{ "updated": "2026-10-26", "description": "Robotics R&D: foundation models, world models, benchmarks, research areas, datasets, compute and the humanoid component stack. Curated from a 2026-10-26 research sweep; every record cites its source.", "method": { "vendor_reported": "Vendor-reported means the number comes from the company or lab that built the model (or that runs the benchmark and also competes on it). Nobody else has reproduced it. It is shown because it is the only figure available, not because it has been checked.", "snippet_only": "Snippet-only means the claim was read from a search-result snippet or a secondary report because the primary page could not be opened. Treat it as a lead with a source attached, not a confirmed fact.", "as_of": "Leaderboards move. Each benchmark carries the date its numbers were read; a paper table and a live board are shown separately even when they share a name." }, "sections": { "models": [ { "id": "pi0-5", "name": "pi0.5", "developer": "Physical Intelligence", "country": "US", "country_label": "US", "released": "2025-04-22", "class": "VLA", "params": null, "weights": "open", "embodiments": "multiple robots incl. mobile manipulators; evaluated in entirely new homes", "training_data": "data from multiple robots, high-level semantic prediction, web data, and other sources (co-training on heterogeneous tasks)", "summary": "First end-to-end learned system shown doing long-horizon kitchen/bedroom cleaning in entirely new homes via heterogeneous co-training.", "results": [], "caveat": null, "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2504.16054", "captured_at": "2026-10-26", "quote": "cleaning a kitchen or bedroom, in entirely new homes" } ] }, { "id": "gemini-4-argon", "name": "Gemini 4 Argon (robotics use)", "developer": "Google DeepMind / Google", "country": "US", "country_label": "US", "released": "2026-10-13", "class": "Embodied-reasoning VLM", "params": null, "weights": "closed", "embodiments": "demo-controlled in sim/real: Boston Dynamics Spot (12-DoF legs + 6-DoF arm), simple 6-DoF arm, bimanual ALOHA 2 video retargeted to 76-DoF humanoid, Apollo 2 humanoid, Franka FR3 in MuJoCo", "training_data": null, "summary": "A general frontier Gemini model rather than a dedicated robot-action model; the launch post leads with an embodied-reasoning score and code-as-policy robot demos.", "results": [ { "benchmark": "ERQA", "value": "87%", "setting": "No tools (\"pure reasoning\"). Benchmark built by the same vendor.", "vendor_reported": true }, { "benchmark": "Internal real-world eval (6-DoF arm self-improvement)", "value": "success 56% to 75%", "setting": "Self-reflection loop on a simple 6-DoF arm; completion time 35 s to 24 s.", "vendor_reported": true } ], "caveat": "Its 87% ERQA score is vendor-reported, on a benchmark Google DeepMind itself built; other models' scores appear only in an image chart, so no comparison is shown.", "status": "verified", "sources": [ { "url": "https://developers.googleblog.com/gemini-4-argon-for-robotics/", "captured_at": "2026-10-26", "quote": "Using pure reasoning, Argon scores 87%" } ] }, { "id": "gemini-robotics-1-5", "name": "Gemini Robotics 1.5 + ER 1.5", "developer": "Google DeepMind", "country": "US", "country_label": "US", "released": "2025-09-25", "class": "VLA", "params": null, "weights": "closed", "embodiments": "ALOHA 2, Apptronik humanoid, bi-arm Franka", "training_data": null, "summary": "Pairs a VLA (1.5, partners only) with an agentic ER VLM (ER 1.5, public via Gemini API) and claims motion transfer across robots without per-robot retraining.", "results": [ { "benchmark": "Aggregate of 15 embodied-reasoning benchmarks (ER 1.5)", "value": "highest score of over 60", "setting": "Chart only; no standalone ERQA number.", "vendor_reported": true } ], "caveat": "The VLA is partner-only; ER 1.5 is public via the Gemini API.", "status": "verified", "sources": [ { "url": "https://deepmind.google/blog/gemini-robotics-15-brings-ai-agents-into-the-physical-world/", "captured_at": "2026-10-26", "quote": "Gemini Robotics 1.5 is currently available to select partners." } ] }, { "id": "gemini-robotics-er-1-6", "name": "Gemini Robotics-ER 1.6", "developer": "Google DeepMind", "country": "US", "country_label": "US", "released": "2026-04-14", "class": "Embodied-reasoning VLM", "params": null, "weights": "closed", "embodiments": "high-level reasoning layer; Boston Dynamics instrument reading", "training_data": null, "summary": "Public ER model that adds agentic-vision instrument reading.", "results": [ { "benchmark": "instrument reading", "value": "86% (93% with agentic vision)", "setting": "vs 67% Gemini 3.0 Flash, 23% ER 1.5", "vendor_reported": true } ], "caveat": null, "status": "verified", "sources": [ { "url": "https://deepmind.google/blog/gemini-robotics-er-1-6", "captured_at": "2026-10-26", "quote": "Starting today, Gemini Robotics-ER 1.6 is available to developers via the Gemini API and Google AI Studio." } ] }, { "id": "gemini-robotics-2-on-device-2", "name": "Gemini Robotics 2 (VLA) + On-Device 2", "developer": "Google DeepMind", "country": "US", "country_label": "US", "released": "2026-07-30", "class": "VLA", "params": null, "weights": "closed", "embodiments": "Apptronik Apollo 2 (Inspire and SharpaWave hands), Franka Duo (Robotiq gripper); On-Device 2 adapts to Dexmate, SO101, Trossen", "training_data": null, "summary": "Whole-body humanoid control plus fine dexterity in one VLA; On-Device 2 claims new bi-arm adaptation in a few hours with under 200 examples.", "results": [ { "benchmark": "Internal real-world eval (whole-body, Apollo 2 Inspire hands)", "value": "shelf pick 76.3%, table pick 68.4%, floor pick 45.7%", "setting": "single checkpoint, per-task success", "vendor_reported": true }, { "benchmark": "Internal real-world eval (multi-finger dexterity, Apollo 2 SharpaWave)", "value": "unscrew bulb 92%, tie trash bag 44%, ziplock 40%, screw bulb 36%, dustpan 32%", "setting": "per-task success", "vendor_reported": true }, { "benchmark": "Internal real-world eval (gripper, Franka Duo)", "value": "precise insertion 89.6%, tool kitting 78.9%, pick-and-place 74.2%", "setting": "per-task success", "vendor_reported": true } ], "caveat": "The VLA and On-Device 2 are early access for partners.", "status": "verified", "sources": [ { "url": "https://deepmind.google/blog/gemini-robotics-2-brings-whole-body-intelligence-to-robots/", "captured_at": "2026-10-26", "quote": "with just a few hours of adaptation time, typically with less than 200 examples" } ] }, { "id": "gemini-robotics-er-2", "name": "Gemini Robotics-ER 2", "developer": "Google DeepMind", "country": "US", "country_label": "US", "released": "2026-07-30", "class": "Embodied-reasoning VLM", "params": null, "weights": "closed", "embodiments": "reasoning layer for Boston Dynamics Spot, Apptronik Apollo 2, Franka Duo", "training_data": null, "summary": "Upgrade over ER 1.6 with safety-instruction and controlling-a-VLA