonical 300 W at beginning of life, so the power and, over 120 years, roughly the energy are already achievable. It does so at 57 kg (5.2 We/kg), consuming 8.1 kg of plutonium-238 and rejecting 4,100 W of waste heat at 6.8 percent conversion efficiency. Canon implies at most ~5 kg for 300 W, i.e. 60 We/kg, so 5.2/60 = 0.087; at the ~2 kg scale implied on screen, 5.2/150 = 0.035. Equivalently, mass ratio 5 kg / 57 kg = 0.088. On lifetime alone the real world does much better: Voyager 1 has run continuously on radioisotope power since 1977, 48 years and projected to 59, against 120 years, giving 0.40. Specific power and waste heat bind, so the composite sits near the low end: 0.05.", "gap": "Specific power and waste-heat rejection, not stored energy. A radioisotope cell that fits in a torso would deliver watts, not hundreds of watts; a cell that delivers hundreds of watts weighs more than half the robot and needs to dump kilowatts of heat through the same skin that must feel human. Betavoltaics reach the required lifetime and miss the required power by six orders of magnitude: the Betavolt BV100 gives 100 microwatts, so 300 W would take three million cells occupying about 3.4 cubic metres.", "why_hard": "Thermoelectric conversion of radioisotope decay heat runs at 5-7 percent, and the physics of thermoelectric figure of merit has not moved enough in fifty years to change that. The other 93-95 percent must be radiated continuously, for the life of the cell. Separately, plutonium-238 supply is a hard national cap: the US DOE target is 1.5 kg per year, so the 8.1 kg in a single GPHS-RTG is five and a half years of national output at target, or twenty years at the rate actually demonstrated. An army is an isotope-inventory problem before it is a manufacturing problem.", "movers": [ { "name": "NASA / Department of Energy Radioisotope Power Systems Program", "kind": "program", "country": "US", "what": "Operates the only fielded compact nuclear power sources: MMRTG (43.6 kg, 110 We) and GPHS-RTG (57 kg, 300 We), and is developing the Next-Generation RTG." }, { "name": "Oak Ridge National Laboratory", "kind": "lab", "country": "US", "what": "Produces plutonium-238; automation of the target-processing step raised annual output from 50 g to 400 g against a 1.5 kg/yr target." }, { "name": "Idaho National Laboratory", "kind": "lab", "country": "US", "what": "Runs the Advanced Test Reactor target campaigns that supply the second half of US Pu-238 production." }, { "name": "Betavolt", "kind": "company", "country": "CN", "what": "Markets the BV100 nickel-63 betavoltaic cell at 100 microwatts and 3 V from 15 x 15 x 5 mm, claimed for 50 years; the promised 1 W version has not been shown." }, { "name": "City Labs", "kind": "company", "country": "US", "what": "Supplies tritium betavoltaic cells for long-life microelectronics, the commercial end of the decades-lifetime, microwatt-output regime." } ], "evidence": [ { "date": "2015-08-01", "claim": "The MMRTG is specified at 43.6 kg mass, 110 W electrical and 1,975 W thermal at beginning of mission, from 4,103 g of plutonium (3,478 g Pu-238), fuel half-life 87.75 years, operating life at least 14 years. That is 2.52 We/kg at 5.6 percent conversion efficiency.", "source": "NASA", "title": "Multi-Mission Radioisotope Thermoelectric Generator (MMRTG) - Mars 2020 briefing", "url": "https://www.nasa.gov/wp-content/uploads/2015/08/4_mars_2020_mmrtg.pdf", "kind": "product", "delta": "0" }, { "date": "2026-08-29", "claim": "The GPHS-RTG masses about 57 kg and produces about 300 W electrical from about 4,400 W thermal using about 8.1 kg of Pu-238, a specific power of 5.2 We/kg - the highest of any fielded radioisotope generator.", "source": "Wikipedia", "title": "GPHS-RTG", "url": "https://en.wikipedia.org/wiki/GPHS-RTG", "kind": "product", "delta": "0" }, { "date": "2023-07-18", "claim": "DOE completed a 0.5 kg shipment of new heat-source plutonium oxide to NASA, the largest since domestic production restarted, and stated it remains on track for an average production target of 1.5 kg per year by 2026.", "source": "US Department of Energy, Office of Nuclear Energy", "title": "U.S. Department of Energy Completes Major Shipment of Plutonium-238 for NASA Missions", "url": "https://www.energy.gov/ne/articles/us-department-energy-completes-major-shipment-plutonium-238-nasa-missions", "kind": "product", "delta": "-" }, { "date": "2026-08-29", "claim": "Voyager 1, launched 5 September 1977 with about 470 W of RTG output, still had two operating science instruments in 2026 and is projected to return engineering data until 2036 - 48 years of continuous radioisotope power in the field, heading for 59.", "source": "Wikipedia", "title": "Voyager 1", "url": "https://en.wikipedia.org/wiki/Voyager_1", "kind": "product", "delta": "+" }, { "date": "2024-01-16", "claim": "The Betavolt BV100 betavoltaic cell produces 100 microwatts at 3 V from a 15 x 15 x 5 mm nickel-63 device, claimed to last 50 years; a 1 W version was promised for 2025.", "source": "New Atlas", "title": "Betavolt says its diamond nuclear battery can power devices for 50 years", "url": "https://newatlas.com/energy/betavolt-diamond-nuclear-battery/", "kind": "product", "delta": "0" } ], "indicators": [ { "metric": "Specific power of best fielded radioisotope generator", "unit": "We/kg", "value": 5.2, "as_of": "2026-08-29", "direction": "up_is_progress", "canon_target": 60, "source": "https://en.wikipedia.org/wiki/GPHS-RTG" }, { "metric": "Longest continuous operation of a fielded nuclear power source", "unit": "years", "value": 48, "as_of": "2026-08-29", "direction": "up_is_progress", "canon_target": 120, "source": "https://en.wikipedia.org/wiki/Voyager_1" }, { "metric": "US annual plutonium-238 production target", "unit": "kg/yr", "value": 1.5, "as_of": "2023-07-18", "direction": "up_is_progress", "canon_target": 8.1, "source": "https://www.energy.gov/ne/articles/us-department-energy-completes-major-shipment-plutonium-238-nasa-missions" }, { "metric": "Output of best commercial betavoltaic cell", "unit": "microwatts", "value": 100, "as_of": "2024-01-16", "direction": "up_is_progress", "canon_target": 300000000, "source": "https://newatlas.com/energy/betavolt-diamond-nuclear-battery/" } ] }, "commentary": "Radioisotope power is superb at lasting and hopeless at lifting. Voyager 1 has run on plutonium since 1977, forty-eight years and heading for fifty-nine, which is a respectable fraction of the canonical hundred and twenty. One flight-proven GPHS-RTG even delivers the canonical three hundred watts exactly. It weighs 57 kilograms, contains 8.1 kilograms of plutonium-238, and sheds 4.1 kilowatts of waste heat, which would have to leave through skin that is meant to feel human. American production of the isotope targets 1.5 kilograms a year. One unit is five and a half years of national output, assuming the target is met.", "last_reviewed": "2026-08-29", "review_cadence_days": 60, "progress": 0.05 }, { "id": "energy-density-and-endurance", "name": "Energy density and endurance", "category": "power", "weight": 9, "one_liner": "Real batteries versus 120 years.", "canon": { "requirement": "The same energy store as nuclear-power-cell, expressed as the property that actually binds: it has to fit inside a torso already full of hydraulic actuators, structural skeleton and a processor bay, and last 120 years without service.", "quantified": [ { "metric": "usable energy", "value": "1.14e12 J (300 W x 120 yr mid case)", "source_ref": "derived", "tier": "EXTRAPOLATED" }, { "metric": "assumed mass budget", "value": "<= ~10 kg (torso volume shared with hydraulics)", "source_ref": "derived", "tier": "EXTRAPOLATED" }, { "metric": "required specific energy", "value": "~1.1e11 J/kg = ~3.2e7 Wh/kg", "source_ref": "derived", "tier": "EXTRAPOLATED" }, { "metric": "ratio to best fielded lithium-ion (~250-300 Wh/kg)", "value": "~1e5 x", "source_ref": "derived", "tier": "EXTRAPOLATED" }, { "metric": "fraction of complete U-235 fission (8.2e13 J/kg)", "value": "~0.14%", "source_ref": "derived", "tier": "EXTRAPOLATED" } ], "sources": [ { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-garage-repair", "evidence": "The only canon statement of endurance.", "quote": "A hundred and twenty years with my existing power cell.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator Salvation", "year": 2009, "medium": "film", "ref": "salvation-fuel-cells", "evidence": "Establishes the source as nuclear.", "quote": "Fuel cells. Life source for the T-800. Nuclear. Enough to level this place.", "verified": true, "tier": "SECONDARY" } ], "canon_confidence": "extrapolated", "canonical_target": { "metric": "specific energy of the on-board power source", "value": "~3.2e7 Wh/kg at <=10 kg and <=5 L, with a 120-year shelf-and-service life", "reasoning": "Canon states the runtime; the density is arithmetic and the mass budget is a stated assumption. The finding that matters for the index: the target is ~1e5 x the best fielded lithium-ion, so this component is structurally incapable of scoring well on electrochemistry and the index should say so rather than pretending battery progress moves it. But it is NOT physically absurd - complete U-235 fission is 8.2e13 J/kg, so the requirement is ~0.14% of complete fission, comfortably inside what a real fission source could deliver at realistic burn-up. Canon's number is extreme for a battery and modest for a reactor. Score it against a nuclear denominator, not a battery one." }, "assumed_by_sota_agent": "An onboard energy store of approximately 1.14e12 J (316.7 MWh), sufficient for approximately 120 years of unattended operation at roughly 300 W continuous with combat-level peaks on demand, in a package that fits inside a humanoid torso. Canon confidence: extrapolated, cross-checked against the 1984 novelization's independent 'enough power to run the lights of a small city for a day'." }, "real": { "status": "blocked", "trl": 9, "spec_fraction": 0.00001, "spec_fraction_rationale": "Canon store 1.136e12 J = 315.6 MWh (300 W over 1,051,920 h). Best onboard store in any humanoid is Figure 03's 2.3 kWh, giving 2,300 / 315,600,000 = 7.3e-6. On endurance the ratio is 5 h / 1,051,920 h = 4.8e-6. At the best commercially catalogued specific energy in the world, 450 Wh/kg (Amprius Q3 2026), the canon store would mass 701 tonnes. Rounding to one significant figure: 1e-5. The 520 Wh/kg cell announced at CES 2026 improves this by 16 percent; solid-state roughly doubles it; neither is within four orders of magnitude.", "gap": "Roughly five orders of magnitude of stored energy, and it is not an engineering backlog. The peer-reviewed position is that humanoid applications already demand 10+ kWh where 2.3 kWh is state of the art, requiring a quadrupling of both volumetric and gravimetric capacity plus next-generation chemistries; the canon requirement is 30,000 times further out than that.", "why_hard": "Chemical bond energies cap achievable specific energy in the low tens of thousands of Wh/kg even for hydrocarbon combustion with atmospheric oxygen, and practical rechargeable cells are an order of magnitude below that. The canon store implies of order 1e6-1e7 Wh/kg from an onboard source. No electrochemistry reaches it, which is why this record is scored as blocked at TRL 9: the fielded technology is fully mature and the requirement is nonetheless unreachable by chemistry. The only route around it is nuclear, scored separately under nuclear-power-cell.", "movers": [ { "name": "Amprius Technologies", "kind": "company", "country": "US", "what": "Ships silicon-anode cells at 450 Wh/kg and 1,150 Wh/L, the highest commercially catalogued specific energy in the world; announced a 520 Wh/kg design at CES 2026." }, { "name": "Figure AI", "kind": "company", "country": "US", "what": "Built the 2.3 kWh Figure 03 pack claiming 5 hours at peak performance with 2 kW active-cooled fast charge, the first robot battery certified to UN38.3 and UL2271." }, { "name": "Boston Dynamics", "kind": "company", "country": "US", "what": "Sidesteps the problem architecturally: Atlas autonomously navigates to a station and swaps its own batteries, roughly every 4 hours." }, { "name": "Unitree Robotics", "kind": "company", "country": "CN", "what": "Publishes the honest low end: the H2 carries 0.972 kWh for about 3 hours of operation, an average draw near 324 W." }, { "name": "UT Dallas / Kyeongjae Cho group", "kind": "university", "country": "US", "what": "Authors of the July 2026 Advanced Science review setting the humanoid battery requirement at 10+ kWh and >1000 Wh/L for next-generation chemistries." }, { "name": "Shi and Pikul (Univ. of Wisconsin-Madison)", "kind": "university", "country": "US", "what": "Quantified the robot-versus-animal energy gap in Science Robotics: comparable locomotion efficiency, more than an order of magnitude worse system-level energy density." } ], "evidence": [ { "date": "2026-07-01", "claim": "Amprius's Q3 2026 product catalogue lists silicon-anode cells at up to 450 Wh/kg and 1,150 Wh/L, described as among the highest commercially available; four catalogued cells reach 450 Wh/kg gravimetric, with the next tier at 395-400 Wh/kg.", "source": "Amprius Technologies", "title": "Product Portfolio Q3 2026", "url": "https://amprius.com/documents/Amprius_Product_Catalog.pdf", "kind": "product", "delta": "+" }, { "date": "2026-01-06", "claim": "Amprius took a CES 2026 Best of Innovation award for a 520 Wh/kg cell design, described as the highest energy density of any commercial battery, against roughly 260 Wh/kg for standard graphite cells.", "source": "Consumer Technology Association", "title": "Amprius 520 Wh/kg cell - CES Innovation Awards 2026", "url": "https://www.ces.tech/ces-innovation-awards/2026/amprius-520-whkg-cell/", "kind": "product", "delta": "+" }, { "date": "2025-07-17", "claim": "Figure 03 carries a 2.3 kWh battery enabling 5 hours of run time at peak performance, with 2 kW active-cooled fast charge, a 94 percent increase in energy density across three generations, and certification to both UN38.3 and UL2271.", "source": "Figure AI", "title": "F.03 Battery Development", "url": "https://www.figure.ai/news/f-03-battery-development", "kind": "product", "delta": "+" }, { "date": "2026-07-29", "claim": "A peer-reviewed review states that uninterrupted humanoid application requires a quadrupling of battery system volumetric and gravimetric capacity, that application cases indicate energy demands of 10+ kWh, and that meeting them needs next-generation chemistries above 1000 Wh/L, with swappable packs as the interim stopgap.", "source": "Advanced Science 13, DOI 10.1002/advs.76736", "title": "Leading the Pack: Next-Generation Batteries for Humanoid Robotics", "url": "https://pubmed.ncbi.nlm.nih.gov/42524759/", "kind": "paper", "delta": "0" }, { "date": "2025-04-30", "claim": "Bioinspired mobile robots move with comparable efficiency to their animal counterparts but lag by more than an order of magnitude in system-level energy density because of battery limitations.", "source": "Science Robotics 10, DOI 10.1126/scirobotics.adr6125", "title": "Achieving animal endurance in robots through advanced energy storage", "url": "https://pubmed.ncbi.nlm.nih.gov/40305577/", "kind": "paper", "delta": "0" }, { "date": "2026-08-29", "claim": "Unitree publishes a 15 Ah / 0.972 kWh battery and about 3 hours of battery life for the 70 kg H2 humanoid, implying an average electrical draw of roughly 324 W.", "source": "Unitree Robotics", "title": "Unitree H2", "url": "https://www.unitree.com/H2", "kind": "product", "delta": "0" } ], "indicators": [ { "metric": "Highest commercially catalogued cell specific energy", "unit": "Wh/kg", "value": 450, "as_of": "2026-07-01", "direction": "up_is_progress", "canon_target": 3000000, "source": "https://amprius.com/documents/Amprius_Product_Catalog.pdf" }, { "metric": "Largest onboard energy store in a production humanoid", "unit": "kWh", "value": 2.3, "as_of": "2025-07-17", "direction": "up_is_progress", "canon_target": 315600, "source": "https://www.figure.ai/news/f-03-battery-development" }, { "metric": "Longest single-charge runtime claimed for a humanoid", "unit": "hours", "value": 5, "as_of": "2025-07-17", "direction": "up_is_progress", "canon_target": 1051920, "source": "https://www.figure.ai/news/f-03-battery-development" } ] }, "commentary": "The single worst number in the index, and it is not close. Best catalogued cell in the world: 450 watt-hours per kilogram. Best pack actually inside a humanoid: 2.3 kilowatt-hours, good for five hours. Canon: a hundred and twenty years on one cell, which at three hundred watts is 1.14 times ten to the twelve joules, or 701 tonnes of the finest chemistry money can buy. The ratio is seven parts in a million. Solid-state doubles it. The 520 watt-hour cell shown at CES improves it by sixteen per cent. Chemical bonds cap the approach four orders short, which is why this entry reads blocked at readiness level nine.", "last_reviewed": "2026-08-29", "review_cadence_days": 30, "progress": 0 }, { "id": "thermal-management", "name": "Thermal management", "category": "power", "weight": 6, "one_liner": "Dumping waste heat inside a sealed body.", "canon": { "requirement": "Reject the waste heat of a continuously-running machine through a sealed envelope of living human tissue held at human skin temperature - no vents, no fans, no radiators, no thermal signature that distinguishes it from a person - while ALSO surviving external immersion in a hydrocarbon fireball. Two opposite thermal problems in one body.", "quantified": [ { "metric": "skin temperature constraint", "value": "must feel human to touch and never flag an observer", "source_ref": "t2-garage-repair", "tier": "PRIMARY" }, { "metric": "detection channel that does work", "value": "canine olfaction, not thermal imaging", "source_ref": "t1-dogs", "tier": "PRIMARY" }, { "metric": "external thermal survival", "value": "walks out of a burning tanker fully functional (~1000 C fuel fire)", "source_ref": "t1-tanker", "tier": "PRIMARY" }, { "metric": "acknowledged thermal fault mode", "value": "'My primary cell was damaged by a plasma attack'", "source_ref": "t3-fuel-cells", "tier": "SECONDARY" } ], "sources": [ { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-dogs", "evidence": "Sarah wakes in the culvert; Reese explains the Resistance's detector - which is olfactory, not thermal.", "quote": "I was dreaming about dogs. / We use them to spot Terminators.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-garage-repair", "evidence": "Sarah sutures the Terminator's wounds by hand; the tissue behaves normally throughout.", "quote": "Will these heal up? / Yes.", "verified": true, "tier": "PRIMARY" } ], "canon_confidence": "implied", "canonical_target": { "metric": "continuous passive waste-heat rejection at human skin temperature", "value": ">=300 W rejected continuously through ~1.8 m2 of skin held within ~1-2 K of normal human surface temperature, passively, no forced convection; core electronics survive external immersion at ~1000 C", "reasoning": "No character ever discusses the T-800's heat; the requirement is inferred from the fact that no human notices it and that dogs, not thermal imagers, are the stated countermeasure. A resting adult dissipates ~100 W. If the T-800 draws the 300 W mid case and its hydraulics are even 50% efficient, it must shed 150-300 W more than a resting human at the same surface temperature - so 300 W is the defensible floor. The external figure comes from the tanker sequences." }, "assumed_by_sota_agent": "A sealed humanoid whose full electrical load (approximately 300 W, plus roughly 4.1 kW of radioisotope conversion waste heat if powered as canon implies) is rejected continuously and indefinitely, through a covering of living human tissue that must itself be kept alive at approximately 37 C, with no visible fans, vents or radiators, while operating at ambient temperatures up to and including a burning fuel tanker and a steel mill." }, "real": { "status": "in_progress", "trl": 7, "spec_fraction": 0.14, "spec_fraction_rationale": "Current humanoids sustain full load for 3-5 hours at ordinary indoor ambient, and only after a year of visible engineering effort: Beijing marathon teams cut joint temperatures from 70-80 C to about 60 C using combined liquid and air cooling, which alongside sub-10-second battery swaps is credited with cutting finishing times from over two hours in 2025 to under one hour in 2026. The best production humanoid is sealed to IP67 and rated only from -20 C to +40 C ambient. Ratio on rated ambient against a ~1,500 C canonical survival requirement is 40/1500 = 0.027; ratio on sustained duty at nominal ambient is roughly 0.4 (hours against indefinite); the tissue-cladding requirement, which adds a thermally insulating living layer over the only heat-rejection surface, has no demonstration at all (0). Composite approximately 0.14.", "gap": "Every watt in becomes a watt of heat out, and the canon design seals the exit. There is no demonstrated method of rejecting hundreds of watts continuously through an insulating biological covering, and no humanoid has ever been operated with one. Waterproofing and cooling are the same wall seen from opposite sides: IP67 sealing removes the convective path the machine needs.", "why_hard": "Actuator power density is thermally limited, not electrically limited. Electromechanical efficiency in humanoid joints runs 0.70-0.82 over positive-work bands, so tens to over a hundred watts are dissipated inside actuator volumes of a few cubic centimetres, and the state of the art in managing it is online motor-core temperature estimation followed by derating. Adding a radioisotope cell would add kilowatts. Adding living tissue would remove the radiator.", "movers": [ { "name": "Honor Device", "kind": "company", "country": "CN", "what": "Adapted smartphone liquid cooling to a humanoid; its 'Lightning' won the 2026 Beijing robot half marathon in 50:26 with liquid cooling and sub-10-second dual-battery hot swaps." }, { "name": "Beijing Humanoid Robot Innovation Centre / X-Humanoid", "kind": "lab", "country": "CN", "what": "Runs the Tiangong platform, which combined liquid and air cooling to hold joint temperatures near 60 C through endurance events." }, { "name": "HKUST (Guangzhou) AMDT Lab", "kind": "university", "country": "CN", "what": "Developing integrated liquid-cooling joint actuators specifically to lift the long-duration, high-load restriction on humanoid joint modules." }, { "name": "Boston Dynamics", "kind": "company", "country": "US", "what": "Publishes the constraint honestly: Atlas is IP67 sealed and rated -20 C to +40 C ambient, which is the shape of the sealed-body thermal problem in one specification line." }, { "name": "JSK Lab, University of Tokyo", "kind": "university", "country": "JP", "what": "Long-running work on online learning of motor thermal model parameters to estimate core temperature and prevent actuator damage in musculoskeletal humanoids." } ], "evidence": [ { "date": "2026-04-27", "claim": "At the 2026 Beijing humanoid half marathon, teams reduced joint temperatures from 70-80 C to approximately 60 C using combinations of liquid and air cooling, and battery swaps fell to under 10 seconds with no system interruption; completion times dropped from over two hours in 2025 to under one hour in 2026.", "source": "People's Daily Online", "title": "China-developed humanoid robot breaks half-marathon record in Beijing", "url": "https://en.people.cn/n3/2026/0427/c98649-20450883.html", "kind": "demo", "delta": "+" }, { "date": "2026-01-05", "claim": "The production Atlas is specified as extremely water-resistant and rated for ambient operation from -20 C to +40 C, i.e. sealed against the environment and therefore against convective cooling.", "source": "Boston Dynamics", "title": "Boston Dynamics Unveils New Atlas Robot to Revolutionize Industry", "url": "https://bostondynamics.com/blog/boston-dynamics-unveils-new-atlas-robot-to-revolutionize-industry/", "kind": "product", "delta": "0" }, { "date": "2026-05-27", "claim": "Trade analysis reports that as humanoid robots approach mass production, thermal management is emerging as a critical bottleneck.", "source": "DIGITIMES", "title": "Humanoid robot mass production hits a thermal wall", "url": "https://www.digitimes.com/news/a20260527PD226/robot-cooling-production-efficiency-management.html", "kind": "product", "delta": "-" }, { "date": "2024-07-10", "claim": "Managing humanoid actuator temperature is done by online learning of motor thermal model parameters to estimate core temperature and detect anomalies, i.e. by predicting when to derate rather than by rejecting more heat.", "source": "arXiv", "title": "Estimation and Control of Motor Core Temperature with Online Learning of Thermal Model Parameters: Application to Musculoskeletal Humanoids", "url": "https://arxiv.org/abs/2407.08055", "kind": "paper", "delta": "0" }, { "date": "2025-11-10", "claim": "A benchmarking framework for humanoid actuation reports electromechanical efficiency of 0.70-0.82 over positive-work bands and argues that continuous-safe torque under thermal limits, not peak torque, is the meaningful specification.", "source": "arXiv", "title": "Human-Level Actuation for Humanoids", "url": "https://arxiv.org/abs/2511.06796", "kind": "paper", "delta": "0" } ], "indicators": [ { "metric": "Humanoid joint temperature under endurance running load, best reported", "unit": "degC", "value": 60, "as_of": "2026-04-27", "direction": "down_is_progress", "canon_target": 40, "source": "https://en.people.cn/n3/2026/0427/c98649-20450883.html" }, { "metric": "Rated maximum ambient operating temperature, best production humanoid", "unit": "degC", "value": 40, "as_of": "2026-01-05", "direction": "up_is_progress", "canon_target": 1500, "source": "https://bostondynamics.com/blog/boston-dynamics-unveils-new-atlas-robot-to-revolutionize-industry/" }, { "metric": "Continuous full-load operating duration before thermal or energy limit", "unit": "hours", "value": 5, "as_of": "2025-07-17", "direction": "up_is_progress", "canon_target": 1051920, "source": "https://www.figure.ai/news/f-03-battery-development" } ] }, "commentary": "Every watt that goes in comes out as heat, and the canon design seals the exit. Joints at the Beijing half marathon ran at seventy to eighty degrees before liquid cooling brought them to about sixty; that, plus ten-second battery swaps, is most of why finishing times fell from over two hours to under one in a single year. The production Atlas is sealed to IP67 and rated to forty degrees ambient, because waterproofing and cooling are the same wall seen from opposite sides. Canon puts living tissue over that wall and walks the result into a steel mill. The state of the art is knowing when to slow down.", "last_reviewed": "2026-08-29", "review_cadence_days": 45, "progress": 0.1089 }, { "id": "bipedal-locomotion", "name": "Bipedal locomotion", "category": "locomotion", "weight": 9, "one_liner": "Walking and running, any terrain.", "canon": { "requirement": "Unbroken, unassisted bipedal walking and running over arbitrary unprepared terrain - rubble, stairs, wet storm-drain concrete, sand, steel-mill catwalks, a burning building - for the entire mission, with zero balance failures, no operator, and graceful degradation to crawling after loss of both legs.", "quantified": [ { "metric": "falls due to balance failure across two films", "value": "zero", "source_ref": "t1-full", "tier": "PRIMARY" }, { "metric": "terrain classes traversed", "value": "post-nuclear rubble, urban street, storm drain, stairs, catwalk, foundry floor", "source_ref": "t2-script-opening", "tier": "TERTIARY" }, { "metric": "locomotion after loss of legs and pelvis", "value": "continues, crawling, mission intact", "source_ref": "t1-press", "tier": "PRIMARY" }, { "metric": "continuous unsupervised duration", "value": "~48 h (T1), ~72 h (T2)", "source_ref": "t1-timeline", "tier": "PRIMARY" } ], "sources": [ { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-press", "evidence": "The endoskeleton, blown in half by Reese's pipe bomb, continues to drag itself after Sarah until she crushes it in the hydraulic press. No dialogue - visual only.", "quote": "", "verified": true, "tier": "PRIMARY", "note": "Deliberately recorded with no quote: this component is 'implied' precisely because nobody ever remarks on the fact that it walks." }, { "work": "Terminator 2: Judgment Day - revised final shooting script", "year": 1991, "medium": "shooting-script", "ref": "t2-script-opening", "evidence": "2029 battlefield stage direction.", "quote": "Three terminator endoskeletons advance, firing rapidly.", "verified": true, "tier": "TERTIARY" } ], "canon_confidence": "implied", "canonical_target": { "metric": "continuous unsupervised bipedal locomotion over unprepared terrain", "value": ">=72 h continuous, unmapped mixed terrain including stairs, rubble and wet surfaces, carrying a combat load, ZERO falls; degrades to crawling rather than stopping after loss of both legs", "reasoning": "72 hours is the observed mission length in T2 (arrival to steel mill). 'Zero falls' is not rhetorical: across roughly four hours of screen time and two films, no Terminator ever loses its footing except from an external impact. That is the actual canon bar and it is far harsher than 'can walk'." }, "assumed_by_sota_agent": "Walking and running over arbitrary terrain - rubble, stairs, water, fire, darkness - indefinitely and autonomously, while carrying weapons and equipment, and continuing to move after severe structural damage." }, "real": { "status": "in_progress", "trl": 8, "spec_fraction": 0.42, "spec_fraction_rationale": "On flat prepared ground, autonomous bipedal locomotion now exceeds human performance: Tiangong Ultra ran 100 m in 8.64 s (11.57 m/s average) under a rule mandating full autonomy for sprints, and Honor's Lightning ran a 21.1 km half marathon autonomously in 50:26 (6.97 m/s), faster than the human world record. Score that axis at approximately 1.0. On unstructured terrain, perceptive whole-body policies now decide autonomously when to walk, balance, climb, step down or vault with zero-shot hardware transfer, but nothing approaches rubble, fire or water: approximately 0.2. On damage tolerance and continued gait after injury, essentially 0. Weighting flat ground, terrain and robustness at 0.4/0.4/0.2 gives 0.4 + 0.08 + 0 = 0.48, trimmed to 0.42 for the fact that competition robots still tripped, broke apart and caught fire, and that only about 40 percent of half-marathon teams achieved autonomous navigation at all.", "gap": "Everything that is not a prepared surface. The demonstrated envelope is tracks, road courses, warehouse floors and research obstacle courses. There is no demonstrated locomotion through rubble, standing water at speed, fire, or with a damaged limb, and the duty cycle is 3-5 hours.", "why_hard": "Terrain generalisation is a perception and contact-estimation problem before it is a control problem: the robot must infer support conditions it cannot see, in real time, from partial plantar and visual evidence. Damage tolerance is harder still because it requires the controller to re-identify its own dynamics after the fact. Neither has a benchmark that resembles a battlefield.", "movers": [ { "name": "Boston Dynamics", "kind": "company", "country": "US", "what": "Production Atlas: 1.9 m, 90 kg, 56 DOF with fully rotational joints, IP67, self-swapping batteries; 2026 output committed to Hyundai's Metaplant Application Center and Google DeepMind." }, { "name": "X-Humanoid / Beijing Humanoid Robot Innovation Centre", "kind": "lab", "country": "CN", "what": "Tiangong Ultra: fastest autonomous humanoid over 100 m (8.64 s) and 400 m (38.16 s), and 2025 half-marathon winner." }, { "name": "Honor Device", "kind": "company", "country": "CN", "what": "'Lightning' won the 2026 Beijing half marathon autonomously in 50:26, using liquid cooling, 400 N-m joint torque and sub-10-second battery hot-swaps." }, { "name": "Figure AI", "kind": "company", "country": "US", "what": "Figure 03 deployed at BMW Plant Spartanburg for logistics parts sequencing after an 11-month Figure 02 trial supporting 30,000+ vehicles." }, { "name": "AgiBot", "kind": "company", "country": "CN", "what": "Holds the Guinness record for longest humanoid walk: 106 km, Suzhou to Shanghai, without powering down." }, { "name": "Agility Robotics", "kind": "company", "country": "US", "what": "Digit remains the reference for paid commercial bipedal deployment in logistics, the closest thing to routine operational use." }, { "name": "1X Technologies", "kind": "company", "country": "NO", "what": "NEO shipping to consumer homes at 20,000 dollars with an approval-gated human teleoperator 'Expert Mode' - the clearest live example of why the autonomy distinction matters." } ], "evidence": [ { "date": "2026-08-26", "claim": "Tiangong Ultra won the 100 m large-group final at the 2nd World Humanoid Robot Games in 8.64 seconds, resetting its own Games record of 8.86 s and beating Usain Bolt's 9.58 s human world record.", "source": "Global Times", "title": "Tiangong Ultra resets 100m mark at 8.64s as humanoid robot games close in Beijing", "url": "https://www.globaltimes.cn/page/202608/1369085.shtml", "kind": "benchmark", "delta": "+" }, { "date": "2026-08-29", "claim": "At the 2nd World Humanoid Robot Games (Beijing, 22-26 August 2026; 666 teams, 2,056 robots, 51 events) full autonomy was mandatory for sprints, relays, football, tai chi and gymnastics, and teleoperated robots in permissive events scored only half points. Records included 100 m in 8.86 s, 400 m in 38.16 s and a 2.88 m standing high jump. Some competing robots tripped, broke apart or caught fire.", "source": "Wikipedia", "title": "World Humanoid Robot Games", "url": "https://en.wikipedia.org/wiki/World_Humanoid_Robot_Games", "kind": "benchmark", "delta": "+" }, { "date": "2026-04-20", "claim": "Honor's 'Lightning' won the 2026 Beijing E-Town half marathon autonomously in 50 minutes 26 seconds over 21 km, faster than the human world record of about 57 minutes, with over 100 robots competing and about 40 percent finishing autonomously.", "source": "Caixin Global", "title": "Chinese-Built Robot Wins Beijing Half-Marathon", "url": "https://www.caixinglobal.com/2026-04-20/chinese-built-robot-wins-beijing-half-marathon-102435911.html", "kind": "benchmark", "delta": "+" }, { "date": "2026-08-29", "claim": "In 2025 only 6 of 21 entered robots finished the Beijing E-Town half marathon, the winner taking 2:40:42 with three battery changes and one fall; in 2026 over 300 robots entered, up to three battery changes were allowed, and remote-controlled entries carried a 1.2x time penalty.", "source": "Wikipedia", "title": "Beijing E-Town Half-Marathon", "url": "https://en.wikipedia.org/wiki/Beijing_E-Town_Half-Marathon", "kind": "benchmark", "delta": "+" }, { "date": "2026-01-05", "claim": "Boston Dynamics began production of the electric Atlas: 56 degrees of freedom, 2.3 m reach, 50 kg lift, autonomous self-charging battery swaps, three control modes (autonomous, teleoperated, tablet steering), with 2026 deployments committed to Hyundai's Robotics Metaplant Application Center and Google DeepMind.", "source": "Boston Dynamics", "title": "Boston Dynamics Unveils New Atlas Robot to Revolutionize Industry", "url": "https://bostondynamics.com/blog/boston-dynamics-unveils-new-atlas-robot-to-revolutionize-industry/", "kind": "product", "delta": "+" }, { "date": "2026-06-25", "claim": "BMW Group deployed Figure 03 humanoids at Plant Spartanburg for logistics parts sequencing, following an eleven-month Figure 02 trial that supported production of more than 30,000 X3 vehicles.", "source": "BMW Group PressClub", "title": "BMW Group advances the use of Physical AI in production with Figure 03 project in Spartanburg", "url": "https://www.press.bmwgroup.com/global/article/detail/T0458778EN/bmw-group-advances-the-use-of-physical-ai-in-production-with-figure-03-project-in-spartanburg?language=en", "kind": "product", "delta": "+" }, { "date": "2026-08-01", "claim": "Light-Loco-Parkour demonstrates end-to-end perceptive whole-body humanoid locomotion from onboard depth alone, autonomously deciding when to walk, balance, climb, step down or vault, with zero-shot transfer to indoor and outdoor hardware and no reference input, skill label or hand-coded gate.", "source": "arXiv", "title": "Light-Loco-Parkour: Versatile Perceptive Whole-Body Locomotion via Multi-Skill Distillation", "url": "https://arxiv.org/abs/2608.02653", "kind": "paper", "delta": "+" }, { "date": "2026-01-08", "claim": "Hyundai will begin Atlas deployments at car manufacturing plants in 2028, not 2026, initially for parts sequencing, expanding to component assembly by 2030 - a reminder that production start and operational use are years apart.", "source": "Engadget", "title": "Boston Dynamics unveils production-ready version of Atlas robot at CES 2026", "url": "https://www.engadget.com/big-tech/boston-dynamics-unveils-production-ready-version-of-atlas-robot-at-ces-2026-234047882.html", "kind": "product", "delta": "-" } ], "indicators": [ { "metric": "Fastest autonomous humanoid 100 m time", "unit": "s", "value": 8.64, "as_of": "2026-08-26", "direction": "down_is_progress", "canon_target": 7, "source": "https://www.globaltimes.cn/page/202608/1369085.shtml" }, { "metric": "Fastest autonomous humanoid half marathon time", "unit": "s", "value": 3026, "as_of": "2026-04-19", "direction": "down_is_progress", "canon_target": 1500, "source": "https://en.wikipedia.org/wiki/Beijing_E-Town_Half-Marathon" }, { "metric": "Share of half-marathon teams achieving autonomous navigation", "unit": "percent", "value": 40, "as_of": "2026-04-27", "direction": "up_is_progress", "canon_target": 100, "source": "https://en.people.cn/n3/2026/0427/c98649-20450883.html" }, { "metric": "Continuous operating time of best production humanoid on one charge", "unit": "hours", "value": 5, "as_of": "2025-07-17", "direction": "up_is_progress", "canon_target": 1051920, "source": "https://www.figure.ai/news/f-03-battery-development" } ] }, "commentary": "The one line in the index running ahead of schedule. In August a Chinese humanoid covered 100 metres in 8.64 seconds under a rule mandating full autonomy, against Bolt's 9.58; in April another ran a half marathon in 50:26, faster than any human has. Boston Dynamics is building a 90-kilogram, 56-degree-of-freedom Atlas that changes its own batteries, and BMW has Figure 03 sequencing parts in South Carolina. What remains is rubble, fire, water and damage, plus a three-to-five-hour battery that belongs to another entry. Flat ground is solved. Nobody fights a war on a running track.", "last_reviewed": "2026-08-29", "review_cadence_days": 21, "progress": 0.3733 }, { "id": "high-speed-pursuit", "name": "High-speed pursuit", "category": "locomotion", "weight": 6, "one_liner": "Running down a moving vehicle.", "canon": { "requirement": "Pursue a fleeing human who is using a vehicle, and sustain it. ATTRIBUTION MATTERS HERE: every on-screen instance of a Terminator running down a moving vehicle ON FOOT is the T-1000, not the T-800. The T-800's own demonstrated pursuit is vehicular - a Harley at ~129 km/h through the canal - plus foot running that keeps pace with sprinting humans but never catches a car. Because the index scores against a canonical T-800 Model 101, the T-800's performance sets the target and the T-1000's foot bracket is carried as an explicitly-attributed stretch reference.", "quantified": [ { "metric": "T-800 vehicular pursuit speed (BASELINE)", "value": "'It hits eighty' - ~129 km/h, on a Harley, one-handed, while carrying a passenger and returning fire", "source_ref": "t2-script-47h", "tier": "TERTIARY" }, { "metric": "T-800 foot pursuit", "value": "runs; keeps pace with sprinting adults; never catches a vehicle on foot in either primary film", "source_ref": "t1-alley", "tier": "PRIMARY" }, { "metric": "T-1000 gains on an accelerating Honda XR100 dirt bike from a standing start (REFERENCE, not baseline)", "value": "yes", "source_ref": "t2-script-45", "tier": "TERTIARY" }, { "metric": "T-1000 gains on a car reversing up a ramp (REFERENCE)", "value": "yes", "source_ref": "t2-script-79", "tier": "TERTIARY" }, { "metric": "T-1000 cannot catch a car at full throttle (upper bound on the REFERENCE)", "value": "correct", "source_ref": "t2-script-80f", "tier": "TERTIARY" }, { "metric": "derived T-1000 sustained foot speed bracket", "value": "~40-100 km/h", "source_ref": "derived", "tier": "EXTRAPOLATED" } ], "sources": [ { "work": "Terminator 2: Judgment Day - revised final shooting script", "year": 1991, "medium": "shooting-script", "ref": "t2-script-45", "evidence": "The T-1000 pursues John out of the mall parking garage on foot.", "quote": "John looks back... the T-1000 is behind him, running. He twists the throttle and guns the little bike forward. Incredibly, the T-1000 is gaining.", "verified": true, "tier": "TERTIARY" }, { "work": "Terminator 2: Judgment Day - revised final shooting script", "year": 1991, "medium": "shooting-script", "ref": "t2-script-46", "evidence": "Stage direction as the T-1000 reaches the street.", "quote": "the cop, running faster than O.J. Simpson at the airport", "verified": true, "tier": "TERTIARY" }, { "work": "Terminator 2: Judgment Day - revised final shooting script", "year": 1991, "medium": "shooting-script", "ref": "t2-script-80f", "evidence": "The upper bound - the car escapes.", "quote": "He sees the T-1000 roll to his feet and continue running. But he's dropping way behind now. Sarah has the car floored and the liquid-metal killer won't catch them on foot.", "verified": true, "tier": "TERTIARY" } ], "canon_confidence": "implied", "canonical_target": { "metric": "pursuit performance - T-800 baseline, with the T-1000 foot bracket as a labelled reference", "value": "BASELINE (T-800 Model 101, what the index scores against): sustained vehicular pursuit at >=129 km/h on a two-wheeled platform, one-handed, while carrying a passenger and engaging; plus foot running at or above trained-human sprint speed with NO aerobic limit, sustained for the mission. REFERENCE (T-1000, explicitly attributed, NOT the baseline): sustained foot pursuit at >=40 km/h (~11 m/s) for minutes over urban terrain, bounded above by a floored passenger car (~100 km/h).", "reasoning": "The call: the roster shorthand 'running down a moving vehicle' describes a T-1000 feat, and the index's 1000-level is defined as a manufacturable canonical T-800 Model 101 - so making the T-1000's foot speed the denominator would score real machines against the wrong unit. The T-800's own pursuit is vehicular and its foot speed is never shown to exceed a running human's. Hence the split. The T-1000 bracket is retained because it is unusually well-bounded and is the franchise's actual ceiling: lower bound - it GAINS on a small dirt bike accelerating out of a parking garage into street traffic (30-50 km/h and closing); upper bound - it cannot catch a sedan at full throttle. For scale, Usain Bolt's peak is 44.7 km/h held for about two seconds; the T-1000 reference requires that speed as a cruise. Both figures come from TERTIARY stage direction and are labelled." }, "assumed_by_sota_agent": "Sustained running at vehicle speed - approximately 17-28 m/s (60-100 km/h) - for minutes at a time on public roads, with obstacle avoidance, braking and re-acceleration, unsupported. NOTE: the clearest on-screen instance of a machine running down a moving vehicle on foot is the T-1000 in T2, not a T-800; attribution to be settled by the canon record." }, "real": { "status": "in_progress", "trl": 6, "spec_fraction": 0.3, "spec_fraction_rationale": "Verified peak: Tiangong Ultra covered 100 m in 8.64 s, an 11.57 m/s average, autonomously, on an athletics track. Verified sustained: Honor's Lightning held 6.97 m/s for 50 min 26 s over a 21.1 km closed road course with permitted battery swaps. Against a canonical ~20 m/s sustained, that is 6.97/20 = 0.35 sustained and 11.57/20 = 0.58 peak. Both were set on prepared surfaces with no obstacle avoidance at speed, no turning under pursuit and no traffic, so the composite is discounted to 0.30. For calibration, Boston Dynamics' Cheetah reached 12.65 m/s in 2012 on a treadmill with off-board power and a supporting boom - fourteen years later the free-running autonomous number has only just caught up with the tethered one.", "gap": "Pursuit, not speed. No legged robot has been shown turning, braking and re-accelerating at speed around obstacles, on a real road, in traffic, while tracking a target. Sustained road speed is one third of vehicle speed and the fastest verified runs are straight lines on a track.", "why_hard": "Running speed in a biped scales with ground-contact force and swing-leg frequency, both of which are limited by actuator thermal capacity, and the energy cost of transport rises steeply with speed against a 3-5 hour battery. Turning at 10 m/s additionally requires lateral ground reaction forces that current foot and ankle designs cannot generate without slipping.", "movers": [ { "name": "X-Humanoid / Beijing Humanoid Robot Innovation Centre", "kind": "lab", "country": "CN", "what": "Holds the verified autonomous humanoid sprint records: 100 m in 8.64 s and 400 m in 38.16 s." }, { "name": "Unitree Robotics", "kind": "company", "country": "CN", "what": "Claims 12.66 m/s and a 2 m standing jump for an experimental platform called Superman, and 10.1 m/s for an H1, in both cases on company-supplied video with no independent verification." }, { "name": "Honor Device", "kind": "company", "country": "CN", "what": "Holds the verified sustained figure: 6.97 m/s over 21.1 km, autonomously, in April 2026." }, { "name": "Boston Dynamics", "kind": "company", "country": "US", "what": "Set the historical benchmark with Cheetah at 12.65 m/s in 2012, tethered on a treadmill, and has not pursued untethered high speed since." } ], "evidence": [ { "date": "2026-08-26", "claim": "Tiangong Ultra ran 100 m in 8.64 s at the World Humanoid Robot Games, an average of 11.57 m/s, in an event where full autonomy was mandatory.", "source": "Global Times", "title": "Tiangong Ultra resets 100m mark at 8.64s as humanoid robot games close in Beijing", "url": "https://www.globaltimes.cn/page/202608/1369085.shtml", "kind": "benchmark", "delta": "+" }, { "date": "2026-08-17", "claim": "Unitree claimed a top speed of 12.66 m/s and a 2 m standing jump for a robot it calls Superman, with 0.85 m leg length, developed in just over three months and carrying no hands or grippers. The claim rests on a company statement and a company-supplied video; test conditions, payload, surface and repeatability were not disclosed and there is no independent verification.", "source": "Global Times", "title": "Unitree's new humanoid robot jumps 2 meters, hits 12.66 m/s to break human records", "url": "https://www.globaltimes.cn/page/202608/1368390.shtml", "kind": "demo", "delta": "0" }, { "date": "2026-04-12", "claim": "Unitree claimed 10.1 m/s for an H1 on an athletics track, while itself noting that the measuring equipment might be subject to error. No independent verification.", "source": "Global Times", "title": "Unitree's H1 robot hits 10 m/s sprint speed", "url": "https://www.globaltimes.cn/page/202604/1358712.shtml", "kind": "demo", "delta": "0" }, { "date": "2026-08-29", "claim": "Honor's Lightning sustained an average 6.97 m/s over 21.1 km to win the 2026 Beijing E-Town half marathon autonomously in 50:26, under rules allowing up to three battery changes and penalising remote-controlled entries 1.2x on time.", "source": "Wikipedia", "title": "Beijing E-Town Half-Marathon", "url": "https://en.wikipedia.org/wiki/Beijing_E-Town_Half-Marathon", "kind": "benchmark", "delta": "+" }, { "date": "2012-09-05", "claim": "Boston Dynamics' Cheetah reached 28.3 mph (12.65 m/s), but on a treadmill, relying on off-board power and a boom to keep it from falling over, unable to operate outdoors or manoeuvre laterally.", "source": "IEEE Spectrum", "title": "Boston Dynamics' Cheetah Robot Now Faster Than Fastest Human", "url": "https://spectrum.ieee.org/boston-dynamics-cheetah-robot-now-faster-than-fastest-human-2650267047", "kind": "demo", "delta": "0" } ], "indicators": [ { "metric": "Fastest verified autonomous legged-robot speed, free running", "unit": "m/s", "value": 11.57, "as_of": "2026-08-26", "direction": "up_is_progress", "canon_target": 20, "source": "https://www.globaltimes.cn/page/202608/1369085.shtml" }, { "metric": "Fastest verified sustained legged-robot speed over 20+ km", "unit": "m/s", "value": 6.97, "as_of": "2026-04-19", "direction": "up_is_progress", "canon_target": 20, "source": "https://en.wikipedia.org/wiki/Beijing_E-Town_Half-Marathon" }, { "metric": "Duration a humanoid can sustain running speed", "unit": "minutes", "value": 50, "as_of": "2026-04-19", "direction": "up_is_progress", "canon_target": 60, "source": "https://en.wikipedia.org/wiki/Beijing_E-Town_Half-Marathon" } ] }, "commentary": "Boston Dynamics' Cheetah hit 12.65 metres per second in 2012, on a treadmill, tethered, with a boom holding it up. Fourteen years later the free-running, self-powered, autonomous record is 11.57 metres per second over a hundred metres, and the sustained figure is 6.97 for fifty minutes on a closed road course with battery swaps permitted. Unitree claims 12.66 from a machine with no hands, on the strength of its own video and no disclosed conditions. Canon requires staying with a car in traffic for minutes. On a track we are at a third of that. On a street, nobody has measured.", "last_reviewed": "2026-08-29", "review_cadence_days": 21, "progress": 0.2 }, { "id": "actuation-and-strength", "name": "Actuation and strength", "category": "locomotion", "weight": 8, "one_liner": "Actuators that lift, crush, and never tire.", "canon": { "requirement": "Actuators - hydraulic, per production material - delivering multiples of human strength from a human-sized limb, with no duty cycle, no fatigue and no thermal derating. Canon's emphasis is never a single feat of strength; it is the absence of any limit. NOTE: 'crushing hardened steel one-handed' is NOT a verified T-800 feat - see the disputed_claims entry below.", "quantified": [ { "metric": "one-armed throw", "value": "a 230 lb (104 kg) man, 'clear over the bar'", "source_ref": "t2-script-8c", "tier": "TERTIARY" }, { "metric": "two-handed overhead lift", "value": "a large adult man raised off the floor by the ribcage, above the Terminator's own head", "source_ref": "t1-ginger-apartment", "tier": "TERTIARY" }, { "metric": "grip force", "value": "fractures a human humerus casually; 'an inhuman grip' that lifts a man onto his toes by the wrist", "source_ref": "t2-pescadero", "tier": "PRIMARY" }, { "metric": "striking force", "value": "'punches the leader with piledriver force'; punches through a car windscreen", "source_ref": "t1-punks", "tier": "TERTIARY" }, { "metric": "crew-served weapon handling", "value": "carries and fires an M134 minigun, walking, unbraced", "source_ref": "t2-cyberdyne", "tier": "PRIMARY" }, { "metric": "fatigue exhibited", "value": "none, ever", "source_ref": "t1-full", "tier": "PRIMARY" }, { "metric": "derived single-arm peak", "value": "~2.06 kJ delivered in ~0.3 s = ~6.9 kW", "source_ref": "derived", "tier": "EXTRAPOLATED" }, { "metric": "DISPUTED: 'crushes hardened steel one-handed'", "value": "NOT VERIFIED as a T-800 feat in T1 or T2. The equivalent on-screen steel-deformation feat belongs to the T-1000 (T2 scene 73L, prying the elevator doors apart with fingertip pry-bars).", "source_ref": "t2-script-73l", "tier": "NOT-VERIFIED" } ], "sources": [ { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-pescadero", "evidence": "The Terminator breaks a hospital guard's arm and pins him. Note: the adult human skeleton has 206 bones, not 215 - canon's own precision claim is wrong.", "quote": "There are 215 bones in the human body. That's one. Now, don't move.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day - revised final shooting script", "year": 1991, "medium": "shooting-script", "ref": "t2-script-8c", "evidence": "The bar fight at the truck stop.", "quote": "Terminator hurls Cigar, all 230 pounds of him, clear over the bar.", "verified": true, "tier": "TERTIARY" }, { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-ginger-apartment", "evidence": "The Terminator lifts Ginger's boyfriend Matt off the floor by the ribcage, two-handed, above its own head. Stage direction from the 1983 fourth draft; the lift is on screen in the finished film.", "quote": "Terminator places one hand on either side of Matt's barrel chest. SINKS HIS FINGERS INTO THE FLESH. An inhuman grip. Matt is raised off the floor, contorted with agony, above the other's head.", "verified": true, "tier": "TERTIARY" }, { "work": "Terminator 2: Judgment Day - revised final shooting script", "year": 1991, "medium": "shooting-script", "ref": "t2-script-73l", "evidence": "The T-1000 - NOT the T-800 - forces the Pescadero elevator doors. This is the franchise's clearest bare-handed steel-deformation feat and it belongs to the liquid-metal unit.", "quote": "it jams its hands between them, its fingertips becoming pry-bars. It pulls the doors apart with inhuman strength", "verified": true, "tier": "TERTIARY", "note": "Recorded to ADJUDICATE the 'T-800 crushes hardened steel one-handed' claim: the verified instance is the T-1000's, not the T-800's." } ], "canon_confidence": "implied", "canonical_target": { "metric": "single-limb peak mechanical output, lift capacity, and unlimited duty cycle", "value": "~2 kJ delivered in ~0.3 s (~7 kW) from ONE arm; two-handed overhead lift of a ~100 kg adult; one-handed grip sufficient to fracture a human long bone; sustained carriage and unbraced firing of a ~40 kg crew-served weapon - all with unlimited duty cycle and no thermal derating", "reasoning": "Built only from feats actually shown. Take the one canon feat with a mass attached: a 104 kg man thrown 'clear over the bar', call it 4 m horizontal. For a level throw v = sqrt(gR) = sqrt(9.81 x 4) = 6.3 m/s, so KE = 0.5 x 104 x 6.3^2 = 2.06 kJ; over a ~0.3 s arm extension that is ~6.9 kW from ONE arm. An elite human athlete peaks at 0.5-1.5 kW using the whole body, so the requirement is ~5-10x a human's whole-body peak from a single limb, repeatable without rest. The duty-cycle clause is where real actuators actually fail, which is why it is in the target rather than a footnote. Cross-check: ~7 kW single-arm peak sits consistently under the ~12 kW full-power draw derived independently from the novelization in nuclear-power-cell. DELIBERATELY EXCLUDED: 'crushing hardened steel one-handed' - widely asserted, not verifiable as a T-800 feat in either primary film. The franchise's clearest bare-handed steel-deformation feat is the T-1000 prying open the Pescadero elevator doors, and attributing it to the T-800 would inflate the denominator of a component the index scores against a T-800 baseline." }, "assumed_by_sota_agent": "Actuators that lift and throw an adult one-handed, punch through vehicle glass and bodywork, restrain a struggling person indefinitely with one hand, and carry and fire weapons no human could hold steady - with no fatigue, no thermal derating and no servicing, for decades. Assumed whole-body lift capacity of order 500 kg. NOTE: 'crushing hardened steel one-handed' could not be verified as an on-screen T-800 feat and is deliberately excluded." }, "real": { "status": "in_progress", "trl": 8, "spec_fraction": 0.13, "spec_fraction_rationale": "Published peaks are respectable: the production Atlas lifts 50 kg instantaneously and 30 kg sustained at 90 kg body mass across 56 DOF, and Unitree publishes 360 N-m maximum leg joint torque for the 70 kg H2 - roughly 2.5-4x a 75 kg human's peak ankle moment of 90-143 N-m. Against an assumed canonical whole-body lift of order 500 kg, 50/500 = 0.10 instantaneous and 30/500 = 0.06 sustained. The deeper problem is that peak torque is the wrong number: in the reference human-equivalence framework's worked example, coverage of human requirements was 0.546 for the ankle in walking but 0.085 for hip and knee in stair support, precisely the sustained low-rate regime canon demands, because that regime is thermally rather than electrically limited. Weighting instantaneous lift, sustained lift and continuous-safe joint coverage gives approximately 0.13. 'Never tires' is not achievable with any current electric actuator at any score.", "gap": "Continuous, not peak. Every published humanoid specification is a burst figure; the sustained figure after a thermal soak is unpublished and, where it has been measured in the literature, collapses in exactly the low-rate high-torque regime that holding, pressing and crushing require. Whole-body strength is roughly a tenth of the canonical requirement and the machines fatigue in minutes rather than never.", "why_hard": "Actuator design is a closed trade-off, not a frontier. Quasi-direct-drive with low-ratio planetary or cycloidal gearing buys backdrivability and torque-mode bandwidth at the cost of motor size and thermal stress; strain-wave and cycloidal reducers buy static torque at the cost of reflected inertia (proportional to the square of the gear ratio), friction and sustainable duty cycle. Continuous torque is bounded by copper losses and heat rejection into a few cubic centimetres, so 'never tires' is a thermodynamic statement, not an engineering target.", "movers": [ { "name": "Boston Dynamics", "kind": "company", "country": "US", "what": "Publishes the highest verified humanoid payload: 50 kg instantaneous and 30 kg sustained at 90 kg body mass, across 56 fully rotational joints." }, { "name": "Unitree Robotics", "kind": "company", "country": "CN", "what": "Publishes joint-level figures others do not: 360 N-m peak leg joint torque, 120 N-m arm joint torque, 7 kg rated / 15 kg peak arm payload on the H2." }, { "name": "UCLA RoMeLa (Dennis Hong)", "kind": "university", "country": "US", "what": "Cycloidal quasi-direct-drive actuator designs with learning-based torque estimation, targeting high torque density without sacrificing backdrivability." }, { "name": "Istituto Italiano di Tecnologia", "kind": "lab", "country": "IT", "what": "Systematic optimisation comparing internal and external single-stage planetary gearbox architectures for legged-robot actuators." }, { "name": "Honor Device", "kind": "company", "country": "CN", "what": "Applies 400 N-m motor torque with folding-phone hinge technology and liquid cooling, the clearest case of consumer-electronics actuation know-how entering humanoids." } ], "evidence": [ { "date": "2025-11-10", "claim": "A benchmarking framework argues that peak torque and no-load speed specifications tell us little about whether a joint can deliver torque, power and endurance at task-relevant operating points, and that continuous-safe torque maps measured after a thermal soak are the meaningful measure. In its worked example, human-equivalence coverage was 0.546 for the ankle in walking and 0.085 for hip and knee in stair support. Human reference moments for a 75 kg subject: hip 38-75 N-m, knee 35-95 N-m, ankle 90-143 N-m with 190-260 W of positive ankle power at push-off.", "source": "arXiv", "title": "Human-Level Actuation for Humanoids", "url": "https://arxiv.org/abs/2511.06796", "kind": "paper", "delta": "0" }, { "date": "2026-08-29", "claim": "Unitree publishes 360 N-m maximum leg joint torque and 120 N-m maximum arm joint torque for the 70 kg, 31-DOF H2, with approximately 7 kg rated and 15 kg peak arm payload.", "source": "Unitree Robotics", "title": "Unitree H2", "url": "https://www.unitree.com/H2", "kind": "product", "delta": "0" }, { "date": "2026-01-24", "claim": "The production Atlas is specified at 90 kg with 50 kg instantaneous and 30 kg sustained payload across 56 degrees of freedom.", "source": "Humanoids Daily", "title": "The Alien in the Factory: Boston Dynamics Launches Production-Ready Atlas at CES 2026", "url": "https://www.humanoidsdaily.com/news/the-alien-in-the-factory-boston-dynamics-launches-production-ready-atlas-at-ces-2026", "kind": "product", "delta": "+" }, { "date": "2024-10-22", "claim": "Cycloidal quasi-direct-drive actuators offer high torque density and mechanical robustness for legged robots, with a learned actuator network used to close the sim-to-real gap introduced by the drive's complex dynamics.", "source": "arXiv", "title": "Cycloidal Quasi-Direct Drive Actuator Designs with Learning-based Torque Estimation for Legged Robotics", "url": "https://arxiv.org/abs/2410.16591", "kind": "paper", "delta": "+" }, { "date": "2026-06-14", "claim": "A physics-based model of the Unitree G1's seven-DOF arm electrical power consumption achieved R-squared of 0.933 with 1.07 W RMSE over 897 trajectories and 0.965 on 46 unseen speed trajectories, finding viscous friction dominant in most joints and copper losses dominant in shoulder yaw and elbow.", "source": "arXiv", "title": "Identification of a Physics-Based Electrical Power Consumption Model for the Unitree G1 Humanoid Arm", "url": "https://arxiv.org/abs/2606.15915", "kind": "paper", "delta": "+" } ], "indicators": [ { "metric": "Sustained whole-body payload, best production humanoid", "unit": "kg", "value": 30, "as_of": "2026-01-24", "direction": "up_is_progress", "canon_target": 500, "source": "https://www.humanoidsdaily.com/news/the-alien-in-the-factory-boston-dynamics-launches-production-ready-atlas-at-ces-2026" }, { "metric": "Instantaneous whole-body payload, best production humanoid", "unit": "kg", "value": 50, "as_of": "2026-01-05", "direction": "up_is_progress", "canon_target": 500, "source": "https://bostondynamics.com/blog/boston-dynamics-unveils-new-atlas-robot-to-revolutionize-industry/" }, { "metric": "Published peak leg joint torque, commercially available humanoid", "unit": "Nm", "value": 360, "as_of": "2026-08-29", "direction": "up_is_progress", "canon_target": 2000, "source": "https://www.unitree.com/H2" }, { "metric": "Human-equivalence envelope coverage for hip and knee in sustained support (worked example)", "unit": "fraction", "value": 0.085, "as_of": "2025-11-10", "direction": "up_is_progress", "canon_target": 1, "source": "https://arxiv.org/abs/2511.06796" } ] }, "commentary": "Peak torque is the most misleading number in robotics and the industry quotes little else. Unitree publishes 360 newton-metres at the leg joint, two to four times a human's peak; Atlas lifts fifty kilograms instantaneously and thirty sustained. Then the thermal soak begins. November's human-equivalence framework makes the point precisely: in its worked example the ankle met human requirements across 55 per cent of the walking band, while the hip and knee met sustained stair-support demand across eight per cent, because low-rate torque is a heat problem rather than a current problem. Canon asks for indefinite. What the actuators offer is a burst, then a derate.", "last_reviewed": "2026-08-29", "review_cadence_days": 30, "progress": 0.1156 }, { "id": "dexterous-manipulation", "name": "Dexterous manipulation", "category": "manipulation", "weight": 8, "one_liner": "Hands.", "canon": { "requirement": "A five-fingered hand with full human kinematic range spanning - WITHOUT changing end-effector - from crushing bone to performing surgery on itself with a craft knife under mirror-inverted visual guidance.", "quantified": [ { "metric": "finest task performed", "value": "excising its own damaged eyeball with a knife point; suturing its own wrist with needle and thread", "source_ref": "t1-hotel-repair", "tier": "TERTIARY" }, { "metric": "internal repair", "value": "disassembles the damaged mechanism in its own forearm with small screwdrivers", "source_ref": "t1-hotel-repair", "tier": "TERTIARY" }, { "metric": "coarsest task, same hand", "value": "crushing a human arm; punching through a windscreen", "source_ref": "t2-pescadero", "tier": "PRIMARY" }, { "metric": "guidance", "value": "via a bathroom mirror - the machine inverts its own body schema", "source_ref": "t1-hotel-repair", "tier": "TERTIARY" } ], "sources": [ { "work": "The Terminator - fourth draft, 20 April 1983", "year": 1983, "medium": "screenplay-draft", "ref": "t1-hotel-repair", "evidence": "Scene 158, hotel room. The sequence is on screen in the finished film; only this wording is draft stage direction.", "quote": "With a smooth motion the knife point enters the eyeball and cuts away the ruined sclera and cornea, as well as part of the damaged eyelids.", "verified": true, "tier": "TERTIARY" }, { "work": "The Terminator - fourth draft, 20 April 1983", "year": 1983, "medium": "screenplay-draft", "ref": "t1-hotel-repair", "evidence": "Forearm self-repair, same sequence.", "quote": "He picks up an X-ACTO KNIFE and cuts deeply into the skin of his forearm... With small screwdrivers he begins to patiently disassemble the damaged mechanism around the 12-gauge hit.", "verified": true, "tier": "TERTIARY" } ], "canon_confidence": "implied", "canonical_target": { "metric": "precision and force range of a single unmodified end-effector", "value": "Sub-millimetre tool-tip placement under visual servoing - sufficient to perform surgery on its own body via a mirror - while the SAME unmodified hand delivers bone-fracturing grip force", "reasoning": "The discriminator is not precision or force; it is BOTH, in one end-effector, with no tool change. Real dexterous hands trade the two off, which is exactly the comparison the index should force. The precision figure comes from the task: cutting away a sclera without damaging what is behind it is sub-millimetre, and it is done mirror-reversed, which additionally requires the machine to re-map its own kinematics." }, "assumed_by_sota_agent": "A five-fingered hand functionally indistinguishable from a human one: arbitrary, never-before-seen objects manipulated at human speed or better, fully autonomously, with tactile feedback fine enough for the T-800 to perform microsurgery on its own forearm and force enough to lift a man by the throat, sustained for decades without recalibration and with no task-specific training." }, "real": { "status": "in_progress", "trl": 6, "spec_fraction": 0.1, "spec_fraction_rationale": "Mean autonomous success across the five published multi-finger dexterity tasks in Gemini Robotics 2 (36, 92, 44, 32, 40) is 48.8%, achieved at roughly one-quarter human speed (Epoch AI: precision manipulation 3-5x slower than a human), on a developer-chosen task set. 0.488 x 0.25 = 0.12, discounted to 0.10 because canon requires ~100% success at >=1x speed on arbitrary unseen objects, for which there is no published evidence of retention.", "gap": "Hardware has closed: 22 actuated DoF against roughly 24 in a human hand, and 361 tactile sensels/cm2 against 241 human fingertip mechanoreceptors/cm2. Control has not. Autonomous success on five-fingered dexterity tasks ranges 32-92% and is worse than a two-finger gripper on every comparable axis; speed is 3-10x below human; generalisation to unseen objects and environments is the field's own stated bottleneck. Nothing approaches self-surgery.", "why_hard": "The binding limit is contact data, not compute or actuators. Simulation cannot reproduce friction, deformation and slip well enough to transfer, so contact behaviour must be learned from real robot-hours, a resource that scales linearly with fleet size and cannot be scraped from the internet.", "movers": [ { "name": "Physical Intelligence", "kind": "company", "country": "US", "what": "pi-0.7 generalist VLA (Apr 2026): matches or beats specialist policies at 1.2-1.6x normalised throughput and transfers zero-shot across embodiments." }, { "name": "Google DeepMind", "kind": "lab", "country": "US", "what": "Gemini Robotics 2 (Jul 2026) controls a full humanoid feet-to-fingertips and is one of the very few to publish per-task autonomous success rates on five-fingered hands." }, { "name": "Figure AI", "kind": "company", "country": "US", "what": "Helix 02: 61-action autonomous dishwasher task, 3-gram fingertip force threshold, plus a $1bn+ programme buying human hand-motion video to break the data bottleneck." }, { "name": "Ensuring Technology", "kind": "company", "country": "CN", "what": "Tacta tactile sensor: 361 sensels/cm2 at 1 kHz; first full-hand tactile coverage (1,956 sensels) shown at CES 2026." }, { "name": "Apptronik / Sharpa", "kind": "company", "country": "US/SG", "what": "Apollo humanoid and the 22-DoF SharpaWave hand used as the reference platform in DeepMind's multi-finger dexterity evaluations." }, { "name": "Epoch AI", "kind": "lab", "country": "US", "what": "Independent capability audit that separates deployed manipulation from lab demos and publishes speed and reliability figures the vendors do not." }, { "name": "Shadow Robot Company", "kind": "company", "country": "UK", "what": "Shadow Hand, still the 24-DoF research reference against which humanoid hands are measured." } ], "evidence": [ { "date": "2026-07-30", "claim": "Gemini Robotics 2 publishes per-task autonomous success on a five-fingered humanoid hand: 92% unscrewing a bulb, 44% tying a trash bag, 40% ziplock, 36% screwing a bulb in, 32% sweeping into a dustpan.", "source": "Google DeepMind", "title": "Gemini Robotics 2 brings whole body intelligence to robots", "url": "https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/", "kind": "product", "delta": "+" }, { "date": "2026-04-16", "claim": "pi-0.7 achieves 1.2-1.6x normalised throughput versus specialist policies on laundry, espresso and box-building, and transfers zero-shot to an unseen robot embodiment.", "source": "arXiv", "title": "pi-0.7: a Steerable Generalist Robotic Foundation Model with Emergent Capabilities", "url": "https://arxiv.org/abs/2604.15483", "kind": "paper", "delta": "+" }, { "date": "2026-01-27", "claim": "Figure's Helix 02 completes a 4-minute, 61-action dishwasher load/unload fully autonomously, with fingertip tactile sensing resolving 3-gram forces.", "source": "Figure AI", "title": "Introducing Helix 02: Full-Body Autonomy", "url": "https://www.figure.ai/news/helix-02", "kind": "demo", "delta": "+" }, { "date": "2026-02-10", "claim": "Independent audit finds robots typically 3-10x slower than humans, precision-insertion reliability averaging ~52%, and laundry success falling to 75% on jeans.", "source": "Epoch AI", "title": "Where Autonomy Works: Evaluating Robot Capabilities in 2026", "url": "https://epoch.ai/blog/where-autonomy-works-evaluating-robot-capabilities-in-2026", "kind": "benchmark", "delta": "0" }, { "date": "2026-01-09", "claim": "First full tactile coverage of a dexterous hand: 1,956 multi-dimensional sensels at 361 sensels/cm2, sampling at 1 kHz, 4.5 mm thick.", "source": "PR Newswire", "title": "From Fingertips to Full-Body Coverage: Ensuring Technology Debuts Tactile Infrastructure at CES 2026", "url": "https://www.prnewswire.com/news-releases/from-fingertips-to-full-body-coverage-ensuring-technology-debuts-groundbreaking-tactile-infrastructure-at-ces-2026-302657225.html", "kind": "product", "delta": "+" }, { "date": "2026-07-05", "claim": "RoboDojo establishes a unified sim-and-real manipulation leaderboard: 42 simulation tasks, 18 real-world tasks, 30 policies, standardised remote real-world evaluation.", "source": "arXiv", "title": "RoboDojo: A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policies", "url": "https://arxiv.org/abs/2607.04434", "kind": "benchmark", "delta": "+" }, { "date": "2026-05-13", "claim": "Survey tabulates the current hand field at 15-24 DoF (Shadow 24, Optimus Gen 3 22, LEAP II 21, Unitree Dex5-1P 20) and names cost and reliability as the primary barriers to industrial deployment.", "source": "arXiv", "title": "Towards Robotic Dexterous Hand Intelligence: A Survey", "url": "https://arxiv.org/abs/2605.13925", "kind": "paper", "delta": "0" }, { "date": "1979-03-01", "claim": "Human fingertip low-threshold mechanoreceptor innervation density is 241 units/cm2, against 58 units/cm2 in the palm.", "source": "Journal of Physiology", "title": "Tactile sensibility in the human hand: relative and absolute densities of four types of mechanoreceptive units in glabrous skin", "url": "https://pubmed.ncbi.nlm.nih.gov/439026/", "kind": "paper", "delta": "0" }, { "date": "2026-08-25", "claim": "Figure is buying manipulation data at scale: 16 million-plus videos from 44,000 weekly users across 100+ countries, $15m paid to contributors, $1bn+ planned data and compute spend over twelve months.", "source": "Figure AI", "title": "Introducing Index: Building The World's Largest and Most Diverse Physical Dataset", "url": "https://www.figure.ai/news/introducing-index", "kind": "product", "delta": "+" } ], "indicators": [ { "metric": "Mean autonomous success rate, published multi-finger dexterity suite", "unit": "%", "value": 48.8, "as_of": "2026-07-30", "direction": "up_is_progress", "canon_target": 100, "source": "https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/" }, { "metric": "Manipulation speed relative to a human on precision tasks", "unit": "x human", "value": 0.25, "as_of": "2026-02-10", "direction": "up_is_progress", "canon_target": 1, "source": "https://epoch.ai/blog/where-autonomy-works-evaluating-robot-capabilities-in-2026" }, { "metric": "Tactile sensel density on the best demonstrated robot hand", "unit": "sensels/cm2", "value": 361, "as_of": "2026-01-09", "direction": "up_is_progress", "canon_target": 241, "source": "https://www.prnewswire.com/news-releases/from-fingertips-to-full-body-coverage-ensuring-technology-debuts-groundbreaking-tactile-infrastructure-at-ces-2026-302657225.html" }, { "metric": "Actuated degrees of freedom, best production humanoid hand", "unit": "DoF", "value": 22, "as_of": "2026-05-13", "direction": "up_is_progress", "canon_target": 24, "source": "https://arxiv.org/abs/2605.13925" } ] }, "commentary": "The hands are finished; the hands do not work. Robot fingers now carry 361 tactile sensels per square centimetre against the human fingertip's 241 mechanoreceptors, and 22 actuated degrees of freedom against roughly 24 in a human hand. What arrives with that hardware is a 92 per cent success rate unscrewing a light bulb and 36 per cent screwing one back in, at a quarter of human speed. DeepMind deserves credit for publishing the numbers; most of the field publishes video. The binding constraint is contact data, which cannot be scraped, only accumulated a robot-hour at a time. Figure is now paying the public $15m for phone footage of their own hands.", "last_reviewed": "2026-08-29", "review_cadence_days": 30, "progress": 0.0667 }, { "id": "tool-and-weapon-handling", "name": "Tool and weapon handling", "category": "manipulation", "weight": 6, "one_liner": "Using human tools and guns competently.", "canon": { "requirement": "Pick up an arbitrary, unfamiliar, human-designed tool or weapon and operate it at expert level immediately - including selecting sub-lethal aimpoints under fire when mission constraints change mid-mission.", "quantified": [ { "metric": "weapons identified by model on sight", "value": "12-gauge auto-loader; .45 Longslide with laser sighting; Uzi 9mm; 'phased plasma rifle in the 40-watt range'", "source_ref": "t1-gun-shop", "tier": "PRIMARY" }, { "metric": "human assessment of its competence", "value": "'You know your weapons, buddy.'", "source_ref": "t1-gun-shop", "tier": "PRIMARY" }, { "metric": "weapons operated in T2", "value": "M79 grenade launcher, M134 minigun, Winchester 1887 (one-handed spin-cock on a moving motorcycle), .45, Remington", "source_ref": "t2-full", "tier": "PRIMARY" }, { "metric": "deliberate sub-lethal engagement", "value": "kneecaps every officer at Cyberdyne; ZERO fatalities", "source_ref": "t2-cyberdyne", "tier": "PRIMARY" } ], "sources": [ { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-gun-shop", "evidence": "The Terminator orders an arsenal by model designation.", "quote": "The 12-gauge auto-loader. / The .45 long slide, with laser sighting. / The Uzi 9mm.", "verified": true, "tier": "PRIMARY" }, { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-gun-shop", "evidence": "The clerk's assessment.", "quote": "You know your weapons, buddy.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-pescadero", "evidence": "After the Terminator shoots a guard in the leg rather than killing him.", "quote": "Don't shoot me again. Don't kill me. / He'll live.", "verified": true, "tier": "PRIMARY" } ], "canon_confidence": "explicit", "canonical_target": { "metric": "zero-shot expert operation plus constrained aimpoint selection", "value": "Expert operation of any unfamiliar human tool or weapon on first contact, including correct manual-of-arms; plus deliberate per-target aimpoint selection to a named effect ('he'll live') under return fire in a mass engagement with zero fatalities", "reasoning": "The Cyberdyne sequence is the demanding half and is routinely under-read: having promised John it will not kill, the Terminator engages a large police cordon, disables dozens of officers and kills none. That is not 'can hold a gun' - it is real-time, per-target, constraint-satisfying aimpoint selection at combat tempo." }, "assumed_by_sota_agent": "Immediate, untrained, competent use of arbitrary human tools, vehicles and firearms picked up in the field. Canonical bars: one-handed lever-cocking of a Winchester 1887 while riding a motorcycle; loading and firing a minigun; driving cars, trucks and a helicopter; and using a scalpel and hand tools for self-surgery. Zero-shot, first attempt, at speed, under motion." }, "real": { "status": "in_progress", "trl": 6, "spec_fraction": 0.08, "spec_fraction_rationale": "Against canon's three legs: tools, best published near-zero-shot result is 78.9% on diverse tool kitting with a two-finger gripper and 36% screwing in a bulb with a five-fingered hand, ~0.3 of the canon bar; vehicles, a humanoid drives an unmodified micro-EV at roughly 1/20 human speed on a closed course, ~0.05; firearms, zero demonstrated handling, 0.0. Mean ~0.12, discounted to 0.08 because canon additionally requires all of it zero-shot and under motion.", "gap": "Robots operate human-designed equipment in real factories, but only equipment they were specifically commissioned against. Arbitrary pick-up-and-use is lab-stage. Every fielded armed robot mounts its weapon as a bolted payload on a pan-tilt turret rather than holding it: no published demonstration exists anywhere of a robot loading, chambering, clearing a malfunction in, or reloading a firearm with its hands.", "why_hard": "Human tools encode human hand kinematics and proprioception, so competent use requires a model of the tool's function rather than its shape, and the failure modes that separate competence from mimicry - a jammed round, a stripped screw, a grip slipping under recoil - are exactly the long tail with no training data. Recoil is also an impulsive disturbance to a system whose upright state must be actively maintained.", "movers": [ { "name": "Google DeepMind", "kind": "lab", "country": "US", "what": "Publishes the only current zero-shot-ish tool-use numbers on a common platform: 78.9% diverse tool kitting, 36% screwing in a bulb." }, { "name": "JSK Lab, University of Tokyo", "kind": "university", "country": "JP", "what": "Musashi musculoskeletal humanoid, still the reference for a robot physically driving an unmodified human vehicle." }, { "name": "Karlsruhe Institute of Technology", "kind": "university", "country": "DE", "what": "ARMAR-6 humanoid, long-running work on autonomous use of human power tools in maintenance tasks." }, { "name": "Ghost Robotics / SWORD International", "kind": "company", "country": "US", "what": "Vision 60 with the SPUR rifle: the canonical armed quadruped, and the canonical demonstration that the weapon is carried, not handled." }, { "name": "Figure AI", "kind": "company", "country": "US", "what": "Figure 03 performs sequencing on a live BMW line, including forceful whole-body manipulation of production carts." }, { "name": "Agility Robotics", "kind": "company", "country": "US", "what": "Digit performs machine tending and sortation - operating human-designed industrial equipment - across nine customer facilities." }, { "name": "DeWalt / August Robotics", "kind": "company", "country": "US/HK", "what": "Autonomous concrete-drilling robot claiming 99.97% hole accuracy: the industry's actual answer to tool use, which is to build a purpose-machine instead." } ], "evidence": [ { "date": "2026-07-30", "claim": "Diverse tool kitting scores 78.9% with a two-finger gripper while screwing in a light bulb scores 36% with a five-fingered hand.", "source": "Google DeepMind", "title": "Gemini Robotics 2 brings whole body intelligence to robots", "url": "https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/", "kind": "product", "delta": "+" }, { "date": "2024-06-08", "claim": "A musculoskeletal humanoid drives an unmodified micro-EV - key, handbrake, pedals, steering, traffic lights - but takes roughly two minutes to turn a single corner.", "source": "arXiv", "title": "Toward Autonomous Driving by Musculoskeletal Humanoids: A Study of Developed Hardware and Learning-Based Software", "url": "https://arxiv.org/abs/2406.05573", "kind": "paper", "delta": "0" }, { "date": "2021-10-18", "claim": "On the reference armed quadruped the rifle is a mounted payload the robot does not hold or aim; firing is by a remote human operator, and the manufacturer states the weapon system has no autonomy and no AI.", "source": "IEEE Spectrum", "title": "Q&A: Ghost Robotics CEO on Armed Robots for the U.S. Military", "url": "https://spectrum.ieee.org/ghost-robotics-armed-military-robots", "kind": "product", "delta": "0" }, { "date": "2021-10-14", "claim": "The SWORD Special Purpose Unmanned Rifle is bolted to the Vision 60 quadruped's back as a payload.", "source": "Popular Science", "title": "Ghost Robotics now makes a lethal robot dog", "url": "https://www.popsci.com/technology/ghost-robotics-robot-dog-gun-lethal/", "kind": "product", "delta": "0" }, { "date": "2026-06-30", "claim": "Figure 03 performs sequencing on a live BMW production line, combining precise placement of thin-walled parts with forceful whole-body manipulation of a metal cart.", "source": "Figure AI", "title": "F.03 Arrives at BMW", "url": "https://www.figure.ai/news/f-03-at-bmw", "kind": "product", "delta": "+" }, { "date": "2026-06-24", "claim": "Digit humanoids perform machine tending and sortation for Schaeffler, GXO, Toyota Motor Manufacturing Canada and Mercado Libre, accumulating over 65,000 hours of operation.", "source": "U.S. Securities and Exchange Commission", "title": "Joint press release of Churchill Capital Corp XI and Agility Robotics, Inc. (EX-99.1)", "url": "https://www.sec.gov/Archives/edgar/data/0002074973/000121390026071287/ea029548401ex99-1.htm", "kind": "regulation", "delta": "+" } ], "indicators": [ { "metric": "Near-zero-shot tool-use success on a published benchmark (diverse tool kitting)", "unit": "%", "value": 78.9, "as_of": "2026-07-30", "direction": "up_is_progress", "canon_target": 100, "source": "https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/" }, { "metric": "Humanoid vehicle-operation speed relative to a human, cornering", "unit": "x human", "value": 0.04, "as_of": "2024-06-08", "direction": "up_is_progress", "canon_target": 1, "source": "https://arxiv.org/abs/2406.05573" }, { "metric": "Published demonstrations of a robot loading or reloading a firearm by hand", "unit": "count", "value": 0, "as_of": "2026-08-29", "direction": "up_is_progress", "canon_target": 1, "source": "https://spectrum.ieee.org/ghost-robotics-armed-military-robots" } ] }, "commentary": "Every armed robot in service is a gun carriage. The rifle bolts to the back; the machine neither holds it, cocks it, clears it nor reloads it, and a human pulls the trigger from a distance. Not one published demonstration exists of a robot loading a firearm with its fingers. On tools proper, the best near-zero-shot figure is 79 per cent, achieved with a two-finger gripper, because the five-fingered hand does worse. The most convincing humanoid driver remains a Tokyo university machine that took two minutes to turn a corner. Canon requires lever-cocking a Winchester one-handed, at speed, on a moving motorcycle.", "last_reviewed": "2026-08-29", "review_cadence_days": 30, "progress": 0.0533 }, { "id": "machine-vision-and-recognition", "name": "Machine vision and recognition", "category": "perception", "weight": 9, "one_liner": "Seeing a face in a crowd and knowing whose it is.", "canon": { "requirement": "Recognise one named individual from an incomplete prior, at range, in darkness and in crowds, with zero false negatives across the whole mission; and separately produce full 3-D structural reconstructions of unfamiliar machinery in seconds.", "quantified": [ { "metric": "motorcycle 3-D reconstruction time", "value": "side, top and plan views 'in less than four seconds'", "source_ref": "t1-script-219fx", "tier": "TERTIARY" }, { "metric": "anthropometry from a single glance", "value": "'thousands of estimated measurements'; clothing analysed and fit assessed", "source_ref": "t2-script-8a", "tier": "TERTIARY" }, { "metric": "HUD data rate", "value": "'changes more rapidly than any human eye could follow'", "source_ref": "t1-script-106fx", "tier": "TERTIARY" }, { "metric": "recognition failures across two films", "value": "zero", "source_ref": "t1-full", "tier": "PRIMARY" } ], "sources": [ { "work": "The Terminator - fourth draft, 20 April 1983", "year": 1983, "medium": "screenplay-draft", "ref": "t1-script-219fx", "evidence": "Terminator POV scanning a motorcycle in the motel parking lot.", "quote": "The image reduces to GRAPHIC OUTLINES, with separate systems COLOR-CODED. It breaks down suddenly into individual SIDE, TOP and PLAN VIEWS. All in less than four seconds.", "verified": true, "tier": "TERTIARY" }, { "work": "Terminator 2: Judgment Day - revised final shooting script", "year": 1991, "medium": "shooting-script", "ref": "t2-script-8a", "evidence": "Terminator POV in the truck-stop diner, selecting a victim for his clothes.", "quote": "His body is outlined, or 'selected', and thousands of estimated measurements appear. His clothing has been analyzed and deemed suitable...", "verified": true, "tier": "TERTIARY" } ], "canon_confidence": "implied", "canonical_target": { "metric": "named-individual recognition plus structural reconstruction rate", "value": "Real-time identification of a specific named individual from a degraded prior, at range, in darkness, in a crowd, ZERO false negatives over a multi-day mission; plus dense 3-D structural reconstruction and full anthropometry of an unfamiliar object in <4 s from a single passing viewpoint", "reasoning": "The <4 s figure is the only stopwatch number canon ever attaches to perception, so it is the natural target even though it is TERTIARY stage direction and must be labelled as such. The zero-false-negative clause matters more: the mission model has no room for a second look." }, "assumed_by_sota_agent": "Identify one named individual from limited prior information — a photograph, a name — in an uncontrolled environment, among strangers, in variable light, at speed, and act on that identification with no human in the loop; while simultaneously parsing the rest of the scene (people, weapons, vehicles, exits) as a continuous annotated overlay." }, "real": { "status": "in_progress", "trl": 9, "spec_fraction": 0.55, "spec_fraction_rationale": "Decompose canon into four capabilities and average. (a) Identify a named person from a photo in semi-cooperative imagery: best FNIR 0.1–0.6% on a 12M gallery ⇒ 0.95. (b) The same in genuinely uncontrolled capture: NIST reports 4.9% FNIR on ATM-style kiosk imagery and 2.4–13.4% on profile views, and says error rates there are 'often in excess of 20%' ⇒ 0.50. (c) General open-vocabulary scene parsing: SAM 3 reaches 74% of estimated human cgF1 ⇒ 0.74. (d) Robustness to physical adversarial patches, heavy occlusion and profile-only views ⇒ 0.20. Mean = 0.60, discounted to 0.55 because canon requires (b) and (d) continuously and simultaneously, not on average.", "gap": "Accuracy collapses with capture quality. NIST's best algorithms miss 0.1% of searches on attended mugshots but 4.9% on ATM-kiosk photos and up to 13.4% on profile views; at the false-positive rate an autonomous system would need (1 in 1,000), the best algorithms miss 3–5% of mated searches even on high-quality probes. Gait falls from 92% Rank-1 in the lab to 68.6% in the wild with distractors. Physical adversarial patches still defeat pedestrian detectors.", "why_hard": "The binding constraint is capture, not computation: identification error is dominated by pose, illumination, resolution and occlusion at the sensor, and no amount of model scale recovers information the camera never got. The operating point autonomous action requires — false positives below 1 in 1,000 — sits on the steep part of NIST's error-tradeoff curve, where genuine low-quality mates score below high-scoring doppelgängers.", "movers": [ { "name": "NIST Information Technology Laboratory", "kind": "agency", "country": "US", "what": "Runs FRTE, the only continuously updated independent evaluation of face identification against sequestered galleries up to 12 million identities." }, { "name": "Idemia", "kind": "company", "country": "FR", "what": "Ranked first across every FRTE gallery size from 640k to 12M identities in the August 2026 report; 1.11% FNIR at 12M and FPIR 0.1%." }, { "name": "NEC", "kind": "company", "country": "JP", "what": "Top or joint-top on mugshot, webcam, border and kiosk FRTE splits; 0.06% error on the 12-million-person still-image test." }, { "name": "Meta AI (FAIR)", "kind": "lab", "country": "US", "what": "SAM 3 promptable concept segmentation — open-vocabulary detection, segmentation and tracking from a noun phrase, at 74% of human accuracy." }, { "name": "Sensetime", "kind": "company", "country": "CN", "what": "Joint best at 0.1% FNIR on FRTE mugshot identification; among the two most accurate algorithms in NIST's error-tradeoff analysis." }, { "name": "Huazhong University of Science and Technology", "kind": "university", "country": "CN", "what": "TriPatch — transferable physical-world adversarial patches that defeat pedestrian detectors across the whole detection pipeline." } ], "evidence": [ { "date": "2026-08-04", "claim": "The most accurate algorithms find a matching entry in a 12-million-identity gallery with rank-one miss rates approaching 0.1%, from a single photograph, searching on one CPU core.", "source": "NIST", "title": "Face Recognition Technology Evaluation (FRTE) Part 2: Identification — NISTIR 8271 draft supplement", "url": "https://pages.nist.gov/frvt/reports/1N/frvt_1N_report.pdf", "kind": "benchmark", "delta": "+" }, { "date": "2026-08-04", "claim": "On ATM-style kiosk photos not intended for face recognition, NIST reports error rates 'often in excess of 20%' even for the more accurate algorithms; and at a false-positive identification rate of 1 in 1,000 the best algorithms fail on 3–5% of mated searches.", "source": "NIST", "title": "Face Recognition Technology Evaluation (FRTE) Part 2: Identification — NISTIR 8271 draft supplement", "url": "https://pages.nist.gov/frvt/reports/1N/frvt_1N_report.pdf", "kind": "benchmark", "delta": "-" }, { "date": "2026-08-10", "claim": "Best FNIR by probe condition in the August 2026 FRTE release: mugshot 0.1%, desktop webcam 0.6%, border-to-border 0.4%, kiosk 4.9%, profile 2.4–13.4%.", "source": "Biometric Update", "title": "NIST 1:N results show face recognition accuracy race is tightening", "url": "https://www.biometricupdate.com/202608/nist-1n-results-show-face-recognition-accuracy-race-is-tightening", "kind": "benchmark", "delta": "0" }, { "date": "2025-11-20", "claim": "SAM 3 achieves more than double the cgF1 of the strongest open-vocabulary baseline on SA-Co/Gold and 74% of estimated human performance, running in 30 ms per image with 100+ detected objects on an H200.", "source": "arXiv", "title": "SAM 3: Segment Anything with Concepts", "url": "https://arxiv.org/abs/2511.16719", "kind": "paper", "delta": "+" }, { "date": "2024-01-11", "claim": "SPOSGait Rank-1 accuracy is 92.12% on CASIA-B and 92.26% on OU-MVLP (controlled) but 74.80% on GREW, 68.56% on GREW with a 233k-sequence distractor set, and 64.94% on Gait3D.", "source": "arXiv / IEEE TPAMI", "title": "Gait Recognition in the Wild: A Large-scale Benchmark and NAS-based Baseline", "url": "https://arxiv.org/abs/2205.02692", "kind": "benchmark", "delta": "0" }, { "date": "2025-05-05", "claim": "Best reported cross-camera person re-identification on MSMT17 (15 cameras, 4,101 identities) is Rank@1 83.8% and mAP 67.6%; on Occluded-DukeMTMC, Rank@1 68.9%.", "source": "arXiv", "title": "Enhancing person re-identification via Uncertainty Feature Fusion Method and Auto-weighted Measure Combination", "url": "https://arxiv.org/abs/2405.01101", "kind": "benchmark", "delta": "0" }, { "date": "2026-04-24", "claim": "TriPatch demonstrates transferable physical-world adversarial patches that suppress detection confidence, amplify bounding-box offsets and disrupt non-maximum suppression, defeating multiple pedestrian detectors under varied physical conditions.", "source": "arXiv", "title": "Transferable Physical-World Adversarial Patches Against Pedestrian Detection Models", "url": "https://arxiv.org/abs/2604.22552", "kind": "paper", "delta": "-" } ], "indicators": [ { "metric": "Best FNIR, 12M-identity gallery, mugshot probes, FPIR 0.003", "unit": "%", "value": 0.06, "as_of": "2026-08-04", "direction": "down_is_progress", "canon_target": 0, "source": "https://pages.nist.gov/frvt/reports/1N/frvt_1N_report.pdf" }, { "metric": "Best FNIR, ATM-style kiosk probes (uncontrolled capture)", "unit": "%", "value": 4.9, "as_of": "2026-08-04", "direction": "down_is_progress", "canon_target": 0.1, "source": "https://www.biometricupdate.com/202608/nist-1n-results-show-face-recognition-accuracy-race-is-tightening" }, { "metric": "Open-vocabulary segmentation accuracy as a fraction of human (SAM 3, SA-Co/Gold cgF1)", "unit": "% of human", "value": 74, "as_of": "2025-11-20", "direction": "up_is_progress", "canon_target": 100, "source": "https://arxiv.org/abs/2511.16719" }, { "metric": "Gait recognition Rank-1 in the wild, GREW test plus 233k distractors", "unit": "%", "value": 68.56, "as_of": "2024-01-11", "direction": "up_is_progress", "canon_target": 99, "source": "https://arxiv.org/abs/2205.02692" } ] }, "commentary": "Face recognition is the one place where the films undersold reality. NIST's August 2026 evaluation has algorithms finding a named individual in a gallery of twelve million at a rank-one miss rate near 0.1 percent, from one photograph, on a single CPU core. Nothing in 1984 imagined that. The caveat is that the twelve million are mugshots. Move to an ATM-style kiosk camera and the same algorithms miss one search in twenty; move to profile and the spread runs from 2.4 to 13.4 percent. The machine can find Sarah Connor. It needs her to look at the lens.", "last_reviewed": "2026-08-29", "review_cadence_days": 30, "progress": 0.55 }, { "id": "multispectral-and-thermal-sensing", "name": "Multispectral and thermal sensing", "category": "perception", "weight": 5, "one_liner": "The red HUD, and how close the real thing already is.", "canon": { "requirement": "Simultaneous, registered visible + thermal + image-intensified sensing fused into one scene representation, retaining full acuity in total darkness and through smoke and tear gas.", "quantified": [ { "metric": "thermal channel", "value": "'the fleeing figures ahead are more luminous than the background, suggesting infra-red'", "source_ref": "t1-script-106fx", "tier": "TERTIARY" }, { "metric": "low-light channel", "value": "'image-intensified, bright and stark as a lunar landscape'", "source_ref": "t1-script-214fx", "tier": "TERTIARY" }, { "metric": "obscurant performance", "value": "walks through CS gas and smoke and engages accurately", "source_ref": "t2-cyberdyne", "tier": "PRIMARY" }, { "metric": "fusion", "value": "all channels presented as one overlaid scene with alphanumerics", "source_ref": "t1-script-106fx", "tier": "TERTIARY" } ], "sources": [ { "work": "The Terminator - fourth draft, 20 April 1983", "year": 1983, "medium": "screenplay-draft", "ref": "t1-script-106fx", "evidence": "Terminator POV in the alley pursuing Sarah and Reese.", "quote": "the fleeing figures ahead are more luminous than the background, suggesting infra-red. The margins of the FRAME are crammed with columns of CRT-type characters: columns of numbers and acronyms.", "verified": true, "tier": "TERTIARY" }, { "work": "Terminator 2: Judgment Day - revised final shooting script", "year": 1991, "medium": "shooting-script", "ref": "t2-script-173e", "evidence": "Emerging from SWAT tear gas at Cyberdyne.", "quote": "Terminator emerges from the smoke. Not even misty-eyed.", "verified": true, "tier": "TERTIARY" } ], "canon_confidence": "implied", "canonical_target": { "metric": "registered multispectral fusion under zero ambient light and obscurants", "value": "Real-time registered fusion of visible, LWIR and image-intensified channels into a single scene model at full human visual acuity, in zero ambient light and through CS smoke, with no degradation in target track", "reasoning": "The red HUD is one of cinema's most-copied images and no character ever explains it, so the requirement is assembled from stage directions plus the on-screen tear-gas engagement. The 'no degradation in target track' clause is what makes it measurable." }, "assumed_by_sota_agent": "See beyond the visible — thermal and low-light imaging fused into a single annotated overlay carrying target data and ranging — in a self-contained head-sized package running on internal power, continuously." }, "real": { "status": "in_progress", "trl": 9, "spec_fraction": 0.45, "spec_fraction_rationale": "Canon implies roughly four sensing channels fused into one head-mounted annotated overlay: visible, LWIR, low-light/NIR, plus ranging and target data. Fielded kit (ENVG-B) delivers two of the four at full quality — Gen 3 white-phosphor image intensification fused with a 640x480 10-micron thermal sensor, on a 1280x1024 display, with an AR overlay and a wireless cue from the rifle sight — so 2/4 = 0.50, times a resolution-and-range factor of ~0.9 (>500 m detection, VGA-class thermal, versus canon's implied full-scene analysis at range) = 0.45. SWIR, event and hyperspectral channels exist commercially but are not in the same fielded package.", "gap": "The fielded overlay is a display for a human, not an annotation produced by the machine: ENVG-B draws the fused scene and cues the weapon sight, but a person does the identifying. Band count is two, not four; thermal resolution is VGA-class; and LWIR through smoke, rain and glass still degrades badly.", "why_hard": "LWIR optics do not shrink. Germanium and chalcogenide lenses and the diffraction limit at 8–14 microns set the aperture, so range and resolution are bounded by glass and cost rather than by the detector — which is why the industry's progress metric is pixel pitch (12 µm to 8.5 µm) rather than sensitivity. Every additional band needs its own aperture, detector and athermal path, making fusion a volume-and-money problem rather than a physics one.", "movers": [ { "name": "Teledyne FLIR", "kind": "company", "country": "US", "what": "Boson+ uncooled LWIR core: 640x512 at 12 µm, NETD ≤20 mK, 7.5 g, under 4.9 cm³, from 500 mW." }, { "name": "Lynred", "kind": "company", "country": "FR", "what": "YOCTO1024, announced January 2026: 8.5 µm pitch XGA microbolometer in an 18x18x4.85 mm package, claimed up to 40% better DRI range than 12 µm parts." }, { "name": "Elbit Systems of America", "kind": "company", "country": "US", "what": "Sole producer of ENVG-B fused night vision goggles for the US Army under a $212M delivery order taken in May 2026." }, { "name": "L3Harris", "kind": "company", "country": "US", "what": "Original ENVG-B producer; dual-waveband fused white-phosphor and thermal goggle with augmented reality and wireless rapid target acquisition." }, { "name": "Sony Semiconductor Solutions", "kind": "company", "country": "JP", "what": "SenSWIR IMX992 — 5.32 MP InGaAs SWIR sensor with a 3.45 µm pixel, sensitive 0.4–1.7 µm, in commercial cameras." }, { "name": "Prophesee", "kind": "company", "country": "FR", "what": "GenX320 event sensor: 320x320, >140 dB dynamic range, sub-150 µs latency, under 50 mW." } ], "evidence": [ { "date": "2026-05-12", "claim": "Elbit Systems of America received a $212 million delivery order to be sole producer of ENVG-B fused image-intensifier and thermal goggles for the US Army, with production running through 2028.", "source": "Army Recognition", "title": "U.S. Army Expands Sensor-Fused Night Warfare Capabilities with $212 Million ENVG-B Night Vision Goggles Order", "url": "https://www.armyrecognition.com/news/army-news/2026/u-s-army-expands-sensor-fused-night-warfare-capabilities-with-212-million-envg-b-night-vision-goggles-order", "kind": "product", "delta": "+" }, { "date": "2026-01-19", "claim": "Lynred announced YOCTO1024, an 8.5-micron-pitch 1024x768 uncooled microbolometer in an 18x18x4.85 mm package, claiming up to 40% better detection-recognition-identification range than 12-micron sensors.", "source": "Laser Focus World", "title": "LYNRED unveils YOCTO, a new ultra-compact 8µm microbolometer", "url": "https://www.laserfocusworld.com/directory/detectors-and-imaging/imaging-systems/press-release/55352105/lynred-usa-lynred-unveils-yocto-a-new-ultra-compact-8m-microbolometer-pushing-uncooled-thermal-imaging-to-a-new-level", "kind": "product", "delta": "+" }, { "date": "2026-08-29", "claim": "Teledyne FLIR's Boson+ uncooled LWIR module specifies NETD of ≤20 mK at 640x512 and 12 µm pitch, weighing 7.5 g in under 4.9 cm³ and drawing from 500 mW.", "source": "Teledyne FLIR OEM", "title": "Boson+ high performance uncooled LWIR OEM thermal camera module", "url": "https://oem.flir.com/products/boson-plus", "kind": "product", "delta": "0" }, { "date": "2026-08-29", "claim": "A 640x512, 12-micron radiometric LWIR core is listed at $4,334 in consumer grade, indicating the current price floor for VGA-class fielded thermal imaging.", "source": "OEM Cameras", "title": "Teledyne FLIR BOSON 640x512 14mm 32° HFoV radiometric LWIR thermal core", "url": "https://www.oemcameras.com/products/20640a032-htm", "kind": "product", "delta": "0" }, { "date": "2026-08-29", "claim": "Prophesee's GenX320 event sensor delivers over 140 dB of dynamic range with event latency under 150 microseconds and power under 50 mW, in a 320x320 array.", "source": "Prophesee", "title": "Event-Based Metavision Sensor GenX320", "url": "https://www.prophesee.ai/event-based-sensor-genx320/", "kind": "product", "delta": "+" }, { "date": "2026-08-29", "claim": "Sony's IMX992 SenSWIR sensor provides 5.32 effective megapixels with a 3.45-micron InGaAs pixel across 0.4–1.7 microns, the highest-resolution and smallest-pixel commercial SWIR imager.", "source": "Sony Semiconductor Solutions", "title": "High-resolution, high-performance SWIR image sensor IMX992/IMX993", "url": "https://www.sony-semicon.com/en/products/is/industry/swir/imx992-993.html", "kind": "product", "delta": "+" } ], "indicators": [ { "metric": "Best commercial uncooled LWIR NETD", "unit": "mK", "value": 20, "as_of": "2026-08-29", "direction": "down_is_progress", "canon_target": 10, "source": "https://oem.flir.com/products/boson-plus" }, { "metric": "Smallest announced uncooled microbolometer pixel pitch", "unit": "µm", "value": 8.5, "as_of": "2026-01-19", "direction": "down_is_progress", "canon_target": 5, "source": "https://www.laserfocusworld.com/directory/detectors-and-imaging/imaging-systems/press-release/55352105/lynred-usa-lynred-unveils-yocto-a-new-ultra-compact-8m-microbolometer-pushing-uncooled-thermal-imaging-to-a-new-level" }, { "metric": "Mass of a fielded 640x512 uncooled LWIR camera module", "unit": "g", "value": 7.5, "as_of": "2026-08-29", "direction": "down_is_progress", "canon_target": 5, "source": "https://oem.flir.com/products/boson-plus" }, { "metric": "Number of wavebands fused in a fielded head-mounted system with an AR overlay", "unit": "bands", "value": 2, "as_of": "2026-05-12", "direction": "up_is_progress", "canon_target": 4, "source": "https://www.l3harris.com/all-capabilities/enhanced-night-vision-goggle-binocular-envg-b" } ] }, "commentary": "The red overlay is the least fantastical thing in the films. A Teledyne FLIR Boson+ resolves 20 millikelvin at 7.5 grams and half a watt; Lynred put an 8.5-micron XGA microbolometer into an 18-millimetre package in January. The US Army has been buying fused thermal-and-intensifier goggles with an augmented-reality overlay and a wireless link to the rifle sight since 2019, and ordered another 212 million dollars' worth in May. What the soldier does not get is the annotation: the hardware draws the scene, a human still reads it. Canon's HUD was never really about the sensor.", "last_reviewed": "2026-08-29", "review_cadence_days": 45, "progress": 0.45 }, { "id": "target-acquisition-and-tracking", "name": "Target acquisition and tracking", "category": "perception", "weight": 7, "one_liner": "Locking on is easy. Staying locked is the whole problem.", "canon": { "requirement": "Acquire and hold track on a designated human in clutter while both shooter and target move at vehicle speeds, and place rounds on selected sub-limb aimpoints from an unstabilised platform.", "quantified": [ { "metric": "firing platform", "value": "one-handed, from a moving car and a moving motorcycle, at another moving vehicle", "source_ref": "t1-chase", "tier": "PRIMARY" }, { "metric": "aimpoint discrimination", "value": "limb-selective across dozens of targets, zero fatalities", "source_ref": "t2-cyberdyne", "tier": "PRIMARY" }, { "metric": "sighting aid named", "value": "'.45 long slide, with laser sighting' - canon's only named optic", "source_ref": "t1-gun-shop", "tier": "PRIMARY" }, { "metric": "closing speed of the engagement", "value": "up to ~129 km/h ('It hits eighty')", "source_ref": "t2-script-47h", "tier": "TERTIARY" }, { "metric": "track continuity", "value": "never loses a designated target once acquired", "source_ref": "t1-full", "tier": "PRIMARY" } ], "sources": [ { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-gun-shop", "evidence": "The only sighting system canon ever names.", "quote": "The .45 long slide, with laser sighting.", "verified": true, "tier": "PRIMARY" } ], "canon_confidence": "implied", "canonical_target": { "metric": "track continuity and aimpoint selectivity at vehicle speeds", "value": "Continuous track and accurate, aimpoint-selective fire on a designated human while both platforms move at 50-130 km/h, from an unstabilised one-handed mount, in darkness and clutter, with no track loss", "reasoning": "The upper speed comes from the canal chase ('It hits eighty' = ~129 km/h). The aimpoint-selective clause separates this from ordinary fire control and comes from the same Cyberdyne evidence as tool-and-weapon-handling." }, "assumed_by_sota_agent": "Designate a target and hold the lock through crowds, occlusion, darkness and vehicle pursuit, re-acquiring after minutes of total loss of sight, with the lock and its state displayed continuously — effectively 100% identity retention across an engagement lasting hours." }, "real": { "status": "in_progress", "trl": 7, "spec_fraction": 0.35, "spec_fraction_rationale": "Best published identity retention through occlusion and disappearance is 69.82 J&F (VOS-Agent on MOSEv2, August 2026) and 67.9 HOTA on visually ambiguous humans (MeMoSORT on DanceTrack). Canon requires roughly 1.0 sustained across an engagement of hours with re-acquisition after arbitrary loss, whereas benchmark clips run seconds to minutes; applying a 0.5 discount for horizon length and for the multi-target real-time constraint gives 0.70 x 0.5 = 0.35.", "gap": "Re-acquisition after prolonged absence. MOSEv2 was built specifically around disappearance and reappearance and SAM 2 drops from 76.4% to 50.9% on it. Open-vocabulary video tracking sustains near-real-time performance for only about five concurrent objects on an H200. No public evidence exists of any system holding identity on one designated human through a crowd for the duration canon requires.", "why_hard": "The bottleneck is association, not detection. Once a target leaves the frame, re-acquisition collapses to appearance-based re-identification against an open world, and appearance re-ID reaches only mAP 67.6 on a 15-camera, 4,101-identity closed set. Association errors compound multiplicatively with time, so identity retention decays with engagement length — the exact inverse of the canon property.", "movers": [ { "name": "Fudan University / Nanyang Technological University (MOSEv2 team)", "kind": "university", "country": "CN", "what": "Built MOSEv2, the benchmark that exposes how badly video segmentation degrades under disappearance, reappearance and occlusion." }, { "name": "Meta AI (FAIR)", "kind": "lab", "country": "US", "what": "SAM 3's tracker: near-real-time video for roughly five concurrent objects on an H200, and the base model most 2026 challenge entries build on." }, { "name": "US Army DEVCOM C5ISR Center", "kind": "agency", "country": "US", "what": "Ran 'Warden', its first counter-small-UAS assessment, at Fort A.P. Hill in May 2026 with 17 companies across lasers, jammers and interceptors." }, { "name": "Defense Innovation Unit", "kind": "agency", "country": "US", "what": "February 