evaluations.", "results": [ { "benchmark": "progress classification", "value": "57.4%", "setting": "ER 2, self-reported on blog.google", "vendor_reported": true }, { "benchmark": "moment finding", "value": "91.3% accuracy, 0.96 s mean absolute distance", "setting": "self-reported", "vendor_reported": true } ], "caveat": "Google gives no ERQA figure for ER 2.", "status": "verified", "sources": [ { "url": "https://blog.google/innovation-and-ai/models-and-research/google-deepmind/gemini-robotics-er-2/", "captured_at": "2026-10-26", "quote": "57.4% accuracy on progress classification tasks" } ] }, { "id": "pi0-6", "name": "pi0.6 / pi*0.6 (Recap)", "developer": "Physical Intelligence", "country": "US", "country_label": "US", "released": "2025-11-17", "class": "VLA", "params": "5B VLM backbone", "weights": "closed", "embodiments": "espresso-making, laundry folding in a new home, factory box assembly (robots not itemised in fetched span)", "training_data": "demonstrations, corrections, and autonomous experience via Recap RL", "summary": "First VLA from PI trained with RL from its own deployment experience (Recap: advantage-conditioned policies).", "results": [ { "benchmark": "Internal real-world eval (hardest tasks)", "value": "throughput more than doubled; failure rates down 2x or more", "setting": "after training on autonomous experience with Recap", "vendor_reported": true }, { "benchmark": "Internal real-world eval (deployment)", "value": "espresso 5:30am-11:30pm; 50 novel laundry items in a new home; 59 boxes assembled in a real factory", "setting": "continuous runs", "vendor_reported": true } ], "caveat": "Weights are not stated as released.", "status": "verified", "sources": [ { "url": "https://www.pi.website/blog/pistar06", "captured_at": "2026-10-26", "quote": "based on a 5B-parameter vision-language model" } ] }, { "id": "pi0-7", "name": "pi0.7", "developer": "Physical Intelligence", "country": "US", "country_label": "US", "released": "2026-04-16", "class": "VLA", "params": null, "weights": "closed", "embodiments": "bimanual UR5e, static bimanual robot, mobile robot, Franka arm", "training_data": "multi-robot data, human data, autonomous rollouts from various policies, diverse multimodal prompts; also open-source DROID", "summary": "Steerable generalist prompted by language, metadata, control modality and visual subgoals; first signs of compositional generalization.", "results": [ { "benchmark": "Internal real-world eval (dexterous tasks)", "value": "same performance as fine-tuned specialists", "setting": "generalist, no numeric score on page", "vendor_reported": true }, { "benchmark": "Internal real-world eval (UR5e laundry folding)", "value": "matches expert teleoperators' zero-shot success", "setting": "zero-shot on that robot", "vendor_reported": true } ], "caveat": "Weights are not stated as released.", "status": "verified", "sources": [ { "url": "https://www.pi.website/blog/pi07", "captured_at": "2026-10-26", "quote": "A steerable robotic foundation model that exhibits a step-change in generalization." } ] }, { "id": "isaac-gr00t-n1-5", "name": "Isaac GR00T N1.5", "developer": "NVIDIA", "country": "US", "country_label": "US", "released": null, "class": "VLA", "params": "3B", "weights": "open-licensed", "embodiments": "cross-embodiment, one encoder per unique embodiment; no robots listed on card", "training_data": "nvidia/PhysicalAI-Robotics-GR00T-X-Embodiment-Sim listed on card", "summary": "Medium-sized cross-embodiment model; license is non-commercial (superseded by N1.7 Open Model License).", "results": [], "caveat": "Non-commercial license; superseded by N1.7.", "status": "verified", "sources": [ { "url": "https://huggingface.co/nvidia/GR00T-N1.5-3B", "captured_at": "2026-10-26", "quote": "This model is ready for non-commercial use." } ] }, { "id": "isaac-gr00t-n1-7", "name": "Isaac GR00T N1.7", "developer": "NVIDIA", "country": "US", "country_label": "US", "released": "2026-03-16", "class": "VLA", "params": "3B", "weights": "open-licensed", "embodiments": "cross-embodiment; eval on GR-1 humanoid sim, RoboCasa mobile manipulator, DexMG; README tags Unitree G1, Panda, WidowX", "training_data": "GR00T Pretraining Data: 21.6 million data points across 13 datasets, from human, robot and simulated sources", "summary": "First GR00T release with commercial licensing; no task-success table on HF card.", "results": [ { "benchmark": "inference latency (not task success)", "value": "27.9 ms / 35.9 Hz end-to-end on H100 with TensorRT", "setting": "4 denoising steps, 1 camera", "vendor_reported": true } ], "caveat": "Commercial license; the model card gives no task-success table.", "status": "verified", "sources": [ { "url": "https://huggingface.co/nvidia/GR00T-N1.7-3B", "captured_at": "2026-10-26", "quote": "This model is released under the NVIDIA Open Model License Agreement." } ] }, { "id": "isaac-gr00t-n2", "name": "Isaac GR00T N2", "developer": "NVIDIA", "country": "US", "country_label": "US", "released": "Previewed 2026-03-16; not released", "class": "World-action model", "params": null, "weights": "closed", "embodiments": null, "training_data": "based on DreamZero research (world-model pretraining)", "summary": "Preview of a world-action successor to the N1.x VLAs, based on NVIDIA's DreamZero research; slated for end of 2026.", "results": [ { "benchmark": "MolmoSpaces and RoboArena", "value": "No. 1 for generalist robot policies", "setting": "NVIDIA claim at GTC 2026 preview", "vendor_reported": true }, { "benchmark": "new tasks in new environments", "value": "more than twice as often as leading VLAs", "setting": "NVIDIA claim", "vendor_reported": true } ], "caveat": "No public checkpoint. The ranking claims are NVIDIA's own.", "status": "verified", "sources": [ { "url": "https://nvidianews.nvidia.com/news/nvidia-and-global-robotics-leaders-take-physical-ai-to-the-real-world", "captured_at": "2026-10-26", "quote": "Slated to be available by the end of the year." } ] }, { "id": "helix", "name": "Helix", "developer": "Figure AI", "country": "US", "country_label": "US", "released": "2025-02-20", "class": "Dual-system VLA", "params": "7B (S2) + 80M (S1)", "weights": "closed", "embodiments": "Figure humanoids (two robots simultaneously)", "training_data": "~500 hours of teleoperated robot behavior data", "summary": "First humanoid VLA with a 7-9 Hz VLM (S2) driving a 200 Hz visuomotor policy (S1).", "results": [ { "benchmark": "Internal real-world eval (novel objects)", "value": "zero-shot to thousands of novel test objects", "setting": "language-prompted household pick-up; qualitative, no rate", "vendor_reported": true } ], "caveat": null, "status": "verified", "sources": [ { "url": "https://www.figure.ai/news/helix", "captured_at": "2026-10-26", "quote": "~500 hours in total" } ] }, { "id": "helix-02", "name": "Helix 02", "developer": "Figure AI", "country": "US", "country_label": "US", "released": "2026-01-27", "class": "Dual-system