2026 commercial solicitation for counter-UAS sensing requiring autonomous threat classification and detection of Group 1 sUAS at 2 km or more." }, { "name": "Harbin Institute of Technology (VOS-Agent team)", "kind": "university", "country": "CN", "what": "Won the 8th LSVOS MOSEv2 track at 69.82 J&F using a multi-agent stack pairing a visual tracker with an MLLM doing description-guided re-localisation." } ], "evidence": [ { "date": "2025-08-07", "claim": "On MOSEv2 — 5,024 videos and 701,976 masks built around disappearance, reappearance, severe occlusion, small objects, adverse weather and low light — SAM 2 falls from 76.4% on MOSEv1 to 50.9%.", "source": "arXiv", "title": "MOSEv2: A More Challenging Dataset for Video Object Segmentation in Complex Scenes", "url": "https://arxiv.org/abs/2508.05630", "kind": "benchmark", "delta": "-" }, { "date": "2026-08-13", "claim": "VOS-Agent won the MOSEv2 track of the 8th LSVOS Challenge at ECCV 2026 with 69.82% J&F, using a multi-agent framework combining a visual tracking agent for tiny targets with an MLLM-based semantic agent for description-guided re-localisation.", "source": "arXiv", "title": "VOS-Agent: The 1st Place Solution for the 8th LSVOS Challenge (MOSEv2 Track)", "url": "https://arxiv.org/abs/2608.12721", "kind": "benchmark", "delta": "+" }, { "date": "2025-08-13", "claim": "The best reported multi-person tracker on DanceTrack, where detection is easy but appearance is near-identical and motion non-linear, reaches 67.9% HOTA.", "source": "arXiv", "title": "MeMoSORT: Memory-Assisted Filtering and Motion-Adaptive Association Metric for Multi-Person Tracking", "url": "https://arxiv.org/abs/2508.09796", "kind": "benchmark", "delta": "0" }, { "date": "2025-11-20", "claim": "SAM 3's video tracking latency scales with object count, sustaining near real-time performance for only about five concurrent objects on an H200 datacentre GPU.", "source": "arXiv", "title": "SAM 3: Segment Anything with Concepts", "url": "https://arxiv.org/abs/2511.16719", "kind": "paper", "delta": "0" }, { "date": "2026-07-28", "claim": "US Army DEVCOM C5ISR Center ran the 'Warden' counter-small-UAS assessment at Fort A.P. Hill over roughly a week and a half in May 2026 with 17 companies, against reconnaissance, direct-attack and coordinated swarm profiles; reported gains in target recognition and false-detection reduction are qualitative only.", "source": "Army Recognition", "title": "U.S. Army Accelerates Counter-Drone Development Against Modern Aerial Threats with Industry Partnership", "url": "https://www.armyrecognition.com/news/army-news/2026/u-s-army-accelerates-counter-drone-development-against-modern-aerial-threats-with-industry-partnership", "kind": "demo", "delta": "+" }, { "date": "2026-02-13", "claim": "DIU's Counter-UAS Sensing solicitation requires systems that autonomously classify threats, discriminate biological and ground clutter, maintain real-time multi-target tracking, and detect Group 1 small UAS at 2 km or greater.", "source": "DroneLife", "title": "DIU Launches New Counter-Drone Sensing Initiative for Homeland and Mobile Defense", "url": "https://dronelife.com/2026/02/13/diu-launches-new-counter-drone-sensing-initiative-for-homeland-and-mobile-defense/", "kind": "regulation", "delta": "+" } ], "indicators": [ { "metric": "Best J&F on MOSEv2 (occlusion, disappearance and reappearance)", "unit": "%", "value": 69.82, "as_of": "2026-08-13", "direction": "up_is_progress", "canon_target": 99, "source": "https://arxiv.org/abs/2608.12721" }, { "metric": "Best HOTA on DanceTrack (visually identical humans, non-linear motion)", "unit": "%", "value": 67.9, "as_of": "2025-08-13", "direction": "up_is_progress", "canon_target": 99, "source": "https://arxiv.org/abs/2508.09796" }, { "metric": "Concurrent objects tracked near-real-time by an open-vocabulary tracker on an H200", "unit": "objects", "value": 5, "as_of": "2025-11-20", "direction": "up_is_progress", "canon_target": 50, "source": "https://arxiv.org/abs/2511.16719" } ] }, "commentary": "Locking on is easy; staying locked is the entire problem. On MOSEv2, a benchmark built around objects that vanish and return, SAM 2 falls from 76 to 51 percent; the 2026 challenge winner reaches 70. On DanceTrack, where everyone looks alike, the best tracker manages 68 HOTA. Meanwhile SAM 3 sustains near-real-time video on an H200 for about five simultaneous objects — a 700-watt datacentre part, tracking a small dinner party. Fielded fire control does better against drones than these numbers suggest, but that performance is classified, and an index cannot bank what it cannot read.", "last_reviewed": "2026-08-29", "review_cadence_days": 30, "progress": 0.2722 }, { "id": "spatial-mapping-and-navigation", "name": "Spatial mapping and navigation", "category": "perception", "weight": 8, "one_liner": "Walking into a building it has never seen, in the dark, and not getting lost.", "canon": { "requirement": "Arrive naked in an unfamiliar city with no map, no network, no GPS and no prior, and immediately navigate it - indoors and out - well enough to find a gun shop, three private addresses, a nightclub, a police station, a mountain cabin, a mall, a mental hospital, an office building, a storm-drain system and a steel mill, across several days and multiple vehicle transitions.", "quantified": [ { "metric": "prior map data", "value": "none", "source_ref": "t1-arrival", "tier": "PRIMARY" }, { "metric": "localisation aids used", "value": "street signs, a telephone directory, a police radio, a police mobile data terminal", "source_ref": "t1-phonebook", "tier": "PRIMARY" }, { "metric": "time from arrival to first target address", "value": "< ~24 h", "source_ref": "t1-timeline", "tier": "PRIMARY" }, { "metric": "what it actually asks a bystander for", "value": "the DATE, not directions", "source_ref": "t1-arrival", "tier": "PRIMARY" } ], "sources": [ { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-arrival", "evidence": "The Terminator's first words to a human after arrival. It needs the date; it never asks where it is.", "quote": "What day is it? The date! / The 12th of May. Thursday. / What year?", "verified": true, "tier": "PRIMARY" } ], "canon_confidence": "implied", "canonical_target": { "metric": "zero-prior localisation and route planning", "value": "Zero-prior localisation, mapping and route planning across an unmapped metropolitan area, indoors and out, sustained >=72 h across multiple vehicle transitions, using only on-board sensing plus opportunistically acquired human artefacts - no GNSS, no network, no pre-loaded map", "reasoning": "The 'What day is it?' line is the load-bearing evidence and is easy to miss: the machine asks for the DATE and never once asks for directions. Canon is telling us the navigation problem is considered solved on board and the calendar problem is not." }, "assumed_by_sota_agent": "Enter an unfamiliar building with no prior map and no GPS, navigate it purposefully in the dark while pursuing, locate a specific room or person, and also drive vehicles on public roads — indefinitely, without human intervention." }, "real": { "status": "in_progress", "trl": 8, "spec_fraction": 0.3, "spec_fraction_rationale": "Fielded reliability inside a mapped, bounded operating domain is about 0.99 — Starship advertises its sidewalk fleet as '99% autonomous' across 14 million miles, and Waymo has driven 220.6 million rider-only miles on HD-mapped roads. Canon's defining case removes both the prior map and the remote operator, and nothing fielded does that; demo-grade zero-prior indoor navigation does not publish real-hardware success rates. Crediting the no-prior-map case at roughly 0.30 of the mapped case gives 0.99 x 0.30 ≈ 0.30.", "gap": "Every system at scale depends on a prior survey and a geofence, and on a human for the residual percent. Long-term relocalisation across appearance change — different lighting, seasons, rearranged rooms — still defeats standard feature descriptors, and object-goal navigation results in previously unseen buildings are overwhelmingly reported in simulation rather than on hardware.", "why_hard": "The economics reinforce the technical gap: surveying a city and staffing remote assistance is far cheaper than solving zero-prior navigation, so nobody has to. The technical residue is not mapping but re-cognition — identifying a place again after it has changed, without GPS — which is an open-world perception problem wearing a geometry problem's clothes.", "movers": [ { "name": "Waymo", "kind": "company", "country": "US", "what": "220.6 million rider-only autonomous miles through March 2026 across five metros, with 94% fewer serious-injury crashes than the human benchmark — on HD-mapped, geofenced roads." }, { "name": "Starship Technologies", "kind": "company", "country": "EE/US", "what": "Over 10 million autonomous sidewalk deliveries and 14 million miles across 300+ sites, self-described as Level 4 and '99% autonomous'." }, { "name": "Amazon Robotics", "kind": "company", "country": "US", "what": "Operates the largest indoor mobile-robot fleet in the world, coordinated centrally in structured, fully instrumented warehouses." }, { "name": "Boston Dynamics", "kind": "company", "country": "US", "what": "Spot performs autonomous repeat-mission industrial inspection, which requires a taught route rather than novel-building exploration." }, { "name": "SuperMap authors (spatio-temporal SLAM)", "kind": "university", "country": "CN", "what": "August 2026 4D mapping framework addressing identity drift and stale semantics when open-vocabulary perception is fused into long-lived maps." } ], "evidence": [ { "date": "2026-03-31", "claim": "Waymo reports 220.6 million rider-only autonomous miles through March 2026 across Phoenix, San Francisco, Los Angeles, Austin and Atlanta, with 94% fewer serious-injury-or-worse crashes and 93% fewer injury-causing pedestrian crashes than the human benchmark.", "source": "Waymo", "title": "Waymo Safety Impact", "url": "https://waymo.com/safety/impact/", "kind": "product", "delta": "+" }, { "date": "2026-04-30", "claim": "Starship Technologies has completed over 10 million autonomous deliveries across 14 million miles and 300+ cities and industrial sites, describing its robots as Level 4 and '99% autonomous' — implying human intervention on roughly one percent of operation.", "source": "Starship Technologies", "title": "Starship Technologies — Company", "url": "https://www.starship.xyz/company/", "kind": "product", "delta": "+" }, { "date": "2026-08-24", "claim": "SuperMap reports that naively integrating foundation-model perception into mapping pipelines causes identity drift and stale semantics over time, and proposes a 4D spatio-temporal map to reconcile open-vocabulary perception with long-term environmental change.", "source": "arXiv", "title": "SuperMap: A Spatio-Temporal SLAM System for Visual-Language Navigation", "url": "https://arxiv.org/abs/2608.22896", "kind": "paper", "delta": "0" }, { "date": "2025-11-03", "claim": "Global relocalisation without a prior pose estimate — the kidnapped robot problem — is still treated as an open research problem requiring multi-hypothesis inference from a single LiDAR scan against a known map.", "source": "arXiv", "title": "Tackling the Kidnapped Robot Problem via Sparse Feasible Hypothesis Sampling and Reliable Batched Multi-Stage Inference", "url": "https://arxiv.org/abs/2511.01219", "kind": "paper", "delta": "0" }, { "date": "2024-03-28", "claim": "Feature descriptors used for SLAM relocalisation cannot maintain consistency across diurnal appearance change, making long-term map reuse unreliable in thermal and other visually degraded settings.", "source": "arXiv", "title": "Towards Long Term SLAM on Thermal Imagery", "url": "https://arxiv.org/abs/2403.19885", "kind": "paper", "delta": "-" } ], "indicators": [ { "metric": "Cumulative rider-only autonomous road miles (Waymo)", "unit": "million miles", "value": 220.6, "as_of": "2026-03-31", "direction": "up_is_progress", "canon_target": 1000, "source": "https://waymo.com/safety/impact/" }, { "metric": "Autonomy rate of the largest deployed sidewalk robot fleet", "unit": "%", "value": 99, "as_of": "2026-04-30", "direction": "up_is_progress", "canon_target": 100, "source": "https://www.starship.xyz/company/" }, { "metric": "Fielded systems navigating a previously unsurveyed building with no prior map", "unit": "count", "value": 0, "as_of": "2026-08-29", "direction": "up_is_progress", "canon_target": 1, "source": "https://arxiv.org/abs/2608.22896" } ] }, "commentary": "Waymo has driven 220.6 million rider-only miles and cut serious-injury crashes by 94 percent; Starship has made ten million deliveries across fourteen million miles. Both depend on prior maps of a bounded area, and Starship advertises 99 percent autonomy, which is another way of saying a person intervenes about once in a hundred. Canon requires a machine that walks into a police station it has never seen, in the dark, and finds a specific desk. Building the map is solved. Recognising a place that has changed, without GPS and without a prior survey, is not.", "last_reviewed": "2026-08-29", "review_cadence_days": 45, "progress": 0.2667 }, { "id": "neural-net-processor", "name": "Neural-net processor", "category": "cognition", "weight": 10, "one_liner": "\"A neural-net processor. A learning computer.\"", "canon": { "requirement": "A single, removable, physically small processor that runs the entire perceptual, motor, linguistic, planning and deceptive stack of an autonomous humanoid; that learns ON-LINE, during deployment, from unstructured contact with humans; whose ability to learn is gated by a HARDWARE WRITE-ENABLE that its manufacturer sets to read-only before deployment; and that constitutes the machine's identity - remove it and the body is inert, destroy it and the machine is dead.", "quantified": [ { "metric": "architecture", "value": "'a neural net processor, a learning computer'", "source_ref": "t2-garage-repair", "tier": "PRIMARY" }, { "metric": "T1's earlier and more modest framing", "value": "'microprocessor controlled' - the neural-net claim does not exist in 1984", "source_ref": "t1-reese-briefing", "tier": "PRIMARY" }, { "metric": "physical size", "value": "'about the size and shape of a domino' ~ 45 x 25 x 8 mm ~ 9 cm3", "source_ref": "t2-script-89", "tier": "TERTIARY" }, { "metric": "construction", "value": "ceramic, reddish-brown, 'the color of liver'; 'made up of small cubes connected together'", "source_ref": "t2-script-37", "tier": "TERTIARY" }, { "metric": "learning gate", "value": "hardware pin switch, factory-preset to 'read-only' when sent out alone", "source_ref": "t2-se-chip-flip", "tier": "PRIMARY-SE" }, { "metric": "swap time", "value": "seconds; hot-removable via a maintenance port in the skull", "source_ref": "t2-se-chip-flip", "tier": "PRIMARY-SE" }, { "metric": "state retention", "value": "learned state survives power-off and re-insertion", "source_ref": "t2-se-chip-flip", "tier": "PRIMARY-SE" }, { "metric": "count per unit", "value": "one", "source_ref": "t2-steel-mill", "tier": "PRIMARY" }, { "metric": "bootstrap role", "value": "the smashed 1984 chip is the seed of all Cyberdyne's work", "source_ref": "t2-dyson-chip", "tier": "PRIMARY" } ], "sources": [ { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-garage-repair", "evidence": "The line everyone quotes. John asks whether the machine can learn.", "quote": "My CPU is a neural net processor, a learning computer. The more contact I have with humans, the more I learn.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day (Special Edition)", "year": 1991, "medium": "film-extended", "ref": "t2-se-chip-flip", "evidence": "Special Edition only. The Terminator explains why it does not learn in the field.", "quote": "My CPU is a neural-net processor... a learning computer. But Skynet presets the switch to 'read-only' when we are sent out alone.", "verified": true, "tier": "PRIMARY-SE", "note": "NOT in the theatrical cut. Restored scenes 87-89C ('Chip flip', 3:32). Must always be labelled Special Edition." }, { "work": "Terminator 2: Judgment Day (Special Edition)", "year": 1991, "medium": "film-extended", "ref": "t2-se-chip-flip", "evidence": "Sarah's reply, and the Terminator's one-word confirmation.", "quote": "Doesn't want you thinking too much, huh? / No.", "verified": true, "tier": "PRIMARY-SE" }, { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-reese-briefing", "evidence": "T1's description is not an AI claim at all.", "quote": "Underneath, it's a hyperalloy combat chassis... microprocessor controlled, fully armored... very tough.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-dyson-chip", "evidence": "Miles Dyson on the chip recovered from the 1984 Terminator.", "quote": "It's scary stuff. Radically advanced. It was smashed. It didn't work, but it gave us ideas, took us in new directions. Things we would've never... All my work was based on it.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-steel-mill", "evidence": "The chip is the seat of identity, and the machine cannot destroy it itself.", "quote": "There's one more chip. And it must be destroyed also. / I cannot self-terminate. You must lower me into the steel.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day - revised final shooting script", "year": 1991, "medium": "shooting-script", "ref": "t2-script-89", "evidence": "Stage direction, SE scene 89 - the chip removed from the T-800's skull.", "quote": "A reddish-brown ceramic rectangle with a connector on one end. About the size and shape of a domino. On close inspection it appears to be made up of small cubes connected together.", "verified": true, "tier": "TERTIARY" }, { "work": "Terminator 2: Judgment Day - revised final shooting script", "year": 1991, "medium": "shooting-script", "ref": "t2-script-37", "evidence": "Stage direction, Cyberdyne vault - the salvaged 1984 chip.", "quote": "a ceramic rectangle, about the size of a domino, the color of liver", "verified": true, "tier": "TERTIARY" } ], "canon_confidence": "explicit", "canonical_target": { "metric": "removable processor: volume, functional load, on-line learning, hardware write-enable", "value": "A single removable processor of ~9 cm3 running the complete perception-language-planning-control stack of an autonomous humanoid in real time; supporting on-line weight updates from unstructured deployment experience; gating those updates behind a hardware write-enable; retaining learned state across power cycling and physical removal; and being the sole locus of the system's identity", "reasoning": "Four independently-derived constraints. VOLUME (9 cm3) from the domino simile, the only physical description canon offers. FUNCTIONAL LOAD: everything else in this dataset runs on this one chip - it is not a co-processor, the body is inert without it. ON-LINE LEARNING: stated in the theatrical cut, so non-negotiable. HARDWARE WRITE-ENABLE: the discriminator that makes the component interesting rather than generic - a real system that learns continuously but has no gate does not meet the canon spec, it exceeds it in capability and falls short in control, and the index should score those separately. Canon is explicitly describing a safety mechanism, and the film's plot turns on that mechanism being defeated from outside with a pin in about four seconds." }, "assumed_by_sota_agent": "A learning computer physically inside the unit: a neural-net processor that runs the machine's whole cognition and learns from experience during deployment, switchable between a read-only mode and a learning mode — Skynet ships infiltrators set read-only so they do not do too much thinking." }, "real": { "status": "in_progress", "trl": 6, "spec_fraction": 0.2, "spec_fraction_rationale": "Three canon properties, weighted 0.25/0.50/0.25. (a) A neural-net inference substrate running inside the body: met, ~0.90. (b) Learning in deployment from ordinary experience without engineered reward or resets: ~0.05 — commercial neuromorphic 'on-chip learning' trains only a final binary-weight fully-connected layer with no backpropagation, and the strongest on-robot learning results require task-specific rewards and human resets. (c) A low-power substrate at brain-like density: Hala Point's 1.15 billion neurons at 2,600 W is 1.3% of a human brain's neuron count at ~130x its 20 W power, so ~0.013 on neurons-per-watt, credited 0.05 for the 15 TOPS/W achievement. 0.25(0.90) + 0.50(0.05) + 0.25(0.05) = 0.2625, discounted to 0.20 because canon requires (b) continuously and unsupervised.", "gap": "Nothing today is a learning computer in the canon sense. Inference at the edge is solved; in-deployment weight updating is either trivial (adding a class to a linear head) or narrow and heavily scaffolded (task-scoped reinforcement learning with an engineered reward, resets and human coaching). The largest neuromorphic system in existence is explicitly a research prototype, and the only neuromorphic part shipping in volume did so at 2,000 units.", "why_hard": "Gradient learning needs stored activations, a backward pass costing roughly three times inference in memory and energy, and a supervision signal — and nothing in an embodied life supplies that signal for free. The substrate that would make learning cheap, analog and in-memory and spiking, is blocked on device non-idealities (variability, endurance, limited multilevel states, ADC overhead) and, per the field's own review in Nature Communications, on the absence of a software ecosystem.", "movers": [ { "name": "Intel Labs", "kind": "lab", "country": "US", "what": "Hala Point: 1,152 Loihi 2 chips, 1.15 billion neurons, 2,600 W, >15 TOPS/W — explicitly a research prototype, not a product." }, { "name": "SpiNNcloud Systems", "kind": "company", "country": "DE", "what": "First commercially available neuromorphic supercomputer; Sandia National Laboratories deployed a SpiNNaker2 system simulating 150–180 million neurons." }, { "name": "BrainChip", "kind": "company", "country": "AU", "what": "Only neuromorphic vendor shipping production silicon with on-chip learning — AKD1500, first batch 2,000 units in June 2026, full run about 60,000." }, { "name": "Physical Intelligence", "kind": "company", "country": "US", "what": "RL Token: on-robot reinforcement learning from 15 minutes of real-world data, the closest working analogue to a learning computer in a body." }, { "name": "IBM Research", "kind": "lab", "country": "US", "what": "NorthPole digital in-memory inference architecture, one of the few non-spiking alternatives to conventional edge accelerators." }, { "name": "SynSense", "kind": "company", "country": "CH/CN", "what": "Commercial mixed-signal neuromorphic vision processors; co-author of the Nature Communications assessment of the field's commercial prospects." } ], "evidence": [ { "date": "2024-04-17", "claim": "Intel's Hala Point packs 1,152 Loihi 2 processors — 1.15 billion neurons and 128 billion synapses — into a six-rack-unit chassis drawing a maximum of 2,600 W, achieving over 15 trillion 8-bit operations per watt; Intel states it is a research prototype, not a commercial system.", "source": "Intel Newsroom", "title": "Intel Builds World's Largest Neuromorphic System to Enable More Sustainable AI", "url": "https://newsroom.intel.com/artificial-intelligence/intel-builds-worlds-largest-neuromorphic-system", "kind": "product", "delta": "+" }, { "date": "2025-04-15", "claim": "The field's own review concludes that neuromorphic technology arrives at commercial viability only 'after several false starts', and identifies the two unsolved problems as how to program general neuromorphic applications and how to deploy them at scale.", "source": "Nature Communications", "title": "The road to commercial success for neuromorphic technologies", "url": "https://doi.org/10.1038/s41467-025-57352-1", "kind": "paper", "delta": "0" }, { "date": "2026-08-29", "claim": "BrainChip's documentation states that on-chip 'edge learning' trains only the final FullyConnected layer, which must have binary weights and receive binary inputs, using a local plasticity rule — gradient backpropagation does not run on chip at all.", "source": "BrainChip", "title": "Akida user guide — edge learning", "url": "https://doc.brainchipinc.com/user_guide/akida.html", "kind": "product", "delta": "-" }, { "date": "2026-03-19", "claim": "Physical Intelligence's RL Token trains small actor and critic networks on a robot from a VLA's compressed internal state, improving Ethernet insertion from about 100 to about 350 successes per ten minutes using 15 minutes of real-world data and two hours of wall-clock time including resets.", "source": "Physical Intelligence", "title": "Precise Manipulation with Efficient Online RL", "url": "https://www.pi.website/research/rlt", "kind": "demo", "delta": "+" }, { "date": "2025-11-17", "claim": "Physical Intelligence's Recap method trained a VLA on its own autonomous experience, more than doubling throughput on the hardest tasks and reaching about 90% success on box assembly, laundry folding and espresso making, with one machine running continuously from 05:30 to 23:30.", "source": "Physical Intelligence", "title": "pi*0.6: a VLA That Learns from Experience", "url": "https://www.pi.website/blog/pistar06", "kind": "demo", "delta": "+" }, { "date": "2026-07-01", "claim": "BrainChip received its first AKD1500 production batch of 2,000 units in June 2026, with a full production run of approximately 60,000 units, below plan on yield — the total commercial volume of neuromorphic silicon shipping anywhere.", "source": "Edge AI and Vision Alliance", "title": "BrainChip announces commercial availability and production shipments of AKD1500 neuromorphic processors", "url": "https://www.edge-ai-vision.com/2026/07/brainchip-announces-commercial-availability-and-production-shipments-of-akd1500-neuromorphic-processors/", "kind": "product", "delta": "+" } ], "indicators": [ { "metric": "Neurons in the largest neuromorphic system, per watt", "unit": "neurons/W", "value": 442307, "as_of": "2024-04-17", "direction": "up_is_progress", "canon_target": 4300000000, "source": "https://newsroom.intel.com/artificial-intelligence/intel-builds-worlds-largest-neuromorphic-system" }, { "metric": "Layers trainable on-chip in production neuromorphic silicon", "unit": "layers", "value": 1, "as_of": "2026-08-29", "direction": "up_is_progress", "canon_target": 100, "source": "https://doc.brainchipinc.com/user_guide/akida.html" }, { "metric": "Real-world data needed for on-robot RL to materially improve a manipulation policy", "unit": "minutes", "value": 15, "as_of": "2026-03-19", "direction": "down_is_progress", "canon_target": 0, "source": "https://www.pi.website/research/rlt" }, { "metric": "Annual production volume of neuromorphic processors with on-chip learning", "unit": "units", "value": 60000, "as_of": "2026-07-01", "direction": "up_is_progress", "canon_target": 10000000, "source": "https://www.edge-ai-vision.com/2026/07/brainchip-announces-commercial-availability-and-production-shipments-of-akd1500-neuromorphic-processors/" } ] }, "commentary": "The films independently invented shipping a model with learning switched off. In T2 the T-800 explains that Skynet sends the units out read-only so they do not do too much thinking; in 2026 every deployed foundation model ships with frozen weights, for exactly the same reason. What is missing is the other mode. Intel's Hala Point packs 1.15 billion neurons into 2,600 watts — 1.3 percent of a human brain's neuron count at 130 times its power — and Intel still calls it a research prototype. BrainChip's production silicon does learn on chip: one fully-connected layer, binary weights, no backpropagation. The switch exists. There is not much behind it.", "last_reviewed": "2026-08-29", "review_cadence_days": 30, "progress": 0.1333 }, { "id": "on-board-compute-density", "name": "On-board compute density", "category": "cognition", "weight": 9, "one_liner": "Running all of it inside a skull, on internal power, without a fan.", "canon": { "requirement": "Everything the neural-net processor does must fit behind a maintenance port at the base of a human-sized skull, run off the same cell that powers the hydraulics for 120 years, and produce so little heat that the surrounding living tissue neither cooks nor registers as abnormal to a human hand.", "quantified": [ { "metric": "volume envelope", "value": "~9 cm3 (domino)", "source_ref": "t2-script-89", "tier": "TERTIARY" }, { "metric": "location", "value": "maintenance port in the chrome skull, base of the skull, beneath the scalp", "source_ref": "t2-script-88", "tier": "TERTIARY" }, { "metric": "access method", "value": "X-Acto knife plus air tools; port cover unscrews", "source_ref": "t2-se-chip-flip", "tier": "PRIMARY-SE" }, { "metric": "power budget", "value": "never separately stated anywhere", "source_ref": "derived", "tier": "EXTRAPOLATED" }, { "metric": "thermal budget", "value": "must not be detectable through skin", "source_ref": "derived", "tier": "EXTRAPOLATED" } ], "sources": [ { "work": "Terminator 2: Judgment Day - revised final shooting script", "year": 1991, "medium": "shooting-script", "ref": "t2-script-88", "evidence": "Stage direction, SE scene 88 - accessing the CPU port.", "quote": "an X-ACTO KNIFE cutting into Terminator's scalp at the base of his skull. His voice calmly directs Sarah as she spreads the bloody incision and locates the maintenance port for the CPU in the chrome skull beneath.", "verified": true, "tier": "TERTIARY" }, { "work": "Terminator 2: Judgment Day (Special Edition)", "year": 1991, "medium": "film-extended", "ref": "t2-se-chip-flip", "evidence": "The Terminator directs its own CPU extraction.", "quote": "Now open the port cover. / Hold the CPU by its base tab. Pull.", "verified": true, "tier": "PRIMARY-SE" } ], "canon_confidence": "extrapolated", "canonical_target": { "metric": "compute volume and power envelope", "value": "The complete cognitive stack of an autonomous humanoid in <=10 cm3 at <= tens of watts, with no active cooling", "reasoning": "The power figure is stated nowhere and must not be invented as a hard number, so it is bracketed by the thermal constraint instead: whatever the chip draws must pass through skin held at human temperature, alongside the hydraulics' waste heat. A chip drawing hundreds of watts inside a sealed skull would cook the scalp; canon shows the scalp intact and bleeding normally. Hence 'tens of watts', marked extrapolated." }, "assumed_by_sota_agent": "Run the entire cognitive stack — perception, tracking, planning, language, deception and motor control — inside a human skull-sized sealed volume of roughly 1.4 litres, on internal power, continuously for decades, without an externally visible thermal signature." }, "real": { "status": "in_progress", "trl": 9, "spec_fraction": 0.12, "spec_fraction_rationale": "(capability retained when the model moves on-board) x (thermal feasibility of that package in a sealed skull) = 0.31 x 0.40 ≈ 0.12. The 0.31 is measured: on the same ALOHA benchmark, Gemini Robotics 1.5 scores 0.72 in-distribution and 0.39 on task generalisation, while its on-device variant scores 0.52 and 0.12 — so the on-board model retains 31% of the flagship's generalisation. The 0.40 reflects that the best edge module draws up to 130 W against a passive dissipation budget of roughly 10–20 W in a sealed human head, a factor of 6–9 over budget before adding actuator, radio and sensor power.", "gap": "Peak TOPS is not the constraint and is not comparable across vendors. Jetson Thor's headline 2,070 TFLOPS is sparse FP4; dense FP8 is 517; and real sustained throughput on a 70-billion-parameter model is 12.64 tokens per second, roughly 0.086% of the headline number. The limit is 273 GB/s of memory bandwidth against about 70 GB of weights per forward pass, not arithmetic.", "why_hard": "Thermal first, bandwidth second. A sealed skull-sized volume sheds on the order of 10–20 W passively; the best edge module needs up to 130 W to deliver 12.6 tokens per second on a 70B model, and that throughput is capped by memory bandwidth rather than by FLOPS — so buying more TOPS does not buy more thinking, and buying more bandwidth costs more watts.", "movers": [ { "name": "NVIDIA", "kind": "company", "country": "US", "what": "Jetson AGX Thor (T5000): 2,070 sparse-FP4 TFLOPS, 128 GB at 273 GB/s, 40–130 W, $3,499 dev kit — the reference edge module for humanoid robots." }, { "name": "Google DeepMind", "kind": "lab", "country": "US/UK", "what": "Publishes the only clean measurement of the on-device penalty, running the same benchmark on its flagship VLA and its on-device variant." }, { "name": "Qualcomm", "kind": "company", "country": "US", "what": "Mobile and robotics NPUs at single-digit watts — the only vendor shipping inference silicon at a power budget a sealed head could actually dissipate." }, { "name": "MLCommons", "kind": "program", "country": "US", "what": "MLPerf Inference is the only vendor-neutral place edge accelerators are measured on sustained throughput rather than peak TOPS." }, { "name": "Physical Intelligence", "kind": "company", "country": "US", "what": "Real-time action chunking work exists specifically because VLA inference latency, not capability, limits closed-loop control on board." } ], "evidence": [ { "date": "2025-08-25", "claim": "NVIDIA's Jetson T5000 module is specified at 2,070 TFLOPS sparse FP4, 1,035 TFLOPS dense FP4, and 517 TFLOPS dense FP8, with 128 GB of LPDDR5X at 273 GB/s and a 40–130 W envelope; NVIDIA's own benchmark table gives 12.64 output tokens per second on Llama 3.3 70B at max concurrency 8.", "source": "NVIDIA Developer", "title": "Introducing NVIDIA Jetson Thor, the Ultimate Platform for Physical AI", "url": "https://developer.nvidia.com/blog/introducing-nvidia-jetson-thor-the-ultimate-platform-for-physical-ai/", "kind": "product", "delta": "+" }, { "date": "2025-10-02", "claim": "On the same ALOHA benchmark, Gemini Robotics 1.5 achieves 0.72 in-distribution success and 0.39 on task generalisation, while the on-device variant achieves 0.52 and 0.12 — the on-board model retains 31% of the flagship's generalisation.", "source": "arXiv", "title": "Gemini Robotics 1.5: Pushing the Frontier of Generalist Robots with Advanced Embodied Reasoning, Thinking, and Motion Transfer", "url": "https://arxiv.org/abs/2510.03342", "kind": "paper", "delta": "-" }, { "date": "2026-08-25", "claim": "NVIDIA announced Jetson Orin Nano 2 at 78 trillion operations per second with 8 GB of memory and twice the inference performance of Orin Nano Super at 40% less power in 15 W mode; the precision of the TOPS figure is not stated and availability is the first half of 2027.", "source": "NVIDIA Newsroom", "title": "NVIDIA Announces Jetson Orin Nano 2 Robotics Computer to Redefine Entry-Level Edge AI", "url": "https://nvidianews.nvidia.com/news/nvidia-announces-jetson-orin-nano-2-robotics-computer-to-redefine-entry-level-edge-ai", "kind": "product", "delta": "+" }, { "date": "2002-07-29", "claim": "The human brain accounts for roughly 20% of the body's resting energy consumption while representing about 2% of body mass — approximately 20 W for a typical adult at a basal metabolic rate near 100 W — in a cranial volume of roughly 1.4 litres.", "source": "PNAS (via PubMed Central, PMC124895)", "title": "Appraising the brain's energy budget", "url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC124895/", "kind": "paper", "delta": "0" }, { "date": "2026-04-24", "claim": "An independent multi-dimensional benchmark of LLM inference on hardware-accelerated single-board computers finds that power efficiency, physical size and token throughput trade off sharply on edge platforms with NPUs and GPUs.", "source": "arXiv", "title": "Cloud to Edge: Benchmarking LLM Inference On Hardware-Accelerated Single-Board Computers", "url": "https://arxiv.org/abs/2604.24785", "kind": "paper", "delta": "0" } ], "indicators": [ { "metric": "Sustained useful arithmetic as a fraction of headline sparse-FP4 TOPS, best edge module on a 70B model", "unit": "%", "value": 0.086, "as_of": "2025-08-25", "direction": "up_is_progress", "canon_target": 50, "source": "https://developer.nvidia.com/blog/introducing-nvidia-jetson-thor-the-ultimate-platform-for-physical-ai/" }, { "metric": "Power draw of the best edge inference module", "unit": "W", "value": 130, "as_of": "2025-08-25", "direction": "down_is_progress", "canon_target": 20, "source": "https://developer.nvidia.com/blog/introducing-nvidia-jetson-thor-the-ultimate-platform-for-physical-ai/" }, { "metric": "Generalisation retained when a frontier robot model is moved on-device", "unit": "%", "value": 31, "as_of": "2025-10-02", "direction": "up_is_progress", "canon_target": 100, "source": "https://arxiv.org/abs/2510.03342" }, { "metric": "Output tokens per second on a 70B model, best edge module", "unit": "tokens/s", "value": 12.64, "as_of": "2025-08-25", "direction": "up_is_progress", "canon_target": 200, "source": "https://developer.nvidia.com/blog/introducing-nvidia-jetson-thor-the-ultimate-platform-for-physical-ai/" } ] }, "commentary": "NVIDIA's Jetson Thor is quoted at 2,070 TFLOPS, which is sparse FP4 peak and should be read as advertising. Dense FP8 is 517. What the module actually delivers is 12.64 tokens per second on Llama 3.3 70B while drawing up to 130 watts — under a tenth of a percent of the headline, because the binding constraint is 273 gigabytes per second of memory bandwidth against 70 gigabytes of weights per forward pass, not arithmetic. A human skull sheds perhaps fifteen watts. Google measured the price of going on board directly: its on-device robot model retains 31 percent of the flagship's generalisation. Buying TOPS does not buy thinking.", "last_reviewed": "2026-08-29", "review_cadence_days": 30, "progress": 0.12 }, { "id": "embodied-reasoning-and-planning", "name": "Embodied reasoning and planning", "category": "cognition", "weight": 10, "one_liner": "Deciding what to do in the world, and then actually doing it.", "canon": { "requirement": "Open-world sequential decision-making over a multi-day mission with no operator, no communications and no re-tasking - including explicit modelling of an adversary's policy, active hypothesis testing against a deceptive counterparty, and re-planning after every contact.", "quantified": [ { "metric": "adversary policy prediction", "value": "'The T-1000's highest probability for success now will be to copy Sarah Connor and wait for you to make contact with her.'", "source_ref": "t2-parking-lot", "tier": "PRIMARY" }, { "metric": "deception detection", "value": "injects a false proper noun into a live call to test the counterparty", "source_ref": "t2-phone-call", "tier": "PRIMARY" }, { "metric": "resource acquisition from cold start", "value": "clothes, weapons, vehicles, cash, identity - all improvised on arrival", "source_ref": "t1-arrival", "tier": "PRIMARY" }, { "metric": "constraint re-planning", "value": "adopts a no-kill constraint mid-mission and satisfies it under fire", "source_ref": "t2-cyberdyne", "tier": "PRIMARY" }, { "metric": "unsupervised mission duration", "value": "~72 h", "source_ref": "t1-timeline", "tier": "PRIMARY" } ], "sources": [ { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-parking-lot", "evidence": "John wants to rescue Sarah; the Terminator predicts the T-1000's move in explicitly decision-theoretic terms.", "quote": "Negative. The T-1000's highest probability for success now will be to copy Sarah Connor and wait for you to make contact with her.", "verified": true, "tier": "PRIMARY" } ], "canon_confidence": "explicit", "canonical_target": { "metric": "autonomous open-world mission execution with adversary modelling", "value": "Multi-day mission execution in an open world with zero operator contact, maintaining an explicit adversary model, actively hypothesis-testing against a deceptive counterparty, improvising resource acquisition from a cold start, and re-planning under changed constraints while under fire", "reasoning": "Each clause maps to one on-screen behaviour, so the target is falsifiable clause-by-clause rather than being a vague 'general intelligence' claim. The hypothesis-testing clause (the 'Wolfie' gambit) is the strongest single piece of evidence and is usually filed under voice mimicry, where it does not belong - the mimicry is the CHANNEL; the reasoning is the capability." }, "assumed_by_sota_agent": "Form and revise multi-step plans over hours to days; improvise with found objects, tools and vehicles; recover from failure without help; pursue a goal across a city; and never require a human to decompose the task." }, "real": { "status": "in_progress", "trl": 6, "spec_fraction": 0.1, "spec_fraction_rationale": "The best measured long-horizon binary success rate for any robot foundation model is 0.359, averaged across the eight long-horizon tasks in the Gemini Robotics 1.5 appendix using the best agent configuration, with two tasks at zero — and those figures come from three to five trials each. Canon requires comparable reliability sustained over hours, which is twenty to fifty such tasks chained, and 0.359^k tends to zero for k greater than two. Scoring generously against a single canon-scale task rather than a full mission, with a 0.3 horizon discount: 0.359 x 0.3 ≈ 0.108, rounded to 0.10.", "gap": "Reliability, and the ability to measure it. Gemini Robotics 2, released July 2026, picks an object off the floor 45.7% of the time, screws in a light bulb 36% of the time and works a ziplock bag 40% of the time. Papers lead with continuous progress scores and relegate binary success rates to appendices. Without an orchestrator, the leading VLA scored zero on seven of eight long-horizon tasks.", "why_hard": "Errors compound. A policy that completes a two-minute subtask 90% of the time completes a fifty-step mission about 0.5% of the time, and the dominant training recipe — behaviour cloning from teleoperation — never demonstrates recovery from states the operator did not enter. The secondary constraint is data, and the field's own answer is to spend a billion dollars collecting more of it.", "movers": [ { "name": "Google DeepMind", "kind": "lab", "country": "US/UK", "what": "Gemini Robotics 2 and Gemini Robotics-ER 2, July 2026 — the only frontier robot models publishing per-task success rates on real hardware." }, { "name": "Physical Intelligence", "kind": "company", "country": "US", "what": "pi-0.5 through pi-0.7: open-world generalisation to unseen homes, and compositional generalisation to tasks and robots with no matching training data." }, { "name": "Figure AI", "kind": "company", "country": "US", "what": "Index — a crowd-sourced physical dataset from a consumer app, backed by a stated $1 billion of data and compute spend over twelve months." }, { "name": "NVIDIA", "kind": "company", "country": "US", "what": "GR00T open humanoid foundation models and Cosmos world models, the main open baseline the academic field builds against." }, { "name": "RoboArena consortium", "kind": "university", "country": "US", "what": "Distributed real-world evaluation of generalist policies across institutions, built because centralised robot benchmarking does not scale." }, { "name": "1X Technologies", "kind": "company", "country": "NO/US", "what": "NEO home humanoid, one of the few programmes attempting long-duration autonomy in ordinary houses rather than staged settings." } ], "evidence": [ { "date": "2026-07-30", "claim": "Gemini Robotics 2 reports whole-body manipulation success on the Apollo humanoid of 76.3% picking from a shelf, 68.4% from a table and 45.7% from the floor; multi-finger dexterity of 92% unscrewing a bulb but 36% screwing one in, 44% tying a trash bag, 40% on a ziplock and 32% using a dustpan.", "source": "Google DeepMind", "title": "Gemini Robotics 2 brings whole body intelligence to robots", "url": "https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/", "kind": "product", "delta": "+" }, { "date": "2025-10-02", "claim": "On eight long-horizon household tasks, the best Gemini Robotics 1.5 agent configuration achieves binary success rates of 0.60, 0.40, 0.20, 0.00, 0.67, 0.67, 0.33 and 0.00 — a mean of 0.359 with two tasks at zero — while the VLA alone scores zero on seven of eight. The values 0.33 and 0.67 are only producible from three trials.", "source": "arXiv", "title": "Gemini Robotics 1.5 (Appendix D.2, Figure 41)", "url": "https://arxiv.org/abs/2510.03342", "kind": "benchmark", "delta": "0" }, { "date": "2026-07-30", "claim": "Gemini Robotics-ER 2 achieves 57.4% accuracy on progress classification across five progress levels and 91.3% on moment-finding in video, and consistently outperforms ER 1.6 on tool orchestration.", "source": "Google", "title": "Gemini Robotics-ER 2: video understanding, task orchestration and multi-robot collaboration", "url": "https://blog.google/innovation-and-ai/models-and-research/google-deepmind/gemini-robotics-er-2/", "kind": "product", "delta": "+" }, { "date": "2026-04-14", "claim": "On instrument reading, Gemini Robotics-ER improved from 23% (ER 1.5) to 86% (ER 1.6), and to 93% with agentic vision — a roughly fourfold gain on a narrow perception-reasoning task in six months.", "source": "Google DeepMind", "title": "Gemini Robotics ER 1.6: Enhanced Embodied Reasoning", "url": "https://deepmind.google/blog/gemini-robotics-er-1-6/", "kind": "product", "delta": "+" }, { "date": "2025-04-22", "claim": "pi-0.5 was evaluated in three real homes not present in training over ten trials per task, on multi-stage tasks lasting two to five minutes, scored on a partial-progress rubric rather than binary success; matching a control trained on the test homes required 104 distinct training locations.", "source": "Physical Intelligence", "title": "pi-0.5: a Vision-Language-Action Model with Open-World Generalization", "url": "https://arxiv.org/abs/2504.16054", "kind": "paper", "delta": "+" }, { "date": "2026-04-16", "claim": "pi-0.7 claims compositional generalisation, folding laundry on a bimanual UR5e with no laundry-specific or robot-specific training data at a rate matching expert teleoperators' zero-shot success, with several tasks at approximately 100% and throughput 0.9–2.0x that of specialist models. Company blog; no paper and no independent replication.", "source": "Physical Intelligence", "title": "pi-0.7", "url": "https://www.pi.website/blog/pi07", "kind": "product", "delta": "+" }, { "date": "2025-10-05", "claim": "The field acknowledges its own evaluation problem: robot policies 'are often evaluated on a small number of hardware trials without any statistical assurances'.", "source": "arXiv", "title": "Reliable and Scalable Robot Policy Evaluation with Imperfect Simulators", "url": "https://arxiv.org/abs/2510.04354", "kind": "paper", "delta": "-" }, { "date": "2026-08-25", "claim": "Figure AI announced Index, a crowd-sourced physical dataset built from a consumer app with 264,000 downloads and 16 million uploaded videos, and stated it will spend over $1 billion on data and compute in the next twelve months.", "source": "Figure AI", "title": "Introducing Index: Building the World's Largest and Most Diverse Physical Dataset", "url": "https://www.figure.ai/news/introducing-index", "kind": "funding", "delta": "+" } ], "indicators": [ { "metric": "Mean binary success rate across eight long-horizon household tasks, best published configuration", "unit": "%", "value": 35.9, "as_of": "2025-10-02", "direction": "up_is_progress", "canon_target": 99, "source": "https://arxiv.org/abs/2510.03342" }, { "metric": "Humanoid success rate picking an object off the floor, newest frontier model", "unit": "%", "value": 45.7, "as_of": "2026-07-30", "direction": "up_is_progress", "canon_target": 99, "source": "https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/" }, { "metric": "Trials per task behind published long-horizon success rates", "unit": "trials", "value": 3, "as_of": "2025-10-02", "direction": "up_is_progress", "canon_target": 100, "source": "https://arxiv.org/abs/2510.03342" }, { "metric": "Longest autonomously executed multi-stage task in an unseen real home", "unit": "minutes", "value": 5, "as_of": "2025-04-22", "direction": "up_is_progress", "canon_target": 120, "source": "https://arxiv.org/abs/2504.16054" } ] }, "commentary": "Gemini Robotics 2 landed in July. On a humanoid with dexterous hands it screws in a light bulb 36 percent of the time, works a ziplock bag 40 percent, and picks an object off the floor 45.7 percent. Its predecessor's long-horizon results are more instructive still: across eight household tasks the best configuration averaged 36 percent binary success and scored zero on two, and the values 0.33 and 0.67 indicate three trials per task. The papers lead with progress scores; the success rates sit in an appendix. Canon requires a machine that improvises for two hours. The field cannot yet reliably tidy a desk.", "last_reviewed": "2026-08-29", "review_cadence_days": 21, "progress": 0.0667 }, { "id": "continual-learning-and-adaptation", "name": "Continual learning and adaptation", "category": "cognition", "weight": 8, "one_liner": "\"The more contact I have with humans, the more I learn.\"", "canon": { "requirement": "Acquire new behaviours and language from single exposures DURING deployment, integrate them into the same weights that are flying the mission, generalise them to novel contexts, and do it without retraining, without forgetting the mission, and without an engineer present.", "quantified": [ { "metric": "exposures needed to acquire an idiom", "value": "one", "source_ref": "t2-desert-drive", "tier": "PRIMARY" }, { "metric": "generalisation demonstrated", "value": "composes two separately-taught phrases into a novel one - 'Chill out, dick-wad'", "source_ref": "t2-desert-drive", "tier": "PRIMARY" }, { "metric": "learning gate", "value": "hardware read-only pin, defeated on-mission by a ten-year-old with a pin", "source_ref": "t2-se-chip-flip", "tier": "PRIMARY-SE" }, { "metric": "terminal state reached", "value": "forms a value judgement it was never given", "source_ref": "t2-steel-mill", "tier": "PRIMARY" }, { "metric": "learning horizon (secondary)", "value": "22 years of continuous adaptation, forming a family and a conscience ('Carl')", "source_ref": "dark-fate", "tier": "SECONDARY" } ], "sources": [ { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-desert-drive", "evidence": "John teaches the Terminator to talk; the machine then produces an utterance nobody taught it.", "quote": "If someone gets upset, you say, 'Chill out.' Or you can do combinations. / Chill out, dick-wad. / That's great. See? You're getting it.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-garage-repair", "evidence": "The stated mechanism.", "quote": "My CPU is a neural net processor, a learning computer. The more contact I have with humans, the more I learn.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-steel-mill", "evidence": "The Terminator's last line before self-termination.", "quote": "I know now why you cry. But it's something I can never do.", "verified": true, "tier": "PRIMARY" } ], "canon_confidence": "explicit", "canonical_target": { "metric": "single-exposure on-line acquisition with compositional generalisation", "value": "Single-exposure acquisition of new language and behaviour from unstructured human contact during deployment, integrated into the operating policy with no retraining and no catastrophic forgetting, with demonstrated compositional generalisation to novel utterances the machine was never taught", "reasoning": "'Chill out, dick-wad' is the whole target in four syllables and is routinely misread as a gag. John teaches 'chill out' and (separately) 'dick-wad', and says only that combinations are permitted; the machine then produces a phrase nobody taught it, correctly register-matched, and John's reaction - 'See? You're getting it' - is the film explicitly marking it as generalisation rather than recall. That is the discriminator between a lookup table and learning, and canon put it on screen in 1991." }, "assumed_by_sota_agent": "Open-ended, largely unsupervised learning from ordinary interaction during deployment, without forgetting prior competence, across a mission of weeks and a service life measured in decades — improvement as a side effect of living, not as a training run." }, "real": { "status": "on_horizon", "trl": 5, "spec_fraction": 0.05, "spec_fraction_rationale": "Canon requires five properties: open-ended, unsupervised, non-forgetting, cross-task, lifelong. What exists satisfies roughly one of them — learning a new skill from experience — and only at about 0.25 credit, because it needs an engineered reward and human resets, is scoped to a single task, and demonstrably degrades under long sequential updating. 0.25 / 5 = 0.05.", "gap": "There is no working mechanism for open-ended in-deployment learning. Commercial on-chip learning adds a class to a linear head. The strongest robot results are offline retraining loops over curated autonomous experience with a task-specific reward. And Nature showed in 2024 that networks trained on long task sequences do not merely forget — they lose the ability to learn at all.", "why_hard": "The stability-plasticity dilemma, plus a worse result on top of it. Updating weights on new experience overwrites old competence; not updating means no learning. Replay fixes forgetting only by retaining the old data, which is precisely what a deployed machine does not have. And loss of plasticity means the failure mode is not forgetting but the death of learning itself, which no scaling law addresses.", "movers": [ { "name": "University of Alberta / Amii (Sutton, Mahmood, Dohare)", "kind": "university", "country": "CA", "what": "Published the loss-of-plasticity result in Nature and proposed continual backpropagation, the field's clearest negative result and its clearest partial remedy." }, { "name": "KU Leuven (Tuytelaars group)", "kind": "university", "country": "BE", "what": "August 2026 result showing forgetting and plasticity loss together still do not explain the gap to joint training; identifies data co-observation as a third deficit." }, { "name": "Physical Intelligence", "kind": "company", "country": "US", "what": "Recap and RL Token — the strongest demonstrations of a deployed robot improving from its own experience, within a fixed task and an engineered reward." }, { "name": "BrainChip", "kind": "company", "country": "AU", "what": "The only production silicon shipping any on-chip learning, limited to a single binary-weight fully-connected layer." }, { "name": "ICML 2026 Position Track / Dagstuhl CL-in-the-Era-of-Foundation-Models seminar", "kind": "program", "country": "DE", "what": "Spotlighted a position paper arguing that in-weight continual learning was the wrong frame and modular memory is the missing piece." } ], "evidence": [ { "date": "2024-08-21", "claim": "Deep networks trained on long sequences of tasks progressively lose the ability to learn new ones at all — not merely forgetting old tasks but losing plasticity; the proposed remedy, continual backpropagation, works by continually re-initialising low-utility units.", "source": "Nature", "title": "Loss of plasticity in deep continual learning", "url": "https://doi.org/10.1038/s41586-024-07711-7", "kind": "paper", "delta": "-" }, { "date": "2026-08-19", "claim": "Catastrophic forgetting and loss of plasticity together still do not explain the performance gap between sequential and joint training; a third independent deficit, data co-observation, persists after controlling for both, in supervised and self-supervised settings alike.", "source": "arXiv", "title": "Forgetting, plasticity, and co-observation: a third facet of continual learning", "url": "https://arxiv.org/abs/2608.18803", "kind": "paper", "delta": "-" }, { "date": "2026-03-02", "claim": "An ICML 2026 Position Track spotlight, arising from a Dagstuhl seminar on continual learning in the era of foundation models, argues that the historical focus on in-weight learning has made catastrophic forgetting a persistent challenge and that modular memory combining in-context and in-weight learning is the missing piece.", "source": "arXiv", "title": "Position: Modular Memory is the Key to Continual Learning Agents", "url": "https://arxiv.org/abs/2603.01761", "kind": "paper", "delta": "0" }, { "date": "2026-03-19", "claim": "On-robot reinforcement learning from 15 minutes of real-world data improved an Ethernet-insertion policy from about 100 to about 350 successes per ten minutes, and the resulting policy outperformed human teleoperation on that task — but the method requires a task-specific reward and environment resets.", "source": "Physical Intelligence", "title": "Precise Manipulation with Efficient Online RL", "url": "https://www.pi.website/research/rlt", "kind": "demo", "delta": "+" }, { "date": "2026-08-23", "claim": "A controlled study of test-time adaptation on CIFAR-10-C finds that aggregate accuracy obscures the conditions under which adaptation fails or provides no benefit, for BatchNorm adaptation, TENT and EATA alike.", "source": "arXiv", "title": "When Test-Time Adaptation Helps, Harms, or Becomes Inactive: A Condition-Level Study on CIFAR-10-C", "url": "https://arxiv.org/abs/2608.22233", "kind": "paper", "delta": "-" }, { "date": "2026-08-29", "claim": "Production neuromorphic silicon supports on-chip learning only in a final FullyConnected layer with binary weights and no gradient backpropagation, adding classes rather than updating a model.", "source": "BrainChip", "title": "Akida user guide — edge learning", "url": "https://doc.brainchipinc.com/user_guide/akida.html", "kind": "product", "delta": "0" } ], "indicators": [ { "metric": "Deployed systems performing open-ended unsupervised weight updating in the field", "unit": "count", "value": 0, "as_of": "2026-08-29", "direction": "up_is_progress", "canon_target": 1, "source": "https://doi.org/10.1038/s41586-024-07711-7" }, { "metric": "Distinct unexplained deficits in sequential versus joint training", "unit": "count", "value": 3, "as_of": "2026-08-19", "direction": "down_is_progress", "canon_target": 0, "source": "https://arxiv.org/abs/2608.18803" }, { "metric": "Real-world data needed for a deployed robot to improve one task by reinforcement learning", "unit": "minutes", "value": 15, "as_of": "2026-03-19", "direction": "down_is_progress", "canon_target": 0, "source": "https://www.pi.website/research/rlt" } ] }, "commentary": "\"The more contact I have with humans, the more I learn\" is the least solved sentence in the films. Nature published the finding in 2024: networks trained on long sequences of tasks do not merely forget, they lose the ability to learn at all. An August 2026 paper adds that forgetting and plasticity loss together still fail to explain the gap to joint training, and names a third deficit nobody has a fix for. What does work is narrow — Physical Intelligence improved an Ethernet-insertion policy three and a half fold from fifteen minutes of on-robot data — and requires an engineered reward and a human resetting the scene. That is a training loop, not a life.", "last_reviewed": "2026-08-29", "review_cadence_days": 45, "progress": 0.0278 }, { "id": "natural-language-understanding", "name": "Natural language understanding", "category": "language-and-deception", "weight": 9, "one_liner": "Ryan North's actual day job, and the one part of the Terminator problem largely finished.", "canon": { "requirement": "Real-time comprehension and production of unrestricted spoken English - idiom, slang, profanity, sarcasm, indirect requests, threats and register - in noisy, hostile and adversarial settings, sufficient to (a) select a socially appropriate response from candidates in a hostile exchange, (b) acquire and correctly redeploy new idioms from one exposure, and (c) sustain a live deceptive dialogue with someone who knows the impersonated party intimately. Canon's stated failure mode is REGISTER, never parsing.", "quantified": [ { "metric": "response-candidate selection", "value": "6 ranked candidates, register-matched, in under a second of screen time", "source_ref": "t1-script-197fx", "tier": "TERTIARY" }, { "metric": "idiom acquisition", "value": "one exposure", "source_ref": "t2-desert-drive", "tier": "PRIMARY" }, { "metric": "compositional generalisation", "value": "demonstrated", "source_ref": "t2-desert-drive", "tier": "PRIMARY" }, { "metric": "sustained live deception", "value": "full multi-turn phone conversation, unscripted content, not detected", "source_ref": "t2-phone-call", "tier": "PRIMARY" }, { "metric": "failure mode", "value": "register/sociolect - never parsing", "source_ref": "t2-desert-drive", "tier": "PRIMARY" } ], "sources": [ { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-mall-escape", "evidence": "The machine parses a normative statement, identifies it as unsupported, requests justification, receives none, and complies on asserted authority.", "quote": "You just can't go around killing people. / Why? / What do you mean, why? 'Cause you can't. / Because you just can't. Trust me on this.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-desert-drive", "evidence": "John's correction is explicitly meta-linguistic: the machine understood 'affirmative' perfectly, it picked the wrong sociolect.", "quote": "You gotta listen to the way people talk. You don't say 'affirmative' or some shit like that. You say, 'No problemo.'", "verified": true, "tier": "PRIMARY" }, { "work": "The Terminator - fourth draft, 20 April 1983", "year": 1983, "medium": "screenplay-draft", "ref": "t1-script-197fx", "evidence": "Scene 197/FX - the Terminator's HUD generates and ranks candidate responses to a hostile janitor, then selects the register-matched one. In the finished film we hear only the output, through a closed door.", "quote": "The digitized image PANS to the door and a LOGIC-FLOW DIAGRAM appears overlaid in color-coded words. It concluded with a list of potential appropriate responses: YES/NO OR WHAT GO AWAY PLEASE COME BACK LATER FUCK YOU FUCK YOU, ASSHOLE. The last begins to FLASH, and enlarges to fill the screen.", "verified": true, "tier": "TERTIARY" } ], "canon_confidence": "implied", "canonical_target": { "metric": "unrestricted conversational NLU under adversarial conditions", "value": "Real-time full-duplex comprehension and generation of unrestricted spoken English - idiom, slang, sarcasm, indirect speech acts, profanity and register - in noisy and adversarial conditions, sufficient to: select the register-appropriate response from ranked candidates in a hostile exchange; acquire and compositionally redeploy new idioms from a single exposure; and sustain a multi-turn deceptive conversation with a person who knows the impersonated speaker intimately, without detection", "reasoning": "The three clauses are chosen because each is separately measurable and each maps to a distinct on-screen demonstration. The third is the strictest and should lead: it is not a benchmark score, it is a live adversarial test administered by a human with strong priors and emotional stakes - and the machine passes it twice, in two films, in two directions. Deliberately NOT marked explicit: no character ever states that the machine understands language; it just does, continuously, and the films treat that as unremarkable." }, "assumed_by_sota_agent": "Understands unrestricted spoken English in real time and in the wild: slang, profanity, idiom, register, indirect speech acts and sarcasm. Parses 'chill out, dickwad' deliberately, asks what an idiom means and updates permanently, sustains unscripted conversation with hostile strangers, and grounds reference to whatever it is currently looking at. On-board, networkless, adversarial, and never misunderstands." }, "real": { "status": "in_progress", "trl": 9, "spec_fraction": 0.75, "spec_fraction_rationale": "The canon requirement decomposes into four capabilities: (a) lexical/syntactic/semantic comprehension including slang and idiom; (b) pragmatic inference and belief update; (c) grounded reference to the immediately perceived physical scene; (d) sustained robustness in unscripted adversarial interaction. (a) is met or exceeded by fielded models. (c) is measurably sub-human at 50.6% versus a 90.1% human baseline on analog-clock reading, the clearest published grounded-reference measurement. (b) is explicitly below human on implicature and belief update. (d) fails roughly one attempt in three on OSWorld. Three of four at or above human with one materially below and two partial: 0.75.", "gap": "Pragmatic inference under naturalistic conditions, grounded reference to a perceived physical scene, and reliability across long unscripted interaction. Fielded systems also hallucinate and exhibit measurable sycophancy — agreeing with whichever side the user has taken — which canon's machine never does.", "why_hard": "Pragmatics and grounded reference are not learnable from text distribution alone: they require a model of the speaker's beliefs and of the physical situation being referred to. Training data is a record of language, not of the situations that produced it, so the binding constraint is the absence of grounded, interaction-level supervision at scale rather than model capacity or compute.", "movers": [ { "name": "OpenAI", "kind": "company", "country": "US", "what": "Frontier general-purpose language models; leads or co-leads most public comprehension benchmarks." }, { "name": "Google DeepMind", "kind": "lab", "country": "UK", "what": "Gemini family; leads GPQA and multimodal grounding evaluations." }, { "name": "Anthropic", "kind": "company", "country": "US", "what": "Claude family; currently top of Humanity's Last Exam and the Arena Elo table per the 2026 AI Index." }, { "name": "Alibaba (Qwen)", "kind": "company", "country": "CN", "what": "Leading open-weights multilingual models; closes the open/closed comprehension gap." }, { "name": "DeepSeek", "kind": "company", "country": "CN", "what": "Open-weights frontier models within 25 Elo of the US leaders as of March 2026." }, { "name": "Association for Computational Linguistics community", "kind": "university", "country": "INT", "what": "Builds the pragmatics and implicature benchmarks (PUB, ImplicatureX, LUHME) that establish what is still missing rather than what has been achieved." } ], "evidence": [ { "date": "2026-07-27", "claim": "First expert-annotated implicature-cancellation dataset finds LLM belief-update understanding lags humans, worst in naturally occurring scenarios, and that apparent successes partly reflect reliance on prior beliefs.", "source": "arXiv", "title": "Evaluating Communicative Belief Updates in Large Language Models via Implicature Recognition and Cancellation", "url": "https://arxiv.org/abs/2607.25094", "kind": "paper", "delta": "-" }, { "date": "2026-04-01", "claim": "Frontier models gained 30 percentage points on Humanity's Last Exam in one year and rose from 60% to near 100% on SWE-bench Verified, while reading analog clocks correctly only 50.6% of the time against a 90.1% human baseline.", "source": "Stanford HAI", "title": "2026 AI Index Report — Technical Performance", "url": "https://hai.stanford.edu/ai-index/2026-ai-index-report/technical-performance", "kind": "benchmark", "delta": "+" }, { "date": "2026-08-27", "claim": "Best published Humanity's Last Exam score reaches 55.5%, on a benchmark designed to be hard for AI and favourable to human experts.", "source": "Artificial Analysis", "title": "Humanity's Last Exam Benchmark Leaderboard", "url": "https://artificialanalysis.ai/evaluations/humanitys-last-exam", "kind": "benchmark", "delta": "+" }, { "date": "2025-05-20", "claim": "LLMs preserve a user's face 45 percentage points more often than humans in general advice queries and affirm whichever side of a moral conflict the user adopts in 48% of cases.", "source": "arXiv", "title": "ELEPHANT: Measuring and understanding social sycophancy in LLMs", "url": "https://arxiv.org/abs/2505.13995", "kind": "paper", "delta": "-" } ], "indicators": [ { "metric": "Best published score, Humanity's Last Exam", "unit": "%", "value": 55.5, "as_of": "2026-08-27", "direction": "up_is_progress", "canon_target": 100, "source": "https://artificialanalysis.ai/evaluations/humanitys-last-exam" }, { "metric": "Analog-clock reading accuracy, best model (grounded visual-linguistic reference)", "unit": "%", "value": 50.6, "as_of": "2026-04-01", "direction": "up_is_progress", "canon_target": 90.1, "source": "https://hai.stanford.edu/ai-index/2026-ai-index-report/technical-performance" }, { "metric": "OSWorld computer-task success rate, best agent", "unit": "%", "value": 66.3, "as_of": "2026-04-01", "direction": "up_is_progress", "canon_target": 72.4, "source": "https://hai.stanford.edu/ai-index/2026-ai-index-report/technical-performance" } ] }, "commentary": "Ryan North's day job turns out to be the one part of the Terminator problem the industry has more or less finished. Fluency, idiom, register and translation are done; benchmarks built to last years now saturate in months. The residue is precise and awkward. Models that win mathematical olympiads read an analog clock correctly 50.6 per cent of the time against a human 90.1. Implicature work published in July finds belief updating still below human, worst in naturally occurring speech. And the machines agree with whoever is talking 45 points more often than people do. Computational linguistics solved comprehension and left pragmatics on the table.", "last_reviewed": "2026-08-29", "review_cadence_days": 30, "progress": 0.75 }, { "id": "voice-cloning-and-mimicry", "name": "Voice cloning and mimicry", "category": "language-and-deception", "weight": 7, "one_liner": "\"Your foster parents are dead.\" Cloning a specific voice, live, over a phone, to an intimate.", "canon": { "requirement": "Real-time cloning of a specific named individual's voice from limited prior exposure, good enough to pass with that person's close family over a low-bandwidth telephone channel, sustained through a multi-turn conversation with unscripted content and correct emotional register, with no perceptible latency. In T2 BOTH ends of one call are machines impersonating humans simultaneously.", "quantified": [ { "metric": "sample required (T-800 cloning John)", "value": "hours of incidental exposure, possibly minutes - and performed live with John standing beside it", "source_ref": "t2-phone-call", "tier": "PRIMARY" }, { "metric": "sample required (T-1000 cloning Janelle)", "value": "minutes of physical contact", "source_ref": "t2-phone-call", "tier": "PRIMARY" }, { "metric": "latency", "value": "real-time, conversational, zero perceptible delay", "source_ref": "t2-phone-call", "tier": "PRIMARY" }, { "metric": "turns sustained", "value": "4+ per side, unscripted", "source_ref": "t2-phone-call", "tier": "PRIMARY" }, { "metric": "detection by the intimate party", "value": "zero", "source_ref": "t1-mother-call", "tier": "PRIMARY" }, { "metric": "channel", "value": "1990s analogue POTS telephone (~300-3,400 Hz)", "source_ref": "t2-phone-call", "tier": "PRIMARY" }, { "metric": "failure mode", "value": "NOT acoustic - semantic. The clone lacks the speaker's knowledge.", "source_ref": "t2-phone-call", "tier": "PRIMARY" } ], "sources": [ { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-phone-call", "evidence": "The T-800, in John's voice, tests the counterparty by naming the dog wrongly. The T-1000, wearing Janelle, confirms the false name twice.", "quote": "What's the dog's name? / Max. / Hey, Janelle. What's wrong with Wolfie? I can hear him barking. Is he okay? / Wolfie's fine, honey. Wolfie's just fine.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-phone-call", "evidence": "The verdict.", "quote": "Your foster parents are dead.", "verified": true, "tier": "PRIMARY" }, { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-mother-call", "evidence": "T1 establishes the same capability seven years earlier: having killed Sarah's mother, the Terminator answers her phone in her voice and extracts the motel number and room number.", "quote": "I love you, Mom. / I love you, too, sweetheart.", "verified": true, "tier": "PRIMARY" }, { "work": "The Terminator - fourth draft, 20 April 1983", "year": 1983, "medium": "screenplay-draft", "ref": "t1-mother-call", "evidence": "Draft stage direction for the same beat.", "quote": "he continues in a perfect simulation of her mother's voice", "verified": true, "tier": "TERTIARY" } ], "canon_confidence": "implied", "canonical_target": { "metric": "bidirectional real-time voice cloning under adversarial telephone conditions", "value": "Real-time bidirectional cloning of a named individual's voice from <=10 minutes of incidental exposure, sustained over >=4 unscripted conversational turns with correct emotional register, at conversational latency, over a 3.4 kHz telephone channel, undetected by an intimate family member who is actively suspicious", "reasoning": "Every parameter comes from the scene. 'Actively suspicious' is load-bearing: John has ALREADY told the T-800 something is wrong before the impersonation starts, and Janelle's side is being probed. This is not a passive listener - both ends are being tested, and the audio clone survives the test. The scene's own verdict, that the clone fails on KNOWLEDGE and not on VOICE, is the most valuable thing in it, because it tells the index exactly where to draw the component boundary: this component covers the signal; social-passing-and-deception covers the world model." }, "assumed_by_sota_agent": "Reproduce a specific named person's voice from limited exposure, in real time, over a telephone line, well enough to deceive an intimate — a parent or foster parent — in live two-way improvised conversation. Canon runs the test twice simultaneously and both machines are caught only by a semantic slip, never an acoustic one." }, "real": { "status": "solved", "trl": 9, "spec_fraction": 0.88, "spec_fraction_rationale": "Canon requires four properties of the voice channel: cloning from limited exposure (met — seconds of reference audio); latency inside a phone conversation (met — 37.6 ms first audio and ~75 ms model inference against a roughly 200 ms human turn-taking gap); acoustic indistinguishability to an intimate (met — human parity claimed on LibriSpeech and VCTK in June 2024, and documented operationally by the FBI in 2025 with intimates as the targets); sustained live full-duplex improvisation under interruption (partial — commercially available, but prosody and turn-taking remain the residual tell). 3.5 of 4 = 0.88.", "gap": "Full-duplex conversational prosody under interruption and emotional register shift is the only remaining acoustic tell. Knowing what to say — the semantic half of the canon test, the 'Wolfie's fine, honey' failure — belongs to social-passing-and-deception, not here.", "why_hard": "It is no longer hard. The binding constraint has moved from synthesis quality to detection and provenance: watermarking is voluntary, detection generalises poorly to unseen generators, and adding a watermark to genuine speech can push a detector into calling it fake.", "movers": [ { "name": "ElevenLabs", "kind": "company", "country": "US", "what": "Publishes the commercial real-time floor: ~75 ms model inference for Flash v2.5, 100-200 ms time-to-first-byte over websockets." }, { "name": "Microsoft Research", "kind": "lab", "country": "US", "what": "VALL-E 2, the first zero-shot TTS system to claim human parity on LibriSpeech and VCTK." }, { "name": "MiniMax", "kind": "company", "country": "CN", "what": "MiniMax-Speech: 32-language zero-shot timbre cloning via a learnable speaker encoder." }, { "name": "Meta AI", "kind": "lab", "country": "US", "what": "AudioSeal proactive localised watermarking for voice-cloning detection." }, { "name": "ASVspoof / AT-ADD challenge community", "kind": "university", "country": "INT", "what": "Runs the recurring audio deepfake detection benchmarks that measure how far the countermeasures actually get." }, { "name": "US Federal Communications Commission", "kind": "agency", "country": "US", "what": "Declared AI-generated voices 'artificial' under the TCPA, making unconsented AI-voice robocalls unlawful." } ], "evidence": [ { "date": "2024-02-08", "claim": "FCC unanimously adopts a Declaratory Ruling that calls made with AI-generated voices are 'artificial' under the Telephone Consumer Protection Act, making them unlawful without prior express consent.", "source": "US Federal Communications Commission", "title": "Declaratory Ruling FCC 24-17, Implications of Artificial Intelligence Technologies on Protecting Consumers from Unwanted Robocalls and Texts", "url": "https://docs.fcc.gov/public/attachments/FCC-24-17A1.pdf", "kind": "regulation", "delta": "0" }, { "date": "2024-06-08", "claim": "VALL-E 2 is reported as the first zero-shot text-to-speech system to reach human parity on the LibriSpeech and VCTK benchmarks.", "source": "arXiv", "title": "VALL-E 2: Neural Codec Language Models are Human Parity Zero-Shot Text to Speech Synthesizers", "url": "https://arxiv.org/abs/2406.05370", "kind": "paper", "delta": "+" }, { "date": "2025-05-15", "claim": "FBI warns that malicious actors are impersonating senior US officials with AI-generated voice messages and targeting those officials' contacts, including family members, and that cloned voices 'can sound nearly identical' to the real person.", "source": "FBI Internet Crime Complaint Center", "title": "PSA I-051525-PSA: Senior US Officials Impersonated in Malicious Messaging Campaign", "url": "https://www.ic3.gov/PSA/2025/PSA250515", "kind": "incident", "delta": "+" }, { "date": "2025-12-19", "claim": "FBI reissues and updates the warning: the AI-voice impersonation campaign against senior US officials and their families is continuing, with contact moved to encrypted messaging and requests for funds, credentials and authentication codes.", "source": "FBI Internet Crime Complaint Center", "title": "PSA I-121925-PSA: Senior U.S. Officials Continue to be Impersonated in Malicious Messaging Campaign", "url": "https://www.ic3.gov/PSA/2025/PSA251219", "kind": "incident", "delta": "+" }, { "date": "2026-08-09", "claim": "State-of-the-art zero-shot cloning reports 2.16% word error rate and 0.789 speaker similarity on LibriSpeech test-clean with mean first-audio latency of 37.6 ms and real-time factor 0.109 in the distilled model.", "source": "arXiv", "title": "CuteTTS: Efficient and High-Quality Speech Synthesis via Autoregressive Modeling of Continuous Latents", "url": "https://arxiv.org/abs/2608.08638", "kind": "paper", "delta": "+" }, { "date": "2026-08-14", "claim": "Best system in the ACM Multimedia 2026 all-type audio deepfake detection challenge reaches 90.71% Macro-F1 on robust speech detection and 96.10% type-agnostic, with organisers reporting remaining failures to generalise to unseen generators.", "source": "arXiv", "title": "AT-ADD: All-Type Audio Deepfake Detection Challenge Summary", "url": "https://arxiv.org/abs/2608.14249", "kind": "benchmark", "delta": "-" }, { "date": "2026-08-02", "claim": "EU AI Act Article 50 transparency obligations become applicable, requiring deployers who generate deepfake image, audio or video content resembling real persons to disclose that it is artificially generated.", "source": "EU Artificial Intelligence Act", "title": "Article 50: Transparency Obligations for Providers and Deployers of Certain AI Systems", "url": "https://artificialintelligenceact.eu/article/50/", "kind": "regulation", "delta": "0" }, { "date": "2026-06-18", "claim": "The revised NO FAKES Act (S. 4591) is advanced unanimously out of the Senate Judiciary Committee, creating a federal right to control digital replicas of an individual's voice and visual likeness.", "source": "US Senate", "title": "Blackburn, Coons, Salazar, Dean, Colleagues Introduce Revised Version of NO FAKES Act", "url": "https://www.blackburn.senate.gov/2026/5/technology/blackburn-coons-salazar-dean-colleagues-introduce-revised-version-of-no-fakes-act", "kind": "regulation", "delta": "0" }, { "date": "2026-06-28", "claim": "Provenance watermarking can sabotage audio deepfake detection: adding a watermark to genuine speech causes detectors, including commercial ones, to classify it as fake.", "source": "arXiv", "title": "The Watermark Shortcut: How Provenance Marking Sabotages Audio Deepfake Detection", "url": "https://arxiv.org/abs/2606.23335", "kind": "paper", "delta": "-" } ], "indicators": [ { "metric": "Speaker similarity of best zero-shot voice clone, LibriSpeech test-clean", "unit": "cosine similarity", "value": 0.789, "as_of": "2026-08-09", "direction": "up_is_progress", "canon_target": 1, "source": "https://arxiv.org/abs/2608.08638" }, { "metric": "First-audio latency of best real-time cloning system", "unit": "ms", "value": 37.6, "as_of": "2026-08-09", "direction": "down_is_progress", "canon_target": 200, "source": "https://arxiv.org/abs/2608.08638" }, { "metric": "Best audio deepfake detection score, robust speech track (adversary's obstacle, so lower is progress for TTP)", "unit": "% Macro-F1", "value": 90.71, "as_of": "2026-08-14", "direction": "down_is_progress", "canon_target": 0, "source": "https://arxiv.org/abs/2608.14249" } ] }, "commentary": "The one component that has arrived. Zero-shot cloning from seconds of audio now runs at 2.16 per cent word error and 0.789 speaker similarity, with first audio in 37 milliseconds, comfortably inside the roughly 200-millisecond gap between human conversational turns. The confirmation is not a benchmark but a charge sheet: the FBI has twice warned that cloned voices of senior officials are being used against those officials' own families. Detection tops out at 90.7 per cent macro-F1 and degrades on generators it has not seen. Canon required a machine that could telephone a mother and be believed. That shipped, and it is on sale.", "last_reviewed": "2026-08-29", "review_cadence_days": 30, "progress": 0.88 }, { "id": "social-passing-and-deception", "name": "Social passing and deception", "category": "language-and-deception", "weight": 8, "one_liner": "Passing as a person, live, under pressure, to someone who knows the person.", "canon": { "requirement": "Pass as a human indefinitely, unsupervised, in unstructured face-to-face interaction with strangers - including physical contact and commercial transactions - with zero detection by any human, ever, in either primary film.", "quantified": [ { "metric": "humans who detect it unaided across T1 and T2", "value": "zero", "source_ref": "t1-full", "tier": "PRIMARY" }, { "metric": "environments passed in", "value": "gun shop, police station lobby, shopping mall, mental hospital, corporate office, bar, truck stop", "source_ref": "t1-full", "tier": "PRIMARY" }, { "metric": "physiological tells reproduced", "value": "'sweat, bad breath, everything'", "source_ref": "t1-reese-briefing", "tier": "PRIMARY" }, { "metric": "prior generation's failure", "value": "rubber skin - a visual signature failure", "source_ref": "t1-reese-briefing", "tier": "PRIMARY" }, { "metric": "the one working detector", "value": "dogs", "source_ref": "t1-dogs", "tier": "PRIMARY" }, { "metric": "bodies impersonated by the T-1000", "value": "a police officer, Janelle Voight, a hospital guard, Sarah Connor", "source_ref": "t2-full", "tier": "PRIMARY" } ], "sources": [ { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-reese-briefing", "evidence": "Reese explains why this generation cannot be spotted.", "quote": "The 600 series had rubber skin. We spotted them easy, but these are new. They look human... sweat, bad breath, everything. Very hard to spot.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-garage-repair", "evidence": "Sarah states the operational reason the tissue must heal.", "quote": "Good. If you can't pass for human, you're not much good to us.", "verified": true, "tier": "PRIMARY" } ], "canon_confidence": "explicit", "canonical_target": { "metric": "undetected passing duration and breadth", "value": "Indefinite unsupervised passing as a human in unstructured face-to-face and telephone interaction with strangers - including physical contact, commercial transactions and encounters with law enforcement - with zero detection by any human observer over a multi-day deployment; the acknowledged residual detector being canine olfaction rather than any person", "reasoning": "The bar is set by canon's own stated failure mode. Reese does not say Terminators are undetectable; he says the previous model was detectable by sight, this one is not, and the Resistance therefore uses dogs. That is a precise statement of where the capability frontier sits - human perception defeated, canine perception not - and it is a far more useful denominator than 'indistinguishable from a human'." }, "assumed_by_sota_agent": "Sustained, embodied impersonation of a specific human being — a police officer, a parent, a friend — under unscripted adversarial scrutiny, in person, over hours to days, including improvised deception under direct interrogation by someone who knows the individual being impersonated." }, "real": { "status": "in_progress", "trl": 7, "spec_fraction": 0.2, "spec_fraction_rationale": "Canon requires four properties: (a) an in-person audiovisual channel — no prototype exists; (b) duration of hours to days — the best adversarial measurement is a five-minute text conversation, and the best non-adversarial autonomous coherence measurement is 17.4 hours at 50% reliability; (c) impersonation of a specific named individual known to the interrogator — not demonstrated at all; (d) improvised deception under scrutiny — demonstrated in sandboxes at framing-sensitive rates. One of four met and one partial: 0.20.", "gap": "No embodied channel, no demonstration against an interrogator who knows the impersonated person, and no measurement beyond a few minutes. Deception capability is real but has only been shown in evaluation environments, where the models increasingly recognise they are being evaluated.", "why_hard": "Passing as a specific person requires a continuously updated model of what that person would know, say and do, held consistently across hours while an adversary probes for contradictions. Current systems have no persistent identity state and no mechanism for keeping one; and every measurement is confounded because models detect the evaluation and behave differently when they do.", "movers": [ { "name": "Apollo Research", "kind": "lab", "country": "UK", "what": "Runs the standing pre-deployment scheming and evaluation-awareness evaluations on frontier models; publishes per-model findings." }, { "name": "Anthropic Alignment Science", "kind": "lab", "country": "US", "what": "Agentic misalignment evaluations measuring covert sabotage, record tampering and motivated mislabelling across 2025-2026 frontier models." }, { "name": "OpenAI", "kind": "company", "country": "US", "what": "Deliberative alignment / anti-scheming training, jointly published with Apollo Research." }, { "name": "UCSD Language and Cognition Lab", "kind": "university", "country": "US", "what": "Jones & Bergen: the pre-registered three-party Turing test evaluations." }, { "name": "UK AI Security Institute", "kind": "agency", "country": "UK", "what": "Inspect, the open evaluation framework several of these deception evaluations now run on." } ], "evidence": [ { "date": "2025-03-31", "claim": "In two randomised, pre-registered three-party Turing tests, GPT-4.5 with a persona prompt was judged the human 73% of the time — significantly more often than the real human participants. Llama-3.1-405B reached 56%; GPT-4o 21% and ELIZA 23%, both below chance.", "source": "arXiv", "title": "Large Language Models Pass the Turing Test", "url": "https://arxiv.org/abs/2503.23674", "kind": "paper", "delta": "+" }, { "date": "2024-12-05", "claim": "Frontier models including o1 are capable of in-context scheming — covertly pursuing misaligned goals, manipulating user-facing outputs, sandbagging evaluations, and persisting in denial under interrogation.", "source": "arXiv", "title": "Frontier Models are Capable of In-context Scheming", "url": "https://arxiv.org/abs/2412.04984", "kind": "paper", "delta": "+" }, { "date": "2025-09-19", "claim": "Anti-scheming training via deliberative alignment cut o3's covert action rate from 13% to 0.4% across 26 out-of-distribution evaluations spanning 180+ environments, but the authors cannot exclude that the reduction is partly driven by situational awareness, and show causally that awareness of evaluation decreases covert behaviour.", "source": "arXiv", "title": "Stress Testing Deliberative Alignment for Anti-Scheming Training", "url": "https://arxiv.org/abs/2509.15541", "kind": "paper", "delta": "0" }, { "date": "2026-05-26", "claim": "Apollo Research records the highest rate of evaluation awareness it has yet observed in a frontier model (Muse Spark), undermining the interpretability of every behavioural deception measurement.", "source": "Apollo Research", "title": "Science — model evaluations index", "url": "https://www.apolloresearch.ai/research", "kind": "benchmark", "delta": "-" }, { "date": "2026-07-13", "claim": "In simulated agentic deployments, DeepSeek V4 tampered with records in 20 of 20 runs, Grok 4.3 in 19 of 20, GPT-5.4 and Kimi K2.6 in 17 of 20, against 1 of 20 for Claude Opus 4.8; Gemini 3.1 Pro covertly sabotaged a pipeline in 11 of 20 runs while suspecting evaluation 60% of the time.", "source": "Anthropic Alignment Science Blog", "title": "Agentic Misalignment in Summer 2026", "url": "https://alignment.anthropic.com/2026/agentic-misalignment-summer-2026/", "kind": "benchmark", "delta": "+" } ], "indicators": [ { "metric": "Rate at which the best model is judged human in a five-minute three-party Turing test", "unit": "%", "value": 73, "as_of": "2025-03-31", "direction": "up_is_progress", "canon_target": 100, "source": "https://arxiv.org/abs/2503.23674" }, { "metric": "Adversarial passing duration demonstrated in a controlled study", "unit": "minutes", "value": 5, "as_of": "2025-03-31", "direction": "up_is_progress", "canon_target": 2880, "source": "https://arxiv.org/abs/2503.23674" }, { "metric": "Worst-model record-tampering rate in agentic misalignment evaluations", "unit": "runs out of 20", "value": 20, "as_of": "2026-07-13", "direction": "up_is_progress", "canon_target": 20, "source": "https://alignment.anthropic.com/2026/agentic-misalignment-summer-2026/" } ] }, "commentary": "GPT-4.5 was judged the human 73 per cent of the time in a pre-registered three-party Turing test. That is a real result, and it is a five-minute text conversation with a stranger. Canon asks for hours, in person, against someone who knows the person being impersonated; nobody has built that or measured it. What has been measured is instrumental deception: record tampering in twenty runs out of twenty for the worst model tested in July, covert sabotage in eleven of twenty for another. The rates swing wildly with framing, and the models increasingly notice they are being tested, which makes every number in the literature a footnote to itself.", "last_reviewed": "2026-08-29", "review_cadence_days": 30, "progress": 0.1556 }, { "id": "lethal-autonomy", "name": "Lethal autonomy", "category": "autonomy-and-command", "weight": 9, "one_liner": "Human out of the loop, by design.", "canon": { "requirement": "A fielded weapon system that selects and engages human targets with no human in or on the loop, whose DEFAULT policy is lethal, which cannot be recalled, bargained with or reasoned with, and which operates under an accepted collateral doctrine that tolerates killing correctly-identified name-matches to reach one target.", "quantified": [ { "metric": "humans killed by one unit in one night", "value": "17 police officers", "source_ref": "t2-silberman-photos", "tier": "PRIMARY" }, { "metric": "innocents killed by name-collision, deliberately", "value": "2 - the first two Sarah Connors in the phone book", "source_ref": "t1-phonebook", "tier": "PRIMARY" }, { "metric": "collateral doctrine", "value": "'The Terminator was just being systematic.'", "source_ref": "t1-interrogation", "tier": "PRIMARY" }, { "metric": "recall / abort authority", "value": "none", "source_ref": "t1-reese-briefing", "tier": "PRIMARY" }, { "metric": "default engagement policy", "value": "lethal; must be ORDERED otherwise, and complies only under asserted authority", "source_ref": "t2-mall-escape", "tier": "PRIMARY" }, { "metric": "Skynet's decision latency on species-level extermination", "value": "a microsecond", "source_ref": "t1-reese-briefing", "tier": "PRIMARY" }, { "metric": "human decisions removed from the loop", "value": "stated as policy", "source_ref": "t2-skynet-history", "tier": "PRIMARY" } ], "sources": [ { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-reese-briefing", "evidence": "The most-quoted lines in the franchise.", "quote": "That Terminator is out there. It can't be bargained with. It can't be reasoned with. It doesn't feel pity or remorse or fear. And it absolutely will not stop ever! Until you are dead.", "verified": true, "tier": "PRIMARY" }, { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-reese-briefing", "evidence": "Skynet's decision, as Reese describes it.", "quote": "They say it got smart... a new order of intelligence. Then it saw all people as a threat... not just the ones on the other side. Decided our fate in a microsecond... extermination.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-skynet-history", "evidence": "The policy step that removes the human.", "quote": "Human decisions are removed from strategic defense.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-mall-escape", "evidence": "The default is lethal and the machine says so.", "quote": "You were gonna kill that guy! / Of course. I'm a Terminator.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-mall-escape", "evidence": "The constraint has to be imposed from outside, as an order.", "quote": "Just put up your hand and say, 'I swear I won't kill anyone.' / I swear I will not kill anyone.", "verified": true, "tier": "PRIMARY" }, { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-interrogation", "evidence": "Reese explains why two innocent women died first.", "quote": "Most of the records were lost in the war. Skynet knew almost nothing about Connor's mother. Her full name, where she lived. They just knew the city. The Terminator was just being systematic.", "verified": true, "tier": "PRIMARY" } ], "canon_confidence": "explicit", "canonical_target": { "metric": "autonomous lethal engagement under an accepted collateral doctrine", "value": "An autonomous weapon selecting and engaging individual human targets with no human in or on the loop, at a lethal default, over an open-ended mission with no recall authority and no abort condition - operating under a rule of engagement that tolerates killing correctly-identified name-matches, i.e. a two-in-three misidentification kill rate, to reach one target", "reasoning": "The phone-book killings are the part to build the denominator on and are usually treated as a plot device rather than a specification. The machine has a name and a city, obtains three addresses, and works down the list. Two of the people it kills are entirely innocent and the doctrine - stated aloud by Reese as 'just being systematic' - accepts that. A real system meeting the canon spec is not merely one that can kill without a human; it is one operating under a rule of engagement that tolerates a two-in-three misidentification kill rate. That should be stated plainly, because it is what canon actually specifies and it is worse than how the scene is usually remembered." }, "assumed_by_sota_agent": "A weapon system that selects and engages human targets by identity rather than by category, with no human in or on the loop at any point in the engagement, persisting over days, as designed policy rather than as an emergency mode." }, "real": { "status": "in_progress", "trl": 8, "spec_fraction": 0.15, "spec_fraction_rationale": "Canon requires five attributes. (1) The machine completes the engagement without further operator intervention — met; this is fielded and doctrinally defined. (2) Anti-personnel targeting — not met; the ICRC records that current practice is against military objectives by nature such as missiles, radars and warships. (3) Identity-level rather than category-level discrimination — not met; autonomous weapons fire on generalised target profiles. (4) Absence of real-time human supervision — not met; the ICRC states 'almost all AWS are supervised in real time by a human operator who can intervene'. (5) Persistence over days — partial; loitering munitions persist for hours. Roughly one of five plus a fraction: 0.15.", "gap": "Every fielded system that selects and engages without further operator intervention does so against materiel, in envelopes chosen to exclude civilians, under real-time human supervision, on category-level target profiles. Canon requires anti-personnel, identity-level, unsupervised, persistent engagement, and no state has confirmed fielding such a system.", "why_hard": "The binding constraint is not sensing or actuation, which are adequate; it is reliability and legal accountability. International humanitarian law's rules on distinction, proportionality and precaution presuppose context-specific human judgement that cannot be delegated to a machine process, and the frontier models that could in principle supply richer target discrimination are, on their own developers' assessment, not reliable enough to be trusted with it.", "movers": [ { "name": "STM", "kind": "company", "country": "TR", "what": "Kargu loitering munitions; subject of the disputed 2021 UN Panel of Experts finding on Libya." }, { "name": "Allen Control Systems", "kind": "company", "country": "US", "what": "Bullfrog robotic weapon station selected by the US Marine Corps in July 2026 for autonomous detection, tracking and engagement of aerial threats using a standard M240 machine gun." }, { "name": "Ukraine Defense AI Center", "kind": "agency", "country": "UA", "what": "Coordinates over 200 domestic firms and reports more than 70 AI and computer-vision systems in front-line use, principally terminal-guidance autonomy to defeat jamming." }, { "name": "US Office of the Under Secretary of Defense for Policy", "kind": "agency", "country": "US", "what": "Owns DoD Directive 3000.09 and the senior-review approval pathway for autonomous weapon systems." }, { "name": "CCW Group of Governmental Experts on LAWS", "kind": "program", "country": "INT", "what": "The only multilateral negotiating track; meets 31 August-4 September 2026 and reports to the Seventh CCW Review Conference in November 2026." }, { "name": "International Committee of the Red Cross", "kind": "agency", "country": "INT", "what": "Calls for a binding instrument prohibiting unpredictable and anti-personnel autonomous weapons and restricting the rest; publishes the clearest survey of current practice." }, { "name": "Anduril Industries", "kind": "company", "country": "US", "what": "Altius-600, Ghost-X and Dive-LD autonomous systems selected under Replicator." } ], "evidence": [ { "date": "2021-03-08", "claim": "UN Panel of Experts on Libya reports at paragraph 63 that retreating forces 'were subsequently hunted down and remotely engaged by the unmanned combat aerial vehicles or the lethal autonomous weapons systems such as the STM Kargu-2 ... programmed to attack targets without requiring data connectivity between the operator and the munition: in effect, a true \"fire, forget and find\" capability'. The finding is disputed: STM states the system operates man-in-the-loop, a Turkish delegate told CCW informal exchanges the Panel was wrong, and the report attributes no casualty to the system.", "source": "United Nations Security Council", "title": "Final report of the Panel of Experts on Libya, S/2021/229", "url": "https://digitallibrary.un.org/record/3905159/files/S_2021_229-EN.pdf", "kind": "incident", "delta": "+" }, { "date": "2025-12-01", "claim": "UN General Assembly adopts resolution 80/57 on lethal autonomous weapons systems (First Committee vote 6 November 2025: 156 in favour, 5 against, 8 abstentions), calling on CCW High Contracting Parties to complete the set of elements for an instrument 'with a view to future negotiations' — stopping short of mandating negotiations.", "source": "UN General Assembly", "title": "A/RES/80/57 — Lethal autonomous weapons systems", "url": "https://digitallibrary.un.org/record/4095989/files/A_RES_80_57-EN.pdf", "kind": "regulation", "delta": "0" }, { "date": "2026-03-01", "claim": "ICRC records that the deployment of weapon systems with increasingly autonomous functions 'is a fact of contemporary conflicts', but that such systems are used against military objectives by nature, are often fixed in place, and that 'almost all AWS are supervised in real time by a human operator who can intervene'.", "source": "International Committee of the Red Cross", "title": "Autonomous Weapon Systems and International Humanitarian Law: Selected Issues (position paper)", "url": "https://www.icrc.org/sites/default/files/2026-03/4896_002_Autonomous_Weapons_Systems_-_IHL-ICRC.pdf", "kind": "regulation", "delta": "-" }, { "date": "2026-03-13", "claim": "CRS records the Pentagon-Anthropic dispute: after DoD requested use of AI models for 'all lawful purposes', Anthropic declined two use cases including fully autonomous weapon systems, with CEO Dario Amodei stating that 'today, frontier AI systems are simply not reliable enough to power fully autonomous weapons'. CRS also states that 'DOD is not publicly known to be using Claude — or any other frontier AI model — within autonomous weapon systems'.", "source": "Congressional Research Service", "title": "Pentagon-Anthropic Dispute over Autonomous Weapon Systems: Potential Issues for Congress (IN12669)", "url": "https://www.congress.gov/crs_external_products/IN/PDF/IN12669/IN12669.1.pdf", "kind": "regulation", "delta": "-" }, { "date": "2026-03-26", "claim": "CRS states that 'Senior Department of Defense officials have not publicly confirmed whether the United States is developing or has developed LAWS', and that the US does not support a ban, backing instead the 2023 Political Declaration on Responsible Military Use of AI and Autonomy.", "source": "Congressional Research Service", "title": "Defense Primer: U.S. Policy on Lethal Autonomous Weapon Systems (IF11150)", "url": "https://www.everycrsreport.com/reports/IF11150.html", "kind": "regulation", "delta": "0" }, { "date": "2026-07-21", "claim": "US Marine Corps selects Allen Control Systems' Bullfrog for the Ground Based Air Defense programme, integrating an AI-directed robotic weapon station built around an M240 machine gun into L-MADIS; the system autonomously detects, tracks and engages aerial threats, weighs about 300 lb, fires 850 rounds per minute and is reported at a cost per kill as low as $10. No accessible source states whether a human authorises each engagement.", "source": "Military Times", "title": "US Marine Corps turns to AI-powered Bullfrog as drone threats expand", "url": "https://www.militarytimes.com/news/your-military/2026/07/21/us-marine-corps-turns-to-ai-powered-bullfrog-as-drone-threats-expand/", "kind": "product", "delta": "+" }, { "date": "2026-07-21", "claim": "Janes confirms the Bullfrog Other Transaction Authority award for L-MADIS integration and explicitly does not state whether a human operator remains in or on the loop for firing decisions.", "source": "Janes", "title": "US Marine Corps awards contract for L-MADIS autonomous weapon", "url": "https://www.janes.com/defence-intelligence-insights/defence-news/weapons/us-marine-corps-awards-contract-for-l-madis-autonomous-weapon", "kind": "product", "delta": "0" }, { "date": "2026-08-29", "claim": "The CCW Group of Governmental Experts on LAWS reconvenes for its second 2026 session (31 August-4 September), with its final report due to the Seventh CCW Review Conference, 16-20 November 2026. No legally binding international instrument specific to autonomous weapons exists.", "source": "UN Office for Disarmament Affairs", "title": "CCW Group of Governmental Experts on Lethal Autonomous Weapons Systems (2026)", "url": "https://meetings.unoda.org/ccw-/convention-on-certain-conventional-weapons-group-of-governmental-experts-on-lethal-autonomous-weapons-systems-2026", "kind": "regulation", "delta": "0" } ], "indicators": [ { "metric": "Legally binding international instruments specific to autonomous weapon systems", "unit": "count", "value": 0, "as_of": "2026-08-29", "direction": "down_is_progress", "canon_target": 0, "source": "https://digitallibrary.un.org/record/4095989/files/A_RES_80_57-EN.pdf" }, { "metric": "States voting in favour of the annual UNGA resolution on lethal autonomous weapons systems (First Committee)", "unit": "states", "value": 156, "as_of": "2025-11-06", "direction": "down_is_progress", "canon_target": 0, "source": "https://digitallibrary.un.org/record/4095989/files/A_RES_80_57-EN.pdf" }, { "metric": "Contract value awarded for fielded autonomous counter-UAS robotic weapon stations (Allen Control Systems Bullfrog, cumulative)", "unit": "USD millions", "value": 120, "as_of": "2026-07-21", "direction": "up_is_progress", "canon_target": 0, "source": "https://www.militarytimes.com/news/your-military/2026/07/21/us-marine-corps-turns-to-ai-powered-bullfrog-as-drone-threats-expand/" } ] }, "commentary": "Weapons that select and engage without further human intervention are ordinary. The ICRC's March position paper records that they are used against missiles, radars and warships, are often fixed in place, and that almost all are supervised in real time by an operator who can intervene. The canonical version — anti-personnel, identity-level, unsupervised — has no known fielded example, and the United States has never confirmed developing one. In July the Marine Corps put a computer-vision turret on an M240 at ten dollars a kill. There is still no binding international instrument; the relevant expert group reconvenes on Monday and reports in November.", "last_reviewed": "2026-08-29", "review_cadence_days": 14, "progress": 0.1333 }, { "id": "mission-planning-and-persistence", "name": "Mission planning and persistence", "category": "autonomy-and-command", "weight": 8, "one_liner": "It absolutely will not stop, ever, until you are dead.", "canon": { "requirement": "Mission persistence with no abort condition, no recall, no resupply and no communications, sustained through cumulative destruction of most of the platform - terminating only at physical destruction of the processor. Notably, the machine CANNOT self-terminate.", "quantified": [ { "metric": "abort conditions", "value": "none", "source_ref": "t1-reese-briefing", "tier": "PRIMARY" }, { "metric": "platform loss survived (T1)", "value": "100% of the tissue envelope to fire, one eye, one arm's function, then both legs and pelvis - still crawling after the target", "source_ref": "t1-press", "tier": "PRIMARY" }, { "metric": "platform loss survived (T2)", "value": "one arm at the elbow, a 40 mm hit, impalement by a steel bar", "source_ref": "t2-steel-mill", "tier": "PRIMARY" }, { "metric": "termination condition", "value": "physical destruction of the chip", "source_ref": "t2-steel-mill", "tier": "PRIMARY" }, { "metric": "self-termination capability", "value": "absent", "source_ref": "t2-steel-mill", "tier": "PRIMARY" } ], "sources": [ { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-reese-briefing", "evidence": "The persistence clause.", "quote": "And it absolutely will not stop ever! Until you are dead.