VLA", "params": "10M (S0); S1/S2 not stated", "weights": "closed", "embodiments": "Figure humanoid (full body, loco-manipulation)", "training_data": "S0: over 1,000 hours of retargeted human motion data", "summary": "Adds System 0 whole-body controller (1 kHz) beneath S1 (200 Hz) and S2.", "results": [ { "benchmark": "Internal real-world eval (dishwasher)", "value": "continuous 4-minute task, 61 loco-manipulation actions", "setting": "no success rate reported", "vendor_reported": true } ], "caveat": null, "status": "verified", "sources": [ { "url": "https://www.figure.ai/news/helix-02", "captured_at": "2026-10-26", "quote": "10M‑parameter neural network" } ] }, { "id": "helix-2-5", "name": "Helix 2.5", "developer": "Figure AI", "country": "US", "country_label": "US", "released": "2026-09-17", "class": "Dual-system VLA", "params": null, "weights": "closed", "embodiments": "Figure humanoid; robot count not stated", "training_data": "pretrained on 'Index' human-experience dataset (post says Index generates 35 minutes of new human experience every second); total hours not stated", "summary": "Claims zero-shot generalization to 30 unseen homes from human-video (Index) pretraining.", "results": [ { "benchmark": "Internal real-world eval (30 unseen homes, zero-shot)", "value": "9% to 56% success", "setting": "30 unseen homes, zero-shot. Baseline is Figure's own policy trained from scratch, not another lab's model.", "vendor_reported": true } ], "caveat": "The 56% is measured against Figure's own policy trained from scratch (9%), not against another lab's model.", "status": "verified", "sources": [ { "url": "https://www.figure.ai/news/helix-2-5-zero-shot-30-home-generalization", "captured_at": "2026-10-26", "quote": "30 unseen homes" } ] }, { "id": "skild-ai-s1", "name": "Skild AI S1 (powers Skild Brain)", "developer": "Skild AI", "country": "US", "country_label": "US", "released": "2026-08-18", "class": "VLA", "params": null, "weights": "closed", "embodiments": "omni-bodied; demonstrations may differ in scene, viewpoint or embodiment", "training_data": "pretrained on episodic data; scaling study used datasets from 1k to 100k hours", "summary": "Built from the ground up as an in-context learner: new tasks from one demonstration video without retraining.", "results": [ { "benchmark": "Internal eval (in-context learning, seen tasks)", "value": "96%", "setting": "ICL after scaling pretraining; per blog", "vendor_reported": true }, { "benchmark": "Internal eval (in-context learning, unseen tasks)", "value": "66%", "setting": "100k-hour dataset size", "vendor_reported": true } ], "caveat": null, "status": "verified", "sources": [ { "url": "https://skild.ai/blogs/s1", "captured_at": "2026-10-26", "quote": "built from the ground up as an in-context learner" } ] }, { "id": "rho", "name": "Rho", "developer": "Microsoft Research", "country": "US", "country_label": "US", "released": "2026", "class": "VLA", "params": "5B", "weights": "open", "embodiments": "bimanual: YAM Box, UR AI Trainer, FR3 Duo", "training_data": "Rho-Tomyum mixture: 3,200+ hours from more than a dozen robots, mostly dual-arm; vision-language data is 10% of updates", "summary": "Microsoft's open-weight 5B bimanual VLA family with in-house embodiment-specific variants.", "results": [ { "benchmark": "BusyBox", "value": "90% overall (Rho-YAM-Box)", "setting": "tied with pi0.5", "vendor_reported": true }, { "benchmark": "LIBERO", "value": "97.9% average", "setting": "Rho-base, strongest aggregate among compared models", "vendor_reported": true }, { "benchmark": "RoboEval", "value": "73% vs 67% (pi0.5); GR00T N1.7 and MolmoAct2 at 61%", "setting": "overall success", "vendor_reported": true } ], "caveat": "The project page names no license.", "status": "verified", "sources": [ { "url": "https://microsoft.github.io/rhobotics/", "captured_at": "2026-10-26", "quote": "open-weights, 5B-parameter vision-language-action models for bimanual robotic manipulation" } ] }, { "id": "gen-0", "name": "GEN-0", "developer": "Generalist AI", "country": "US", "country_label": "US", "released": "2025-11-04", "class": "VLA", "params": "10B+", "weights": "closed", "embodiments": null, "training_data": "raw high-fidelity physical interaction data", "summary": "Scaling-law claim: power-law link between pretraining data and post-training performance, phase transition near 7B.", "results": [], "caveat": null, "status": "verified", "sources": [ { "url": "https://generalistai.com/blog/gen-0", "captured_at": "2026-10-26", "quote": "November 4, 2025" } ] }, { "id": "gen-1", "name": "GEN-1", "developer": "Generalist AI", "country": "US", "country_label": "US", "released": "2026-04-02", "class": "VLA", "params": null, "weights": "closed", "embodiments": null, "training_data": "over half a million hours of high-fidelity physical interaction data (no robot data in pretraining); about one hour of robot data per task", "summary": "Claims production-grade reliability (99%) on simple physical tasks from ~1 hour of per-task robot data.", "results": [ { "benchmark": "Internal real-world eval (6 reliability tasks)", "value": "average 99%", "setting": "fine-tuned on ~1 hour robot data per task; GEN-0 fine-tuned averaged 64%, from-scratch 19%", "vendor_reported": true }, { "benchmark": "box folding speed", "value": "12.1 s (about 2.8x faster than GEN-0 and pi0 at ~34 s)", "setting": "self-reported", "vendor_reported": true } ], "caveat": "Early-access partners only.", "status": "verified", "sources": [ { "url": "https://generalistai.com/blog/gen-1", "captured_at": "2026-10-26", "quote": "average 99%" } ] }, { "id": "gen-1-5", "name": "GEN-1.5", "developer": "Generalist AI", "country": "US", "country_label": "US", "released": "2026-08-19", "class": "VLA", "params": null, "weights": "closed", "embodiments": "transfer to new embodiments, hands, and from human demonstrations (no list)", "training_data": "spans from Generalist's data engine across homes, warehouses, factories; no simulation data", "summary": "Adds one-shot in-context task learning ('learn a new task in seconds, from a single example').", "results": [ { "benchmark": "Internal real-world eval (one-shot in-context)", "value": "59% (+/-10%) average success", "setting": "single example, no gradient updates", "vendor_reported": true }, { "benchmark": "Internal real-world eval (few-shot)", "value": "83% (+/-9%)", "setting": "10 gradient steps on 5 minutes of data per task", "vendor_reported": true } ], "caveat": null, "status": "verified", "sources": [ { "url": "https://generalistai.com/blog/gen-1.5", "captured_at": "2026-10-26", "quote": "learn a new task in seconds, from a single example, without gradient updates or fine-tuning" } ] }, { "id": "natura-1", "name": "Natura-1", "developer": "Nature Will", "country": "China", "country_label": "China", "released": "2026-10-13", "class": "VLA", "params": "50B mixture-of-experts", "weights": "unknown", "embodiments": "20-DoF dexterous hands (Wuji hands