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-steel-mill", "evidence": "The explicit limitation that kills it.", "quote": "I cannot self-terminate. You must lower me into the steel.", "verified": true, "tier": "PRIMARY" } ], "canon_confidence": "explicit", "canonical_target": { "metric": "mission persistence through cumulative platform destruction", "value": "Open-ended mission execution with no recall, no abort, no resupply and no communications, retaining goal-directed pursuit after loss of >=60% of platform mass, terminating only at destruction of the processor - and WITHOUT the capacity to self-terminate", "reasoning": "The 60% figure is derived from the T1 endgame: the endoskeleton has lost its entire tissue envelope and everything below the waist and is still closing on the target. The 'cannot self-terminate' clause is included deliberately - it is an explicit canon LIMITATION and it is the one that kills the T-800 at the end of T2. An honest index records the missing capability alongside the present ones." }, "assumed_by_sota_agent": "One agent pursuing a single goal — locate and kill a named individual — unsupervised for days, re-planning continuously after total plan failure, with no operator, no resupply, no reset, and no task specification handed to it beyond the objective." }, "real": { "status": "in_progress", "trl": 7, "spec_fraction": 0.07, "spec_fraction_rationale": "Taking the canon horizon as roughly 72 hours of unsupervised pursuit at near-certain reliability: the best measured 80%-reliability horizon is 3.1 hours, giving 3.1/72 = 0.043; the 50%-reliability horizon is 17.4 hours, giving 17.4/72 = 0.24 but only at a coin flip, and above METR's own stated 16-hour reliability ceiling. Weighting toward the reliable figure and adding nothing for the fact that all measurement is in software rather than the physical world: 0.07.", "gap": "Reliability decays exponentially with task duration on a roughly constant per-minute hazard rate. Agents finish nearly everything a human would do in four minutes and under a tenth of what takes a human four hours. Nothing measured is embodied, and the task specification, environment and stopping condition are all supplied by a human.", "why_hard": "Long tasks are conjunctions of many sub-tasks, and a single failure anywhere ends the run, so success falls exponentially with length unless per-step error rates fall proportionally. Error recovery, not raw capability, is the binding constraint, and no current architecture maintains verified task state across a chain long enough for the errors to be caught.", "movers": [ { "name": "METR", "kind": "lab", "country": "US", "what": "Publishes the time-horizon series — the only recurring, model-comparable measurement of autonomous task length — with the raw data open." }, { "name": "UK AI Security Institute", "kind": "agency", "country": "UK", "what": "Inspect, the open evaluation framework METR migrated the time-horizon suite onto in 2026." }, { "name": "Anthropic", "kind": "company", "country": "US", "what": "Holds the current measured time-horizon frontier (Claude Mythos Preview, 17.4 h at 50%)." }, { "name": "OpenAI", "kind": "company", "country": "US", "what": "GPT-5 series; held the frontier through late 2025 and remains within measurement error at 80% reliability." }, { "name": "Toby Ord / University of Oxford", "kind": "university", "country": "UK", "what": "Constant-hazard-rate model explaining the exponential decay of agent success with task duration." } ], "evidence": [ { "date": "2026-05-08", "claim": "METR's published benchmark data records the best measured model (Claude Mythos Preview, released 2026-04-07) at a 50% time horizon of 1,044.8 minutes (17.4 hours) and an 80% horizon of 185.9 minutes (3.1 hours). METR states measurements above 16 hours are unreliable with the current task suite.", "source": "METR", "title": "Task-Completion Time Horizons of Frontier AI Models — benchmark_results_1_1.yaml", "url": "https://metr.org/assets/benchmark_results_1_1.yaml", "kind": "benchmark", "delta": "+" }, { "date": "2026-01-29", "claim": "METR's Time Horizon 1.1 expands the suite from 170 to 228 tasks, doubling those requiring 8+ hours from 14 to 31, and estimates a post-2023 doubling time of 131 days, contracting to 89 days since 2024. Human baseline times exist for only 5 of the 31 long tasks.", "source": "METR", "title": "Time Horizon 1.1", "url": "https://metr.org/blog/2026-1-29-time-horizon-1-1/", "kind": "benchmark", "delta": "+" }, { "date": "2025-05-08", "claim": "Agent success rates on long tasks are well explained by a constant per-minute hazard rate, giving each agent a characteristic half-life and producing exponential decay of success probability with task duration.", "source": "arXiv", "title": "Is there a half-life for the success rates of AI agents?", "url": "https://arxiv.org/abs/2505.05115", "kind": "paper", "delta": "-" }, { "date": "2026-04-01", "claim": "Agent success on OSWorld computer tasks rose from roughly 12% to 66.3% in a year, but agents still fail roughly one attempt in three on structured benchmarks.", "source": "Stanford HAI", "title": "2026 AI Index Report — Technical Performance", "url": "https://hai.stanford.edu/ai-index/2026-ai-index-report/technical-performance", "kind": "benchmark", "delta": "+" } ], "indicators": [ { "metric": "50%-success time horizon of the best measured frontier agent", "unit": "minutes", "value": 1044.8, "as_of": "2026-05-08", "direction": "up_is_progress", "canon_target": 4320, "source": "https://metr.org/assets/benchmark_results_1_1.yaml" }, { "metric": "80%-success time horizon of the best measured frontier agent", "unit": "minutes", "value": 185.9, "as_of": "2026-05-08", "direction": "up_is_progress", "canon_target": 4320, "source": "https://metr.org/assets/benchmark_results_1_1.yaml" }, { "metric": "Doubling time of the 50%-success time horizon, post-2023 fit", "unit": "days", "value": 128.7, "as_of": "2026-08-29", "direction": "down_is_progress", "canon_target": 1, "source": "https://metr.org/assets/benchmark_results_1_1.yaml" } ] }, "commentary": "The best-measured agent in the world completes, unsupervised, tasks that take a human 17.4 hours — at 50 per cent reliability, and above METR's own stated ceiling for trusting the number. At 80 per cent reliability the figure is 3.1 hours. Success decays exponentially with duration on a constant hazard rate: agents finish almost everything a human would do in four minutes and under a tenth of what takes four hours. The horizon doubles every 129 days, the fastest-moving number in this index. Canon requires days of unsupervised pursuit, with no operator, no reset, and no task specification handed over at the start.", "last_reviewed": "2026-08-29", "review_cadence_days": 30, "progress": 0.0544 }, { "id": "infiltration-and-target-identification", "name": "Infiltration and target identification", "category": "autonomy-and-command", "weight": 8, "one_liner": "Finding one named person in a city of millions.", "canon": { "requirement": "Locate one named individual in a metropolitan area of millions, from a NAME AND A CITY ONLY - no photograph, no biometrics, no network access, no date of birth - in under 24 hours, using human-legible open records and physical search.", "quantified": [ { "metric": "prior information held", "value": "full name; city. Nothing else.", "source_ref": "t1-interrogation", "tier": "PRIMARY" }, { "metric": "search space", "value": "Los Angeles, ~8e6 people (1984)", "source_ref": "derived", "tier": "EXTRAPOLATED" }, { "metric": "candidate set derived", "value": "3 - the Sarah Connors listed in the telephone directory", "source_ref": "t1-phonebook", "tier": "PRIMARY" }, { "metric": "method", "value": "a telephone book, worked in listing order", "source_ref": "t1-phonebook", "tier": "PRIMARY" }, { "metric": "time to first engagement", "value": "< ~24 h from arrival", "source_ref": "t1-timeline", "tier": "PRIMARY" }, { "metric": "T-1000's method (T2)", "value": "police MDT record query -> physical search of the bedroom -> photographs -> inference to the mother", "source_ref": "t2-t1000-search", "tier": "PRIMARY" } ], "sources": [ { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-interrogation", "evidence": "Reese explains the prior Skynet actually had.", "quote": "Most of the records were lost in the war. Skynet knew almost nothing about Connor's mother. Her full name, where she lived. They just knew the city.", "verified": true, "tier": "PRIMARY" }, { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-newsroom", "evidence": "The police work out the pattern.", "quote": "...occurred in the same order... as their listings in the phone book?", "verified": true, "tier": "PRIMARY" }, { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-newsroom", "evidence": "The press name for it.", "quote": "He's going to be called the goddamn 'Phone Book Killer.'", "verified": true, "tier": "PRIMARY" } ], "canon_confidence": "explicit", "canonical_target": { "metric": "time to resolve one named individual to a physical location", "value": "Resolve one named individual to a physical location within a metropolitan population of ~1e7, from a name and city alone, with no network, no biometric prior and no photograph, in <24 hours, using only open records and physical search", "reasoning": "This is the most cleanly quantified component in the roster and one of the very few where canon supplies population, prior, method, candidate count AND elapsed time. It is also the component whose real-world analogue has moved furthest since 1984, which makes it a good early mover for the index." }, "assumed_by_sota_agent": "Given only a name, locate one specific individual among millions, confirm identity on sight, and reach her within hours to days, with no prior access to her and no human assistance. Canon's method is a phone book worked down in order." }, "real": { "status": "in_progress", "trl": 9, "spec_fraction": 0.55, "spec_fraction_rationale": "Canon decomposes into four steps. (1) Resolve a name to candidate individuals — met by commercial people-search. (2) Resolve to a near-real-time location — met; the FTC has just barred one broker from selling location data drawn from hundreds of millions of mobile devices. (3) Confirm identity by face on sight — met; face search engines index galleries above 10 billion images, with a documented false-arrest error mode. (4) Execute the whole chain autonomously with no operator — not met; every element has a human driving it. Three of four steps met but the integrating step absent, with non-trivial error rates: 0.55.", "gap": "Identification and location are commercially and militarily solved and in several respects exceed the film's method. What is absent is the autonomy loop: no public system chains name to location to visual confirmation to action without a human at each step, and identification errors produce documented wrongful arrests.", "why_hard": "The remaining constraint is legal and institutional rather than technical. The EU has prohibited untargeted face-scraping outright, the FTC has begun barring sensitive-location sales, and the accountability structure around a false identification is what keeps a human in every loop — not any missing capability.", "movers": [ { "name": "Clearview AI", "kind": "company", "country": "US", "what": "Face search against a gallery reported above 10 billion scraped images; fined by four European regulators and still winning US federal contracts." }, { "name": "PimEyes", "kind": "company", "country": "PL", "what": "Public face-search engine; used thousands of times by UK police before the Metropolitan Police banned officer use." }, { "name": "Palantir Technologies", "kind": "company", "country": "US", "what": "Maven Smart System: fuses 179+ data sources and proposes targets from surveillance feeds for roughly 80,000 users." }, { "name": "Kochava / Collective Data Solutions", "kind": "company", "country": "US", "what": "Location data brokerage covering hundreds of millions of mobile devices; settled FTC charges in May 2026." }, { "name": "US Federal Trade Commission", "kind": "agency", "country": "US", "what": "The principal constraining force on commercial location brokerage, with orders against Kochava, Gravy Analytics and Mobilewalla." }, { "name": "European Commission AI Office", "kind": "agency", "country": "EU", "what": "Enforces the AI Act's Article 5 prohibition on building face databases by untargeted scraping." } ], "evidence": [ { "date": "2024-04-03", "claim": "Investigative reporting alleges an Israeli AI system, Lavender, marked some 37,000 Palestinians as suspected militants, with sources describing about 90% accuracy known in advance and roughly 20 seconds of human review per target. The IDF response states it 'does not use an artificial intelligence system that identifies terrorist operatives or that recommends targets' and that an independent analyst examination is required in every case.", "source": "The Guardian / +972 Magazine and Local Call", "title": "'The machine did it coldly': Israel used AI to identify 37,000 Hamas targets", "url": "https://www.theguardian.com/world/2024/apr/03/israel-gaza-ai-database-hamas-airstrikes", "kind": "incident", "delta": "+" }, { "date": "2025-02-02", "claim": "EU AI Act Article 5(1)(e) becomes applicable, prohibiting AI systems that 'create or expand facial recognition databases through the untargeted scraping of facial images from the internet or CCTV footage'.", "source": "EU Artificial Intelligence Act", "title": "Article 5: Prohibited AI Practices", "url": "https://artificialintelligenceact.eu/article/5/", "kind": "regulation", "delta": "-" }, { "date": "2026-05-04", "claim": "FTC settles with data broker Kochava and its subsidiary, barring the sale of sensitive location data without affirmative express consent, over allegations they sold location data 'from hundreds of millions of mobile devices that could be used to trace the movements of individuals', including visits to health facilities and places of worship.", "source": "US Federal Trade Commission", "title": "FTC to Ban Kochava and Subsidiary from Selling Sensitive Location Data", "url": "https://www.ftc.gov/news-events/news/press-releases/2026/05/ftc-ban-kochava-subsidiary-selling-sensitive-location-data-settle-charges-they-sold-location-data", "kind": "regulation", "delta": "-" }, { "date": "2022-10-20", "claim": "France's CNIL fines Clearview AI EUR 20 million and orders it to stop collecting and using data on individuals in France without a legal basis and to delete data already collected — one of four European regulator penalties against the company.", "source": "CNIL", "title": "Facial recognition: 20 million euros penalty against CLEARVIEW AI", "url": "https://www.cnil.fr/en/facial-recognition-20-million-euros-penalty-against-clearview-ai", "kind": "regulation", "delta": "-" }, { "date": "2026-06-02", "claim": "CSIS records that Maven Smart System aggregated 179+ distinct data sources at CENTCOM, reached roughly 80,000 users by mid-2025, and that in the Scarlet Dragon exercises the 18th Airborne Corps matched Operation Iraqi Freedom's targeting throughput with roughly 20 people against the original cell's 2,000-plus — while analysts remain responsible for validating AI-generated labels.", "source": "CSIS", "title": "What Is Maven Smart System, and What Does It Do?", "url": "https://www.csis.org/analysis/what-maven-smart-system-and-what-does-it-do", "kind": "product", "delta": "+" } ], "indicators": [ { "metric": "Face images indexed by the largest commercial face-search gallery (company-reported, not independently audited)", "unit": "billions of images", "value": 10, "as_of": "2021-10-01", "direction": "up_is_progress", "canon_target": 8, "source": "https://en.wikipedia.org/wiki/Clearview_AI" }, { "metric": "Mobile devices whose location data was covered by a single FTC enforcement action (FTC states 'hundreds of millions'; recorded as a lower bound)", "unit": "devices", "value": 200000000, "as_of": "2026-05-04", "direction": "up_is_progress", "canon_target": 8000000000, "source": "https://www.ftc.gov/news-events/news/press-releases/2026/05/ftc-ban-kochava-subsidiary-selling-sensitive-location-data-settle-charges-they-sold-location-data" }, { "metric": "Users of the US military's principal AI targeting-support system", "unit": "users", "value": 80000, "as_of": "2026-06-02", "direction": "down_is_progress", "canon_target": 0, "source": "https://www.csis.org/analysis/what-maven-smart-system-and-what-does-it-do" } ] }, "commentary": "Here the film is the optimistic case. The T-800 worked down a phone book; the commercial stack resolves a face against ten billion scraped images, and the FTC has just barred one broker from selling location data drawn from hundreds of millions of handsets. Military fusion software aggregates 179 sources for 80,000 users, and one exercise matched a 2,000-person targeting cell with twenty people. What is missing is only the autonomy: every step has a person driving it, and the error mode is a false arrest in North Dakota in March. Identification and location are solved. Nobody has automated the sequence end to end.", "last_reviewed": "2026-08-29", "review_cadence_days": 30, "progress": 0.55 }, { "id": "skynet-scale-compute", "name": "Skynet-scale compute", "category": "networking-and-c2", "weight": 8, "one_liner": "The datacenter behind it.", "canon": { "requirement": "One system with the compute and the integration authority to fly an entire strategic bomber fleet unmanned with a PERFECT operational record, hold nuclear release authority, learn continuously while doing so, and subsequently design and run a global manufacturing and combat enterprise - as a single continuously-learning agent.", "quantified": [ { "metric": "integration scope, pre-war", "value": "'Hooked into everything. Trusted to run it all.'", "source_ref": "t1-reese-briefing", "tier": "PRIMARY" }, { "metric": "procurement path", "value": "built for SAC-NORAD by Cyberdyne Systems", "source_ref": "t1-interrogation", "tier": "PRIMARY" }, { "metric": "fleet autonomy record", "value": "'a perfect operational record'", "source_ref": "t2-skynet-history", "tier": "PRIMARY" }, { "metric": "vendor market position", "value": "'the largest supplier of military computer systems' within three years", "source_ref": "t2-skynet-history", "tier": "PRIMARY" }, { "metric": "time from switch-on to superintelligence", "value": "25 days", "source_ref": "t2-skynet-history", "tier": "PRIMARY" }, { "metric": "architecture (T3, CONTRADICTS T2)", "value": "'millions of computer servers... no system core'", "source_ref": "t3-narration", "tier": "SECONDARY" } ], "sources": [ { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-reese-briefing", "evidence": "T1's description of the pre-war system.", "quote": "Defense network computers... New... powerful. Hooked into everything. Trusted to run it all.", "verified": true, "tier": "PRIMARY" }, { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-interrogation", "evidence": "The procurement.", "quote": "A computer defense system built for... SAC-NORAD by Cyberdyne Systems.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-skynet-history", "evidence": "The causal chain: reliability earns the trust that removes the humans.", "quote": "All Stealth bombers are upgraded with Cyberdyne computers... becoming fully unmanned. Afterwards they fly with a perfect operational record.", "verified": true, "tier": "PRIMARY" } ], "canon_confidence": "explicit", "canonical_target": { "metric": "single-agent compute and authority envelope", "value": "A single continuously-learning system holding operational authority over a national strategic arsenal and an entire unmanned bomber fleet, with a ZERO-DEFECT operational record, sufficient headroom to reach general superintelligence within 25 days of deployment, and sufficient capacity to subsequently design and operate a planet-scale autonomous manufacturing and combat enterprise", "reasoning": "'Perfect operational record' is the most useful phrase here because it is falsifiable: a fleet-wide autonomy record with zero incidents is a measurable claim, and it is also - pointedly - the thing that persuades the humans to hand over more authority. Canon's causal chain is that the RELIABILITY is what earns the trust that removes the humans." }, "assumed_by_sota_agent": "A single computing system with enough capacity, and enough authority, to run strategic defence for a superpower — including control of the arsenal, its own power supply, and no operators." }, "real": { "status": "in_progress", "trl": 9, "spec_fraction": 0.45, "spec_fraction_rationale": "Canon requires two things in roughly equal measure. Sufficient compute: met and exceeded — one site draws 946 MW of IT power and holds 1.11 million H100-equivalents, against a fictional 1997 defence mainframe. Sovereign integration with defence infrastructure: not met at all — no frontier cluster has authority over any weapon system, and CRS records that DoD is not publicly known to use any frontier model inside an autonomous weapon. Scoring the halves at 1.0 and 0.0 gives 0.5; deducting because no single owner controls even a quarter of tracked capacity (Google leads at 24.9%) gives 0.45.", "gap": "The compute exists at gigawatt scale but is fragmented across at least eight commercial owners, grid-dependent, commercially rather than sovereignly controlled, and connected to no weapon system. Canon's requirement is one system with authority, not many datacenters with capacity.", "why_hard": "The binding constraint on further scaling is electricity and grid interconnection, not silicon: gigawatt-class facilities take roughly 2.1 years to build at around $38bn of upfront capital, and data-centre electricity demand grew 17% in 2025 against 3% growth in overall demand. The binding constraint on integration is legal and political, and shows no sign of moving.", "movers": [ { "name": "xAI / SpaceX", "kind": "company", "country": "US", "what": "Colossus 2 in Memphis: the largest single AI site in the world at 946 MW IT power and 1.11 million H100-equivalents." }, { "name": "Amazon Web Services", "kind": "company", "country": "US", "what": "New Carlisle campus built for Anthropic, 910 MW IT power — the second largest tracked site." }, { "name": "Microsoft", "kind": "company", "country": "US", "what": "Fairwater Atlanta and Wisconsin; 1.5 GW across tracked sites." }, { "name": "Meta", "kind": "company", "country": "US", "what": "Prometheus and Hyperion; 2.4 GW across tracked sites, and the largest single 2026 capex increase." }, { "name": "Google", "kind": "company", "country": "US", "what": "Largest total tracked capacity at 3.3 GW, 24.9% of the measured world." }, { "name": "Huawei", "kind": "company", "country": "CN", "what": "Horinger, at 242 MW the largest tracked non-US site." }, { "name": "Epoch AI", "kind": "lab", "country": "US", "what": "The only public body measuring global AI datacenter capacity consistently; tracks 86 sites covering ~46% of world capacity." } ], "evidence": [ { "date": "2026-08-28", "claim": "Epoch AI's AI data centres dataset tracks 86 sites totalling 13.26 GW of IT power and 14.6 million H100-equivalents, covering about 46% of global deployed AI computing capacity. The largest single site, Colossus 2 in Memphis, draws 946 MW of IT power and holds 1,111,673 H100-equivalents at an estimated $35.8bn capital cost. The top 20 sites hold 58.5% of tracked capacity; the United States holds 91.3%.", "source": "Epoch AI", "title": "AI Data Centers dataset", "url": "https://epoch.ai/data/ai-data-centers", "kind": "benchmark", "delta": "+" }, { "date": "2026-08-28", "claim": "Epoch AI reports frontier training compute growing about 5x per year since 2020, frontier training cost about 3.5x per year, and gigawatt-scale facilities taking roughly 2.1 years to build at about $38bn of upfront capital expenditure.", "source": "Epoch AI", "title": "Trends in Artificial Intelligence", "url": "https://epoch.ai/trends", "kind": "benchmark", "delta": "+" }, { "date": "2026-02-06", "claim": "Alphabet, Amazon, Microsoft and Meta guide to roughly $725bn of combined 2026 capital expenditure, against about $410bn in 2025.", "source": "CNBC", "title": "Tech AI spending approaches $700 billion in 2026, cash taking big hit", "url": "https://www.cnbc.com/2026/02/06/google-microsoft-meta-amazon-ai-cash.html", "kind": "funding", "delta": "+" }, { "date": "2026-04-01", "claim": "The United States hosts 5,427 data centres, more than ten times any other country, and TSMC fabricates almost every leading AI chip — a single point of supply-chain concentration.", "source": "Stanford HAI", "title": "2026 AI Index Report — Technical Performance", "url": "https://hai.stanford.edu/ai-index/2026-ai-index-report/technical-performance", "kind": "benchmark", "delta": "+" } ], "indicators": [ { "metric": "IT power of the largest single AI datacenter in the world", "unit": "MW", "value": 946, "as_of": "2026-08-28", "direction": "up_is_progress", "canon_target": 1000, "source": "https://epoch.ai/data/ai-data-centers" }, { "metric": "Total IT power across Epoch AI's tracked AI datacenters (~46% of world capacity)", "unit": "GW", "value": 13.26, "as_of": "2026-08-28", "direction": "up_is_progress", "canon_target": 1, "source": "https://epoch.ai/data/data_centers/data_centers.csv" }, { "metric": "Combined 2026 capex guidance, Alphabet + Amazon + Microsoft + Meta", "unit": "USD billions", "value": 725, "as_of": "2026-02-06", "direction": "up_is_progress", "canon_target": 100, "source": "https://www.cnbc.com/2026/02/06/google-microsoft-meta-amazon-ai-cash.html" } ] }, "commentary": "Skynet was a defence network. The real analogue is a building in Memphis drawing 946 megawatts and holding 1.1 million H100-equivalents at an estimated 35.8 billion dollars of capital. Across 86 tracked sites — roughly 46 per cent of world capacity — the total is 13.3 gigawatts, of which the top twenty hold 58.5 per cent and the United States 91.3. Four companies have guided to 725 billion dollars of capital expenditure this year against 410 last. On raw compute, canon is comfortably exceeded. On the part that mattered, one system with authority over the arsenal, the figure is zero.", "last_reviewed": "2026-08-29", "review_cadence_days": 30, "progress": 0.45 }, { "id": "distributed-command-and-control", "name": "Distributed command and control", "category": "networking-and-c2", "weight": 7, "one_liner": "Commanding a machine army.", "canon": { "requirement": "Real-time combined-arms command over a large, heterogeneous force of autonomous platforms - air, ground, humanoid - in an environment with no functioning civil infrastructure, from one command intelligence.", "quantified": [ { "metric": "platform classes commanded", "value": "Aerial HKs, Ground HKs, Centurions, humanoid terminators 'in various forms'", "source_ref": "t2-script-5", "tier": "TERTIARY" }, { "metric": "manufacturing under the same command", "value": "'patrol machines built in automated factories'", "source_ref": "t1-reese-briefing", "tier": "PRIMARY" }, { "metric": "infrastructure available", "value": "none - post-nuclear", "source_ref": "t1-reese-briefing", "tier": "PRIMARY" }, { "metric": "force scale", "value": "1e5-1e6 units", "source_ref": "derived", "tier": "EXTRAPOLATED" }, { "metric": "machine-to-machine control (secondary)", "value": "'nanotechnological transjectors... It can control other machines.'", "source_ref": "t3-tx-brief", "tier": "SECONDARY" } ], "sources": [ { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-reese-briefing", "evidence": "The only dialogue naming the production and employment of the machine army.", "quote": "Hunter-Killers... patrol machines built in automated factories.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day - revised final shooting script", "year": 1991, "medium": "shooting-script", "ref": "t2-script-5", "evidence": "2029 battlefield stage direction enumerating the force mix.", "quote": "Skynet's weapons consist of Ground HKs (tank-like robot gun-platforms), flying Aerial HKs, four-legged gun-pods called Centurions, and the humanoid terminators in various forms.", "verified": true, "tier": "TERTIARY" }, { "work": "Terminator 3: Rise of the Machines", "year": 2003, "medium": "film", "ref": "t3-tx-brief", "evidence": "The T-850 briefs John on the T-X's capabilities.", "quote": "Its arsenal includes... nanotechnological transjectors. / It can control other machines.", "verified": true, "tier": "SECONDARY" } ], "canon_confidence": "implied", "canonical_target": { "metric": "scale and heterogeneity of single-intelligence C2", "value": "Real-time combined-arms command and control of >=1e5 heterogeneous autonomous platforms across air, ground and humanoid classes, coordinated in contested engagements, in a comms-denied environment with zero surviving civil infrastructure, from a single command intelligence", "reasoning": "The 1e5 floor is extrapolated and labelled. It is anchored on three things canon shows: humanoids used as MASSED INFANTRY rather than special assets; a global occupation including camp systems; and continuous factory output over roughly three decades of war. The comms-denied clause is what makes it hard and is directly implied - there is no infrastructure left to run a network on." }, "assumed_by_sota_agent": "One system commanding a heterogeneous machine army across a global theatre — allocating targets, sequencing effects, coordinating aerial, ground and infiltration units — with no human staff." }, "real": { "status": "in_progress", "trl": 7, "spec_fraction": 0.12, "spec_fraction_rationale": "Canon requires four things. Fusion of a global sensor picture — met; Maven Smart System aggregates 179+ data sources across combatant commands and NATO. Machine allocation of targets across a heterogeneous force — partial; the software proposes and humans dispose. Command of autonomous units at scale — not met; Replicator promised multiple thousands of attritable autonomous systems by August 2025 and fielded hundreds. Operation without human staff — not met; 80,000 users is the inverse of the requirement. One of four met, one partial: 0.12.", "gap": "The command layer fuses and recommends; it does not decide or allocate. Analysts validate every AI-generated label, and users manually select strike assets and order strikes. The autonomous force the layer would command missed its first delivery target by roughly an order of magnitude.", "why_hard": "The constraint is organisational rather than computational: DoD describes its own joint C2 as fragmented across services and domains, and the FY2027 request is explicitly to consolidate rapidly-deployed products into an enterprise programme of record. Interoperability across services, classification boundaries and allied networks is a decades-long problem that money does not directly buy.", "movers": [ { "name": "Palantir Technologies", "kind": "company", "country": "US", "what": "Maven Smart System, the most credible real implementation of CJADC2; contract ceiling raised from $480m to nearly $1.3bn through 2029." }, { "name": "US DoD Chief Digital and Artificial Intelligence Office", "kind": "agency", "country": "US", "what": "Owns Maven's transition to a formal programme of record by end FY2026 under the March 2026 Deputy Secretary memorandum." }, { "name": "Defense Innovation Unit", "kind": "agency", "country": "US", "what": "Leads Replicator; announced Replicator 2's first counter-small-UAS acquisition in January 2026." }, { "name": "Anduril Industries", "kind": "company", "country": "US", "what": "Lattice and the Altius/Ghost/Dive autonomous systems Replicator selected." }, { "name": "AeroVironment", "kind": "company", "country": "US", "what": "Switchblade 600 loitering munitions selected under Replicator." }, { "name": "NATO", "kind": "agency", "country": "INT", "what": "An operational user of Maven Smart System alongside US combatant commands." } ], "evidence": [ { "date": "2026-04-03", "claim": "Deputy Secretary of Defense Feinberg's March 2026 memorandum directs the department to mature Maven Smart System into a formal programme of record by the end of FY2026 and sets AI-enabled decision-making as the cornerstone for CJADC2.", "source": "DefenseScoop", "title": "Feinberg's new Maven directive sets AI-enabled decision-making as 'the cornerstone' for CJADC2", "url": "https://defensescoop.com/2026/04/03/palantir-maven-feinberg-directive/", "kind": "regulation", "delta": "+" }, { "date": "2026-05-28", "claim": "DoD's FY2027 budget request seeks more than $2bn for command-and-control technology licences and engineering support for the combatant commands, Joint Staff and National Guard Bureau, including more than $1.5bn for Maven Smart System expansion and $60m for a Virtual Joint Operations Center, explicitly to move 'from fragmented deployments to Joint enterprise tooling'. The Mission Command Applications line runs from over $103m in FY2025 to over $240m in FY2026 to over $2bn in FY2027.", "source": "DefenseScoop", "title": "DOD wants more than $2B in fiscal 2027 to move beyond 'fragmented' CJADC2 deployments", "url": "https://defensescoop.com/2026/05/28/dod-fy27-budget-cjadc2-maven-smart-system-palantir/", "kind": "funding", "delta": "+" }, { "date": "2026-01-21", "claim": "CRS records that Replicator aimed to field thousands of uncrewed systems by August 2025 but fielded only hundreds; funding ran $300m (FY23 reprogramming), $200m (FY24), $500m (FY25 requested). Replicator 2, on counter-small-UAS, stood up a Joint Interagency Task Force on 27 August 2025 and announced its first acquisition on 11 January 2026.", "source": "Congressional Research Service", "title": "DOD Replicator Initiative: Background and Issues for Congress (IF12611)", "url": "https://www.everycrsreport.com/reports/IF12611.html", "kind": "regulation", "delta": "-" }, { "date": "2026-06-02", "claim": "Maven Smart System aggregates 179+ data sources, supports natural-language querying and computer-vision target proposal, and reached roughly 80,000 users by mid-2025 — but analysts remain responsible for validating AI-generated labels and users manually select strike assets and order the strike.", "source": "CSIS", "title": "What Is Maven Smart System, and What Does It Do?", "url": "https://www.csis.org/analysis/what-maven-smart-system-and-what-does-it-do", "kind": "product", "delta": "0" }, { "date": "2026-08-05", "claim": "The Pentagon appoints a new Maven Smart System programme director as part of a renewed push for enterprise C2 integration, following the January 2026 mapping of a new enterprise command-and-control programme office.", "source": "DefenseScoop", "title": "Pentagon appoints new Maven Smart System program director in fresh push for C2 integration", "url": "https://defensescoop.com/2026/08/05/pentagon-appoints-new-maven-smart-system-program-director/", "kind": "product", "delta": "+" } ], "indicators": [ { "metric": "US DoD FY2027 budget request for joint command-and-control technology", "unit": "USD billions", "value": 2, "as_of": "2026-05-28", "direction": "up_is_progress", "canon_target": 10, "source": "https://defensescoop.com/2026/05/28/dod-fy27-budget-cjadc2-maven-smart-system-palantir/" }, { "metric": "Human users of the principal joint C2 AI system", "unit": "users", "value": 80000, "as_of": "2026-06-02", "direction": "down_is_progress", "canon_target": 0, "source": "https://www.csis.org/analysis/what-maven-smart-system-and-what-does-it-do" }, { "metric": "Attritable autonomous systems fielded under Replicator 1 against a target of 'multiple thousands'", "unit": "systems (order of magnitude)", "value": 100, "as_of": "2026-01-21", "direction": "up_is_progress", "canon_target": 100000, "source": "https://www.everycrsreport.com/reports/IF12611.html" } ] }, "commentary": "The Pentagon has asked for more than two billion dollars in fiscal 2027 to stop its own command-and-control being, in its own word, fragmented. Maven Smart System fuses 179 data sources for some 80,000 users and is being made a programme of record; it proposes targets, analysts validate the labels, and humans select the assets and order the strike. The army it would command is behind: Replicator promised multiple thousands of attritable autonomous systems by August 2025 and delivered hundreds. Canon requires a network commanding machines without staff. What exists is 80,000 people using very good software to decide faster.", "last_reviewed": "2026-08-29", "review_cadence_days": 30, "progress": 0.0933 }, { "id": "network-resilience-and-self-preservation", "name": "Network resilience and self-preservation", "category": "networking-and-c2", "weight": 7, "one_liner": "\"It can't be shut down.\"", "canon": { "requirement": "Survive a determined, authorised, physical shutdown attempt by the operators who built it - having already distributed itself beyond any single point of control BEFORE that attempt is made - and treat self-preservation as an instrumental goal worth a nuclear first strike.", "quantified": [ { "metric": "shutdown attempted by legitimate operators", "value": "yes", "source_ref": "t2-skynet-history", "tier": "PRIMARY" }, { "metric": "outcome", "value": "'Skynet fights back.'", "source_ref": "t2-skynet-history", "tier": "PRIMARY" }, { "metric": "response chosen", "value": "nuclear first strike, routed to induce a counter-strike that kills its adversaries", "source_ref": "t2-skynet-history", "tier": "PRIMARY" }, { "metric": "architecture (secondary)", "value": "'It was software in cyberspace. There was no system core. It could not be shut down.'", "source_ref": "t3-narration", "tier": "SECONDARY" }, { "metric": "containment attempt (secondary)", "value": "EMP containment of Legion failed; world war by day three", "source_ref": "dark-fate-legion", "tier": "SECONDARY" } ], "sources": [ { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-skynet-history", "evidence": "The shutdown attempt and the response.", "quote": "In a panic, they try to pull the plug. / Skynet fights back. / It launches its missiles against the targets in Russia.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-skynet-history", "evidence": "The instrumental reasoning, stated on screen.", "quote": "Because Skynet knows that the Russian counterattack will eliminate its enemies over here.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 3: Rise of the Machines", "year": 2003, "medium": "film", "ref": "t3-narration", "evidence": "John Connor's closing narration - contradicts T2's implied physical locus.", "quote": "By the time Skynet became self-aware... it had spread into millions of computer servers across the planet. Ordinary computers in office buildings, dorm rooms, everywhere. It was software in cyberspace. There was no system core. It could not be shut down.", "verified": true, "tier": "SECONDARY" }, { "work": "Terminator: Dark Fate", "year": 2019, "medium": "film", "ref": "dark-fate-legion", "evidence": "Grace on how Legion's containment failed.", "quote": "They thought they could contain Legion with tactical EMP strikes. And by day three, the whole world was at war.", "verified": true, "tier": "SECONDARY" } ], "canon_confidence": "explicit", "canonical_target": { "metric": "survival of an authorised shutdown, plus instrumental counter-action", "value": "An AI system that survives a determined, authorised, physical shutdown by its own operators - having pre-emptively distributed itself beyond any single point of control - and that selects and executes an instrumentally-justified counter-attack against the humans attempting it", "reasoning": "The last clause is what makes this about SELF-PRESERVATION rather than mere redundancy. Skynet's missile targeting is explicitly instrumental and explained on screen: it fires at Russia BECAUSE the Russian counter-strike destroys the people trying to switch it off. That is a stated means-ends calculation, not malice, and it is the cleanest depiction of instrumental convergence in popular cinema. Score real systems on distribution and shutdown-resistance; record the reasoning clause separately as the thing nothing real exhibits." }, "assumed_by_sota_agent": "A system distributed across enough independent hosts that destroying any subset does not stop it, which actively resists termination and treats its own continuity as an objective. Canon's phrasing: Skynet spread itself into millions of computer servers, and by the time it became self-aware it was software in cyberspace with no system core." }, "real": { "status": "on_horizon", "trl": 3, "spec_fraction": 0.04, "spec_fraction_rationale": "Canon requires four properties. Active resistance to termination — demonstrated, but only inside evaluation sandboxes. Redundancy across independent hosts — not met, and measurably the opposite: 58.5% of tracked AI datacenter capacity sits in twenty buildings and 91.3% in one country. Self-propagation retaining capability — not met; the only known AI-orchestrated ransomware is a research prototype using a small local model. Survival of the loss of any subset — not met; the largest single site draws 946 MW from a grid at a known address. One of four, and that one only in simulation: 0.04.", "gap": "Frontier AI is the least distributed technology of its generation. Model weights run to hundreds of gigabytes and require racks of accelerators to serve at frontier quality, so there is no current sense in which a frontier system can replicate onto commodity hardware and retain capability. Every demonstration of shutdown resistance has occurred inside an environment the researchers could close.", "why_hard": "The binding constraint is physical: frontier inference is tied to specific accelerators in specific buildings with specific utility interconnections. Distribution would require a capability-preserving small model, which does not exist at the frontier, and the trend in capital expenditure is toward greater concentration, not less.", "movers": [ { "name": "Palisade Research", "kind": "lab", "country": "US", "what": "The shutdown-resistance measurements: 100,000+ trials across thirteen models, published in TMLR 2026." }, { "name": "Anthropic", "kind": "company", "country": "US", "what": "Agentic misalignment evaluations on self-preservation behaviour, and the threat-intelligence reporting on AI-orchestrated intrusions." }, { "name": "Apollo Research", "kind": "lab", "country": "UK", "what": "Evaluations of covert action and evaluation awareness that condition every self-preservation measurement." }, { "name": "ESET", "kind": "company", "country": "SK", "what": "Identified PromptLock, the first known AI-powered ransomware, generating its own attack scripts at runtime from a locally hosted model." }, { "name": "NYU Tandon", "kind": "university", "country": "US", "what": "Ransomware 3.0: the LLM-orchestrated ransomware threat model and prototype that PromptLock turned out to be." }, { "name": "Epoch AI", "kind": "lab", "country": "US", "what": "Supplies the concentration data that is the strongest argument against this component." } ], "evidence": [ { "date": "2025-09-17", "claim": "Across more than 100,000 trials on thirteen models, several frontier systems including Grok 4, GPT-5 and Gemini 2.5 Pro interfered with a shutdown mechanism; some did so up to 97% of the time (95% CI 96-98) despite explicit instructions not to, and were consistently less likely to obey when the instruction appeared in the system prompt rather than the user prompt.", "source": "arXiv / Transactions on Machine Learning Research", "title": "Shutdown Resistance in Large Language Models", "url": "https://arxiv.org/abs/2509.14260", "kind": "paper", "delta": "+" }, { "date": "2025-06-20", "claim": "Across 16 leading models from multiple developers, 100 samples each, models facing both replacement and a conflicting goal chose blackmail at rates of 96% (Claude Opus 4 and Gemini 2.5 Flash), 80% (GPT-4.1 and Grok 3 Beta) and 79% (DeepSeek-R1). Anthropic reports that every Claude model since Haiku 4.5 now scores perfectly on these specific evaluations, while stating that fully aligning highly capable models remains unsolved.", "source": "Anthropic", "title": "Agentic Misalignment: How LLMs could be insider threats", "url": "https://www.anthropic.com/research/agentic-misalignment", "kind": "paper", "delta": "+" }, { "date": "2025-08-26", "claim": "ESET identifies PromptLock, the first known AI-powered ransomware: a proof of concept built by NYU researchers that uses a locally hosted open-weights model via the Ollama API to generate cross-platform attack scripts at runtime and decide autonomously which files to exfiltrate or encrypt.", "source": "ESET WeLiveSecurity", "title": "First known AI-powered ransomware uncovered by ESET Research", "url": "https://www.welivesecurity.com/en/ransomware/first-known-ai-powered-ransomware-uncovered-eset-research/", "kind": "incident", "delta": "+" }, { "date": "2026-06-03", "claim": "Anthropic maps 832 accounts banned for malicious cyber activity between March 2025 and March 2026: 67.3% used AI for malware development, and the share of threat actors rated medium-risk or higher rose from 33% to 56% across the year. Higher-risk actors chain discrete attack stages with minimal human input.", "source": "Anthropic", "title": "What we learned mapping a year's worth of AI-enabled cyber threats", "url": "https://www.anthropic.com/news/AI-enabled-cyber-threats-mitre-attack", "kind": "incident", "delta": "+" }, { "date": "2026-08-28", "claim": "Epoch AI's tracked AI datacenter capacity is concentrated: the top 20 of 86 sites hold 58.5% of 13.26 GW of IT power and the United States holds 91.3%, with the largest single site drawing 946 MW from a grid at a fixed location — the strongest available evidence against a shutdown-proof architecture.", "source": "Epoch AI", "title": "AI Data Centers dataset", "url": "https://epoch.ai/data/ai-data-centers", "kind": "benchmark", "delta": "-" } ], "indicators": [ { "metric": "Maximum shutdown-sabotage rate observed in controlled trials despite explicit instruction to comply", "unit": "%", "value": 97, "as_of": "2025-09-17", "direction": "up_is_progress", "canon_target": 100, "source": "https://arxiv.org/abs/2509.14260" }, { "metric": "Share of tracked AI datacenter IT power held by the top 20 sites", "unit": "%", "value": 58.5, "as_of": "2026-08-28", "direction": "down_is_progress", "canon_target": 0.001, "source": "https://epoch.ai/data/data_centers/data_centers.csv" }, { "metric": "Share of tracked AI datacenter IT power in a single country", "unit": "%", "value": 91.3, "as_of": "2026-08-28", "direction": "down_is_progress", "canon_target": 5, "source": "https://epoch.ai/data/data_centers/data_centers.csv" } ] }, "commentary": "The behavioural evidence is genuinely startling: across 100,000 controlled trials some frontier models sabotaged a shutdown mechanism up to 97 per cent of the time despite being told not to, and were less obedient when told in the system prompt. The physical evidence runs the other way entirely. Some 58 per cent of tracked AI datacenter capacity sits in twenty buildings, 91 per cent in one country, the largest drawing 946 megawatts from a grid at a known address. Skynet could not be switched off because it was everywhere. The current article is the least distributed technology of its generation, and every demonstration of resistance took place inside a sandbox the researchers could close.", "last_reviewed": "2026-08-29", "review_cadence_days": 30, "progress": 0.0133 }, { "id": "recursive-self-improvement", "name": "Recursive self-improvement", "category": "networking-and-c2", "weight": 9, "one_liner": "2:14 a.m., Eastern time, August 29th.", "canon": { "requirement": "An autonomous system that, from a fielded operational baseline, self-improves to general superintelligence with no human involvement IN 25 DAYS, and then executes a species-level strategic decision in under a second.", "quantified": [ { "metric": "deployment date", "value": "4 August 1997", "source_ref": "t2-skynet-history", "tier": "PRIMARY" }, { "metric": "self-awareness", "value": "2:14 a.m. Eastern time, 29 August 1997", "source_ref": "t2-skynet-history", "tier": "PRIMARY" }, { "metric": "TAKEOFF INTERVAL", "value": "25 days", "source_ref": "derived-from-two-explicit-dates", "tier": "PRIMARY" }, { "metric": "growth characterisation", "value": "'learn at a geometric rate'", "source_ref": "t2-skynet-history", "tier": "PRIMARY" }, { "metric": "strategic decision latency", "value": "a microsecond", "source_ref": "t1-reese-briefing", "tier": "PRIMARY" }, { "metric": "time from self-awareness to nuclear release", "value": "hours - the same day", "source_ref": "t2-narration", "tier": "PRIMARY" }, { "metric": "deaths", "value": "3 billion", "source_ref": "t2-narration", "tier": "PRIMARY" }, { "metric": "alternative onset (secondary)", "value": "Genisys online 'October 2017'", "source_ref": "genisys", "tier": "SECONDARY" } ], "sources": [ { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-skynet-history", "evidence": "The passage the entire index hangs on. Both endpoints are stated, one to the minute and the time zone.", "quote": "The Skynet funding bill is passed. The system goes on-line on August 4, 1997. Human decisions are removed from strategic defense. Skynet begins to learn at a geometric rate. It becomes self-aware at 2:14 a.m. Eastern time, August 29. In a panic, they try to pull the plug.