on ROKAE AR5); gripper platform ARX AC One for baselines", "training_data": "more than one million hours of egocentric human experience: 61,702,471 episodes, 317,842 tasks, 14,286 scenes", "summary": "Claims first 50B-parameter model for 20-DoF dexterous hands, pretrained on over a million hours of human egocentric data, 100 ms control loop.", "results": [ { "benchmark": "Internal dexterous-manipulation benchmark (12 capability dimensions)", "value": "53.4 (Base), 59.1 (Agentic)", "setting": "dexterous hands; baselines on grippers: pi0.5 13.5, GPT-6-Astra 16.7; baselines produced no score on the dexterous platform; scale not named", "vendor_reported": true } ], "caveat": "Weights status unknown: the page does not say. The benchmark is the developer's own; baselines were run on grippers because they produced no score on the dexterous platform.", "status": "verified", "sources": [ { "url": "https://www.naturewill.ai/blog/natura-1?lang=en", "captured_at": "2026-10-26", "quote": "50-billion-parameter mixture-of-experts (MoE) model" } ] }, { "id": "dyna-1", "name": "Dyna-1 (Dynamism v1)", "developer": "Dyna Robotics", "country": "US", "country_label": "US", "released": "2025-06-24", "class": "VLA", "params": null, "weights": "closed", "embodiments": null, "training_data": null, "summary": "First Dyna foundation model: continuous 24-hour autonomous napkin folding.", "results": [ { "benchmark": "Internal real-world eval (24-hour napkin folding)", "value": "99.4% success, 850+ napkins, ~60% of human speed", "setting": "no human intervention", "vendor_reported": true } ], "caveat": "Earlier press reported 700+ napkins at 99%; the primary page says 850+ at 99.4%.", "status": "verified", "sources": [ { "url": "https://www.dyna.co/research/dyna-1", "captured_at": "2026-10-26", "quote": "99.4% success rate" } ] }, { "id": "dyna-2", "name": "Dyna-2", "developer": "Dyna Robotics", "country": "US", "country_label": "US", "released": "2026-08-15", "class": "World-action model", "params": null, "weights": "closed", "embodiments": "stationary bimanual 6-DOF YAM arms with parallel grippers (11 tasks), same arms with dexterous hands (2 tasks), early semi-humanoid prototype (1 task)", "training_data": "over one million hours of human video (mostly head-mounted first-person)", "summary": "Claims a 1-million-hour scaling law for world-action models; one-step video sampling 10,203 ms -> 110 ms on one H100.", "results": [ { "benchmark": "Internal customer acceptance tests", "value": "Dyna-2 87% vs Dyna-1 46%", "setting": "zero-shot at customer sites (stationary robot)", "vendor_reported": true }, { "benchmark": "Internal 14-task mean normalized score", "value": "20% -> 28% -> 45% -> 53% of attainable maximum", "setting": "data ladder up to 1M hours", "vendor_reported": true }, { "benchmark": "WAM vs VLA", "value": "1.55x VLA success rate", "setting": "early Dyna-2, pooled comparison", "vendor_reported": true } ], "caveat": null, "status": "verified", "sources": [ { "url": "https://www.dyna.co/dyna-2", "captured_at": "2026-10-26", "quote": "A scaling law for world-action models exists up to one million hours of human data." } ] }, { "id": "dyna-2-1", "name": "Dyna-2.1", "developer": "Dyna Robotics", "country": "US", "country_label": "US", "released": "2026-09-28", "class": "World-action model", "params": null, "weights": "closed", "embodiments": "Taku semi-humanoid: two 7-DoF arms, 21-DoF upper body, mobile base", "training_data": "one million hours of human video (pre-training)", "summary": "Full-stack physical agent (100 Hz whole-body controller + ~5 Hz Dyna-2 policy + VLM workflow orchestrator) doing hour-long loco-dexterous laundry workflows.", "results": [ { "benchmark": "Internal customer acceptance tests", "value": "87% (Dyna-2) vs 46% (Dyna-1)", "setting": "zero-shot, stationary robot", "vendor_reported": true } ], "caveat": "Dyna's own arithmetic: 95% per-step success over about 79 steps leaves a 1.7% chance of finishing unaided.", "status": "verified", "sources": [ { "url": "https://www.dyna.co/dyna-2.1", "captured_at": "2026-10-26", "quote": "the first physical agent to perform hour-long, non-linear loco-dexterous manipulation workflows" } ] }, { "id": "act-2", "name": "ACT-2", "developer": "Sunday Robotics", "country": "US", "country_label": "US", "released": "2026-07-17", "class": "VLA", "params": null, "weights": "closed", "embodiments": "Memo wheeled home robot", "training_data": "pretrained on a high-quality, high-diversity, sensorized human dataset (proprietary glove hardware); post-training used one demonstration per folding technique", "summary": "Claims 99.1% zero-shot laundry folding in unseen homes with a single checkpoint; socks/underwear out of scope.", "results": [ { "benchmark": "Internal real-world eval (laundry folding)", "value": "99.1% (+/-0.3% SE), n=785 attempts, 9 garment types", "setting": "zero-shot in unseen homes, zero per-home adaptation, weights fixed; number of homes not stated", "vendor_reported": true } ], "caveat": "A preview; the number of evaluation homes is not given.", "status": "verified", "sources": [ { "url": "https://www.sunday.ai/blog/act-2-preview", "captured_at": "2026-10-26", "quote": "The result is 99.1% zero-shot success on laundry folding across diverse unseen environments." } ] }, { "id": "flux-3-action", "name": "FLUX 3 Action", "developer": "Black Forest Labs", "country": "Other", "country_label": "Germany", "released": "2026-09-23", "class": "World-action model", "params": "7B", "weights": "open-licensed", "embodiments": "Franka arm (DROID tasks), SO-101; midtraining on teleoperation across 14 embodiments", "training_data": "pretraining on image, video and audio with video over 95% of tokens; midtraining added gaming, egocentric human-hand, handheld gripper, and teleoperation data", "summary": "Open-weight 7B joint video+action flow-matching model; policies for DROID and low-cost SO-101.", "results": [ { "benchmark": "RoboLab-120 (NVIDIA)", "value": "38.3% (base), 42.2% (guidance-distilled)", "setting": "Simulation; same table lists Cosmos 3 Nano 36.8% and pi0.5 28.0%.", "vendor_reported": true } ], "caveat": "The page gives 42.2% in its text and 42.92% in its table. Weights carry a custom license.", "status": "verified", "sources": [ { "url": "https://bfl.ai/blog/flux-3-action", "captured_at": "2026-10-26", "quote": "7B" } ] }, { "id": "redwood-ai", "name": "Redwood AI", "developer": "1X Technologies", "country": "US", "country_label": "US", "released": "2025-06-10", "class": "VLA", "params": "160M", "weights": "closed", "embodiments": "NEO humanoid", "training_data": null, "summary": "Small 160M model running at about 5 Hz on NEO's onboard GPU.", "results": [], "caveat": "Weights are not stated as released.", "status": "verified", "sources": [ { "url": "https://www.1x.tech/discover/redwood-ai", "captured_at": "2026-10-26", "quote": "160M parameters" } ] }, { "id": "tri-large-behavior-model", "name": "TRI Large