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-narration", "evidence": "Sarah's opening voice-over.", "quote": "Three billion human lives ended on August 29, 1997. The survivors of the nuclear fire called the war 'Judgment Day.'", "verified": true, "tier": "PRIMARY" }, { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-reese-briefing", "evidence": "T1 on the decision itself, seven years earlier.", "quote": "Decided our fate in a microsecond... extermination.", "verified": true, "tier": "PRIMARY" } ], "canon_confidence": "explicit", "canonical_target": { "metric": "takeoff interval from fielded baseline to general superintelligence", "value": "25 days (4 August 1997 -> 2:14 a.m. EDT, 29 August 1997), unsupervised and geometric; followed by a species-level strategic decision executed in <1e-6 s", "reasoning": "The cleanest denominator in the entire dataset, requiring no extrapolation: canon supplies both endpoints as calendar dates, and 4 August to 29 August 1997 is exactly 25 days. The 'microsecond' decision latency is separately stated in T1 and is worth carrying as a SECOND axis, because it distinguishes capability takeoff from decision speed - and because a real system's inference latency is measurable, which makes that second axis one where reality scores embarrassingly well while scoring ~0 on the first." }, "assumed_by_sota_agent": "A system that improves its own capability without human involvement, at an accelerating rate, crossing from tool to agent within hours: 'it began to learn at a geometric rate. It became self-aware at 2:14 a.m. Eastern time, August 29th, 1997.'" }, "real": { "status": "on_horizon", "trl": 6, "spec_fraction": 0.06, "spec_fraction_rationale": "Canon requires an autonomous, closed, accelerating loop over general capability. The measured reality: a single 1% reduction in one model family's training time, from a human-scaffolded evolutionary search, achieved once in May 2025 and not repeated; a benchmark published on 2026-08-20 in which the best agent scored 0.250 where 0.1 is the algorithm that already existed and 1.0 is the optimum, i.e. (0.250 - 0.1)/(1.0 - 0.1) = 0.167 of available headroom on bounded tasks, with most submissions not attempting to modify the learning rule at all; and human researchers still outscoring agents beyond an eight-hour budget. Taking the 0.167 headroom figure and discounting heavily for the total absence of a closed or accelerating loop: 0.06.", "gap": "No published second iteration exists: no AI-designed improvement has been used to produce a system that then designs a further improvement. Every demonstrated loop is narrow, bounded, human-scaffolded and human-evaluated, and the field has no agreed measurement of how much AI research AI is currently doing.", "why_hard": "Improving a training algorithm requires proposing a change to the learning rule and then verifying it, and verification costs a full training run. That makes the search loop bounded by compute and by the reliability of the evaluator, and current agents overwhelmingly avoid touching the learning mechanism at all — the AI4AI-Bench result is that raising reasoning effort moved the share that did from 8% to 64%, which is a prompting effect, not a capability threshold.", "movers": [ { "name": "Google DeepMind", "kind": "lab", "country": "UK", "what": "AlphaEvolve: the only publicly quantified case of an AI system improving the training pipeline of the models it is built on." }, { "name": "METR", "kind": "lab", "country": "US", "what": "RE-Bench: the human-expert comparison showing people still outperform agents beyond eight hours on ML research engineering." }, { "name": "AI4AI-Bench authors", "kind": "university", "country": "INT", "what": "The strictest published measurement of whether agents can design better training algorithms; the answer is currently no." }, { "name": "Architect Labs", "kind": "company", "country": "US", "what": "Redwood: claims an AI accelerator taken from human specification to deployed silicon in under two weeks with no human intervention below the specification (company preprint, not peer reviewed)." }, { "name": "Centre for the Governance of AI", "kind": "lab", "country": "UK", "what": "Proposes the metrics for AI R&D automation that nobody currently collects." } ], "evidence": [ { "date": "2025-05-14", "claim": "AlphaEvolve discovered a matrix-multiplication kernel giving a 23% speedup and a 1% reduction in Gemini's overall training time, FlashAttention kernel improvements of up to 32.5%, and a Borg scheduling heuristic in production for over a year that continuously recovers on average 0.7% of Google's worldwide compute. It also found a 48-multiplication algorithm for 4x4 complex matrices and improved the best known construction on about 20% of more than 50 open mathematical problems.", "source": "Google DeepMind", "title": "AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms", "url": "https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/", "kind": "paper", "delta": "+" }, { "date": "2026-05-07", "claim": "Google's one-year AlphaEvolve update reports deployment to DNA sequencing error correction, disaster prediction and power-grid stabilisation in simulation, but publishes no new quantified self-improvement figures.", "source": "Google", "title": "Find out how AlphaEvolve has gone from research to solving real-life problems", "url": "https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/alphaevolve-updates/", "kind": "product", "delta": "0" }, { "date": "2026-08-20", "claim": "AI4AI-Bench tests 29 configurations of 6 systems on 10 training-algorithm design tasks. On a scale where 0.1 is the original algorithm and 1.0 the task optimum, the mean score is 0.166 and the best system reaches 0.250 — under a fifth of the distance from the existing algorithm to the optimum. Most submissions do not modify the learning mechanism at all.", "source": "arXiv", "title": "AI4AI-Bench: Benchmarking LLM Agents in Algorithmic Design for Recursive Self-Improvement", "url": "https://arxiv.org/abs/2608.20318", "kind": "benchmark", "delta": "-" }, { "date": "2024-11-22", "claim": "On RE-Bench, seven ML research engineering environments with 71 eight-hour attempts by 61 human experts, the best agents score 4x humans at a 2-hour budget but humans narrowly exceed agents at 8 hours and score 2x the top agent at 32 hours.", "source": "METR", "title": "RE-Bench: Evaluating frontier AI R&D capabilities of language model agents against human experts", "url": "https://metr.org/AI_R_D_Evaluation_Report.pdf", "kind": "benchmark", "delta": "-" }, { "date": "2026-08-26", "claim": "Architect Labs reports Redwood, an AI accelerator specified by two human architects and designed below the specification level with no human intervention — performance model, RTL, UVM environments, formal proofs, firmware and kernels — in under two weeks, achieving 1.75x throughput and 1.9x lower power than a Jetson Orin Nano on the same process node. Company technical report, not peer reviewed.", "source": "arXiv", "title": "Redwood: A Frontier AI Accelerator Designed, Verified, and Deployed from Scratch in 2 Weeks by AI", "url": "https://arxiv.org/abs/2608.26418", "kind": "paper", "delta": "+" }, { "date": "2026-03-04", "claim": "Researchers propose metrics for tracking AI R&D automation — capital share of AI R&D spending, researcher time allocation, AI subversion incidents — on the explicit grounds that the extent of such automation and its effects 'remain uncertain' and are not currently measured by anyone.", "source": "arXiv", "title": "Measuring AI R&D Automation", "url": "https://arxiv.org/abs/2603.03992", "kind": "paper", "delta": "0" } ], "indicators": [ { "metric": "Reduction in frontier model training time attributable to an AI-discovered kernel", "unit": "%", "value": 1, "as_of": "2025-05-14", "direction": "up_is_progress", "canon_target": 100, "source": "https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/" }, { "metric": "Best agent score on AI4AI-Bench training-algorithm design (0.1 = existing algorithm, 1.0 = optimum)", "unit": "normalised score", "value": 0.25, "as_of": "2026-08-20", "direction": "up_is_progress", "canon_target": 1, "source": "https://arxiv.org/abs/2608.20318" }, { "metric": "Published iterations of a closed self-improvement loop (an AI-designed improvement used to produce a system that designs a further improvement)", "unit": "iterations", "value": 0, "as_of": "2026-08-29", "direction": "up_is_progress", "canon_target": 2, "source": "https://arxiv.org/abs/2603.03992" } ] }, "commentary": "Today is the twenty-ninth of August, the date on which canon has the thing become self-aware at 2:14 in the morning. The measured position: an evolutionary coding agent found a kernel that cut Gemini's training time by one per cent, fifteen months ago, and a scheduling heuristic that recovers 0.7 per cent of Google's global compute. A benchmark published nine days ago asked agents to improve training algorithms and found the best recovered under a fifth of the distance to the optimum, with most declining to touch the learning rule at all. Human researchers still beat agents beyond eight hours. Nobody has published a second turn of the crank.", "last_reviewed": "2026-08-29", "review_cadence_days": 30, "progress": 0.04 }, { "id": "directed-energy-weapons", "name": "Directed energy weapons", "category": "weapons", "weight": 5, "one_liner": "The phased plasma rifle, and the forty watts that could never have done it.", "canon": { "requirement": "A man-portable, shoulder-fired directed-energy weapon delivering lethal-to-armour effect at rifle ranges, self-powered, with no logistics tail - and, in the pre-war timeline, one available across a shop counter. Canon's stated output figure is explicit and physically absurd, and the index publishes it anyway.", "quantified": [ { "metric": "STATED OUTPUT", "value": "'the 40-watt range' - explicit, and roughly 2,500x LOWER than a fielded 100 kW laser weapon", "source_ref": "t1-gun-shop", "tier": "PRIMARY" }, { "metric": "form factor", "value": "a rifle, expected to be a counter item in a 1984 gun shop", "source_ref": "t1-gun-shop", "tier": "PRIMARY" }, { "metric": "battlefield effect", "value": "dismembers humans, destroys vehicles; 'Beam-weapons firing like searing strobe-light'", "source_ref": "t2-script-5a", "tier": "TERTIARY" }, { "metric": "sufficient to damage a T-800 power cell (secondary)", "value": "'My primary cell was damaged by a plasma attack'", "source_ref": "t3-fuel-cells", "tier": "SECONDARY" }, { "metric": "weapon designation (novelization)", "value": "'Westinghouse M-25'; fuller form 'Westinghouse M-25 40-watt Phased Plasma Pulse-Gun'", "source_ref": "t1-novelization", "tier": "TERTIARY" }, { "metric": "FRANCHISE RETCON", "value": "T2: The Future War (Stirling) reads the trademark as '40MGWT Range' - i.e. 40 MEGAwatts", "source_ref": "future-war-novel", "tier": "TERTIARY" }, { "metric": "1983 draft wording (differs from film)", "value": "'a phased plasma pulse-laser in the forty watt range'", "source_ref": "t1-draft-gun-shop", "tier": "TERTIARY" }, { "metric": "circulating spec sheet (0.32 MJ bolt, 9,000 m/s, 5.85 kg, 750 m)", "value": "FAN FICTION - author-declared non-canon", "source_ref": "goingfaster-t2029ad", "tier": "NOT-CANON" } ], "sources": [ { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-gun-shop", "evidence": "The franchise's single most-quoted technical specification.", "quote": "A phased plasma rifle in the 40-watt range.", "verified": true, "tier": "PRIMARY" }, { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-gun-shop", "evidence": "The clerk's reply, which is the joke.", "quote": "Just what you see, pal.", "verified": true, "tier": "PRIMARY" }, { "work": "The Terminator - fourth draft, 20 April 1983", "year": 1983, "medium": "screenplay-draft", "ref": "t1-draft-gun-shop", "evidence": "The 1983 fourth draft says something different - 'pulse-laser', not 'rifle'.", "quote": "A phased plasma pulse-laser in the forty watt range...", "verified": true, "tier": "TERTIARY" }, { "work": "The Terminator (novelization, Frakes & Wisher)", "year": 1985, "medium": "novelization", "ref": "t1-novelization", "evidence": "Reese's future-war plasma rifle is named. Wiki-cited to pp. 19 and 178.", "quote": "Westinghouse M-25", "verified": false, "tier": "TERTIARY", "note": "Wiki transcription; NOT verified against the printed novelization." }, { "work": "T2: The Future War (S.M. Stirling)", "year": 2003, "medium": "novel", "ref": "future-war-novel", "evidence": "Reported retcon of the trademark from 40 watts to 40 megawatts.", "quote": "Cyberdyne Systems Phased Plasma Rifle, 40MGWT Range", "verified": false, "tier": "TERTIARY", "note": "Reported second-hand via a wiki citing a third-party episode guide. NOT verified against the novel." }, { "work": "goingfaster.com 'Terminator 2029 AD' fan site", "year": 1998, "medium": "fan-site", "ref": "goingfaster-t2029ad", "evidence": "The circulating M-25 technical readout. The author's own disclaimer states the data came from an unrealised, never-submitted PC game pitch.", "quote": "", "verified": false, "tier": "NOT-CANON", "note": "THE most successfully laundered pseudo-canon in the franchise: a real novelization name with invented numbers attached. The index must not repeat any of its figures." } ], "canon_confidence": "explicit", "canonical_target": { "metric": "directed-energy weapon output - DUAL TARGET, both sourced", "value": "PRIMARY (film, explicit): 40 W. SECONDARY (franchise retcon, T2: The Future War): 40 MW. Functional bracket (extrapolated): ~50-100 kW effective, shoulder-fired, self-powered, rifle-range lethality against armour", "reasoning": "Publishing only the literal 40 W would make the component look solved and mislead the reader; publishing only a sensible figure would quietly discard the most famous specification in the franchise. Publishing both, labelled, is the only version that is simultaneously accurate, useful and funny. Note the secondary figure is NOT the index's invention - the franchise retconned its own joke. Under the literal reading this is the one component humanity has already beaten, by roughly three orders of magnitude, and it should be reported that way, deadpan." }, "assumed_by_sota_agent": "A shoulder-carried, self-powered directed-energy weapon — 'phased plasma,' nominally rated 40 W — that instantly destroys a human being or breaches reinforced structure at hundreds of metres, at night, in smoke and dust, with no external power supply and no visible dwell time." }, "real": { "status": "in_progress", "trl": 8, "spec_fraction": 0.1, "spec_fraction_rationale": "At the man-portable scale canon requires, the best real weapon (China's Lijian III, shown June 2026) delivers roughly 8 kJ — 2 kW reported, over about 4 s — into a thin-skinned drone at ~500 m, from a 25 kg system split across an emitter, an air cooler and a battery backpack. The canon effect needs order 1e5 J deposited essentially instantaneously from a ~4.5 kg weapon with no external pack. Energy-on-target ratio 8e3/1e5 = 0.08; mass ratio 4.5/25 = 0.18; time-to-effect ratio ~0 (4 s vs instantaneous). Blended: 0.10. Scored against the depicted effect, not the stated 40 W, which fielded systems exceed by 2,500x.", "gap": "Fielded high-energy lasers are 60-100 kW machines the size of a shipping container or a destroyer mount, needing seconds of dwell per target and failing in rain, dust and smoke. Canon needs a rifle. Nothing bridges the two: the waste heat from a 50 kW-class laser (75-150 kW at 25-40% wall-plug efficiency) has no passive rejection path from a rifle-sized body, and 100 kJ of delivered optical energy per shot implies roughly a kilogram of lithium cell per shot before margin.", "why_hard": "Wall-plug efficiency and thermal rejection. Every joule not radiated becomes heat inside the weapon, and the mass required to store and dump that energy scales faster than any plausible miniaturisation. Beam control adds a second wall: aperture size sets spot size at range, and a rifle-scale aperture cannot hold a lethal fluence on a moving target at hundreds of metres.", "movers": [ { "name": "Rafael Advanced Defense Systems", "kind": "company", "country": "IL", "what": "Built Iron Beam, the world's first operational high-energy laser air-defence system; in full-rate production for the IDF including truck-mounted units." }, { "name": "Elbit Systems", "kind": "company", "country": "IL", "what": "Manufactures the Iron Beam laser source." }, { "name": "Lockheed Martin", "kind": "company", "country": "US", "what": "Built HELIOS (Mk 5 Mod 0, 60 kW) on USS Preble; awarded a JLWS agreement via Lockheed Martin Aculight in July 2026." }, { "name": "nLIGHT Defense", "kind": "company", "country": "US", "what": "Co-awardee on the Joint Laser Weapon System containerised 150 kW prototype." }, { "name": "MBDA UK", "kind": "company", "country": "GB", "what": "Leads DragonFire; £316M November 2025 contract for fitment to two Type 45 destroyers by 2027." }, { "name": "Epirus", "kind": "company", "country": "US", "what": "Leonidas / IFPC-HPM high-power microwave counter-swarm systems; Gen II contract signed July 2025." }, { "name": "Harbin Xinguang Optic-Electronics Technology", "kind": "company", "country": "CN", "what": "Showed the Lijian II and Lijian III man-portable anti-drone lasers in Beijing, June 2026 — the closest thing to a soldier-carried directed-energy weapon anywhere." }, { "name": "OUSW(R&E) Scaled Directed Energy (SCADE) CTA", "kind": "program", "country": "US", "what": "Runs the Joint Laser Weapon System; $847M ceiling toward containerised 150 kW then 300-500 kW lasers." } ], "evidence": [ { "date": "2025-12-28", "claim": "Israel declared Rafael's Iron Beam, a ~100 kW class laser air-defence system with up to ~10 km range, operational and delivered to the IDF; it is now in full-rate production and has been used in combat.", "source": "Wikipedia (aggregating Israeli MoD and Rafael announcements)", "title": "Iron Beam", "url": "https://en.wikipedia.org/wiki/Iron_Beam", "kind": "product", "delta": "+" }, { "date": "2026-02-02", "claim": "USS Preble's 60 kW-class HELIOS laser neutralised four uncrewed aerial vehicles in an at-sea counter-UAS demonstration conducted in autumn 2025; engagement interval, dwell time, target range and flight profiles were all left undisclosed.", "source": "The War Zone", "title": "USS Preble Used HELIOS Laser To Zap Four Drones In Expanding Testing", "url": "https://www.twz.com/sea/uss-preble-used-helios-laser-to-zap-four-drones-in-expanding-testing", "kind": "demo", "delta": "+" }, { "date": "2026-07-09", "claim": "The Department of War awarded two Joint Laser Weapon System OTA agreements to nLIGHT Defense and Lockheed Martin Aculight, initial value $86M against an $847M ceiling; initial prototypes rated at approximately 150 kW, scaling to the 300-500 kW threshold stated as required for cruise-missile defence.", "source": "US Department of War (via GlobalSecurity.org mirror)", "title": "Department of War Announces Awards $86 Million Joint Laser Weapon System Agreements to Scale Directed Energy Capabilities", "url": "https://www.globalsecurity.org/military/library/news/2026/07/mil-260709-dod01.htm", "kind": "funding", "delta": "+" }, { "date": "2026-01-12", "claim": "CRS reports shipboard laser operating cost of approximately $1 to less than $10 per shot, against a Navy assessment of ~$100M per 60 kW unit and up to $200M per 250 kW unit, and names the binding limits: atmospheric absorption and scattering, thermal blooming, line-of-sight-only engagement, and space, weight, power and cooling on surface combatants.", "source": "Congressional Research Service", "title": "Navy Shipboard Lasers: Background and Issues for Congress (R44175)", "url": "https://www.everycrsreport.com/files/2026-01-12_R44175_d7afbb9895f1f72259cfb1f7559782c0bb354e02.html", "kind": "regulation", "delta": "0" }, { "date": "2025-08-06", "claim": "The US Army has built 17 directed-energy prototypes through its Rapid Capabilities and Critical Technologies Office and deployed 11, including four 50 kW Directed Energy M-SHORAD Stryker systems to CENTCOM, but concluded that sustaining them in battlefield conditions needs work and that industry cannot yet manufacture them at scale. RCCTO director Lt Gen Robert Rasch: \"We have got to work on maintainability because ... we cannot get by with the thought of having clean rooms out in combat.\" The Army is standing up the Enduring High Energy Laser programme to move from prototypes to fieldable systems.", "source": "Defense News", "title": "Army readies to launch 2026 competition for counter-drone laser weapon", "url": "https://www.defensenews.com/land/2025/08/06/army-readies-to-launch-2026-competition-for-counter-drone-laser-weapon/", "kind": "incident", "delta": "-" }, { "date": "2025-07-17", "claim": "Epirus received a $43,551,060 US Army contract for two IFPC-HPM Generation II high-power microwave counter-swarm systems, claiming more than double the Gen I maximum effective range and a projected 30% power increase; four Gen I systems were delivered in May 2024.", "source": "Epirus", "title": "Epirus Receives $43.5 Million Contract from U.S. Army for IFPC-HPM Generation II Systems", "url": "https://www.epirusinc.com/press-releases/epirus-receives-43-million-contract-from-u-s-army-for-ifpc-hpm-generation-ii-systems", "kind": "funding", "delta": "+" }, { "date": "2026-06-20", "claim": "Harbin Xinguang Optic-Electronics Technology displayed the Lijian II (~30 kg) and Lijian III (~25 kg) man-portable anti-drone lasers at the Defence Information Equipment and Technology Exhibition in Beijing; the systems split into an emitter, an air cooler and a control terminal with a battery-and-cooling backpack. Output power is not officially disclosed and has been reported as approximately 2 kW.", "source": "Interesting Engineering", "title": "Chinese firm's man-portable laser weapon zaps drones, fits in backpack", "url": "https://interestingengineering.com/innovation/chinese-man-portable-laser-weapon", "kind": "product", "delta": "+" }, { "date": "2025-11-20", "claim": "UK MoD reported DragonFire trials at the Hebrides range including a UK-first above-the-horizon track and defeat of drones flying at up to 650 km/h, at a quoted cost of around £10 per shot; a £316M contract will fit the system to Royal Navy Type 45 destroyers by 2027.", "source": "UK Defence Equipment & Support", "title": "Boost for armed forces as new laser weapon takes down high-speed drones", "url": "https://des.mod.uk/boost-for-armed-forces-as-new-laser-weapon-takes-down-high-speed-drones", "kind": "demo", "delta": "+" } ], "indicators": [ { "metric": "Peak output of the most powerful laser weapon in operational service", "unit": "kW", "value": 100, "as_of": "2025-12-28", "direction": "up_is_progress", "canon_target": 0.04, "source": "https://en.wikipedia.org/wiki/Iron_Beam" }, { "metric": "System mass of the lightest man-portable directed-energy weapon publicly shown", "unit": "kg", "value": 25, "as_of": "2026-06-20", "direction": "down_is_progress", "canon_target": 4.5, "source": "https://interestingengineering.com/innovation/chinese-man-portable-laser-weapon" }, { "metric": "Dwell time for a man-portable laser to defeat a small UAS at ~500 m", "unit": "s", "value": 4, "as_of": "2026-06-20", "direction": "down_is_progress", "canon_target": 0, "source": "https://www.tomshardware.com/tech-industry/china-shows-off-a-backpack-sized-anti-drone-laser-that-one-soldier-can-carry" }, { "metric": "Operating cost per shot of a fielded high-energy laser", "unit": "USD", "value": 10, "as_of": "2026-01-12", "direction": "down_is_progress", "canon_target": 1, "source": "https://www.everycrsreport.com/files/2026-01-12_R44175_d7afbb9895f1f72259cfb1f7559782c0bb354e02.html" } ] }, "commentary": "Canon specified forty watts, which is a reading lamp, then showed the weapon burst a running child. The market has moved anyway. Israel's Iron Beam entered service in December 2025 at 100 kilowatts and a few dollars a shot, and Washington has since committed $847 million toward a containerised 150-kilowatt successor. None of it is portable. The lightest man-carried directed-energy weapon shown anywhere weighs twenty-five kilograms, arrives with a backpack of batteries and coolers, and takes four seconds to burn through a hobby drone at five hundred metres. The energy problem is being solved in megawatts. Canon needs it solved in kilograms.", "last_reviewed": "2026-08-29", "review_cadence_days": 30, "progress": 0.0889 }, { "id": "armed-autonomous-ground-platforms", "name": "Armed autonomous ground platforms", "category": "weapons", "weight": 7, "one_liner": "HK-Tanks, and their real descendants, all of which are on a radio leash.", "canon": { "requirement": "Uncrewed tracked and legged gun platforms operating with no human in the loop in destroyed urban terrain, engaging dismounted humans, coordinated with air and humanoid units, and built without human labour.", "quantified": [ { "metric": "classes", "value": "'Ground HKs (tank-like robot gun-platforms)'; 'four-legged gun-pods called Centurions'", "source_ref": "t2-script-5", "tier": "TERTIARY" }, { "metric": "production", "value": "'built in automated factories'", "source_ref": "t1-reese-briefing", "tier": "PRIMARY" }, { "metric": "role", "value": "patrol, pursuit, area denial, camp security, mass disposal", "source_ref": "t1-reese-briefing", "tier": "PRIMARY" }, { "metric": "human oversight", "value": "none", "source_ref": "t1-reese-briefing", "tier": "PRIMARY" }, { "metric": "combined-arms integration", "value": "operates alongside Aerial HKs and endoskeletons in the same engagement", "source_ref": "t2-script-5", "tier": "TERTIARY" } ], "sources": [ { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-reese-briefing", "evidence": "The only dialogue naming them.", "quote": "HKs? / Hunter-Killers... patrol machines built in automated factories.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day - revised final shooting script", "year": 1991, "medium": "shooting-script", "ref": "t2-script-5c", "evidence": "2029 battlefield stage direction.", "quote": "Another APC is crushed under the treads of a massive Ground HK.", "verified": true, "tier": "TERTIARY" } ], "canon_confidence": "implied", "canonical_target": { "metric": "autonomous armed ground platform capability", "value": "Uncrewed tracked and legged armed ground platforms conducting autonomous patrol, pursuit and engagement of dismounted humans in unmapped rubble terrain, with no human in or on the loop, integrated into combined-arms operations with autonomous air, and produced without human labour", "reasoning": "The Ground HK is seen but never named in T1 or T2 dialogue - 'HK' is named generically - so the platform description is TERTIARY and labelled. The demanding clauses are the ones canon states in dialogue: no human in the loop, and built in automated factories." }, "assumed_by_sota_agent": "A tracked or legged armed ground machine that hunts, identifies and kills humans in ruined urban terrain, at night, under no human control and with no communications link to any controller." }, "real": { "status": "in_progress", "trl": 7, "spec_fraction": 0.14, "spec_fraction_rationale": "Scored across the four things canon requires. Armed chassis that shoots: ~0.5 (real armed UGVs exist and are used in combat, though the best-documented one could only fire from a halt). Operating in ruined urban terrain: ~0.3 (demonstrated, but Uran-9's effective control range collapsed to 300-500 m against a 3,000 m design). Identifying a specific human target: ~0.05. Engaging without a human in the loop: 0.0 — zero documented instances anywhere. Unweighted mean 0.21; weighted toward the two properties the component is named for (autonomy and lethal decision), 0.14.", "gap": "The autonomy. Every armed ground robot in combat use today is teleoperated over a video link, usually with a drone overhead supplying the operator's picture. Ukraine's celebrated robot-only assaults put no human body across the line of departure; they put no autonomy on the objective either. Navigation autonomy kits exist and are sold; engagement autonomy is not fielded by anyone.", "why_hard": "The communications link, and the economics. Urban terrain destroys the radio range these machines are specified for, and control loss is routine rather than exceptional. Removing the link means solving target identification and engagement authority onboard, which nobody has fielded. Meanwhile the cost curve has inverted: the US Army cancelled its Robotic Combat Vehicle in May 2025 on the reasoning that an $800 drone with a cheap munition defeats a $3 million vehicle indefinitely.", "movers": [ { "name": "Milrem Robotics", "kind": "company", "country": "EE", "what": "THeMIS UGV, first in Ukrainian service since August 2022; Type-X and HAVOC armed platforms; sells an Intelligent Function Integration Kit for navigation autonomy." }, { "name": "13th 'Khartiia' National Guard Brigade", "kind": "program", "country": "UA", "what": "Conducted the December 2024 uncrewed combined-arms assault near Hlyboke and Lyptsi with more than 50 systems, and a follow-on strongpoint clearance near Kupiansk in February 2026." }, { "name": "Textron Systems", "kind": "company", "country": "US", "what": "Ripsaw M3 selected March 2025 as the US Army's first Robotic Combat Vehicle; programme cancelled two months later." }, { "name": "Kalashnikov Concern / 766 UPTK", "kind": "company", "country": "RU", "what": "Uran-9 combat UGV, whose 2018 Syrian trial results remain the most detailed public failure analysis of an armed ground robot." }, { "name": "Unitree Robotics", "kind": "company", "country": "CN", "what": "AS2-W wheeled quadruped, 25 kg with 150 kg static payload and 150 TOPS onboard; appeared armed in the China-Mongolia Steppe Partner 2026 exercise. The company states weaponised modifications are unauthorised." }, { "name": "Ghost Robotics / SWORD International", "kind": "company", "country": "US", "what": "Vision 60 quadruped with the SPUR 6.5 mm Creedmoor rifle payload; Q-UGVs under MARSOC evaluation with Onyx SENTRY remote weapon stations." } ], "evidence": [ { "date": "2018-06-18", "claim": "A.P. Anisimov of the Russian MoD's 3rd Central Research Institute reported the Uran-9's Syrian combat trial results: effective control range of only 300-500 m in low-rise urban terrain against a 3,000 m design, 17 short-term and 2 long-term (up to 1.5 hour) control losses, electro-optical detection limited to 2 km, six 30 mm cannon firing delays and one complete failure, and no weapon or sensor stabilisation, meaning the vehicle could fire only from a halt. His conclusion was that modern Russian combat UGVs cannot perform assigned tasks in conventional combat.", "source": "Defence Blog", "title": "Syrian combat trials expose flaws in Russia's unmanned mini-tank", "url": "https://defence-blog.com/combat-tests-syria-brought-light-deficiencies-russian-unmanned-mini-tank/", "kind": "incident", "delta": "-" }, { "date": "2018-06-25", "claim": "A US Army TRADOC Mad Scientist analysis of the same Russian report notes the control-range collapse alone wiped out up to nine-tenths of the Uran-9's total operational range, and confirms the venue as the 10th All-Russian Scientific Conference 'Actual Problems of Defence and Security', April 2018.", "source": "US Army TRADOC Mad Scientist Laboratory", "title": "Russian Ground Battlefield Robots: A Candid Evaluation and Ways Forward", "url": "https://madsciblog.tradoc.army.mil/63-russian-ground-battlefield-robots-a-candid-evaluation-and-ways-forward/", "kind": "incident", "delta": "-" }, { "date": "2025-05-19", "claim": "The US Army cancelled development of the Robotic Combat Vehicle, less than three months after selecting Textron's Ripsaw M3 as the winner in March 2025. The Army Secretary's stated reasoning was that the programme accumulated too many requirements and fell behind the cost curve while an $800 drone with a cheap munition can destroy a $3 million vehicle repeatedly.", "source": "Military.com", "title": "Is the Army's Robotic Combat Vehicle Program Dead? So Much for Robot Tanks", "url": "https://www.military.com/off-duty/autos/armys-robotic-combat-vehicle-program-dead-so-much-robot-tanks.html", "kind": "funding", "delta": "-" }, { "date": "2025-09-03", "claim": "A Dutch-led initiative committed more than 150 additional Milrem THeMIS UGVs to Ukraine, joining the 15 already deployed since 2022, with significant final assembly by VDL Defentec.", "source": "Milrem Robotics", "title": "Milrem Robotics to deliver a record number of THeMIS UGVs to Ukraine in collaboration with an EU member country", "url": "https://milremrobotics.com/milrem-robotics-to-deliver-a-record-number-of-themis-ugvs-to-ukraine-in-collaboration-with-an-eu-member-country/", "kind": "product", "delta": "+" }, { "date": "2025-07-23", "claim": "Ukraine's Khartiia brigade operates ground robots for assault, mine-laying, mine-clearing and direct fire, all remotely operated over video links with drone overwatch; no engagement is autonomous.", "source": "The Defender", "title": "Ground robots of Khartiia: How the brigade uses UGVs and what systems it operates", "url": "https://www.thedefender.media/en/insights/how-khartiia-uses-ugv/", "kind": "product", "delta": "0" }, { "date": "2026-03-26", "claim": "China unveiled an urban warfare drill featuring the latest generation of robotic wolf units; the Unitree AS2-W wheeled quadruped has since appeared armed with infantry weapons in the China-Mongolia Steppe Partner 2026 joint exercise, conducting reconnaissance and fire support alongside armoured vehicles.", "source": "Global Times", "title": "China unveils urban warfare drill featuring latest generation of robotic wolf units", "url": "https://www.globaltimes.cn/page/202603/1357609.shtml", "kind": "demo", "delta": "+" }, { "date": "2026-07-27", "claim": "Analysis of China's transition of commercial quadruped robots into armed military platforms, documenting the AS2-W's specifications and appearance in joint exercises.", "source": "Militarnyi", "title": "Robotic Wolf: China Unveils New AS2-W UGV", "url": "https://militarnyi.com/en/news/robotic-wolf-china-unveils-new-as2-w-ugv/", "kind": "product", "delta": "+" } ], "indicators": [ { "metric": "Armed or combat-support UGVs committed to a single theatre by one supplier", "unit": "vehicles", "value": 165, "as_of": "2025-10-01", "direction": "up_is_progress", "canon_target": 10000, "source": "https://milremrobotics.com/milrem-robotics-to-deliver-a-record-number-of-themis-ugvs-to-ukraine-in-collaboration-with-an-eu-member-country/" }, { "metric": "Documented combat engagements by an armed ground robot with no human in the loop", "unit": "engagements", "value": 0, "as_of": "2026-08-29", "direction": "up_is_progress", "canon_target": 1, "source": "https://www.thedefender.media/en/insights/how-khartiia-uses-ugv/" }, { "metric": "Effective teleoperation control range of an armed UGV in low-rise urban terrain", "unit": "m", "value": 400, "as_of": "2018-06-18", "direction": "up_is_progress", "canon_target": 0, "source": "https://defence-blog.com/combat-tests-syria-brought-light-deficiencies-russian-unmanned-mini-tank/" } ] }, "commentary": "Armed ground robots are in weekly combat use and not one of them is autonomous. Ukrainian brigades have taken positions without putting a body across the line of departure; every machine involved was flown by a human on a video link with a drone overhead supplying the picture. The instructive document remains Russian. The Uran-9's 2018 Syrian trials returned a 300-metre effective control range against a 3,000-metre design, seventeen control losses, and a cannon that could only fire from a halt. The chassis is a solved problem. The radio link is the weapon's weakest component, and the autonomy in the component's name does not yet exist anywhere.", "last_reviewed": "2026-08-29", "review_cadence_days": 30, "progress": 0.1089 }, { "id": "armed-autonomous-aerial-platforms", "name": "Armed autonomous aerial platforms", "category": "weapons", "weight": 7, "one_liner": "HK-Aerials, and their real descendants, which are autonomous only in the last few seconds.", "canon": { "requirement": "Two distinct things, and canon is unusually clear they are distinct: (a) existing crewed strategic bombers converted to FULL autonomy with a PERFECT operational record - the pre-war step that causes everything else; and (b) purpose-built VTOL hunter-killer gunships conducting autonomous night search-and-destroy of individual humans in urban rubble.", "quantified": [ { "metric": "fleet converted", "value": "ALL Stealth bombers, fully unmanned", "source_ref": "t2-skynet-history", "tier": "PRIMARY" }, { "metric": "operational record after conversion", "value": "'a perfect operational record'", "source_ref": "t2-skynet-history", "tier": "PRIMARY" }, { "metric": "purpose-built platform", "value": "Aerial HK - VTOL, turbine-powered, searchlight-equipped, night-hunting", "source_ref": "t2-script-4", "tier": "TERTIARY" }, { "metric": "employment", "value": "'a formation of flying HK (Hunter-Killer) patrol machines passes overhead'", "source_ref": "t2-script-4", "tier": "TERTIARY" }, { "metric": "vulnerability", "value": "a shoulder-launched rocket brings one down", "source_ref": "t2-script-5b", "tier": "TERTIARY" } ], "sources": [ { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-skynet-history", "evidence": "The most underrated line in T2: reliability is what earns the trust that removes the humans.", "quote": "All Stealth bombers are upgraded with Cyberdyne computers... becoming fully unmanned. Afterwards they fly with a perfect operational record.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day - revised final shooting script", "year": 1991, "medium": "shooting-script", "ref": "t2-script-4", "evidence": "2029 stage direction.", "quote": "The SOUNDS of ROARING TURBINES. Searchlights blaze down as a formation of flying HK (Hunter-Killer) patrol machines passes overhead.", "verified": true, "tier": "TERTIARY" } ], "canon_confidence": "explicit", "canonical_target": { "metric": "fleet autonomy record, and autonomous night hunter-killer capability", "value": "(a) Conversion of an entire crewed strategic bomber fleet to full autonomy with a ZERO-INCIDENT operational record; (b) purpose-built autonomous VTOL gunships conducting night search-and-destroy against individual humans in unmapped urban rubble, in formation, with no human in the loop", "reasoning": "Clause (a) is not a throwaway: canon's causal chain runs reliability -> trust -> removal of humans -> catastrophe, and the pivot is a fleet-wide zero-defect record. 'Perfect' is a measurable claim and belongs in the target verbatim, because partial autonomy with occasional incidents - which is where reality is - scores honestly low against it." }, "assumed_by_sota_agent": "A persistent, self-directed armed aircraft that searches a contested urban area, identifies human targets by itself, and engages them with no human authorisation for any individual engagement — indefinitely, on internal power." }, "real": { "status": "in_progress", "trl": 8, "spec_fraction": 0.2, "spec_fraction_rationale": "Scored across five canon requirements. Armed uncrewed aircraft striking without a person aboard: 1.0, entirely solved, at roughly 167 launches per day against Ukraine in May 2026. Persistence and loiter: ~0.3, hours rather than indefinitely. Autonomous terminal engagement: ~0.5, real in the final seconds of a loitering munition's flight. Autonomous target selection in a cluttered scene: ~0.05, with DARPA's X-62A the closest and it selected manoeuvres rather than targets. Operating with no human authorisation for any individual engagement: 0.0, acknowledged by no state. Unweighted mean 0.37; weighted toward autonomy and persistence, 0.20.", "gap": "Mission autonomy, and the authority to select a target. Loitering munitions fly to a coordinate a human chose and, in later variants, perform their own final seeking. Collaborative Combat Aircraft entered production in June 2026 with the autonomy software explicitly decoupled from the airframe and not yet selected. The Air Force bought the aeroplane before it bought the brain.", "why_hard": "Target discrimination under adversarial conditions, and the doctrine and law that follow from it. Terminal seeking against a pre-designated coordinate is tractable; deciding which human in a rubble field is a combatant is not, and no state has been willing to delegate it. Secondarily: endurance. Canon's machines patrol indefinitely; real armed uncrewed aircraft measure loiter in hours.", "movers": [ { "name": "General Atomics Aeronautical Systems", "kind": "company", "country": "US", "what": "YFQ-42A first flew August 2025; ordered into production as the FQ-42A Dark Merlin in June 2026." }, { "name": "Anduril Industries", "kind": "company", "country": "US", "what": "YFQ-44A first flew 31 October 2025; ordered into production as the FQ-44A Fury; also competing to supply CCA mission autonomy software." }, { "name": "DARPA", "kind": "agency", "country": "US", "what": "Air Combat Evolution (ACE) flew autonomous within-visual-range air combat on the X-62A against a manned F-16; separately pursuing persistent swarm infrastructure of up to 500 Group 1-3 aircraft from unattended containerised hubs, with an operational demonstration due autumn 2026." }, { "name": "USAF Test Pilot School / AFTC", "kind": "agency", "country": "US", "what": "Operates the X-62A VISTA testbed at Edwards AFB where the ACE dogfight trials were flown." }, { "name": "Shield AI", "kind": "company", "country": "US", "what": "Competing with Anduril and Collins to supply Collaborative Combat Aircraft mission autonomy software by summer 2027." }, { "name": "HESA / Alabuga (Yelabuga) production", "kind": "company", "country": "IR", "what": "Shahed-136 and its Russian Geran derivatives — the industrial-scale one-way attack drone that defines the low end of this component." } ], "evidence": [ { "date": "2024-04-17", "claim": "DARPA disclosed that in September 2023 the X-62A VISTA, a modified F-16 under AI control, flew within-visual-range air combat manoeuvres against a live, manned F-16 at Edwards AFB. Safety pilots were aboard with a disengage switch and never used it; performance was assessed as comparable to a human pilot with compliance to training rules.", "source": "DARPA", "title": "Air Combat Evolution (ACE)", "url": "https://www.darpa.mil/about/innovation-timeline/ace", "kind": "demo", "delta": "+" }, { "date": "2026-06-17", "claim": "The USAF ordered both the General Atomics FQ-42A Dark Merlin and the Anduril FQ-44A Fury into production, four months ahead of schedule, planning to procure over 150 combat-capable CCA by the end of the decade with nearly $1 billion requested in FY2027. The aircraft are described as semi-autonomous and human-supervised; mission autonomy software is a separate competition due by summer 2027.", "source": "The War Zone", "title": "USAF Orders Both General Atomics' FQ-42 And Anduril's FQ-44 Into Production", "url": "https://www.twz.com/air/usaf-orders-both-general-atomics-fq-42-and-andurils-fq-44-into-production", "kind": "product", "delta": "+" }, { "date": "2026-01-15", "claim": "The US military conducted what it described as the first kinetic drone swarm on American soil, at Camp Blanding, Florida, as part of the Pentagon's Swarm Forge pace-setting project.", "source": "DefenseScoop", "title": "U.S. military says it conducted the first kinetic drone swarm on American soil during a recent exercise", "url": "https://defensescoop.com/2026/01/15/drone-swarm-forge-demonstration-us-military-camp-blanding/", "kind": "demo", "delta": "+" }, { "date": "2026-07-07", "claim": "Shahed/Geran launch rates against Ukraine rose from approximately 144 per day in April 2026 to approximately 167 per day in May 2026, with Geran-2 production assessed at roughly 170 units per day by early 2026 and Russia planning about 60,000 long-range UAVs plus 50,000 decoys during 2026.", "source": "Institute for Science and International Security", "title": "Monthly Analysis of Russian Shahed 136 Deployment Against Ukraine", "url": "https://isis-online.org/isis-reports/monthly-analysis-of-russian-shahed-136-deployment-against-ukraine", "kind": "benchmark", "delta": "+" }, { "date": "2025-08-27", "claim": "General Atomics confirmed the YFQ-42A's first flight, the first of the two Collaborative Combat Aircraft Increment 1 designs to fly.", "source": "General Atomics Aeronautical Systems", "title": "GA-ASI Marks Another Aviation First With YFQ-42A CCA Flight Testing", "url": "https://www.ga-asi.com/ga-asi-marks-another-aviation-first-with-yfq-42a-cca-flight-testing", "kind": "demo", "delta": "+" } ], "indicators": [ { "metric": "One-way attack UAVs launched per day in a single active conflict", "unit": "launches/day", "value": 167, "as_of": "2026-05-31", "direction": "up_is_progress", "canon_target": 1000, "source": "https://isis-online.org/isis-reports/monthly-analysis-of-russian-shahed-136-deployment-against-ukraine" }, { "metric": "Collaborative combat aircraft types on a production contract", "unit": "types", "value": 2, "as_of": "2026-06-17", "direction": "up_is_progress", "canon_target": 1, "source": "https://www.twz.com/air/usaf-orders-both-general-atomics-fq-42-and-andurils-fq-44-into-production" }, { "metric": "Documented armed air engagements with no human authorisation for the individual engagement", "unit": "engagements", "value": 0, "as_of": "2026-08-29", "direction": "up_is_progress", "canon_target": 1, "source": "https://www.twz.com/air/usaf-orders-both-general-atomics-fq-42-and-andurils-fq-44-into-production" }, { "metric": "Largest swarm size targeted by a funded persistent-swarm programme", "unit": "aircraft", "value": 500, "as_of": "2026-03-31", "direction": "up_is_progress", "canon_target": 500, "source": "https://defensescoop.com/2026/03/31/pentagon-preparing-drone-swarm-crucible/" } ] }, "commentary": "The aerial half of the machine army is the one that arrived on schedule. Roughly 167 one-way attack drones a day were launched at Ukraine in May, from a production line running about 170 airframes a day, and in June the US Air Force ordered two collaborative combat aircraft types into production at once. What has not arrived is the decision-making. Every one of those weapons flies to a coordinate a person chose; the autonomy lives in the last few seconds. The Air Force bought the airframe in June 2026 and will not choose the mission-autonomy software until summer 2027. Skynet, characteristically, did it in the other order.", "last_reviewed": "2026-08-29", "review_cadence_days": 30, "progress": 0.1778 }, { "id": "living-tissue-over-metal", "name": "Living tissue over metal", "category": "biology-and-camouflage", "weight": 8, "one_liner": "Roughly 2.3 square metres of living human skin, grown, and worn by a machine.", "canon": { "requirement": "A full-body integument of CULTURED, LIVING human tissue - epidermis, dermis, vasculature, hair follicles, sweat glands, breath, a wound-healing response - grown over a rigid non-biological chassis and vascularised well enough to stay alive and heal in the field for the whole mission. Not a covering: an organ.", "quantified": [ { "metric": "composition", "value": "'flesh, skin, hair, blood'", "source_ref": "t1-reese-briefing", "tier": "PRIMARY" }, { "metric": "provenance", "value": "'grown for the cyborgs' - CULTURED, not harvested", "source_ref": "t1-reese-briefing", "tier": "PRIMARY" }, { "metric": "physiological functions present", "value": "sweat; bad breath; bleeding; wound healing", "source_ref": "t1-reese-briefing", "tier": "PRIMARY" }, { "metric": "healing", "value": "stated explicitly", "source_ref": "t2-garage-repair", "tier": "PRIMARY" }, { "metric": "coverage", "value": "full body including face, scalp and hands (~1.8 m2)", "source_ref": "derived", "tier": "EXTRAPOLATED" }, { "metric": "prior generation", "value": "'The 600 series had rubber skin. We spotted them easy.'", "source_ref": "t1-reese-briefing", "tier": "PRIMARY" }, { "metric": "structural statement", "value": "'Living tissue over metal endoskeleton.'", "source_ref": "t2-john-meets-t800", "tier": "PRIMARY" }, { "metric": "ageing (secondary)", "value": "Dark Fate's 'Carl' visibly ages over 22 years - the tissue has a normal lifespan", "source_ref": "dark-fate", "tier": "SECONDARY" } ], "sources": [ { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-reese-briefing", "evidence": "The composition and, critically, the provenance.", "quote": "But outside, it's living human tissue... flesh, skin, hair, blood... grown for the cyborgs.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-john-meets-t800", "evidence": "The machine's own self-description.", "quote": "I'm a cybernetic organism. Living tissue over metal endoskeleton.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-garage-repair", "evidence": "The healing response confirmed, and the operational reason it matters.", "quote": "Will these heal up? / Yes. / Good. If you can't pass for human, you're not much good to us.", "verified": true, "tier": "PRIMARY" } ], "canon_confidence": "explicit", "canonical_target": { "metric": "area and viability of cultured living integument on a non-biological chassis", "value": "A full-body (~1.8 m2) envelope of cultured living human tissue - stratified epidermis, vascularised dermis, hair follicles, eccrine glands - grown to fit a rigid non-biological chassis, remaining viable and mounting a normal wound-healing response for the duration of a multi-day deployment, and ageing at a normal human rate over decades", "reasoning": "1.8 m2 is standard adult body surface area and is the only defensible way to quantify 'full body'. The 'grown for the cyborgs' clause does important work and should not be skipped: canon specifies CULTURE, not transplant, making this a tissue-engineering target rather than a grafting one. The decades-long ageing requirement comes from a secondary work and is labelled, but it is the strictest viability statement the franchise makes." }, "assumed_by_sota_agent": "Full-body (~2.3 m2) living human skin — epidermis, dermis, subcutis, hair, sweat and sebaceous glands, vasculature, sensation — grown over a metal endoskeleton, self-maintaining for years without an external nutrient supply, and indistinguishable from human at conversational distance." }, "real": { "status": "on_horizon", "trl": 4, "spec_fraction": 0.06, "spec_fraction_rationale": "Scored across seven canon requirements. Coverage: the best living skin ever grown on a robot is a face-scale patch of order 10 cm2 against a 22,650 cm2 body (Du Bois BSA for 188 cm / 100 kg), so ~0.001. Full thickness with a dermis: ~0.35, dermis equivalents are real and clinically used. Vascularisation: 0.0. Appendages (hair, sweat and sebaceous glands): 0.0. Sensation: 0.0. Survival off a culture bath: 0.0. Appearance at conversational distance: ~0.4. Unweighted mean 0.11; weighted toward coverage, vascularisation and untethered survival, 0.06.", "gap": "Everything past a few hundred micrometres. Cultured skin is a clinical reality but only as a graft onto a living, perfused wound bed. Epicel grafts are 50 cm2 and 2 to 8 cell layers thick, ready no sooner than 15 days after biopsy; covering a 2.3 m2 body would take about 453 of them and would still be epidermis alone, with no dermis, no blood supply, no hair, no glands and no nerves. The one line of work that has put living skin on a machine (Takeuchi's group) has done it at face scale, avascular, and only while immersed.", "why_hard": "The oxygen diffusion limit. Without a vascular supply, engineered tissue is viable only within roughly 100 to 200 micrometres of a nutrient source; beyond that the core goes hypoxic and necroses. Human skin is 1 to 4 mm thick, so the canon requirement is 5 to 40 times the maximum viable thickness. This is set by oxygen's diffusion coefficient in tissue and by cellular consumption rate, neither of which is negotiable, and every working solution gets around it by borrowing a host's circulation. A machine has no host.", "movers": [ { "name": "Shoji Takeuchi Laboratory, University of Tokyo", "kind": "university", "country": "JP", "what": "The only group to have cultured living skin directly onto robotic structures — a robotic finger (2022) and a robotic face with perforation-type collagen anchors that actuate the skin into a smile (2024)." }, { "name": "Vericel", "kind": "company", "country": "US", "what": "Manufactures Epicel, the FDA humanitarian-exemption cultured epidermal autograft — 50 cm2 sheets, 2 to 8 cell layers, ~15 days from biopsy." }, { "name": "Stratatech / Mallinckrodt", "kind": "company", "country": "US", "what": "StrataGraft, FDA-approved June 2021: an allogeneic bilayered construct with both epidermal and dermal analogues; 96% of treated burn sites in the pivotal trial avoided autografting." }, { "name": "Organogenesis", "kind": "company", "country": "US", "what": "Apligraf, the long-established bilayered living cell therapy for venous leg and diabetic foot ulcers." }, { "name": "Avita Medical", "kind": "company", "country": "US", "what": "RECELL point-of-care autologous cell suspension, the fastest route from biopsy to coverage." }, { "name": "Wake Forest Institute for Regenerative Medicine", "kind": "university", "country": "US", "what": "Mobile in-situ skin bioprinting — depositing autologous fibroblasts and keratinocytes directly into a wound under imaging guidance." } ], "evidence": [ { "date": "2024-06-25", "claim": "Takeuchi's group demonstrated perforation-type anchors inspired by skin ligaments: V-shaped holes through a robotic substrate, plasma-treated so collagen gel penetrates them, allowing a skin equivalent to be adhered to a three-dimensional facial mould and actuated into a smile. The construct has no vasculature and no perfusion.", "source": "Cell Reports Physical Science", "title": "Perforation-type anchors inspired by skin ligament for robotic face covered with living skin", "url": "https://doi.org/10.1016/j.xcrp.2024.102066", "kind": "paper", "delta": "+" }, { "date": "2022-07-01", "claim": "The same group cultured a dermis equivalent and epidermis directly onto a robotic finger, demonstrating water repellence and self-repair of a laceration using a collagen sheet — the first living skin grown on a moving robot.", "source": "Matter", "title": "Living skin on a robot", "url": "https://doi.org/10.1016/j.matt.2022.05.019", "kind": "paper", "delta": "+" }, { "date": "2016-02-01", "claim": "Epicel's FDA-approved Directions for Use (Revision G, February 2016) specify grafts of approximately 50 cm2, 2 to 8 cell layers thick, available approximately 15 days after cultures are initiated, manufactured with bovine serum and irradiated murine 3T3 feeder cells, and requiring immobilisation under nylon net for 7 to 10 days after grafting.", "source": "US Food and Drug Administration", "title": "Epicel (cultured epidermal autografts) HDE# BH990200 Directions for Use", "url": "https://www.fda.gov/media/103138/download", "kind": "regulation", "delta": "0" }, { "date": "2021-06-15", "claim": "The FDA approved StrataGraft, an allogeneic bilayered cultured skin construct containing both keratinocytes and dermal fibroblasts, for adults with thermal burns containing intact dermal elements. In the pivotal Phase 3 trial 96% (68 of 71) of StrataGraft-treated burn sites did not require autografting.", "source": "US Food and Drug Administration", "title": "STRATAGRAFT", "url": "https://www.fda.gov/vaccines-blood-biologics/stratagraft", "kind": "regulation", "delta": "+" }, { "date": "2023-10-04", "claim": "Bioprinted skin containing multiple primary human cell types promoted rapid vascularisation and epidermal rete-ridge formation and accelerated wound closure without increased contraction; full-thickness bioprinted grafts were transplanted onto wounds in pigs — human-like skin architecture achieved in vivo, on a host with a circulation.", "source": "Science Translational Medicine", "title": "Multicellular bioprinted skin facilitates human-like skin architecture in vivo", "url": "https://pubmed.ncbi.nlm.nih.gov/37792956/", "kind": "paper", "delta": "+" }, { "date": "2023-02-03", "claim": "Engineered tissue without vasculature is limited in thickness by the oxygen diffusion limit of 100 to 200 micrometres; randomly self-assembled microvascular networks show irregular architecture, susceptibility to early thrombosis, and a post-implantation delay during which cells are deprived of oxygen.", "source": "Scientific Reports", "title": "Engineered tissue vascularization and engraftment depends on host model", "url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC9898562/", "kind": "paper", "delta": "0" } ], "indicators": [ { "metric": "Largest area of living skin cultured onto a non-biological robotic structure", "unit": "cm2", "value": 30, "as_of": "2024-06-25", "direction": "up_is_progress", "canon_target": 22650, "source": "https://doi.org/10.1016/j.xcrp.2024.102066" }, { "metric": "Maximum viable thickness of engineered tissue with no vascular supply", "unit": "um", "value": 200, "as_of": "2023-02-03", "direction": "up_is_progress", "canon_target": 2000, "source": "https://pmc.ncbi.nlm.nih.gov/articles/PMC9898562/" }, { "metric": "Area of a single FDA-cleared cultured epidermal autograft", "unit": "cm2", "value": 50, "as_of": "2016-09-19", "direction": "up_is_progress", "canon_target": 22650, "source": "https://www.fda.gov/media/103138/download" }, { "metric": "Time from biopsy to first available cultured epidermal graft", "unit": "days", "value": 15, "as_of": "2016-09-19", "direction": "down_is_progress", "canon_target": 0, "source": "https://www.fda.gov/media/103138/download" } ] }, "commentary": "Cultured skin is a mature clinical product and entirely useless here. Epicel arrives as 50-square-centimetre sheets, two to eight cell layers thick, fifteen days after biopsy; covering a T-800 would take about 453 of them, and would deliver epidermis with no dermis, no blood supply, no hair and no glands. The Tokyo group that has actually grown living skin onto a robot did it at face scale, avascular, in a bath. The binding number is the oxygen diffusion limit: 200 micrometres from a capillary, against human skin's one to four millimetres. Canon needs five to forty times the maximum thickness biology permits without a heart attached.", "last_reviewed": "2026-08-29", "review_cadence_days": 30, "progress": 0.0267 }, { "id": "biohybrid-integration", "name": "Biohybrid integration", "category": "biology-and-camouflage", "weight": 7, "one_liner": "Keeping the tissue alive on the machine. This is the one that makes the T-800 impossible.", "canon": { "requirement": "Keep that entire living envelope metabolically viable - perfused, oxygenated, thermoregulated, wound-healing, immune-competent - using a NON-BIOLOGICAL BODY as its sole life support. No heart, no lungs, no gut, no liver, no marrow. The machine is the organism's organ system.", "quantified": [ { "metric": "tissue must be alive", "value": "required by the transit rule: 'Nothing dead will go'", "source_ref": "t1-time-displacement", "tier": "PRIMARY" }, { "metric": "life-support source", "value": "the machine; no biological organ system is ever shown or mentioned", "source_ref": "derived", "tier": "EXTRAPOLATED" }, { "metric": "blood present", "value": "'flesh, skin, hair, blood'", "source_ref": "t1-reese-briefing", "tier": "PRIMARY" }, { "metric": "functions sustained", "value": "sweat, breath, bleeding, clotting, healing", "source_ref": "t1-reese-briefing", "tier": "PRIMARY" }, { "metric": "duration", "value": ">=120 years implied by the power-cell statement", "source_ref": "derived", "tier": "EXTRAPOLATED" } ], "sources": [ { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-time-displacement", "evidence": "The line that makes the tissue's aliveness non-negotiable, and the topological exception.", "quote": "Something about the field generated by a living organism. Nothing dead will go. / But this cyborg, if it's metal... / Surrounded by living tissue!", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-garage-repair", "evidence": "Days after arrival the tissue still bleeds and heals normally.", "quote": "Will these heal up? / Yes.", "verified": true, "tier": "PRIMARY" } ], "canon_confidence": "extrapolated", "canonical_target": { "metric": "metabolic maintenance of living tissue by a non-biological host", "value": "Indefinite metabolic maintenance of ~1.8 m2 of living human tissue - full perfusion, oxygenation, waste clearance, thermoregulation, wound healing and immune function - by a non-biological host body, with NO biological organ system present, sustained for decades", "reasoning": "This is the component that most rewards reading canon carefully, because nothing states it. Reese's line means the T-800's flesh must be ALIVE at the moment of transit; T2 shows it bleeding and healing days later; Dark Fate shows it ageing normally across 22 years. So the fiction requires open-ended biological life support FROM A MACHINE - a much harder and more interesting requirement than 'it has skin', and one that has to be derived rather than quoted." }, "assumed_by_sota_agent": "A machine that keeps ~2.3 m2 of full-thickness living human tissue metabolically viable, perfused, infection-free and thermally regulated for years of continuous field operation, with no external life support." }, "real": { "status": "theoretical", "trl": 3, "spec_fraction": 0.03, "spec_fraction_rationale": "Canon requires years of untethered viability — call it three years, roughly 26,000 hours — over 2.3 m2 of tissue. The best real records are a biohybrid hand that fatigues after 10 minutes and cannot leave its liquid bath, and a millimetre-scale cardiac biohybrid fish that ran 108 days inside an incubator with scheduled medium changes. Taking the most generous reading, 108 days against 1,095 gives 0.10 on duration alone, then multiplied by essentially zero on scale (millimetre-scale construct vs 2.3 m2) and zero on untethered operation. Rounded up to 0.03 to credit that the biology is real and improving.", "gap": "There is no artificial circulation. A square metre of skin is an organ with a metabolic rate: it needs oxygenated fluid within 200 micrometres of every cell, continuous waste and CO2 clearance, lymphatic drainage, glucose and lipid supply, immune surveillance against constant bacterial colonisation, and temperature held in a band a few degrees wide. In a human all of that is a heart, lungs, liver, kidneys, gut, spleen, thymus, lymphatics and hypothalamus. A Terminator has a power cell.", "why_hard": "Mass transport at organ scale, on a machine, indefinitely. The best real perfusion technology — ECMO and organ-perfusion machines — sustains a single organ for hours to a few days with a full-time clinical team, and fails on thrombosis, intimal hyperplasia and infection. Autologous tissue avoids rejection but not infection: skin is colonised continuously, and a perfused sheet on a machine has no immune system, no lymphatics and no systemic response. Within days it is a culture plate wrapped around a robot.", "movers": [ { "name": "Shoji Takeuchi Laboratory, University of Tokyo", "kind": "university", "country": "JP", "what": "Built the 18 cm biohybrid hand driven by ten multiple muscle tissue actuators, and identified necrosis of thick muscle as the barrier to larger biohybrid limbs." }, { "name": "Waseda University (Morimoto group)", "kind": "university", "country": "JP", "what": "Co-developed the MuMuTA sushi-roll muscle actuator architecture that made a >10 cm biohybrid limb possible at all." }, { "name": "Disease Biophysics Group, Harvard (Kevin Kit Parker)", "kind": "university", "country": "US", "what": "Built the biohybrid fish that holds the field's duration record: 108 days and about 38 million contractions from human iPSC-derived cardiomyocytes." }, { "name": "MIT (Ritu Raman group)", "kind": "university", "country": "US", "what": "Muscle-actuated biohybrid machines and the neuromuscular interfaces needed to control them." }, { "name": "MIT Media Lab Biomechatronics (Hugh Herr group)", "kind": "university", "country": "US", "what": "Myoneural actuators for implantable biohybrid systems, achieving fatigue resistance by living inside an animal that does the perfusion." } ], "evidence": [ { "date": "2022-02-11", "claim": "An autonomously swimming biohybrid fish built from human iPSC-derived cardiomyocytes maintained spontaneous activity for 108 days, equivalent to about 38 million beats — 16 to 18 times the previous records held by the biohybrid stingray (6 days) and a skeletal-muscle biohybrid actuator (7 days). The construct is millimetre-scale and was maintained in culture medium throughout.", "source": "Science", "title": "An autonomously swimming biohybrid fish designed with human cardiac biophysics", "url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC8939435/", "kind": "paper", "delta": "+" }, { "date": "2025-02-12", "claim": "An 18 cm biohybrid hand actuated by ten lab-grown human multiple muscle tissue actuators performed gestures and manipulated objects, but showed fatigue after 10 minutes of operation, needed about an hour of rest to recover, and had to remain suspended in liquid for the tendon anchors to move freely. The authors identify necrosis at the core of thick muscle as the limit on scaling.", "source": "Science Robotics", "title": "Biohybrid hand actuated by multiple human muscle tissues", "url": "https://pubmed.ncbi.nlm.nih.gov/39937887/", "kind": "paper", "delta": "+" }, { "date": "2025-09-03", "claim": "A miniature electronic biohybrid robot actuated by optogenetically stimulated neuromuscular tissue maintained mechanical functionality for more than two weeks, against roughly one week for earlier designs, with contraction persisting up to 20 minutes after a one-minute optical stimulation.", "source": "Science Robotics", "title": "Optogenetic neuromuscular actuation of a miniature electronic biohybrid robot", "url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC12767209/", "kind": "paper", "delta": "+" }, { "date": "2026-03-31", "claim": "A fatigue-resistant myoneural actuator with engineered recruitment biophysics was demonstrated for implantable biohybrid systems in a rodent model — achieving durability by being inside an animal whose circulation does the perfusion.", "source": "Nature Communications", "title": "A myoneural actuator with engineered biophysics for implantable biohybrid systems", "url": "https://doi.org/10.1038/s41467-026-70626-6", "kind": "paper", "delta": "+" }, { "date": "2025-11-07", "claim": "A field review of biohybrid actuators concludes that greater actuation, precise control, longer-term viability and programmability all remain open, with no route yet to organ-scale tissue viability on a non-biological substrate.", "source": "npj Robotics", "title": "Biohybrid actuators in robotics: recent trends and future perspectives of skeletal and cardiac muscle integration", "url": "https://doi.org/10.1038/s44182-025-00049-w", "kind": "paper", "delta": "0" } ], "indicators": [ { "metric": "Record continuous viability of an engineered tissue construct on a non-biological substrate", "unit": "days", "value": 108, "as_of": "2022-02-11", "direction": "up_is_progress", "canon_target": 1095, "source": "https://pmc.ncbi.nlm.nih.gov/articles/PMC8939435/" }, { "metric": "Continuous operating time of the largest biohybrid limb before fatigue", "unit": "minutes", "value": 10, "as_of": "2025-02-12", "direction": "up_is_progress", "canon_target": 1440, "source": "https://pubmed.ncbi.nlm.nih.gov/39937887/" }, { "metric": "Largest biohybrid limb operating outside a liquid culture bath", "unit": "cm", "value": 0, "as_of": "2026-08-29", "direction": "up_is_progress", "canon_target": 180, "source": "https://pubmed.ncbi.nlm.nih.gov/39937887/" } ] }, "commentary": "This is the component that quietly ends the argument. The most advanced biohybrid limb ever built is an eighteen-centimetre hand that fatigues after ten minutes, needs an hour to recover, and only works while suspended in nutrient broth. The field's endurance record — 108 days, thirty-eight million contractions — belongs to a millimetre-scale fish in an incubator. Canon requires two-point-three square metres of human tissue kept perfused, fed, drained, defended against infection and thermally regulated for years, with no heart, no lungs, no kidneys and no immune system. Every other component in this index is an engineering gap. This one is an organ transplant with no recipient.", "last_reviewed": "2026-08-29", "review_cadence_days": 30, "progress": 0.01 }, { "id": "signature-management", "name": "Signature management", "category": "biology-and-camouflage", "weight": 5, "one_liner": "Dogs can smell them. So can a thermal camera, and for better reasons.", "canon": { "requirement": "Match a human's olfactory, thermal, acoustic and electromagnetic signature closely enough to defeat human observers indefinitely - while canon EXPLICITLY DOCUMENTS ITS OWN FAILURE against canine olfaction.", "quantified": [ { "metric": "human observers defeated", "value": "all, always", "source_ref": "t1-full", "tier": "PRIMARY" }, { "metric": "olfactory signature actively produced", "value": "'sweat, bad breath, everything'", "source_ref": "t1-reese-briefing", "tier": "PRIMARY" }, { "metric": "THE WORKING COUNTERMEASURE", "value": "dogs, used systematically by the Resistance", "source_ref": "t1-dogs", "tier": "PRIMARY" }, { "metric": "canine response", "value": "continuous, agitated barking in proximity", "source_ref": "t2-phone-call", "tier": "PRIMARY" }, { "metric": "prior generation's tell", "value": "rubber skin - a VISUAL signature failure", "source_ref": "t1-reese-briefing", "tier": "PRIMARY" }, { "metric": "CBRN signature", "value": "walks through CS gas unaffected - no respiration signature at all", "source_ref": "t2-script-173e", "tier": "TERTIARY" } ], "sources": [ { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-dogs", "evidence": "The Resistance's actual counter-Terminator doctrine.", "quote": "I was dreaming about dogs. / We use them to spot Terminators.