Behavior Model", "developer": "Toyota Research Institute", "country": "US", "country_label": "US", "released": "2025-07-07", "class": "VLA", "params": null, "weights": "closed", "embodiments": "bimanual manipulation", "training_data": "multi-task teleoperated + pretraining at scale", "summary": "Large controlled study finding performance predictably increases with pretraining scale and diversity.", "results": [], "caveat": null, "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2507.05331", "captured_at": "2026-10-26", "quote": "performance predictably increases as pretraining scale and diversity grows" } ] }, { "id": "atlas-lbm", "name": "Atlas LBM", "developer": "Boston Dynamics / Toyota Research Institute", "country": "US", "country_label": "US", "released": "2025-08-20", "class": "VLA", "params": null, "weights": "closed", "embodiments": "Atlas humanoid, whole-body", "training_data": null, "summary": "Single LBM controls hands and feet of Atlas through packing/sorting with recovery from disruptions; company says new skills need no new code.", "results": [], "caveat": "Press page only; no model card or paper.", "status": "snippet-only", "sources": [ { "url": "https://pressroom.toyota.com/ai-powered-robot-by-boston-dynamics-and-toyota-research-institute-takes-a-key-step-towards-general-purpose-humanoids/", "captured_at": "2026-10-26" } ] }, { "id": "smolvla", "name": "SmolVLA", "developer": "Hugging Face", "country": "US", "country_label": "US/France", "released": "2025-06-02", "class": "VLA", "params": null, "weights": "open", "embodiments": "affordable platforms such as SO-100/SO-101", "training_data": "community-collected data from affordable robotic platforms", "summary": "Compact open VLA with an asynchronous inference stack decoupling action prediction from execution.", "results": [], "caveat": null, "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2506.01844", "captured_at": "2026-10-26", "quote": "SmolVLA achieves performance comparable to VLAs that are 10x larger." } ] }, { "id": "openvla-oft", "name": "OpenVLA-OFT", "developer": "Stanford University", "country": "US", "country_label": "US", "released": "2025-02-27", "class": "VLA", "params": null, "weights": "open", "embodiments": "Franka/LIBERO sim; bimanual ALOHA", "training_data": "fine-tuning recipe on OpenVLA", "summary": "Fine-tuning recipe (parallel decoding, action chunking, continuous actions) lifting OpenVLA LIBERO from 76.5% to 97.1% with 26x higher throughput.", "results": [ { "benchmark": "LIBERO", "value": "97.1% (from 76.5%)", "setting": "fine-tuned, new SOTA at the time", "vendor_reported": true } ], "caveat": null, "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2502.19645", "captured_at": "2026-10-26", "quote": "from 76.5% to 97.1%" } ] }, { "id": "covariant-rfm-1", "name": "Covariant RFM-1", "developer": "Covariant (licensed to Amazon, 2024)", "country": "US", "country_label": "US", "released": "2024-03 (RFM-1)", "class": "VLA", "params": null, "weights": "closed", "embodiments": "warehouse picking arms", "training_data": null, "summary": "Warehouse-picking foundation model; Amazon took a non-exclusive license to RFM-1 in August 2024 and hired the founders.", "results": [], "caveat": "No later Covariant model was found.", "status": "snippet-only", "sources": [ { "url": "https://www.aboutamazon.com/news/company-news/amazon-covariant-ai-robots", "captured_at": "2026-10-26" } ] }, { "id": "rldx-1", "name": "RLDX-1", "developer": "RLWRLD", "country": "Other", "country_label": "Korea", "released": "2026-05-05", "class": "VLA", "params": null, "weights": "unknown", "embodiments": "ALLEX humanoid (high-DoF); dexterous manipulation", "training_data": "includes data synthesis for rare manipulation scenarios", "summary": "Dexterous-manipulation policy built on a Multi-Stream Action Transformer, tested on a high-DoF humanoid.", "results": [ { "benchmark": "Internal real-world eval (ALLEX humanoid tasks)", "value": "86.8%", "setting": "ALLEX humanoid tasks, per the arXiv paper; pi0.5 and GR00T N1.6 scored around 40% on the same tasks.", "vendor_reported": true } ], "caveat": "Only the arXiv figure is shown. Press claims that it beat every rival across eight benchmarks are not in the paper. Weight release is not stated.", "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2605.03269", "captured_at": "2026-10-26", "quote": "RLDX-1, a general-purpose robotic policy for dexterous manipulation built on the Multi-Stream Action Transformer (MSAT)" } ] }, { "id": "gr-3", "name": "GR-3", "developer": "ByteDance Seed", "country": "China", "country_label": "China", "released": "2025-07-21", "class": "VLA", "params": null, "weights": "closed", "embodiments": "ByteMini bi-manual mobile robot", "training_data": "web-scale vision-language data, human trajectories via VR devices, and robot trajectories (no quantities)", "summary": "Large VLA co-trained on VL data, VR human trajectories, and robot data on ByteDance's own ByteMini.", "results": [], "caveat": null, "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2507.15493", "captured_at": "2026-10-26", "quote": "GR-3 is a large-scale vision-language-action (VLA) model." } ] }, { "id": "agibot-go-1", "name": "AgiBot GO-1 (Genie Operator-1)", "developer": "AgiBot (Zhiyuan)", "country": "China", "country_label": "China", "released": "2025-03-09", "class": "VLA", "params": null, "weights": "open", "embodiments": "AgiBot G1", "training_data": "AgiBot World: over 1 million trajectories across 217 tasks in five deployment scenarios", "summary": "Generalist policy using latent action representations, released with the large AgiBot World dataset.", "results": [ { "benchmark": "pre-training vs Open X-Embodiment", "value": "30% average improvement", "setting": "policies pre-trained on AgiBot World vs Open X-Embodiment", "vendor_reported": true } ], "caveat": null, "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2503.06669", "captured_at": "2026-10-26", "quote": "achieve an average performance improvement of 30% over those trained on Open X-Embodiment" } ] }, { "id": "ace-brain-0-6", "name": "ACE-Brain-0.6", "developer": "ACE Robotics (Daxiao Robotics)", "country": "China", "country_label": "China", "released": "2026-10-19", "class": "Embodied-reasoning VLM", "params": "9B (fine-tune of Qwen3.5-9B)", "weights": "closed", "embodiments": null, "training_data": "about 1 million samples from 30+ open robotics datasets plus proprietary data", "summary": "SenseTime co-founder Xiaogang Wang's company; 9B spatial-grounding VLM claiming to beat GPT-6 Astra on grounding.", "results": [ { "benchmark": "seven public grounding benchmarks (average)", "value": "74.53 vs GPT-6 Astra 73.68", "setting": "company-reported; beat Astra on 6 of 8 grounding benchmarks", "vendor_reported": true }, { "benchmark": "RobotGrounding", "value": "94.13 vs GPT-6 Astra 91.10", "setting": "real-robot-scenario grounding eval, static images", "vendor_reported": true } ], "caveat": "Company-reported via secondary press; no primary page or weights.", "status": "snippet-only", "sources": [ { "url": "https://technode.com/2026/10/20/sensetime-co-founders-ace-robotics-releases-model-for-robot-spatial-perception/", "captured_at": "2026-10-26", "quote": "an average of 74.53 across seven public benchmarks" } ] }, { "id": "graspvla-astrabrain", "name": "GraspVLA / AstraBrain", "developer": "Galbot", "country": "China", "country_label": "China", "released": "2025 (GraspVLA)", "class": "VLA", "params": null, "weights": "partly open", "embodiments": "Galbot G1 / S1 wheeled humanoids", "training_data": "GraspVLA pretrained on synthetic data (per roboticscenter.ai)", "summary": "GraspVLA is the only Galbot model with released weights; AstraBrain is described as an end-to-end brain/cerebellum model.", "results": [], "caveat": "Release dates conflict across sources. GraspVLA weights are non-commercial (CC BY-NC 4.0).", "status": "snippet-only", "sources": [ { "url": "https://www.roboticscenter.ai/institutions/models/galbot", "captured_at": "2026-10-26" } ] }, { "id": "unifolm-vla-0", "name": "UnifoLM-VLA-0", "developer": "Unitree Robotics", "country": "China", "country_label": "China", "released": "2026-01-29", "class": "VLA", "params": null, "weights": "open-licensed", "embodiments": "Unitree G1 humanoid", "training_data": "UnifoLM-VLM-Base backbone fine-tuned on general image-text VQA data and open-source robot datasets; hours not stated", "summary": "Open humanoid-manipulation VLA with a LIBERO fine-tune; 12 manipulation categories from a single policy per HF page.", "results": [], "caveat": "A LIBERO score circulates in press but is not on the project page, so it is not shown. Non-commercial license (CC BY-NC-SA 4.0).", "status": "verified", "sources": [ { "url": "https://github.com/unitreerobotics/unifolm-vla", "captured_at": "2026-10-26", "quote": "Jan 29, 2026: We released the training and inference code, along with the model weights" } ] }, { "id": "wall-oss-wall-oss-0-5", "name": "WALL-OSS / WALL-OSS-0.5", "developer": "X Square Robot", "country": "China", "country_label": "China", "released": "2025-09-15 (0.5: 2026-05)", "class": "VLA", "params": null, "weights": "open", "embodiments": "real-robot manipulation (unspecified)", "training_data": "large-scale multimodal pretraining; Gradient-Bridged Pretraining for 0.5", "summary": "Open VLA family; 0.5 positioned as deployment-ready; Apache-2.0 per repo.", "results": [], "caveat": "A press figure versus pi0.5 is not shown because it could not be traced to a primary page.", "status": "verified", "sources": [ { "url": "https://github.com/X-Square-Robot/wall-x", "captured_at": "2026-10-26", "quote": "[May 2026] We introduce Wall-OSS-0.5: A Deployment-Ready VLA with Gradient-Bridged Pretraining." } ] }, { "id": "rdt2", "name": "RDT2", "developer": "Tsinghua University (TSAIL)", "country": "China", "country_label": "China", "released": "2026-02-03", "class": "VLA", "params": null, "weights": "open", "embodiments": "zero-shot on unseen platforms: bimanual UR5e and bimanual Franka Research 3", "training_data": "over 10,000 hours of UMI demonstrations in diverse families", "summary": "Scales UMI handheld-gripper data to 10,000+ hours for zero-shot cross-embodiment transfer.", "results": [], "caveat": "The arXiv page does not itself state that weights are released; the open license comes from secondary reporting.", "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2602.03310", "captured_at": "2026-10-26", "quote": "zero-shot generalizes to unseen objects, scenes, instructions, and even robotic platforms" } ] }, { "id": "neuron-ai", "name": "Neuron AI", "developer": "NEURA Robotics", "country": "Other", "country_label": "Germany", "released": "2026-10-14", "class": "VLA", "params": null, "weights": "closed", "embodiments": "five NEURA types: 4NE1 Gen 3, 4NE1 Gen 4, wheeled 4NE1, two further bimanual platforms", "training_data": null, "summary": "Perception-Language-Action model fusing depth, tactile, force-torque and sound with memory; runs on-robot; 'time-to-skill' named as key metric.", "results": [], "caveat": "No benchmarks or weights statement.", "status": "verified", "sources": [ { "url": "https://www.implicator.ai/neura-robotics-neuron-ai-touch-sound-memory/", "captured_at": "2026-10-26", "quote": "NEURA has published no benchmark for the model." } ] }, { "id": "flux-mimic", "name": "FLUX-mimic", "developer": "mimic robotics (with Black Forest Labs)", "country": "Other", "country_label": "Switzerland", "released": "2026-07-23", "class": "World-action model", "params": null, "weights": "closed", "embodiments": "mimic hand M1 (tendon-driven dexterous hand); U1 wearable data glove; Audi factory tests", "training_data": "built on FLUX 3 video model; fine-tune with ~30 minutes of robot data (company claim)", "summary": "Video-action model that attaches an action decoder to a generative video model; 30 minutes of robot data to fine-tune.", "results": [ { "benchmark": "fine-tuning data requirement", "value": "as little as 30 minutes of robot data vs 30+ hours", "setting": "company claim, per press", "vendor_reported": true } ], "caveat": "Release date differs across sources (July 23 vs August 13, 2026).", "status": "snippet-only", "sources": [ { "url": "https://www.businesswire.com/news/home/20260723711096/en/", "captured_at": "2026-10-26" } ] }, { "id": "xiaomi-robotics-1", "name": "Xiaomi-Robotics-1", "developer": "Xiaomi Robotics", "country": "China", "country_label": "China", "released": "2026-08-05", "class": "VLA", "params": "5B (per repo name)", "weights": "open", "embodiments": "mobile manipulation; cross-embodiment; not named", "training_data": "initialized from a model pre-trained on over 100K hours of embodiment-free UMI trajectories, post-trained on over 10K hours of cross-embodiment data", "summary": "Apache-2.0 open 5B VLA on a Qwen3-VL-4B-Instruct backbone pre-trained on 100K hours of UMI data.", "results": [], "caveat": "Xiaomi says it ranked first on RoboCasa365 and RoboDojo; the model card gives no numbers, so none is shown.", "status": "verified", "sources": [ { "url": "https://huggingface.co/XiaomiRobotics/Xiaomi-Robotics-1-5B", "captured_at": "2026-10-26", "quote": "pre-trained on over 100K hours of embodiment-free UMI trajectories" } ] }, { "id": "qwen-robotmanip", "name": "Qwen-RobotManip", "developer": "Alibaba (Qwen)", "country": "China", "country_label": "China", "released": "2026-06-16", "class": "VLA", "params": null, "weights": "closed", "embodiments": null, "training_data": "more than 38,100 hours incl. synthetic demonstrations", "summary": "Part of Qwen's first robot suite (RobotNav, RobotManip, RobotWorld).", "results": [ { "benchmark": "RoboChallenge Table30 v1 (universal models track)", "value": "Ranked first (universal-models track)", "setting": "Company claim; no score published.", "vendor_reported": true } ], "caveat": "Enterprise pilot only; no