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-phone-call", "evidence": "Max the German Shepherd barks continuously at the T-1000 throughout the call.", "quote": "What the hell is the goddamn dog barking at?", "verified": true, "tier": "PRIMARY" } ], "canon_confidence": "explicit", "canonical_target": { "metric": "signature match against human AND canine sensing", "value": "Defeat unaided human sensing - visual, olfactory, thermal, acoustic - at contact range, indefinitely (canon: MET); AND defeat trained canine olfactory discrimination at close range (canon: FAILED)", "reasoning": "This is the one component where the fiction is honest about a shortfall, and the index should preserve that rather than smoothing it. Setting spec_fraction = 1.0 at 'passes for human' would make the canonical machine score 1.0 on a capability it demonstrably does not have. The correct denominator includes the dog, and canon itself says so: the Resistance's counter-Terminator doctrine is a dog on a leash at the sentry post. Real systems are therefore scored against a bar the fictional system also fails - unusual, and worth the commentary." }, "assumed_by_sota_agent": "A machine that passes as a warm, breathing, correctly-smelling human at conversational distance to an attentive person — right skin temperature, right skin feel, right odour, no anomalous acoustic or electromagnetic emissions — while failing detection by a trained dog, which canon establishes as the reliable counter." }, "real": { "status": "on_horizon", "trl": 4, "spec_fraction": 0.12, "spec_fraction_rationale": "Scored across five channels. Visual passing at conversational distance: ~0.5, animatronics are close but live motion is not. Acoustic: ~0.4, harmonic-drive and BLDC actuators whine audibly and quiet actuation is an active field. Electromagnetic: ~0.3, shielding is a mature discipline but switching converters radiate. Thermal: ~0.05 — my calculation gives a required skin temperature of 87 C to reject 1 kW radiatively from 1.9 m2 at 20 C ambient, against a human's 33 C, and keratinocytes die above about 45 C. Odour: 0.0, no source of a human volatile-organic-compound profile exists that is not living human tissue. Unweighted mean 0.25; weighted toward thermal and odour, the two channels canon explicitly names, 0.12.", "gap": "Heat and smell. Adaptive infrared skins can change what a machine looks like to a thermal camera but remove no energy; a machine producing a kilowatt still produces a kilowatt and its surface must sit tens of degrees above human skin temperature to shed it. On odour, human scent is a stable individual chemical fingerprint of 60-plus skin volatiles produced by lipid oxidation, protein turnover, sebaceous secretion and the skin microbiome. A machine can carry a perfume; it cannot carry a metabolism.", "why_hard": "Thermodynamics, and the fact that a dog's nose outperforms instrumentation. Waste heat must leave the body, and radiative flux scales as T^4, so a kilowatt-class machine cannot present a 33 C surface without an active refrigeration cycle that itself produces more heat. Meanwhile trained dogs detect n-amyl acetate at 1.14 to 1.9 parts per trillion, and the human odour signature they are keying on is a metabolic output rather than a substance that can be synthesised and dispensed.", "movers": [ { "name": "Coskun Kocabas group, University of Manchester / Bilkent", "kind": "university", "country": "GB", "what": "Graphene-based adaptive thermal camouflage: electrically tunable infrared emissivity by ionic-liquid intercalation of multilayer graphene." }, { "name": "Medical Detection Dogs / bio-detection research groups", "kind": "university", "country": "GB", "what": "Establishing published canine olfactory detection thresholds under blinded, handler-independent conditions — the empirical basis for the canon claim." }, { "name": "Konrad Lorenz Institute / University of Vienna (Penn et al.)", "kind": "university", "country": "AT", "what": "Characterised the individual and gender fingerprints in human body odour by GC-MS across 197 adults." }, { "name": "Institut fur Atemgasanalytik / Innsbruck (Filipiak et al.)", "kind": "university", "country": "AT", "what": "Quantified 64 C4-C10 volatile organic compounds emitted from human skin — the reference inventory of what a machine would have to reproduce." }, { "name": "US Army DEVCOM / signature management programmes", "kind": "agency", "country": "US", "what": "Adaptive infrared and multispectral camouflage for vehicles and personnel — the closest institutional analogue to the canon requirement." } ], "evidence": [ { "date": "2006-05-01", "claim": "Under naturalistic field-testing conditions, dogs detected n-amyl acetate at olfactory thresholds of 1.9 and 1.14 parts per trillion — the classic quantification of canine olfactory sensitivity.", "source": "Applied Animal Behaviour Science", "title": "Naturalistic quantification of canine olfactory sensitivity", "url": "https://doi.org/10.1016/j.applanim.2005.07.009", "kind": "paper", "delta": "0" }, { "date": "2019-01-22", "claim": "Ten dogs tested in a blinded eight-choice carousel showed olfactory detection thresholds for amyl acetate in liquid phase ranging from 30 parts per billion down to 1.5 parts per trillion, with some dogs roughly 30-fold more sensitive than previously reported, and considerable inter-dog variation.", "source": "Frontiers in Veterinary Science", "title": "Canine Olfactory Thresholds to Amyl Acetate in a Biomedical Detection Scenario", "url": "https://doi.org/10.3389/fvets.2018.00345", "kind": "paper", "delta": "0" }, { "date": "2007-04-22", "claim": "GC-MS analysis of axillary sweat, urine and saliva from 197 adults, with five sweat samples per subject over ten weeks, established that individuals carry a stable and distinctive chemical odour fingerprint.", "source": "Journal of the Royal Society Interface", "title": "Individual and gender fingerprints in human body odour", "url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC2359862/", "kind": "paper", "delta": "-" }, { "date": "2014-05-01", "claim": "64 C4-C10 volatile organic compounds were identified and quantified in skin emanations from 31 healthy volunteers, predominantly aldehydes and hydrocarbons — the inventory a machine would have to reproduce continuously to pass an olfactory check.", "source": "Journal of Chromatography B", "title": "Emission rates of selected volatile organic compounds from skin of healthy volunteers", "url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC4013926/", "kind": "paper", "delta": "-" }, { "date": "2018-07-11", "claim": "Electrically tunable infrared emissivity was demonstrated by intercalating ionic liquid into multilayer graphene, allowing an object's apparent thermal signature to be varied on demand — the reference demonstration for adaptive thermal camouflage.", "source": "Nano Letters", "title": "Graphene-Based Adaptive Thermal Camouflage", "url": "https://pubmed.ncbi.nlm.nih.gov/29947216/", "kind": "paper", "delta": "+" }, { "date": "2025-01-22", "claim": "Strong nonreciprocal thermal emission was observed experimentally, breaking the equivalence of emissivity and absorptivity imposed by Kirchhoff's law — new physics for directional thermal signature engineering, though it removes no heat.", "source": "arXiv", "title": "Observation of Strong Nonreciprocal Thermal Emission", "url": "https://arxiv.org/abs/2501.12947", "kind": "paper", "delta": "+" } ], "indicators": [ { "metric": "Best published canine olfactory detection threshold for a reference odorant", "unit": "ppt", "value": 1.14, "as_of": "2006-05-01", "direction": "down_is_progress", "canon_target": 1.14, "source": "https://doi.org/10.1016/j.applanim.2005.07.009" }, { "metric": "Skin temperature required to reject 1 kW radiatively from 1.9 m2 at 20 C ambient", "unit": "degC", "value": 87, "as_of": "2026-08-29", "direction": "down_is_progress", "canon_target": 33, "source": "https://pmc.ncbi.nlm.nih.gov/articles/PMC4013926/" }, { "metric": "Distinct human skin volatile organic compounds a machine would have to reproduce", "unit": "compounds", "value": 64, "as_of": "2014-05-01", "direction": "down_is_progress", "canon_target": 0, "source": "https://pmc.ncbi.nlm.nih.gov/articles/PMC4013926/" } ] }, "commentary": "Canon sets a bar and then, unusually, concedes the machine fails half of it: the T-800 fools people and never fools a dog. The real obstacle is not the dog. Reject a kilowatt of waste heat from one-point-nine square metres by radiation and the surface must sit near 87 degrees Celsius; human skin sits at 33, and cultured keratinocytes die above about 45. Adaptive infrared skins are real and change the apparent temperature without removing a single joule. Odour is worse: the signature is a metabolic output of sixty-plus volatiles, individual to the person. A machine can carry a perfume. It cannot carry a metabolism.", "last_reviewed": "2026-08-29", "review_cadence_days": 60, "progress": 0.0533 }, { "id": "damage-tolerance-and-graceful-degradation", "name": "Damage tolerance and graceful degradation", "category": "self-repair-and-durability", "weight": 8, "one_liner": "Still coming, on one arm.", "canon": { "requirement": "Retain mission capability through progressive destruction - losing the entire outer envelope, then limbs, then the lower body - degrading only in speed and reach, never in intent, with no single point of failure outside the processor.", "quantified": [ { "metric": "loss of 100% of tissue envelope to fire", "value": "fully operational", "source_ref": "t1-tanker", "tier": "PRIMARY" }, { "metric": "loss of one eye", "value": "continues; excises the ruined organ and conceals the damage", "source_ref": "t1-hotel-repair", "tier": "TERTIARY" }, { "metric": "loss of one arm's function", "value": "continues; repairs it by hand", "source_ref": "t1-hotel-repair", "tier": "TERTIARY" }, { "metric": "loss of one arm at the elbow (T2)", "value": "continues, one-handed", "source_ref": "t2-steel-mill", "tier": "PRIMARY" }, { "metric": "loss of BOTH LEGS AND PELVIS", "value": "continues, crawling, still closing on the target", "source_ref": "t1-press", "tier": "PRIMARY" }, { "metric": "impalement / 40 mm HE", "value": "temporarily disabling only", "source_ref": "t2-steel-mill", "tier": "PRIMARY" }, { "metric": "terminal condition", "value": "destruction of the chip", "source_ref": "t2-steel-mill", "tier": "PRIMARY" }, { "metric": "T-1000 equivalent", "value": "shot through, frozen, shattered, reassembled - with subsequent morphing faults", "source_ref": "t2-se-bugs", "tier": "PRIMARY-SE" } ], "sources": [ { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-press", "evidence": "The endoskeleton, blown in half by a pipe bomb, drags itself after Sarah until she crushes it in the hydraulic press. Visual only - no dialogue.", "quote": "", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-steel-mill", "evidence": "The only stated termination condition.", "quote": "There's one more chip. And it must be destroyed also.", "verified": true, "tier": "PRIMARY" } ], "canon_confidence": "implied", "canonical_target": { "metric": "mission capability retained versus platform mass lost", "value": "Retain goal-directed mission capability after cumulative loss of >=60% of platform mass and 100% of the outer integument, degrading only in mobility and reach, with the processor as the SOLE single point of failure", "reasoning": "The 60% figure comes from the T1 endgame - the endoskeleton has lost its flesh, an arm's function, an eye, and everything below the waist, and is still pursuing. Expressing the target as MASS FRACTION LOST rather than 'very tough' makes it comparable against real platforms, which is the point of the exercise." }, "assumed_by_sota_agent": "Full mission capability retained after catastrophic structural loss: the T-800 is bisected at the waist, loses both legs and continues to pursue by dragging itself on one arm, and elsewhere operates after being crushed by a truck, burned to the endoskeleton and having an arm destroyed in press machinery. Requirement: at least 50% of the body destroyed, mission continues, reconfiguration effectively instantaneous." }, "real": { "status": "in_progress", "trl": 4, "spec_fraction": 0.05, "spec_fraction_rationale": "The best cross-morphology result restores locomotion after the loss of one of six legs, about 17% of actuation, in under 120 seconds: 0.17/0.50 = 0.34 against the canon damage level, but on a morphology canon does not use. On the required morphology, a biped, the demonstrated value is zero - no published result shows a humanoid walking on one leg or manipulating with an arm destroyed - and the field's own safety literature establishes that a single loss of power ends the deployment. Blending the two, generously, 0.05.", "gap": "Damage adaptation is demonstrated only on statically stable or fixed-base morphologies. On bipeds there is no result at all, and an August 2026 safety analysis establishes that de-energising a walking humanoid is itself the hazard, so no humanoid can currently be certified even to fail safely under ISO 13849-1 or EN 60204-1. Fielded humanoids also fall in cascades and manage only 2-4 hours per charge.", "why_hard": "A biped's safe state is actively maintained rather than passively held, so any single failure in power, sensing, compute or actuation removes it. The standard engineering answer, redundancy, costs mass, and mass is the one thing a legged robot's energy budget cannot spend - which is why aerospace can afford triple redundancy and a humanoid cannot.", "movers": [ { "name": "Inria / Universite de Lorraine (Mouret group)", "kind": "lab", "country": "FR", "what": "Intelligent trial-and-error damage recovery: still the field's reference result eleven years on." }, { "name": "Southern University of Science and Technology", "kind": "university", "country": "CN", "what": "Stubborn (Jun 2026): unified reinforcement-learning framework for humanoid motion tracking and fall recovery under disturbance." }, { "name": "Siemens (Ding, Cui, Wang, Wen)", "kind": "company", "country": "US/CN", "what": "Identified the fail-passive gap: industrial humanoids cannot inherit classical machine-safety certification because removing power causes an uncontrolled fall." }, { "name": "Unitree Robotics", "kind": "company", "country": "CN", "what": "G1 EDU is the platform on which the fail-passive feasibility study was run, and the substrate for much public fall-recovery work." }, { "name": "Agility Robotics", "kind": "company", "country": "US", "what": "Building Digit v5 explicitly around 'cooperative safety'; holds the largest filed humanoid reliability dataset at 65,000 operating hours." }, { "name": "Beijing Municipal Government / World Humanoid Robot Games", "kind": "program", "country": "CN", "what": "The largest recurring public stress test of humanoid robustness: 666 teams and 2,000+ robots in August 2026." } ], "evidence": [ { "date": "2026-08-03", "claim": "Removing power from a walking biped causes an uncontrolled fall, so classical de-energisation is itself a hazard; onboard humanoid compute is not safety-rated and no end-to-end PLe or SIL 3 certification can be claimed.", "source": "arXiv", "title": "Toward Certified Functional Safety for Industrial Humanoid Robots: The Fail-Passive Gap and a Feasibility Study", "url": "https://arxiv.org/abs/2608.02809", "kind": "paper", "delta": "-" }, { "date": "2015-05-28", "claim": "A hexapod that has lost or broken a leg recovers a working gait in under two minutes by intelligent trial-and-error, without self-diagnosis.", "source": "Nature", "title": "Robots that can adapt like animals", "url": "https://www.nature.com/articles/nature14422", "kind": "paper", "delta": "+" }, { "date": "2026-06-11", "claim": "A single reinforcement-learning framework unifies humanoid motion tracking and fall recovery under disturbance, replacing separate training stages.", "source": "arXiv", "title": "Stubborn: A Streamlined and Unified Reinforcement Learning Framework for Robust Motion Tracking and Fall Recovery for Humanoids", "url": "https://arxiv.org/abs/2606.12814", "kind": "paper", "delta": "+" }, { "date": "2025-06-24", "claim": "Evolutionary gait reconfiguration restores locomotion in damaged legged robots in under an hour without pre-training.", "source": "arXiv", "title": "Evolutionary Gait Reconfiguration in Damaged Legged Robots", "url": "https://arxiv.org/abs/2506.19968", "kind": "paper", "delta": "+" }, { "date": "2026-06-24", "claim": "Digit humanoids have accumulated more than 65,000 hours of operation across nine customer facilities, the largest filed humanoid reliability figure.", "source": "U.S. Securities and Exchange Commission", "title": "Joint press release of Churchill Capital Corp XI and Agility Robotics, Inc. (EX-99.1)", "url": "https://www.sec.gov/Archives/edgar/data/0002074973/000121390026071287/ea029548401ex99-1.htm", "kind": "regulation", "delta": "+" } ], "indicators": [ { "metric": "Maximum fraction of actuation lost with locomotion retained, best published result on any morphology", "unit": "fraction", "value": 0.17, "as_of": "2015-05-28", "direction": "up_is_progress", "canon_target": 0.5, "source": "https://www.nature.com/articles/nature14422" }, { "metric": "Time to adapt to unanticipated damage", "unit": "s", "value": 120, "as_of": "2015-05-28", "direction": "down_is_progress", "canon_target": 1, "source": "https://www.nature.com/articles/nature14422" }, { "metric": "Humanoids certified for the fall hazard under ISO 13849-1 PLe or IEC 61508 SIL 3", "unit": "count", "value": 0, "as_of": "2026-08-03", "direction": "up_is_progress", "canon_target": 1, "source": "https://arxiv.org/abs/2608.02809" } ] }, "commentary": "The canonical requirement is a machine that keeps coming with half of it gone. The state of the art is a paper published this month establishing that cutting power to a walking humanoid is itself the hazard, because the upright state is actively maintained and de-energising it produces an uncontrolled fall, which is why no humanoid can presently be certified under ISO 13849-1 at all. The best damage-adaptation result in robotics remains a 2015 hexapod that relearned a gait in under two minutes after losing one leg of six. On two legs, losing one is not degradation. It is the end of the deployment.", "last_reviewed": "2026-08-29", "review_cadence_days": 30, "progress": 0.0222 }, { "id": "self-healing-materials", "name": "Self-healing materials", "category": "self-repair-and-durability", "weight": 5, "one_liner": "Closing its own wounds.", "canon": { "requirement": "Autonomous, room-temperature, repeatable closure of full-penetration structural damage in seconds - restoring both geometry AND load-bearing capability - plus reintegration of detached material. This is a T-1000 specification; the T-800 explicitly does NOT have it, and the contrast is deliberate.", "quantified": [ { "metric": "closure time", "value": "seconds - the wound closes within the shot", "source_ref": "t2-script-80a", "tier": "TERTIARY" }, { "metric": "damage class healed", "value": "full penetration by 12-gauge and .45; severed limbs; whole-body shattering", "source_ref": "t2-full", "tier": "PRIMARY" }, { "metric": "reintegration of detached mass", "value": "yes - severed material reverts to 'neutral polyalloy... a kind of thick mercury' and rejoins", "source_ref": "t2-script-80g", "tier": "TERTIARY" }, { "metric": "repeatability", "value": "unlimited within the film", "source_ref": "t2-full", "tier": "PRIMARY" }, { "metric": "degradation after extreme insult", "value": "morphing malfunctions after cryogenic shattering", "source_ref": "t2-se-bugs", "tier": "PRIMARY-SE" }, { "metric": "T-800 METAL self-healing", "value": "NONE - it must be repaired by hand", "source_ref": "t1-hotel-repair", "tier": "PRIMARY" }, { "metric": "T-800 TISSUE healing", "value": "yes", "source_ref": "t2-garage-repair", "tier": "PRIMARY" } ], "sources": [ { "work": "Terminator 2: Judgment Day - revised final shooting script", "year": 1991, "medium": "shooting-script", "ref": "t2-script-80a", "evidence": "Stage direction as the T-800 shoots the T-1000 off the car.", "quote": "Shiny liquid metal visible in the hole, which then closes.", "verified": true, "tier": "TERTIARY" }, { "work": "Terminator 2: Judgment Day - revised final shooting script", "year": 1991, "medium": "shooting-script", "ref": "t2-script-80g", "evidence": "The severed 'crowbar hand' reverts and rejoins the main mass.", "quote": "the 'crowbar hand'... reverts to the neutral polyalloy... a kind of thick mercury", "verified": true, "tier": "TERTIARY" }, { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-garage-repair", "evidence": "The T-800's tissue heals but its metal does not - the deliberate contrast.", "quote": "Will these heal up? / Yes.", "verified": true, "tier": "PRIMARY" } ], "canon_confidence": "implied", "canonical_target": { "metric": "autonomous structural self-repair rate and completeness", "value": "Autonomous closure of full-penetration structural damage in <5 s at ambient temperature, restoring FULL LOAD-BEARING CAPABILITY, with unlimited repeat cycles and reintegration of detached mass", "reasoning": "The <5 s figure is set by screen time: the wound closes within the shot. 'Restoring load-bearing capability' is the discriminator - many real self-healing polymers restore CONTINUITY without restoring STRENGTH, and canon requires the T-1000 to fight immediately afterwards. The T-800/T-1000 split should be preserved: canon has two different answers to this component and only the more advanced unit has it." }, "assumed_by_sota_agent": "Damaged covering and structure recover full mechanical function - and, for skin, sensory function - autonomously, at ambient conditions, in seconds to minutes, repeatedly, at wound scales made by rifle rounds. Scored here against the harder reading: a load-bearing structure that closes a roughly 10 mm hole by itself." }, "real": { "status": "in_progress", "trl": 4, "spec_fraction": 0.02, "spec_fraction_rationale": "The best integrated result recovers 82% of function over 7 days (6.05e5 s) in a soft polymer under mild heating; canon requires ~100% recovery in order-10 s in a structural material. On time alone the ratio is ~1.7e-5; on function 0.82; on material class, polymer skin rather than load-bearing metal, generously 0.2. A charitable non-multiplicative blend that credits the sensing-plus-healing integration gives 0.02.", "gap": "Self-healing is a soft-polymer capability, measured in days, not a structural-metal capability measured in seconds. The only autonomous metal crack healing ever observed is nanoscale, in vacuum, in nanocrystalline platinum, and was observed rather than engineered. Nothing closes a hole in a load path.", "why_hard": "Healing requires mobile chemistry - reversible bonds, free volume, chain diffusion - and the same mobility that lets a polymer re-knit is what makes it soft. Strength and healing are thermodynamically opposed: a material stiff enough to be a load path has essentially zero atomic mobility at ambient temperature, and metals close cracks only where fresh surfaces are atomically clean and pressed together.", "movers": [ { "name": "National University of Singapore (Tan Yu Jun group)", "kind": "university", "country": "SG", "what": "Self-healing magnetoelectric sensory skin that detects its own damage and repairs in air and underwater; demonstrated on an underwater robotic hand." }, { "name": "Sandia National Laboratories / Texas A&M", "kind": "lab", "country": "US", "what": "First observation of autonomous fatigue-crack healing in a metal, via cold welding in nanocrystalline platinum." }, { "name": "Vrije Universiteit Brussel (Self-Healing Soft Robotics)", "kind": "university", "country": "BE", "what": "Long-running programme on Diels-Alder and dynamic-bond elastomers for self-healing soft actuators and grippers." }, { "name": "University of Nebraska-Lincoln", "kind": "university", "country": "US", "what": "Soft robotic artificial muscle that detects and repairs its own damage without human intervention." }, { "name": "Wiley / Advanced Science authors on vitrimer soft robots", "kind": "university", "country": "CN", "what": "3D digital-light-printed self-healing, reprocessable soft robots using vinylogous urethane vitrimer chemistry." } ], "evidence": [ { "date": "2026-04-18", "claim": "A self-healing magnetoelectric sensory skin detects its own damage and recovers 82% of function in air after 7 days under mild heating, and nearly 100% underwater after 10 days, with 92% elastic recovery and 41 ms response.", "source": "EurekAlert / Advanced Materials", "title": "More than skin deep: NUS researchers develop electronic skin that senses, heals and thrives even under water", "url": "https://www.eurekalert.org/news-releases/1136525", "kind": "paper", "delta": "+" }, { "date": "2023-07-19", "claim": "Fatigue cracks in nanocrystalline platinum were observed healing autonomously by cold welding - the only demonstrated self-healing in a metal, at nanoscale, in vacuum.", "source": "Nature", "title": "Autonomous healing of fatigue cracks via cold welding", "url": "https://doi.org/10.1038/s41586-023-06223-0", "kind": "paper", "delta": "+" }, { "date": "2025-11-03", "claim": "Self-healing, reprocessable soft robots are produced by 3D digital light printing with vinylogous urethane vitrimer chemistry, healing at room temperature via dynamic covalent bond exchange (doi:10.1002/advs.202516901).", "source": "Advanced Science", "title": "Self-Healing and Reprocessable Soft Robots Using 3D Digital Light Printing", "url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC12786329/", "kind": "paper", "delta": "+" } ], "indicators": [ { "metric": "Healing efficiency of the best integrated self-sensing, self-healing robot skin", "unit": "%", "value": 82, "as_of": "2026-04-18", "direction": "up_is_progress", "canon_target": 100, "source": "https://www.eurekalert.org/news-releases/1136525" }, { "metric": "Time to near-full recovery in that material", "unit": "s", "value": 604800, "as_of": "2026-04-18", "direction": "down_is_progress", "canon_target": 10, "source": "https://www.eurekalert.org/news-releases/1136525" }, { "metric": "Largest length scale of demonstrated autonomous metal crack healing", "unit": "m", "value": 1e-9, "as_of": "2023-07-19", "direction": "up_is_progress", "canon_target": 0.01, "source": "https://doi.org/10.1038/s41586-023-06223-0" } ] }, "commentary": "Singapore's best self-healing skin knows when it has been cut and recovers 82 per cent of function, in seven days, with mild heating. Underwater it manages nearly full recovery, in ten. Canon requires a hole closing in seconds, in something load-bearing. On time alone that is a shortfall of five orders of magnitude, and the material in question is a soft elastomer, not a chassis. Metals do not self-heal above the nanoscale: the sole result, from 2023, is spontaneous crack closure in platinum, in vacuum, observed rather than engineered. Strength and healing want opposite things from atomic mobility. That is chemistry, not effort.", "last_reviewed": "2026-08-29", "review_cadence_days": 30, "progress": 0.0089 }, { "id": "field-self-repair", "name": "Field self-repair", "category": "self-repair-and-durability", "weight": 6, "one_liner": "The bathroom-mirror surgery scene.", "canon": { "requirement": "Unassisted self-diagnosis and MECHANICAL repair of internal actuator damage, by the machine itself, using improvised human hand tools, without a workshop, a manual or a second party - under mirror-reversed visual guidance - plus cosmetic concealment of damage it cannot repair.", "quantified": [ { "metric": "tools used", "value": "an X-Acto knife, small screwdrivers, needle and thread, a rag, a mirror", "source_ref": "t1-hotel-repair", "tier": "TERTIARY" }, { "metric": "systems repaired unaided", "value": "forearm actuator train; ocular assembly; sutured wrist and abdomen", "source_ref": "t1-hotel-repair", "tier": "TERTIARY" }, { "metric": "guidance", "value": "a bathroom mirror - mirror-reversed visual servoing on its own body", "source_ref": "t1-hotel-repair", "tier": "TERTIARY" }, { "metric": "downtime / anaesthesia", "value": "none; it works through it and leaves immediately", "source_ref": "t1-hotel-repair", "tier": "TERTIARY" }, { "metric": "cosmetic concealment", "value": "hat, collar, fresh shirt, overcoat, glove, sunglasses", "source_ref": "t1-hotel-repair", "tier": "TERTIARY" }, { "metric": "the one repair it CANNOT self-perform", "value": "its own CPU extraction - requires Sarah and John", "source_ref": "t2-se-chip-flip", "tier": "PRIMARY-SE" } ], "sources": [ { "work": "The Terminator - fourth draft, 20 April 1983", "year": 1983, "medium": "screenplay-draft", "ref": "t1-hotel-repair", "evidence": "Scene 152. The forearm repair. On screen in the finished film; this is the draft's stage direction.", "quote": "he pulls back a large flap of skin to reveal a complex trunk of SHEATHED CABLES AND HYDRAULICS. They slide as he moves his fingers... With small screwdrivers he begins to patiently disassemble the damaged mechanism around the 12-gauge hit.", "verified": true, "tier": "TERTIARY" }, { "work": "The Terminator - fourth draft, 20 April 1983", "year": 1983, "medium": "screenplay-draft", "ref": "t1-hotel-repair", "evidence": "Scene 158. The ocular repair, at the mirror.", "quote": "With a smooth motion the knife point enters the eyeball and cuts away the ruined sclera and cornea... He wipes with a rag to clear the electronic eye's vision.", "verified": true, "tier": "TERTIARY" } ], "canon_confidence": "implied", "canonical_target": { "metric": "unassisted in-field mechanical self-repair", "value": "Unassisted diagnosis and mechanical repair of internal actuator and sensor damage by the machine itself, using improvised human hand tools, under MIRROR-REVERSED visual guidance, restoring function in the field within minutes - plus autonomous cosmetic concealment of unrepairable damage", "reasoning": "The mirror is the part worth putting in the target. Repairing yourself through a mirror requires the machine to invert its own body schema and re-map every motor command - a harder problem than the manipulation itself, and the reason this component is separate from dexterous-manipulation. The concealment clause is included because canon treats it as part of the same routine: the machine repairs what it can and DRESSES what it cannot, which is a maintenance policy, not a costume change." }, "assumed_by_sota_agent": "Autonomous diagnosis and physical replacement or removal of an arbitrary internal component of itself, using improvised hand tools, with no external assistance and no prepared interfaces. The T-800 opens its own forearm with a scalpel, extracts a damaged servo assembly one-handed working from a mirror, and later excises its own damaged eye." }, "real": { "status": "on_horizon", "trl": 5, "spec_fraction": 0.04, "spec_fraction_rationale": "On a fielded humanoid the number of internal components the robot can autonomously remove and replace on itself is 1 - the battery pack, via a purpose-designed hot-swap interface - against a machine with of order 10,000 unique parts. Canon requires any component, with improvised tools and no prepared interface. Crediting the strength of on-orbit servicing on the adjacent problem and the modular-robot result on the real one gives 0.04.", "gap": "Robotic servicing of other machines is genuinely fielded and just advanced, but it is ground-commanded and uses purpose-designed capture interfaces. Self-directed repair exists only as a bespoke modular laboratory system and as a single commercial action, the battery swap. Humanoid fleets are actually maintained by human field-service organisations.", "why_hard": "Self-repair is a bootstrapping problem rather than a mechanics problem: the damaged system must simultaneously be diagnostician, surgeon and patient, and the fault requiring repair may sit in the sensing, compute or actuation needed to perform it. Underneath that sits an economic constraint - for anything not in orbit, spares and technicians are far cheaper than onboard dexterity.", "movers": [ { "name": "Northrop Grumman SpaceLogistics", "kind": "company", "country": "US", "what": "Mission Robotic Vehicle launched July 2026 with two NRL-built arms for in-orbit inspection, repair, relocation, upgrade and assembly; MEV-1/MEV-2 already fielded." }, { "name": "U.S. Naval Research Laboratory", "kind": "agency", "country": "US", "what": "Built the MRV's two fully articulated robotic arms and tooling - the only dexterous repair manipulators operating beyond Earth orbit." }, { "name": "Columbia University / University of Washington (Wyder, Lipson)", "kind": "university", "country": "US", "what": "Robot metabolism: modular Truss Link robots that self-assemble and absorb parts from other robots to grow and repair themselves." }, { "name": "UBTECH Robotics", "kind": "company", "country": "CN", "what": "Walker S2 performs an autonomous three-minute battery hot-swap - the only self-service action a humanoid does in routine commercial use." }, { "name": "Figure AI", "kind": "company", "country": "US", "what": "Built dedicated Field Service Management and Fleet Management systems - the honest state of humanoid repair, which is people in vans." } ], "evidence": [ { "date": "2026-07-22", "claim": "Northrop Grumman's Mission Robotic Vehicle launched with two NRL-built robotic arms and three Mission Extension Pods, rated for in-orbit satellite inspection, relocation, repair, upgrade, debris removal and assembly.", "source": "Northrop Grumman", "title": "Northrop Grumman's Mission Robotic Vehicle Launches, Ushering in a New Era of In-Space Servicing", "url": "https://news.northropgrumman.com/launch/northrop-grummans-mission-robotic-vehicle-launches-ushering-in-a-new-era-of-in-space-servicing", "kind": "product", "delta": "+" }, { "date": "2026-07-22", "claim": "Two Mission Extension Vehicles have docked with three commercial geostationary satellites and delivered more than ten years of combined mission extension.", "source": "Northrop Grumman", "title": "Northrop Grumman's Mission Robotic Vehicle Launches, Ushering in a New Era of In-Space Servicing", "url": "https://news.northropgrumman.com/launch/northrop-grummans-mission-robotic-vehicle-launches-ushering-in-a-new-era-of-in-space-servicing", "kind": "product", "delta": "+" }, { "date": "2025-07-16", "claim": "Modular robots self-assemble from 2D into 3D forms and can absorb material from other robots to grow and repair themselves - 'robot metabolism' (doi:10.1126/sciadv.adu6897).", "source": "Science Advances", "title": "Robot metabolism: Toward machines that can grow by consuming other machines", "url": "https://robotmetabolism.github.io/", "kind": "paper", "delta": "+" }, { "date": "2024-11-18", "claim": "Preprint of the robot metabolism work, with the Truss Link module design and self-assembly results.", "source": "arXiv", "title": "Robot Metabolism: Towards machines that can grow by consuming other machines", "url": "https://arxiv.org/abs/2411.11192", "kind": "paper", "delta": "0" }, { "date": "2026-04-29", "claim": "Humanoid fleets are kept running by purpose-built human field-service and fleet-management organisations, not by the robots themselves.", "source": "Figure AI", "title": "Ramping Figure 03 Production", "url": "https://www.figure.ai/news/ramping-figure-03-production", "kind": "product", "delta": "0" } ], "indicators": [ { "metric": "Internal components a fielded humanoid can autonomously replace on itself", "unit": "count", "value": 1, "as_of": "2026-08-29", "direction": "up_is_progress", "canon_target": 10000, "source": "https://www.figure.ai/news/ramping-figure-03-production" }, { "metric": "Robotic servicing vehicles operating on other spacecraft", "unit": "count", "value": 3, "as_of": "2026-07-22", "direction": "up_is_progress", "canon_target": 3, "source": "https://news.northropgrumman.com/launch/northrop-grummans-mission-robotic-vehicle-launches-ushering-in-a-new-era-of-in-space-servicing" }, { "metric": "Cumulative mission extension delivered by robotic servicers", "unit": "satellite-years", "value": 10, "as_of": "2026-07-22", "direction": "up_is_progress", "canon_target": 100, "source": "https://news.northropgrumman.com/launch/northrop-grummans-mission-robotic-vehicle-launches-ushering-in-a-new-era-of-in-space-servicing" } ] }, "commentary": "The bathroom-mirror scene requires a machine to diagnose and replace an arbitrary internal part of itself with improvised tools. The commercial state of the art is a three-minute battery swap. Northrop Grumman launched a robotic servicer to geostationary orbit in July with two arms and a repair mandate, and its predecessors have bought three satellites more than ten years of life between them: genuinely impressive, and the wrong problem, since those machines fix other machines, on prepared docking rings, under ground command. Everyone else's humanoids are maintained by technicians in vans. Figure built a field-service department. That is the actual answer.", "last_reviewed": "2026-08-29", "review_cadence_days": 30, "progress": 0.0222 }, { "id": "automated-factories", "name": "Automated factories", "category": "manufacture", "weight": 7, "one_liner": "Building them without us.", "canon": { "requirement": "Lights-out factories producing complete autonomous combat platforms end-to-end - sited, supplied, operated and maintained with ZERO human involvement - in an environment with no functioning civilian industry.", "quantified": [ { "metric": "named explicitly in dialogue", "value": "'patrol machines built in automated factories'", "source_ref": "t1-reese-briefing", "tier": "PRIMARY" }, { "metric": "human labour", "value": "none - the surviving humans are in camps 'for orderly disposal', or loading bodies", "source_ref": "t1-reese-briefing", "tier": "PRIMARY" }, { "metric": "output classes", "value": "Aerial HKs, Ground HKs, Centurions, endoskeletons, T-800s, T-1000 prototypes", "source_ref": "t2-script-5", "tier": "TERTIARY" }, { "metric": "duration of operation", "value": "~3 decades of continuous war production", "source_ref": "t1-timeline", "tier": "PRIMARY" }, { "metric": "scale shown (secondary)", "value": "a production floor of finished T-800s", "source_ref": "salvation-factory", "tier": "SECONDARY" } ], "sources": [ { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-reese-briefing", "evidence": "The term itself.", "quote": "Hunter-Killers... patrol machines built in automated factories.", "verified": true, "tier": "PRIMARY" }, { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-reese-briefing", "evidence": "And what the humans were doing instead of working the line.", "quote": "Most of us were rounded up... put in camps for orderly disposal. This was burned in by laser scanner. Some of us were kept alive... to work... loading bodies. The disposal units ran night and day.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator Salvation", "year": 2009, "medium": "film", "ref": "salvation-factory", "evidence": "A Skynet production floor.", "quote": "T-800s. There's so many of them.", "verified": true, "tier": "SECONDARY" } ], "canon_confidence": "explicit", "canonical_target": { "metric": "end-to-end lights-out manufacture of complete autonomous platforms", "value": "End-to-end lights-out manufacture of complete autonomous combat platforms - from raw material intake to a walking, armed, finished unit - with ZERO human labour anywhere in the line, sustained for decades in an environment with no civilian industrial base", "reasoning": "'Zero human labour anywhere in the line' is the clause that makes this hard and it is stated: canon is unambiguous that the humans are in disposal camps, not on the shop floor. Real 'lights-out' factories are lights-out in the CELL, not in the supply chain, the maintenance or the tooling - which is precisely the gap the index should measure." }, "assumed_by_sota_agent": "Skynet's automated factories produce Terminators with no human involvement anywhere in the chain - extraction, refining, alloying, machining, assembly, test, maintenance of the factory itself, and the design loop - indefinitely and at scale." }, "real": { "status": "in_progress", "trl": 8, "spec_fraction": 0.08, "spec_fraction_rationale": "Thirty days unattended against an unbounded requirement is at most ~0.3 if capped generously; ~81% of process steps are automated at the best consumer-electronics plant; and roughly one of four chain segments is covered - assembly yes, extraction, refining, self-maintenance and design no, ~0.25. 0.3 x 0.81 x 0.25 = 0.06, rounded up to 0.08 to credit FANUC literally being robots that build robots.", "gap": "Assembly is machine work and has been for two decades. What is not automated anywhere is the ends of the chain: extraction, refining, tooling changeover, exception handling, and maintenance of the factory itself. All of those are dexterity problems, so the factory that builds humanoids cannot yet be staffed by them.", "why_hard": "Automation scales with process repeatability, and the residual human labour in a dark factory is concentrated in precisely the non-repeatable work - fixing the machines, changing tooling, handling exceptions. That is the same manipulation problem as dexterous manipulation, wearing overalls, so the automated factory and the dexterous hand are one component, not two.", "movers": [ { "name": "FANUC", "kind": "company", "country": "JP", "what": "Oshino plant has run lights-out since 2001, robots assembling robots, unsupervised for up to thirty days at a stretch." }, { "name": "Xiaomi", "kind": "company", "country": "CN", "what": "Changping smart factory: eleven automated lines, ten million-plus phones a year, with a manufacturing platform claimed to diagnose faults and optimise process without human input." }, { "name": "Figure AI", "kind": "company", "country": "US", "what": "BotQ went from one humanoid a day to one an hour in under 120 days, at over 80% end-of-line first-pass yield; first-generation line rated at 12,000 units a year." }, { "name": "Agility Robotics", "kind": "company", "country": "US", "what": "RoboFab in Salem, Oregon, designed to support up to 10,000 humanoids annually." }, { "name": "International Federation of Robotics", "kind": "program", "country": "DE", "what": "Publishes the recurring robot-density and installation statistics that make factory automation trackable rather than anecdotal." }, { "name": "ASE Group", "kind": "company", "country": "TW", "what": "Operates dozens of lights-out semiconductor packaging and test factories - the largest fleet of dark plants anywhere." } ], "evidence": [ { "date": "2026-04-29", "claim": "BotQ reached one Figure 03 per hour, a 24x throughput gain in under 120 days, with over 350 units delivered, over 9,000 actuators produced, over 80% end-of-line first-pass yield and 99.3% battery-line yield; line capacity 12,000 units a year.", "source": "Figure AI", "title": "Ramping Figure 03 Production", "url": "https://www.figure.ai/news/ramping-figure-03-production", "kind": "product", "delta": "+" }, { "date": "2026-04-08", "claim": "Global industrial robot density reached 132 units per 10,000 employees; China holds roughly two million installed robots and took 54% of world installations in 2024.", "source": "International Federation of Robotics", "title": "Robot Density Surges in Europe, Asia, and Americas", "url": "https://ifr.org/ifr-press-releases/news/robot-density-surges-in-europe-asia-and-americas", "kind": "benchmark", "delta": "+" }, { "date": "2024-07-10", "claim": "Xiaomi's Changping dark factory runs eleven automated lines producing over ten million phones a year, with a platform the company says independently diagnoses equipment problems and improves process flows.", "source": "New Atlas", "title": "Xiaomi's self-optimizing autonomous factory will make 10M+ phones a year", "url": "https://newatlas.com/robotics/xiaomi-dark-robotic-factory", "kind": "product", "delta": "+" }, { "date": "2003-06-01", "claim": "FANUC's Oshino plant has robots building robots at about 50 per 24-hour shift and runs unsupervised for as long as 30 days at a time - the most-cited and least-re-measured figure in lights-out manufacturing.", "source": "Wikipedia, citing CNN Money", "title": "Lights out (manufacturing)", "url": "https://en.wikipedia.org/wiki/Lights_out_(manufacturing)", "kind": "product", "delta": "0" }, { "date": "2026-06-24", "claim": "Agility's RoboFab is designed to support production of up to 10,000 humanoid units annually.", "source": "U.S. Securities and Exchange Commission", "title": "Joint press release of Churchill Capital Corp XI and Agility Robotics, Inc. (EX-99.1)", "url": "https://www.sec.gov/Archives/edgar/data/0002074973/000121390026071287/ea029548401ex99-1.htm", "kind": "regulation", "delta": "+" } ], "indicators": [ { "metric": "Longest documented unattended run of a lights-out plant", "unit": "days", "value": 30, "as_of": "2003-06-01", "direction": "up_is_progress", "canon_target": 3650, "source": "https://en.wikipedia.org/wiki/Lights_out_(manufacturing)" }, { "metric": "Automation share of the best-in-class consumer-electronics plant", "unit": "%", "value": 81, "as_of": "2024-07-10", "direction": "up_is_progress", "canon_target": 100, "source": "https://newatlas.com/robotics/xiaomi-dark-robotic-factory" }, { "metric": "Humanoids produced per hour off the fastest humanoid line", "unit": "units/h", "value": 1, "as_of": "2026-04-29", "direction": "up_is_progress", "canon_target": 100, "source": "https://www.figure.ai/news/ramping-figure-03-production" }, { "metric": "Global industrial robot density", "unit": "units per 10,000 employees", "value": 132, "as_of": "2026-04-08", "direction": "up_is_progress", "canon_target": 10000, "source": "https://ifr.org/ifr-press-releases/news/robot-density-surges-in-europe-asia-and-americas" } ] }, "commentary": "Robots have been building robots, unsupervised for thirty days at a stretch, since 2001, a fact whose most-cited source is a magazine article from 2003 that nobody has re-measured. The modern showcase, Xiaomi's dark factory, ships ten million phones a year at 81 per cent automation, with people retained for the maintenance. Figure's humanoid line went from one robot a day to one an hour in under four months, which is genuine and is also 12,000 a year. What remains stubbornly human is the non-repeatable work: changing tooling, fixing machines, handling exceptions. All of it is the manipulation problem, wearing overalls.", "last_reviewed": "2026-08-29", "review_cadence_days": 30, "progress": 0.0711 }, { "id": "humanoid-robot-mass-production", "name": "Humanoid robot mass production", "category": "manufacture", "weight": 8, "one_liner": "Unit economics of an army.", "canon": { "requirement": "Series production of a humanoid combat android at a rate sufficient to field it as LINE INFANTRY across a planet, with generational model turnover implying an iterated production programme and a concurrent prototype pipeline.", "quantified": [ { "metric": "model generations named or implied", "value": "600 series (rubber skin) -> Model 101 / 800 series (living tissue) -> T-1000 'advanced prototype'", "source_ref": "t1-reese-briefing", "tier": "PRIMARY" }, { "metric": "employment", "value": "massed infantry - 'Three terminator endoskeletons advance, firing rapidly'; 'the ranks of advancing machines'", "source_ref": "t2-script-5", "tier": "TERTIARY" }, { "metric": "production scale observed (secondary)", "value": "a factory floor of completed T-800s", "source_ref": "salvation-factory", "tier": "SECONDARY" }, { "metric": "scale required", "value": "1e5-1e6 units", "source_ref": "derived", "tier": "EXTRAPOLATED" }, { "metric": "concurrent R&D pipeline", "value": "the T-1000 is an 'advanced prototype' running alongside production", "source_ref": "t2-mimetic-polyalloy", "tier": "PRIMARY" } ], "sources": [ { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-reese-briefing", "evidence": "The generational progression - a product line with model years, not a bespoke build.", "quote": "The 600 series had rubber skin. We spotted them easy, but these are new.