weights release found.", "status": "snippet-only", "sources": [ { "url": "https://forklog.com/en/news/alibaba-unveils-qwen-robot-suite-for-navigation-manipulation-and-world-modeling", "captured_at": "2026-10-26" } ] }, { "id": "spirit-v1-5", "name": "Spirit v1.5", "developer": "Spirit AI (Qianxun Intelligence)", "country": "China", "country_label": "China", "released": "2026-01-14", "class": "VLA", "params": null, "weights": "open", "embodiments": "RoboChallenge real-robot 30-task benchmark", "training_data": null, "summary": "Model that topped the RoboChallenge real-robot leaderboard in January 2026, per company statements to state media.", "results": [ { "benchmark": "RoboChallenge (real-robot, 30 tasks)", "value": "total score 66.09, 50.33% task success, ahead of pi0.5", "setting": "company-reported to CGTN/China Daily", "vendor_reported": true } ], "caveat": "A third-party catalog says Spirit was later removed from the RoboChallenge leaderboard after a methodology change. This could not be confirmed.", "status": "snippet-only", "sources": [ { "url": "https://news.cgtn.com/news/2026-01-14/Chinese-AI-startup-tops-global-embodied-intelligence-benchmark-1JV32RBb7QQ/p.html", "captured_at": "2026-10-26", "quote": "Chinese AI startup tops global embodied intelligence benchmark" } ] }, { "id": "galaxea-g0plus", "name": "Galaxea G0Plus", "developer": "Galaxea AI", "country": "China", "country_label": "China", "released": "2026-01-04", "class": "Dual-system VLA", "params": "3B (G0Tiny: 250M)", "weights": "partly open", "embodiments": "Galaxea R1Lite and R1 Pro mobile manipulators", "training_data": "G0Plus_3B_base pre-trained on 2k+ hours real-world robot data; Galaxea Open-World Dataset 500+ hours real-world mobile manipulation", "summary": "Open-weight 3B VLA released under a non-commercial G0 PLUS Community License (post-2026-01-04 commits).", "results": [], "caveat": "Newer releases use a non-commercial license.", "status": "verified", "sources": [ { "url": "https://raw.githubusercontent.com/OpenGalaxea/GalaxeaVLA/13a16a9049aee8f1d799b56fccc0c5832a75fc2f/README.md", "captured_at": "2026-10-26", "quote": "[Jan 4, 2026] We are releasing G0Plus, our latest pre-trained VLA model for multi-task robot manipulation." } ] }, { "id": "thinker-4b", "name": "Thinker-4B", "developer": "UBTech Robotics", "country": "China", "country_label": "China", "released": "2025-12", "class": "Embodied-reasoning VLM", "params": "4B", "weights": "open-licensed", "embodiments": null, "training_data": null, "summary": "Open 4B VLM on Qwen3-VL-4B-Instruct; license is CC BY-NC-SA 4.0 (non-commercial).", "results": [], "caveat": "The model card claims records on seven benchmarks but names none. Non-commercial license.", "status": "verified", "sources": [ { "url": "https://huggingface.co/UBTECH-Robotics/Thinker-4B", "captured_at": "2026-10-26", "quote": "Thinker sets new records across 7 embodied AI benchmarks in Task Planning, Visual Grounding and Spatial Understanding" } ] }, { "id": "genie-operator-2", "name": "Genie Operator-2 (GO-2)", "developer": "AgiBot", "country": "China", "country_label": "China", "released": "2026-04-17", "class": "Dual-system VLA", "params": null, "weights": "closed", "embodiments": "fleets of robots; platforms not named", "training_data": "tens of thousands of hours of interaction data", "summary": "Action chain-of-thought plus asynchronous dual-system architecture; core work accepted to CVPR 2026 and ACL 2026.", "results": [ { "benchmark": "LIBERO", "value": "98.5% average", "setting": "Spatial/Object/Goal/Long", "vendor_reported": true }, { "benchmark": "LIBERO-Plus", "value": "86.6% zero-shot", "setting": "under disturbances", "vendor_reported": true }, { "benchmark": "VLABench", "value": "47.4 average score", "setting": "cross-category and texture generalization", "vendor_reported": true }, { "benchmark": "Genie Sim 3.0 sim-to-real", "value": "82.9% real-world success", "setting": "trained on simulation data only", "vendor_reported": true } ], "caveat": "Weights availability is not stated.", "status": "verified", "sources": [ { "url": "https://agibot.com/article/231/detail/56.html", "captured_at": "2026-10-26", "quote": "average success rate of 98.5%" } ] }, { "id": "hy-embodied-0-5", "name": "HY-Embodied-0.5", "developer": "Tencent Hunyuan / Robotics X", "country": "China", "country_label": "China", "released": "2026-04-09", "class": "Embodied-reasoning VLM", "params": "2B and 32B", "weights": "partly open", "embodiments": null, "training_data": null, "summary": "Only the 2B edge model's weights are stated as open; 32B claimed comparable to Gemini 3.0 Pro.", "results": [ { "benchmark": "ERQA", "value": "54.5 (2B model)", "setting": "Also CV-Bench 89.2, DA-2K 92.3, MindCube 66.3.", "vendor_reported": true } ], "caveat": "Only the 2B model's weights are open. ERQA scores across labs use different model sizes and setups, so they are not directly comparable.", "status": "verified", "sources": [ { "url": "https://ai.gitcode.com/tencent_hunyuan/HY-Embodied-0.5", "captured_at": "2026-10-26", "quote": "[2026-04-09] 我们正式发布HY-Embodied-0.5" } ] }, { "id": "xiaomi-robotics-0", "name": "Xiaomi-Robotics-0", "developer": "Xiaomi Robotics", "country": "China", "country_label": "China", "released": "2026-02-13", "class": "VLA", "params": null, "weights": "open", "embodiments": "real robots on consumer-grade GPU", "training_data": null, "summary": "Predecessor to Xiaomi-Robotics-1; real-time execution on consumer GPU claimed.", "results": [], "caveat": "Press names LIBERO, CALVIN and SimplerEnv; the abstract gives no numbers.", "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2602.12684", "captured_at": "2026-10-26", "quote": "Submitted on 13 Feb 2026" } ] } ], "world_models": [ { "id": "praxis-1", "name": "Praxis-1", "developer": "Runway", "country": "US", "country_label": "US", "released": "2026-09", "type": "world-action model", "role": "policy", "role_detail": "policy pretraining", "weights": "closed", "summary": "Announced as an open-weight world-action model built on Runway's video pretraining; weights not yet downloadable. Early partners (Noble Machines, Standard Bots, Ultra) run it on their own hardware. Runway reports 0.95 sim-to-real correlation when evaluating policies inside the model.", "caveat": "Figures are vendor-reported.", "status": "verified", "sources": [ { "url": "https://runway.com/research/introducing-praxis-1", "captured_at": "2026-10-26", "quote": "Today we're announcing Praxis-1, our first open-weight world action model." } ] }, { "id": "cosmos-3", "name": "Cosmos 3", "developer": "NVIDIA", "country": "US", "country_label": "US", "released": "2026-06-01", "type": "video world model", "role": "both", "role_detail": "data generation / policy pretraining / planning / evaluation", "weights": "open", "summary": "Open \"omnimodel\" family (text, image, video, sound, actions) in three sizes: Super 64B, Nano 16B, Edge 4B (Edge listed as