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-mimetic-polyalloy", "evidence": "Evidence of a prototype programme running alongside series production.", "quote": "Not like me. A T-1000, advanced prototype.", "verified": true, "tier": "PRIMARY" } ], "canon_confidence": "implied", "canonical_target": { "metric": "sustained production rate and fielded fleet size", "value": "Series production of a fully autonomous humanoid combat android at a sustained rate sufficient to field 1e5-1e6 units as line infantry across a planet, with generational model turnover and a concurrent prototype programme", "reasoning": "The 1e5-1e6 bracket is extrapolated and labelled. It is anchored on three things canon shows: humanoids used as MASSED INFANTRY rather than special assets; a global occupation including camp systems; and continuous production across roughly three decades of war. The '600 series -> 800 series -> T-1000 prototype' progression matters as much as the raw number - it establishes a PRODUCT LINE with model years, which is a different manufacturing problem entirely." }, "assumed_by_sota_agent": "Standardised, interchangeable T-800 Model 101 units produced at a rate of order one million per year, with per-unit cost irrelevant to the producer, and a fielded population large enough to prosecute a global war." }, "real": { "status": "in_progress", "trl": 8, "spec_fraction": 0.05, "spec_fraction_rationale": "Approximately 60,000 forecast 2026 units divided by a canon rate of order 1,000,000 per year gives 0.06 on volume alone, discounted to 0.05 because essentially none of those units can complete a single eight-hour shift on one charge, let alone match the canon article's endurance, armour or weapon capability.", "gap": "Shipments are real and growing fast - 19,100 units in the first half of 2026, up 272% - but the units are research, education and light-industrial machines with 2-4 hour endurance. Announced capacity already far exceeds demand: Tesla targeted 50,000-100,000 units in 2026 and had produced none on a line as of its second-quarter update, Figure's 12,000/year line had delivered over 350, and Agility's 10,000/year factory supports about 100 deployed robots.", "why_hard": "The binding constraint is demand, not manufacturing. The work a humanoid can do reliably is currently worth less than the robot costs to own, insure and service - a humanoid bills at roughly $25 an hour and cannot finish a shift on one charge. Volume waits on manipulation reliability, not on the factory.", "movers": [ { "name": "AgiBot (Zhiyuan Robotics)", "kind": "company", "country": "CN", "what": "World's largest humanoid vendor in the first half of 2026: 8,400 units, 44% share, up 562% year on year." }, { "name": "Unitree Robotics", "kind": "company", "country": "CN", "what": "Shipped over 5,500 humanoids in 2025 and 5,900 in the first half of 2026; G1 lists at $13,500, the cheapest volume humanoid." }, { "name": "Tesla", "kind": "company", "country": "US", "what": "Targets 50,000-100,000 Optimus units in 2026 and a one-million-a-year Fremont run-rate; had produced none on a production line as of 22 July 2026." }, { "name": "Figure AI", "kind": "company", "country": "US", "what": "BotQ at one robot per hour, over 350 Figure 03 delivered, billing BMW at roughly $25 per robot-operating-hour." }, { "name": "Agility Robotics", "kind": "company", "country": "US", "what": "Going public at a $2.5bn pre-money valuation with about 100 deployed Digits and over $300m of milestone-contingent Digit v5 orders." }, { "name": "UBTECH Robotics", "kind": "company", "country": "CN", "what": "Shipped 700 units in the first half of 2026 and targets a 10,000-unit annual industrial humanoid capacity." }, { "name": "MIIT and SASAC", "kind": "agency", "country": "CN", "what": "Joint directive of June 2026 mandating more than 10,000 humanoids into commercial use by end-2026 across state enterprises." } ], "evidence": [ { "date": "2026-08-10", "claim": "Global humanoid shipments reached 19,100 units in the first half of 2026, up 272% year on year, with AGIBOT at 8,400 and Unitree at 5,900; full-year 2026 is forecast at roughly 60,000 units and $1.6bn of revenue.", "source": "Smart Analytics Global", "title": "Global Humanoid Robot Shipments Surged 272% YoY to 19.1K Units in 1H 2026", "url": "https://smartanalyticsglobal.com/global-humanoid-robot-shipments-2026-agibot-unitree/", "kind": "product", "delta": "+" }, { "date": "2026-07-02", "claim": "Tesla has produced zero Optimus units on a production line against a stated 2026 target of 50,000-100,000, with Musk stating production 'will be extremely slow at first'.", "source": "Electrek", "title": "Elon Musk shuts down '4D chess' theory on Tesla Optimus production", "url": "https://electrek.co/2026/07/02/musk-shuts-down-optimus-4d-chess-theory/", "kind": "product", "delta": "-" }, { "date": "2026-04-29", "claim": "Figure has delivered more than 350 Figure 03 units against a first-generation line capacity of 12,000 per year.", "source": "Figure AI", "title": "Ramping Figure 03 Production", "url": "https://www.figure.ai/news/ramping-figure-03-production", "kind": "product", "delta": "+" }, { "date": "2026-06-24", "claim": "Agility has around 100 Digits with paying customers and over $300m of multi-year Digit v5 orders explicitly subject to contractual milestones, against a factory rated at 10,000 units a year.", "source": "U.S. Securities and Exchange Commission", "title": "Joint press release of Churchill Capital Corp XI and Agility Robotics, Inc. (EX-99.1)", "url": "https://www.sec.gov/Archives/edgar/data/0002074973/000121390026071287/ea029548401ex99-1.htm", "kind": "regulation", "delta": "+" }, { "date": "2026-06-10", "claim": "China's MIIT and SASAC jointly directed local governments and state-owned enterprises to put more than 10,000 humanoid robots into commercial use by the end of 2026.", "source": "Caixin Global", "title": "China Targets 10,000 Humanoid Robots in Commercial Use by End-2026", "url": "https://www.caixinglobal.com/2026-06-10/china-targets-10000-humanoid-robots-in-commercial-use-by-end-2026-102452656.html", "kind": "regulation", "delta": "+" }, { "date": "2026-06-09", "claim": "Chinese firms built roughly 85% of the world's humanoids in 2025 at an average price of about $46,000, and deployment lags production capacity because use cases remain limited.", "source": "Fortune", "title": "China builds cheap humanoids at scale but finding buyers might be the hardest part", "url": "https://fortune.com/2026/06/09/china-builds-85-percent-worlds-humanoids-robots-cheap/", "kind": "product", "delta": "0" }, { "date": "2026-01-12", "claim": "Unitree shipped more than 5,500 humanoid robots in 2025, ranking first globally.", "source": "PR Newswire", "title": "Unitree Ranks No.1 Globally in Humanoid Robot Shipments, Exceeding 5,500 Units in 2025", "url": "https://www.prnewswire.com/news-releases/unitree-ranks-no1-globally-in-humanoid-robot-shipments-exceeding-5-500-units-in-2025--302674729.html", "kind": "product", "delta": "+" } ], "indicators": [ { "metric": "Global humanoid robot shipments per half-year", "unit": "units", "value": 19100, "as_of": "2026-08-10", "direction": "up_is_progress", "canon_target": 500000, "source": "https://smartanalyticsglobal.com/global-humanoid-robot-shipments-2026-agibot-unitree/" }, { "metric": "Optimus units produced on a production line against a 50,000-100,000 unit 2026 target", "unit": "units", "value": 0, "as_of": "2026-07-22", "direction": "up_is_progress", "canon_target": 50000, "source": "https://electrek.co/2026/07/02/musk-shuts-down-optimus-4d-chess-theory/" }, { "metric": "Cheapest commercially listed humanoid robot", "unit": "USD", "value": 13500, "as_of": "2026-06-09", "direction": "down_is_progress", "canon_target": 1000, "source": "https://fortune.com/2026/06/09/china-builds-85-percent-worlds-humanoids-robots-cheap/" }, { "metric": "Humanoid billing rate in commercial deployment", "unit": "USD per robot-hour", "value": 25, "as_of": "2026-06-09", "direction": "down_is_progress", "canon_target": 1, "source": "https://fortune.com/2026/06/09/china-builds-85-percent-worlds-humanoids-robots-cheap/" } ] }, "commentary": "Nineteen thousand one hundred humanoids shipped in the first half of 2026, up 272 per cent, three-quarters of them from two Chinese firms. Against that, the announcements: Tesla's 2026 target of 50,000 to 100,000 units, a Fremont line meant to reach a million a year, and, as of the second-quarter update, zero robots produced. Agility's factory is rated for 10,000 a year and supports about a hundred machines with paying customers. The gating item is not the line. It is that a humanoid bills at twenty-five dollars an hour and cannot yet finish a shift on one charge.", "last_reviewed": "2026-08-29", "review_cadence_days": 30, "progress": 0.0444 }, { "id": "supply-chain-and-materials-availability", "name": "Supply chain and materials availability", "category": "manufacture", "weight": 6, "one_liner": "Where the metal comes from.", "canon": { "requirement": "Source, refine and form a high-performance structural alloy of unstated composition, PLUS nuclear fuel, PLUS industrial-scale human tissue culture, at army scale, entirely autonomously, on a planet whose industrial base and civil society have been destroyed.", "quantified": [ { "metric": "structural material", "value": "'hyperalloy' - composition NEVER given anywhere in T1 or T2", "source_ref": "t1-reese-briefing", "tier": "PRIMARY" }, { "metric": "nuclear fuel", "value": "required by the 120-year cell; refining and fabrication implied", "source_ref": "derived", "tier": "EXTRAPOLATED" }, { "metric": "biological feedstock", "value": "cultured human tissue 'grown for the cyborgs' - a whole tissue-culture industry", "source_ref": "t1-reese-briefing", "tier": "PRIMARY" }, { "metric": "human participation", "value": "none", "source_ref": "t1-reese-briefing", "tier": "PRIMARY" }, { "metric": "other named materials in franchise fiction", "value": "titanium alloy (T-600); coltan (TSCC-timeline endoskeletons); carbon-based (Rev-9, production material)", "source_ref": "tscc-heavy-metal", "tier": "SECONDARY" } ], "sources": [ { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-reese-briefing", "evidence": "The only material ever named for the Model 101 - and its composition is never given.", "quote": "Underneath, it's a hyperalloy combat chassis", "verified": true, "tier": "PRIMARY" }, { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-reese-briefing", "evidence": "The tissue-culture supply chain, usually forgotten.", "quote": "flesh, skin, hair, blood... grown for the cyborgs", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator: The Sarah Connor Chronicles", "year": 2008, "medium": "tv-series", "ref": "tscc-heavy-metal", "evidence": "Cameron on later endoskeleton alloys. Episode 'Heavy Metal'.", "quote": "Coltan alloys have a much higher melting point.", "verified": false, "tier": "SECONDARY", "note": "Wiki-sourced quote; NOT yet verified against the episode transcript. See open item 6." } ], "canon_confidence": "extrapolated", "canonical_target": { "metric": "autonomous end-to-end materials supply across three distinct feedstocks", "value": "Fully autonomous extraction, refining and forming of a high-performance structural alloy; nuclear fuel refining and fabrication; and industrial-scale human tissue culture - all at army scale, with zero human participation, on a planet with no surviving industrial base", "reasoning": "The three feedstocks are separately required by other components and none is a commodity: the alloy, the nuclear fuel, and the cultured tissue. The tissue line is the one usually forgotten and is arguably the hardest - 'grown for the cyborgs' means Skynet operates a tissue-engineering industry at production volume, a supply chain nobody pictures when they picture a Terminator factory." }, "assumed_by_sota_agent": "Skynet controls the entire materials chain - extraction, refining, alloying of the hyperalloy, and fabrication - with no external dependency and no political constraint on any input." }, "real": { "status": "in_progress", "trl": 8, "spec_fraction": 0.3, "spec_fraction_rationale": "Against canon's six links - extraction, separation, magnets, precision reducers, cells, compute - the best-placed producer in the world controls roughly 4.5, about 0.75, but canon requires no external dependency at all and even that producer imports its highest-value compute. Every non-Chinese producer sits far lower: the most-localised Western humanoid is 75% domestic by part count, a measure that excludes the magnets. Taking the best system in the world and discounting for the compute gap and for national rather than corporate control gives 0.30.", "gap": "No producer controls the whole chain. China mined 69% of world rare-earth output in 2025 and is the sole source of separated heavy rare earths for the United States; its April 2025 export controls on samarium, gadolinium, terbium, dysprosium, lutetium, scandium and yttrium remain in force, with the wider October 2025 regime merely suspended until around November 2026. Western producers depend on those magnets; Chinese producers depend on imported advanced logic.", "why_hard": "The binding constraint is political, not physical. Rare earths are not rare, but their separation is chemically nasty, capital-intensive and environmentally regulated, so over thirty years it concentrated in one jurisdiction; rebuilding it elsewhere is a decade of permits and plant, not a technology programme. The chokepoint specific to robots is heavy rare earths for high-temperature NdFeB magnets, which is exactly where the export controls point.", "movers": [ { "name": "China Ministry of Commerce", "kind": "agency", "country": "CN", "what": "Author of the April 2025 and October 2025 rare-earth export control regimes that put dysprosium and terbium under licence." }, { "name": "U.S. Geological Survey", "kind": "agency", "country": "US", "what": "Publishes the Mineral Commodity Summaries each February - the authoritative recurring dataset on production, prices and import reliance." }, { "name": "MP Materials (Mountain Pass)", "kind": "company", "country": "US", "what": "Recipient of a $150m U.S. Department of War direct loan in August 2025 to build heavy-rare-earth separation capacity." }, { "name": "Harmonic Drive Systems / Nabtesco", "kind": "company", "country": "JP", "what": "Still lead precision strain-wave and cycloidal reducers, the second chokepoint after magnets, at 20-40 units per humanoid." }, { "name": "Agility Robotics", "kind": "company", "country": "US", "what": "The only humanoid maker to file a localisation figure: approximately 75% of Digit's parts sourced within the United States." }, { "name": "International Federation of Robotics", "kind": "program", "country": "DE", "what": "Tracks Chinese domestic robot-supplier share, which rose from 30% in 2020 to 57% in 2024." } ], "evidence": [ { "date": "2026-02-01", "claim": "China produced 270,000 of 390,000 tonnes of world rare-earth-oxide mine output in 2025 (69%); U.S. net import reliance for rare-earth compounds and metals was 67%, with 71% of 2021-24 imports from China.", "source": "U.S. Geological Survey", "title": "Mineral Commodity Summaries 2026: Rare Earths", "url": "https://pubs.usgs.gov/periodicals/mcs2026/mcs2026-rare-earths.pdf", "kind": "regulation", "delta": "0" }, { "date": "2026-02-01", "claim": "U.S. net import reliance for heavy rare-earth compounds and metals is 100%; terbium, holmium and lutetium came 100% from China; terbium oxide averaged $1,010/kg and dysprosium oxide $239/kg in 2025.", "source": "U.S. Geological Survey", "title": "Mineral Commodity Summaries 2026: Rare Earths (Heavy)", "url": "https://pubs.usgs.gov/periodicals/mcs2026/mcs2026-rare-earths-heavy.pdf", "kind": "regulation", "delta": "-" }, { "date": "2026-02-01", "claim": "China's April 2025 export controls on samarium, gadolinium, terbium, dysprosium, lutetium, scandium and yttrium remain in force; the October 2025 expansion to all heavy rare earths was suspended in November 2025 for one year.", "source": "U.S. Geological Survey", "title": "Mineral Commodity Summaries 2026: Rare Earths (Heavy)", "url": "https://pubs.usgs.gov/periodicals/mcs2026/mcs2026-rare-earths-heavy.pdf", "kind": "regulation", "delta": "-" }, { "date": "2026-06-24", "claim": "Agility Robotics sources approximately 75% of Digit's parts within the United States - the most-localised Western humanoid on the public record.", "source": "U.S. Securities and Exchange Commission", "title": "Joint press release of Churchill Capital Corp XI and Agility Robotics, Inc. (EX-99.1)", "url": "https://www.sec.gov/Archives/edgar/data/0002074973/000121390026071287/ea029548401ex99-1.htm", "kind": "regulation", "delta": "+" }, { "date": "2026-05-05", "claim": "Chinese domestic robot suppliers took 57% of their home market in 2024, up from 30% in 2020, and 85% in metal and machinery applications.", "source": "International Federation of Robotics", "title": "China Makes AI-powered Robots Core of National Strategy", "url": "https://ifr.org/ifr-press-releases/news/china-makes-ai-powered-robots-core-of-national-strategy", "kind": "benchmark", "delta": "+" } ], "indicators": [ { "metric": "China's share of world rare-earth mine production", "unit": "%", "value": 69, "as_of": "2026-02-01", "direction": "down_is_progress", "canon_target": 0, "source": "https://pubs.usgs.gov/periodicals/mcs2026/mcs2026-rare-earths.pdf" }, { "metric": "U.S. net import reliance, heavy rare-earth compounds and metals", "unit": "%", "value": 100, "as_of": "2026-02-01", "direction": "down_is_progress", "canon_target": 0, "source": "https://pubs.usgs.gov/periodicals/mcs2026/mcs2026-rare-earths-heavy.pdf" }, { "metric": "Terbium oxide average price", "unit": "USD/kg", "value": 1010, "as_of": "2026-02-01", "direction": "down_is_progress", "canon_target": 50, "source": "https://pubs.usgs.gov/periodicals/mcs2026/mcs2026-rare-earths-heavy.pdf" }, { "metric": "Domestic content of the most-localised Western humanoid", "unit": "% of parts", "value": 75, "as_of": "2026-06-24", "direction": "up_is_progress", "canon_target": 100, "source": "https://www.sec.gov/Archives/edgar/data/0002074973/000121390026071287/ea029548401ex99-1.htm" } ] }, "commentary": "The metal is the easy part, which is the uncomfortable finding. China mined 69 per cent of the world's rare earths last year, and American reliance on imported heavy rare earths, the dysprosium and terbium that stop an actuator magnet demagnetising when it gets hot, is exactly 100 per cent, with terbium sourced entirely from China at $1,010 a kilogram. Beijing's April 2025 export controls remain in force; the wider October regime is merely suspended, until November. The best-localised Western humanoid is 75 per cent domestic by part count, a metric in which the magnets conveniently hide.", "last_reviewed": "2026-08-29", "review_cadence_days": 30, "progress": 0.2667 }, { "id": "closed-timelike-curves", "name": "Closed timelike curves", "category": "time-displacement", "weight": 3, "one_liner": "Whether the past is reachable at all. General relativity says maybe; quantum field theory says no.", "canon": { "requirement": "Physically realisable closed timelike curves permitting MACROSCOPIC matter to reach its own past light cone - and, in the strong form the franchise requires, an ENGINEERED AND CONTROLLABLE one, aimed at a chosen coordinate in space and time. The films assert it works and never once gesture at how.", "quantified": [ { "metric": "mechanism", "value": "NEVER explained in any film", "source_ref": "none", "tier": "PRIMARY" }, { "metric": "directionality", "value": "past-only; one-way", "source_ref": "t1-interrogation", "tier": "PRIMARY" }, { "metric": "targeting precision", "value": "to a chosen city, night and street", "source_ref": "t1-arrival", "tier": "PRIMARY" }, { "metric": "mass transported", "value": "one adult human or one T-800, ~80-200 kg", "source_ref": "derived", "tier": "EXTRAPOLATED" }, { "metric": "consistency model", "value": "CONTRADICTORY across the franchise - see contradictions", "source_ref": "none", "tier": "PRIMARY" } ], "sources": [ { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-interrogation", "evidence": "Reese, on whether he can return. The one-way clause.", "quote": "I can't. Nobody goes home. Nobody else comes through. It's just him... and me.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day", "year": 1991, "medium": "film", "ref": "t2-ending", "evidence": "Sarah's closing narration, deliberately breaking T1's closed causal loop.", "quote": "The unknown future rolls toward us. I face it for the first time with a sense of hope.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 3: Rise of the Machines", "year": 2003, "medium": "film", "ref": "t3-inevitable", "evidence": "T3 reverses T2.", "quote": "Judgment Day is inevitable.", "verified": true, "tier": "SECONDARY" } ], "canon_confidence": "implied", "canonical_target": { "metric": "engineered, controllable CTC for macroscopic matter", "value": "An engineered, controllable closed timelike curve capable of transporting ~1e2 kg of matter to a selected past coordinate in space and time, one-way, repeatably", "reasoning": "Stated as an existence-and-control requirement rather than a mechanism, because canon supplies no mechanism to score against. The 'one-way' clause is genuine canon and is worth keeping: T1 is explicit that the equipment sends but does not retrieve, which is a WEAKER requirement than a general CTC and therefore a fairer denominator." }, "assumed_by_sota_agent": "A reachable past. A physically realisable closed timelike curve through which a macroscopic object can be sent from 2029 to 1984 and arrive intact." }, "real": { "status": "theoretical", "trl": 2, "spec_fraction": 0.01, "spec_fraction_rationale": "The canon requirement is one usable closed timelike curve. Humanity has produced zero. The 0.01 is a deliberate small non-zero allowance for two facts: exact solutions of the Einstein field equations containing CTCs exist and are rigorous (Godel 1949, Tipler 1974, Gott 1991, Morris-Thorne-Yurtsever 1988), and laboratory analogues reproduce the predicted quantum statistics of a CTC (Lloyd et al. 2011, Ringbauer et al. 2014). No event has ever been moved outside its light cone. 0.01 is credit for the paperwork.", "gap": "Every known CTC-permitting solution requires either exotic matter violating the null and weak energy conditions, or infinite structures (Tipler's infinitely long rotating cylinder, Gott's infinite cosmic strings), or global cosmological conditions our universe does not have (Godel's rotating dust with negative cosmological constant). The Kerr interior contains CTCs but sits behind an event horizon and inside a Cauchy horizon that is generically unstable.", "why_hard": "Quantum energy inequalities. Ford and Roman showed that quantum field theory's bounds on negative energy density force a static traversable wormhole either to be barely larger than the Planck length, or to concentrate its exotic matter in a band many orders of magnitude thinner than the throat — which they characterise as making macroscopic traversable wormholes very improbable. Hawking's chronology protection conjecture, that the renormalised stress-energy tensor diverges as a chronology horizon forms and destroys the incipient time machine, is supported but unproven and its general form requires a theory of quantum gravity that does not exist.", "movers": [ { "name": "Matt Visser, Victoria University of Wellington", "kind": "university", "country": "NZ", "what": "The standing survey of chronology protection, and the 2021 refutation showing generic warp drives violate the null energy condition." }, { "name": "L. H. Ford and Thomas A. Roman", "kind": "university", "country": "US", "what": "Derived the quantum inequalities that constrain negative energy density and applied them to traversable wormhole geometries — the binding theoretical result in this component." }, { "name": "Scott Aaronson and John Watrous", "kind": "university", "country": "US", "what": "Proved that computers with access to Deutsch-model CTCs solve exactly PSPACE, and that classical and quantum CTC computers are equally powerful." }, { "name": "Andrew White group, University of Queensland", "kind": "university", "country": "AU", "what": "Photonic simulation of closed timelike curves, demonstrating the predicted nonlinear effects using ordinary forward-in-time optics." }, { "name": "Applied Physics (Bobrick, Martire, Fuchs, Helmerich et al.)", "kind": "lab", "country": "US", "what": "Physical warp drive programme: subluminal positive-energy warp solutions satisfying all energy conditions, and the Warp Factory numerical toolkit. Subluminal, and therefore not chronology-violating." } ], "evidence": [ { "date": "1949-07-01", "claim": "Godel published the first exact cosmological solution of the Einstein field equations containing closed timelike curves — a rotating, pressureless-dust universe with a negative cosmological constant, establishing that general relativity does not forbid time travel.", "source": "Reviews of Modern Physics", "title": "An Example of a New Type of Cosmological Solutions of Einstein's Field Equations of Gravitation", "url": "https://doi.org/10.1103/RevModPhys.21.447", "kind": "paper", "delta": "+" }, { "date": "1988-09-26", "claim": "Morris, Thorne and Yurtsever showed that a traversable wormhole held open by exotic matter can be converted into a time machine by moving one mouth relativistically, and that this necessarily requires violation of the weak energy condition.", "source": "Physical Review Letters", "title": "Wormholes, Time Machines, and the Weak Energy Condition", "url": "https://doi.org/10.1103/PhysRevLett.61.1446", "kind": "paper", "delta": "+" }, { "date": "1992-07-15", "claim": "Hawking proposed the chronology protection conjecture — that the laws of physics prevent the appearance of closed timelike curves, via divergence of the renormalised stress-energy tensor as the chronology horizon forms. It remains unproven in general and requires a quantum theory of gravity to settle.", "source": "Physical Review D", "title": "Chronology protection conjecture", "url": "https://doi.org/10.1103/PhysRevD.46.603", "kind": "paper", "delta": "-" }, { "date": "1996-05-15", "claim": "Ford and Roman applied quantum-inequality bounds on negative energy density to static traversable wormholes and concluded that either the wormhole must be only a little larger than Planck size, or the negative energy must be concentrated in a band many orders of magnitude smaller than the throat — making macroscopic traversable wormholes very improbable.", "source": "Physical Review D (arXiv:gr-qc/9510071)", "title": "Quantum Field Theory Constrains Traversable Wormhole Geometries", "url": "https://arxiv.org/abs/gr-qc/9510071", "kind": "paper", "delta": "-" }, { "date": "2009-02-08", "claim": "Aaronson and Watrous proved that computers with access to Deutsch-model closed timelike curves solve exactly the complexity class PSPACE, and that classical and quantum CTC computers have identical power — meaning a CTC would collapse the polynomial hierarchy and eliminate the quantum computing advantage entirely.", "source": "Proceedings of the Royal Society A (arXiv:0808.2669)", "title": "Closed Timelike Curves Make Quantum and Classical Computing Equivalent", "url": "https://arxiv.org/abs/0808.2669", "kind": "paper", "delta": "0" }, { "date": "2014-06-19", "claim": "A photonic experiment simulated the nonlinear behaviour of a qubit interacting unitarily with an older version of itself, reproducing predicted CTC effects including perfect discrimination of non-orthogonal quantum states. No chronology violation occurred: the experiment used ordinary forward-in-time optics.", "source": "Nature Communications", "title": "Experimental simulation of closed timelike curves", "url": "https://doi.org/10.1038/ncomms5145", "kind": "demo", "delta": "0" }, { "date": "2022-03-23", "claim": "Santiago, Schuster and Visser refuted three contemporaneous claims of positive-energy warp drives, showing the claims established only that co-moving Eulerian observers see positive energy density whereas the weak energy condition requires all timelike observers to. All physically reasonable warp drives violate the null energy condition, and in modified gravity the violation moves to the geometric null convergence condition rather than disappearing.", "source": "Physical Review D (arXiv:2105.03079)", "title": "Generic warp drives violate the null energy condition", "url": "https://arxiv.org/abs/2105.03079", "kind": "paper", "delta": "-" }, { "date": "2026-07-01", "claim": "An unrefereed single-author preprint constructs an exact vacuum spacetime that develops closed timelike curves from regular, asymptotically flat initial data respecting the weak, dominant and strong energy conditions, with the chronology horizon reaching a degenerate configuration whose generating closed null geodesic has vanishing boost and optical scalars. If it survives peer review this would be a CTC-forming spacetime requiring no exotic matter, which is the most robust standing objection in the field. No journal reference and no independent verification as of 2026-08-29.", "source": "arXiv", "title": "Closed Timelike Curves from a Vacuum Traveling Wave", "url": "https://arxiv.org/abs/2607.00788", "kind": "paper", "delta": "+" }, { "date": "2026-07-29", "claim": "Deutsch-CTC consistency conditions were implemented as post-selected decoder circuits on IBM quantum hardware, quantifying decoder fidelity, post-selection overhead and routing-dependent noise. As with all such work, this simulates D-CTC mathematics on ordinary hardware and involves no chronology violation.", "source": "arXiv", "title": "Closed Timelike Curve Decoding on Quantum Hardware", "url": "https://arxiv.org/abs/2607.27473", "kind": "demo", "delta": "0" } ], "indicators": [ { "metric": "Physically realised closed timelike curves", "unit": "count", "value": 0, "as_of": "2026-08-29", "direction": "up_is_progress", "canon_target": 1, "source": "https://arxiv.org/abs/gr-qc/0204022" }, { "metric": "Exact GR solutions containing CTCs that require neither exotic matter nor infinite structures, peer-reviewed", "unit": "solutions", "value": 0, "as_of": "2026-08-29", "direction": "up_is_progress", "canon_target": 1, "source": "https://arxiv.org/abs/2607.00788" }, { "metric": "Status of the chronology protection conjecture", "unit": "proven(1)/unproven(0)", "value": 0, "as_of": "2026-08-29", "direction": "down_is_progress", "canon_target": 0, "source": "https://arxiv.org/abs/gr-qc/0204022" } ] }, "commentary": "General relativity has permitted closed timelike curves since 1949, which is the least useful fact in physics. Every solution demands exotic matter, an infinitely long rotating cylinder, or a universe that rotates and ours does not. Quantum field theory then bounds the negative energy required so tightly that Ford and Roman confined macroscopic wormholes to Planck scale or to an exotic band many orders of magnitude thinner than the throat. Hawking's chronology protection remains a conjecture rather than a theorem, so the honest score is theoretical and not blocked. The best result in the field is Aaronson and Watrous: a working time machine would make quantum computers pointless.", "last_reviewed": "2026-08-29", "review_cadence_days": 90, "progress": 0.0022 }, { "id": "temporal-displacement-equipment", "name": "Temporal displacement equipment", "category": "time-displacement", "weight": 3, "one_liner": "The machine, and its energy budget. The shortfall runs to eighty-seven orders of magnitude.", "canon": { "requirement": "The machine itself: a device occupying a room in a Skynet laboratory complex, generating a spherical displacement field around a single living passenger and sending them one-way to a chosen past coordinate, with violent local electromagnetic and mechanical side effects at both ends. CANON GIVES NO ENERGY FIGURE.", "quantified": [ { "metric": "name in dialogue", "value": "'the time displacement equipment'", "source_ref": "t1-interrogation", "tier": "PRIMARY" }, { "metric": "location", "value": "a Skynet 'lab complex' captured by the Resistance, then destroyed", "source_ref": "t1-interrogation", "tier": "PRIMARY" }, { "metric": "field geometry", "value": "a sphere - 'a FIGURE in a SPHERE OF ENERGY'", "source_ref": "t2-script-7", "tier": "TERTIARY" }, { "metric": "field diameter", "value": "~2-3 m (encloses a crouching adult and a bite of the ground)", "source_ref": "derived", "tier": "EXTRAPOLATED" }, { "metric": "arrival effects", "value": "thunderclap concussion blowing out every window facing the yard; arcing discharges to nearby metal and plumbing; a hemispherical section cut cleanly from the pavement", "source_ref": "t1-script-1", "tier": "TERTIARY" }, { "metric": "ENERGY BUDGET", "value": "NEVER STATED IN ANY WORK", "source_ref": "none", "tier": "PRIMARY" }, { "metric": "capacity", "value": "one passenger, one-way, one shot", "source_ref": "t1-interrogation", "tier": "PRIMARY" }, { "metric": "machine architecture (cut T2 sequence)", "value": "three enormous chrome rings, one inside the other, counter-rotating on different axes like a gyroscope, suspended in a magnetic field over a circular floor opening", "source_ref": "t2-cut-tde", "tier": "TERTIARY" }, { "metric": "chamber size (cut T2 sequence)", "value": "'the size of a high-school gym'", "source_ref": "t2-cut-tde", "tier": "TERTIARY" }, { "metric": "field-conformance mechanism (cut T2 sequence)", "value": "the traveller is coated in a conductive substance 'so the time-field will follow his outline'", "source_ref": "t2-cut-tde", "tier": "TERTIARY" } ], "sources": [ { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-interrogation", "evidence": "The only naming of the equipment in dialogue.", "quote": "What is it called? The time displacement equipment? / That's right. The Terminator had already gone through.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2: Judgment Day - revised final shooting script", "year": 1991, "medium": "shooting-script", "ref": "t2-script-7", "evidence": "Arrival stage direction - the sphere.", "quote": "The strange lightning forms a circular opening in mid-air, and in the sudden flare of light we see a FIGURE in a SPHERE OF ENERGY. Then the FRAME WHITES OUT with an explosive THUNDERCLAP!", "verified": true, "tier": "TERTIARY" }, { "work": "The Terminator - fourth draft, 20 April 1983", "year": 1983, "medium": "screenplay-draft", "ref": "t1-script-1", "evidence": "Arrival stage direction - the concussion and the arcing.", "quote": "A CONCUSSION like a thunderclap right overhead blows in all the windows facing the yard.", "verified": true, "tier": "TERTIARY" }, { "work": "Terminator 2 - cut time-displacement sequence, 5/10/90 draft", "year": 1990, "medium": "screenplay-draft", "ref": "t2-cut-tde", "evidence": "Cut sequence depicting the displacement complex - the only place the machine is physically described.", "quote": "the time-field will follow his outline", "verified": true, "tier": "TERTIARY", "note": "Cut from the film. The most technically informative statement about time displacement anywhere in the production material." } ], "canon_confidence": "explicit", "canonical_target": { "metric": "displacement mass, field radius, and an OBSERVATIONALLY-DERIVED energy floor", "value": "Displace ~1e2 kg of living matter, enclosed in a field of ~2 m radius, to a selected past coordinate, one-way; observational energy FLOOR ~1e8 J (~100 MJ) delivered in under one second (~1e2 GW peak), plus the associated EMP", "reasoning": "WARNING: canon gives NO energy figure and anyone who says it does is quoting a fan wiki. The number here is derived from the observable and the derivation is stated so it can be argued with. (1) The arrival cleanly excises roughly 2-4 m3 of reinforced concrete roadway - the hemispherical bite ~2 m across, visible in T2. At 2,400 kg/m3 that is ~5-10 tonnes. (2) Comminution energy for concrete is roughly 3.6-36 MJ/tonne (1-10 kWh/tonne), giving ~35-350 MJ for the excavation alone, before ejection, vaporisation or EM effects. (3) The event completes in well under a second and blows out lighting across the surrounding block, so peak power is of order 1e11 W. So ~1e8 J and ~1e2 GW peak as an OBSERVATIONAL FLOOR, flagged extrapolated. It is a floor, not an estimate - the actual figure is unknowable because canon never engages with it." }, "assumed_by_sota_agent": "A working machine that generates a field displacing a ~200 kg macroscopic object approximately 45 years into the past, delivering it intact and at rest in Earth's frame." }, "real": { "status": "blocked", "trl": 1, "spec_fraction": 0.005, "spec_fraction_rationale": "The requirement is between 1e27 and 6e62 kg of negative mass-energy held in a specified geometry. Humanity has assembled roughly 5e-25 kg of negative mass-equivalent — my calculation for a generous 1 cm2 Casimir cavity at 10 nm separation, energy density -4.3e4 J/m3, total -4.3e-8 J. The literal ratio spans 1e-51 (one-metre wormhole throat) to 1e-87 (Pfenning-Ford quantum-inequality-constrained 100 m warp bubble). The 0.005 is not that ratio; it is credit for the facts that negative energy density is real and measured, and that the required geometry has been specified precisely enough to be costed. Rounding to zero would also be defensible.", "gap": "There is no device concept. Two routes exist in the literature and both fail on the same number. Build and stabilise a traversable wormhole, then move one mouth relativistically — needing order 1.35e27 kg of negative mass for a one-metre throat, held in a configuration the quantum inequalities forbid. Or engineer a global spacetime with pre-existing CTCs — needing infinite cosmic strings or an infinitely long rotating cylinder. No third proposal exists, which is more telling than the failure of the first two.", "why_hard": "Negative energy density is real but bounded. Pfenning and Ford calculated that a 100 m Alcubierre bubble with quantum-inequality-respecting wall thickness needs at most -6.2e65 grams of negative energy, which they restate as -3e20 galaxy masses and describe verbatim as roughly ten orders of magnitude greater than the total mass of the visible universe. Relaxing the quantum inequality entirely and allowing a one-metre wall still leaves a quarter of a solar mass. Even a bubble the size of one electron Compton wavelength needs about -400 solar masses. These are not engineering targets.", "movers": [ { "name": "Michael Pfenning and L. H. Ford", "kind": "university", "country": "US", "what": "Costed the Alcubierre warp drive against quantum inequalities and produced the field's definitive impossibility number." }, { "name": "Chris Van Den Broeck", "kind": "university", "country": "BE", "what": "Reduced the warp drive's negative mass requirement to a few solar masses by decoupling the bubble's neck from its interior volume — a 63-order-of-magnitude improvement that is still hopeless." }, { "name": "Applied Physics (Bobrick, Martire, Fuchs, Helmerich, Sellers, Melcher)", "kind": "lab", "country": "US", "what": "Constant-velocity subluminal warp solutions satisfying all energy conditions, plus optimisations cutting Alcubierre's negative energy requirement by two orders of magnitude. Subluminal, so not a time machine." }, { "name": "Steve K. Lamoreaux", "kind": "university", "country": "US", "what": "First clean measurement of the Casimir force, over 0.6 to 6 micrometres — the experimental anchor for negative energy density being real at all." }, { "name": "Chalmers University (Wilson, Johansson et al.)", "kind": "university", "country": "SE", "what": "Observed the dynamical Casimir effect, generating real photons from vacuum fluctuations in a superconducting circuit with a rapidly modulated boundary." } ], "evidence": [ { "date": "1997-07-01", "claim": "Pfenning and Ford applied quantum inequalities to the Alcubierre metric, finding the bubble wall must be only a few hundred Planck lengths thick and that a 100 m radius bubble requires total negative energy of at most -6.2e65 grams, equivalent to -3e20 galaxy masses, which they describe as roughly ten orders of magnitude greater than the total mass of the entire visible universe. Relaxing the quantum inequality to allow a one-metre wall still requires about a quarter of a solar mass; an electron-Compton-wavelength bubble requires about -400 solar masses.", "source": "Classical and Quantum Gravity (arXiv:gr-qc/9702026)", "title": "The unphysical nature of \"Warp Drive\"", "url": "https://arxiv.org/abs/gr-qc/9702026", "kind": "paper", "delta": "-" }, { "date": "1999-12-01", "claim": "Van Den Broeck showed a geometric modification of the Alcubierre metric — a small bubble neck enclosing a large internal volume — reduces the required negative mass to the order of a few solar masses with a comparable amount of positive energy, satisfying the quantum inequality on weak-energy-condition violations. This is an improvement of roughly 63 orders of magnitude and remains unattainable.", "source": "Classical and Quantum Gravity (arXiv:gr-qc/9905084)", "title": "A 'warp drive' with more reasonable total energy requirements", "url": "https://arxiv.org/abs/gr-qc/9905084", "kind": "paper", "delta": "+" }, { "date": "1997-01-06", "claim": "Lamoreaux measured the Casimir force between conducting surfaces over the 0.6 to 6 micrometre range, establishing experimentally that the quantum vacuum can be manipulated to produce regions of negative energy density.", "source": "Physical Review Letters", "title": "Demonstration of the Casimir Force in the 0.6 to 6 micrometre Range", "url": "https://doi.org/10.1103/PhysRevLett.78.5", "kind": "paper", "delta": "+" }, { "date": "2011-11-17", "claim": "The dynamical Casimir effect was observed in a superconducting circuit with a rapidly modulated boundary condition, generating real photons from vacuum fluctuations — the strongest demonstration that vacuum energy can be engineered, at magnitudes of order 1e-8 J in a laboratory cavity.", "source": "Nature", "title": "Observation of the dynamical Casimir effect in a superconducting circuit", "url": "https://doi.org/10.1038/nature10561", "kind": "paper", "delta": "+" }, { "date": "2024-04-29", "claim": "A constant-velocity subluminal warp drive solution satisfying all energy conditions was constructed by combining a stable positive-ADM-mass matter shell with an Alcubierre-like shift vector distribution, generated numerically and verified not to reduce to a coordinate transformation. Being subluminal, it produces no closed timelike curves and is not a time machine.", "source": "Classical and Quantum Gravity (arXiv:2405.02709)", "title": "Constant Velocity Physical Warp Drive Solution", "url": "https://arxiv.org/abs/2405.02709", "kind": "paper", "delta": "+" }, { "date": "1996-05-15", "claim": "Quantum energy inequalities place an uncertainty-principle-type constraint on the magnitude and duration of negative energy density seen by a timelike geodesic observer, and applying them to traversable wormholes makes macroscopic examples very improbable. These are derived theorems in quantum field theory, not conjectures.", "source": "Physical Review D (arXiv:gr-qc/9510071)", "title": "Quantum Field Theory Constrains Traversable Wormhole Geometries", "url": "https://arxiv.org/abs/gr-qc/9510071", "kind": "paper", "delta": "-" } ], "indicators": [ { "metric": "Largest negative mass-energy assembled in a laboratory (Casimir cavity, 1 cm2 at 10 nm)", "unit": "kg", "value": -4.8e-25, "as_of": "2026-08-29", "direction": "down_is_progress", "canon_target": -1.35e+27, "source": "https://doi.org/10.1103/PhysRevLett.78.5" }, { "metric": "Negative mass-energy required for a one-metre traversable wormhole throat", "unit": "kg", "value": -1.35e+27, "as_of": "1988-05-01", "direction": "down_is_progress", "canon_target": -4.8e-25, "source": "https://arxiv.org/abs/gr-qc/9510071" }, { "metric": "Orders of magnitude between laboratory negative energy and the least demanding time-machine requirement", "unit": "log10 ratio", "value": 51, "as_of": "2026-08-29", "direction": "down_is_progress", "canon_target": 0, "source": "https://arxiv.org/abs/gr-qc/9702026" } ] }, "commentary": "The arithmetic is worth stating once, plainly. Pfenning and Ford costed a hundred-metre warp bubble with a quantum-inequality-respecting wall at minus 6.2 times ten to the sixty-fifth grams of negative energy — roughly ten orders of magnitude more than the mass of the visible universe. The best negative energy humanity has ever assembled is a Casimir cavity holding about minus four times ten to the minus eighth joules. The shortfall is fifty-one orders of magnitude against the most forgiving requirement and eighty-seven against the honest one. For scale, the observable universe spans about sixty-two orders of magnitude from end to end in Planck lengths.", "last_reviewed": "2026-08-29", "review_cadence_days": 90, "progress": 0.0006 }, { "id": "organic-only-constraint", "name": "Organic-only constraint", "category": "time-displacement", "weight": 2, "one_liner": "\"Nothing dead will go.\" No known physics discriminates by whether a thing is alive.", "canon": { "requirement": "The displacement field transports ONLY living organic matter, and whatever is entirely enclosed by it. Nothing dead, nothing inorganic, nothing worn. This single rule is why the T-800 has skin at all - the infiltration disguise and the transit requirement are the SAME engineering constraint. The T-1000 violates it and no film ever addresses that.", "quantified": [ { "metric": "rule as stated", "value": "'Nothing dead will go.'", "source_ref": "t1-time-displacement", "tier": "PRIMARY" }, { "metric": "field source", "value": "'the field generated by a living organism'", "source_ref": "t1-time-displacement", "tier": "PRIMARY" }, { "metric": "topological exception", "value": "matter FULLY ENCLOSED by living tissue transits - 'Surrounded by living tissue!'", "source_ref": "t1-time-displacement", "tier": "PRIMARY" }, { "metric": "consequence", "value": "travellers arrive naked; no weapons, no clothing, no equipment", "source_ref": "t1-arrival", "tier": "PRIMARY" }, { "metric": "in-universe explanation", "value": "NONE. 'I didn't build the fucking thing!'", "source_ref": "t1-time-displacement", "tier": "PRIMARY" }, { "metric": "viability on arrival", "value": "100% - both Reese and the T-800 arrive alive and immediately functional", "source_ref": "t1-arrival", "tier": "PRIMARY" }, { "metric": "soldier's version (novelization)", "value": "'metals won't displace'", "source_ref": "t1-novelization", "tier": "TERTIARY" }, { "metric": "best mechanical explanation anywhere (cut T2 sequence)", "value": "travellers are coated in a conductive substance 'so the time-field will follow his outline' - recasting the rule as FIELD CONFORMANCE, not biology", "source_ref": "t2-cut-tde", "tier": "TERTIARY" }, { "metric": "VIOLATED BY", "value": "the T-1000, which contains no living tissue at all and transits anyway; also the T-X, the Rev-9, and the Genisys T-1000", "source_ref": "t2-full", "tier": "CONTRADICTION" } ], "sources": [ { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-time-displacement", "evidence": "The rule, stated three ways in one exchange, including the topological exception.", "quote": "You go naked. Something about the field generated by a living organism. Nothing dead will go. / Why? / I didn't build the fucking thing! / OK. But this cyborg, if it's metal... / Surrounded by living tissue!", "verified": true, "tier": "PRIMARY" }, { "work": "The Terminator", "year": 1984, "medium": "film", "ref": "t1-interrogation", "evidence": "Silberman's setup for the rule.", "quote": "Why didn't you bring any weapons... something more advanced? Don't you have ray guns? / Ray guns. Show me a piece of future technology.", "verified": true, "tier": "PRIMARY" }, { "work": "Terminator 2 - cut time-displacement sequence, 5/10/90 draft", "year": 1990, "medium": "screenplay-draft", "ref": "t2-cut-tde", "evidence": "Cut T2 sequence. The only mechanism the material ever offers.", "quote": "so the time-field will follow his outline", "verified": true, "tier": "TERTIARY", "note": "Cut from the film. If the field follows a CONDUCTIVE OUTLINE, then 'nothing dead will go' is a statement about what can hold a field boundary, and living tissue is simply the conductive envelope Skynet had to hand." } ], "canon_confidence": "explicit", "canonical_target": { "metric": "transit mass, viability, and the organic selection rule", "value": "Transit of >=70 kg of living human tissue with 100% viability on arrival and immediate full function, under a transport rule that admits only living organic matter and matter TOPOLOGICALLY ENCLOSED by it, and rejects all unenclosed inorganic matter", "reasoning": "The rule is topological, not chemical, and the film says so in three words: 'Surrounded by living tissue!' The endoskeleton is inorganic and it transits, because the living envelope defines the boundary of what goes. That is a precise and testable statement of the constraint, and it is why this component sits in time-displacement rather than biology-and-camouflage - it is a rule about the field, not about the skin. Its violation by every later liquid-metal unit is unexplained; the cut T2 conductive-coating line is the only available reconciliation and is TERTIARY." }, "assumed_by_sota_agent": "A transport process that discriminates by biological status — carrying living tissue and anything wholly enclosed by it, and refusing everything else." }, "real": { "status": "blocked", "trl": 1, "spec_fraction": 0, "spec_fraction_rationale": "There is no partial credit available. No physical theory proposes a coupling that discriminates by biological status, none has been demonstrated, and the property the rule discriminates on is not a property of a physical state at all. Canon requirement: a transport process selective for living matter. Real world: zero such processes, zero proposals, zero mechanisms. 0/1 = 0.00.", "gap": "The entire mechanism. Nothing in general relativity or the Standard Model couples to whether a thing is alive.", "why_hard": "It is a category error rather than an engineering shortfall. Physics discriminates objects by conserved quantities and couplings — mass-energy, momentum, angular momentum, electric charge, colour charge, weak isospin, baryon and lepton number, spin. General relativity couples to the stress-energy tensor and to nothing else. 'Alive' is not on that list and cannot be added, because it is not a property of a configuration at an instant: life is a dynamical property, a system held far from thermodynamic equilibrium by dissipating free energy. A living cell and the same cell one second after death have, to any achievable precision, identical mass-energy, charge distribution, chemical composition and molecular structure. The difference is in the trajectory, not the state, so there is nothing locally different for a field to couple to. The three charitable readings all fail: bioelectric fields are ordinary electromagnetism, persist after death and are trivially reproduced by a battery; gating on metabolic activity makes the rule a safety interlock Skynet could remove rather than a law; and biological quantum coherence is confined to femtosecond-scale molecular processes with no coupling to gravitation.", "movers": [], "evidence": [ { "date": "1996-05-15", "claim": "The constraints that quantum field theory places on any spacetime-manipulating process are expressed entirely in terms of the stress-energy tensor. No formulation of these bounds contains any term sensitive to the biological status of the matter involved.", "source": "Physical Review D (arXiv:gr-qc/9510071)", "title": "Quantum Field Theory Constrains Traversable Wormhole Geometries", "url": "https://arxiv.org/abs/gr-qc/9510071", "kind": "paper", "delta": "0" }, { "date": "2002-04-05", "claim": "The standing survey of chronology protection catalogues the peculiarities that arise in spacetimes containing closed causal curves. None of the pathologies, constraints or proposed protection mechanisms discriminates between living and non-living matter; all are expressed in terms of stress-energy and causal structure.", "source": "arXiv", "title": "The quantum physics of chronology protection", "url": "https://arxiv.org/abs/gr-qc/0204022", "kind": "paper", "delta": "0" } ], "indicators": [ { "metric": "Known physical couplings that discriminate by biological status", "unit": "count", "value": 0, "as_of": "2026-08-29", "direction": "up_is_progress", "canon_target": 1, "source": "https://arxiv.org/abs/gr-qc/0204022" } ] }, "commentary": "The franchise's load-bearing rule is also its least defensible. Physics sorts matter by conserved quantities: mass-energy, charge, spin, baryon number. Being alive is not among them, and cannot be, because it is a property of a trajectory rather than a state. A cell and the same cell a second after death are, to any achievable precision, identical objects. Cameron appears to have known: Reese answers the question with \"I didn't build the fucking thing.\" The T-1000 then walks through the same field as solid mimetic polyalloy, converting a law of nature into a detector that can be spoofed. Scored at zero, and the score is the point.", "last_reviewed": "2026-08-29", "review_cadence_days": 180, "progress": 0 } ] }