coming soon). Used for synthetic data, policy pretraining, planning and evaluation.", "caveat": "NVIDIA says it ranks first among open models on four leaderboards but gives no numbers.", "status": "verified", "sources": [ { "url": "https://investor.nvidia.com/news/press-release-details/2026/NVIDIA-Launches-Cosmos-3-the-Open-Frontier-Foundation-Model-for-Physical-AI/default.aspx", "captured_at": "2026-10-26", "quote": "TAIPEI, Taiwan, June 01, 2026 ... NVIDIA today launched NVIDIA Cosmos 3, an open world foundation model for physical AI" } ] }, { "id": "motubrain", "name": "MotuBrain", "developer": "ShengShu Technology", "country": "China", "country_label": "China", "released": "2026-04-30", "type": "world-action model", "role": "policy", "role_detail": "direct control", "weights": "closed", "summary": "World-action model that jointly models video and action. The paper reports 95.8% (clean) and 96.1% (randomized) average success on RoboTwin 2.0.", "caveat": "Paper result from the developer; weight release not confirmed.", "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2604.27792", "captured_at": "2026-10-26", "quote": "Motubrain achieves 95.8% and 96.1% average success on RoboTwin 2.0 under clean and randomized settings, respectively" } ] }, { "id": "1x-world-model", "name": "1X World Model", "developer": "1X Technologies", "country": "US", "country_label": "US", "released": "2026-01-12", "type": "video world model", "role": "policy", "role_detail": "direct control", "weights": "closed", "summary": "Video world model inside the NEO home humanoid: it imagines the next actions from the robot camera and a prompt, and an inverse-dynamics model turns them into motor commands.", "caveat": "No success rates or benchmark numbers published.", "status": "verified", "sources": [ { "url": "https://www.aap.com.au/aapreleases/globenewswire9623335/", "captured_at": "2026-10-26", "quote": "1X is excited to announce the new 1X World Model, a groundbreaking AI update for NEO" } ] }, { "id": "vpp2", "name": "VPP2", "developer": "Robotera", "country": "China", "country_label": "China", "released": "2026-10-09", "type": "world-action model", "role": "policy", "role_detail": "direct control", "weights": "open", "summary": "Open-sourced world-action model that stages video prediction before action learning. On the live RoboDojo board it scores 39.26 / 32.26% success (entry added 2026-10-08), the best entry outside the agent track as of 2026-10-26.", "caveat": "Developer-contributed leaderboard entry. Code at github.com/roboterax/video-prediction-policy-2.", "status": "verified", "sources": [ { "url": "https://technode.global/prnasia/robotera-tops-embodied-ai-benchmark-robodojo-without-additional-data-or-agent-rsi/", "captured_at": "2026-10-26", "quote": "VPP2 achieved an average success rate of 32.26% and an average score of 39.26," }, { "url": "https://robodojo-benchmark.com/", "captured_at": "2026-10-26", "quote": "Add VPP2 (39.26 / 32.26%, contributed by RobotEra)" } ] }, { "id": "daimon-twm", "name": "Daimon-TWM", "developer": "Daimon Robotics", "country": "China", "country_label": "China", "released": "2026-08", "type": "tactile world model", "role": "policy", "role_detail": "planning / direct control", "weights": "closed", "summary": "Tactile world model for force, friction, deformation and slip, aimed at planning and real-time control. Shown at IROS 2026.", "caveat": "Launch date disputed in trade press; performance claims are company-reported.", "status": "snippet-only", "sources": [ { "url": "https://www.robotics247.com/article/iros-2026-daimon-robotics-showcases-tactile-intelligence-infrastructure-for-physical-ai/technologies", "captured_at": "2026-10-26", "quote": "IROS 2026: Daimon Robotics showcases tactile intelligence infrastructure for physical AI" } ] }, { "id": "genesis-world-1-0", "name": "Genesis World 1.0", "developer": "Genesis AI", "country": "US", "country_label": "US", "released": "2026-05", "type": "simulation platform", "role": "simulator", "role_detail": "evaluation / data generation / policy pretraining", "weights": "open", "summary": "Open-source robotics simulator with its own renderer (Nyx) and GPU compute layer (Quadrants).", "caveat": "Company claims one hour of real-world testing equals 100 days in simulation; not independently checked.", "status": "snippet-only", "sources": [ { "url": "https://app.dealroom.co/news/note/genesis-ai-open-sources-world-1-0-simulation-platform-for-robotics", "captured_at": "2026-10-26", "quote": "Genesis AI open-sources Genesis World 1.0 simulation platform for robotics" } ] }, { "id": "general-intuition", "name": "General Intuition (MIRA)", "developer": "General Intuition", "country": "US", "country_label": "US", "released": "2025-10-16", "type": "interactive/generative environment", "role": "policy", "role_detail": "policy pretraining", "weights": "closed", "summary": "Trains agents on gameplay video (2 billion clips a year from Medal); says the same model adapted to a quadruped with 8 minutes of real robot data.", "caveat": "Robotics is a secondary application; the robot claim is company-reported.", "status": "snippet-only", "sources": [ { "url": "https://techcrunch.com/2025/10/16/general-intuition-lands-134m-seed-to-teach-agents-spatial-reasoning-using-video-game-clips", "captured_at": "2026-10-26", "quote": "General Intuition lands $134M seed to teach agents spatial reasoning using video game clips" } ] }, { "id": "oasis-3", "name": "Oasis 3", "developer": "Decart", "country": "US", "country_label": "US", "released": "2026-06", "type": "interactive/generative environment", "role": "simulator", "role_detail": "evaluation / data generation / policy pretraining", "weights": "api", "summary": "API-accessible interactive world model that generates action-responsive simulation for robot learning, starting with autonomous vehicles.", "caveat": "Latency and speed claims are vendor-reported.", "status": "snippet-only", "sources": [ { "url": "https://wandb.ai/site/customers/decart/", "captured_at": "2026-10-26", "quote": "How Decart built Oasis 3, the first API-accessible world model for faster, more effective, and scalable robotic learning" } ] }, { "id": "ge-act-2-0", "name": "GE-Act 2.0", "developer": "AgiBot (Zhiyuan)", "country": "China", "country_label": "China", "released": "2026-09-04", "type": "world-action model", "role": "policy", "role_detail": "direct control / policy pretraining", "weights": "closed", "summary": "Scaling co-training data from 300 to 30,000 hours raised zero-shot success from 17.1% to 44.1% on one AgiBot platform and 13.4% to 31.1% on another, across 100 tasks.", "caveat": "Developer's own paper; weight release not confirmed.", "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2609.05588", "captured_at": "2026-10-26", "quote": "Scaling co-training data from 300 to 30,000 hours raises success from 17.1% to 44.1% on G1-OP" } ] }, { "id": "dreamdojo", "name": "DreamDojo", "developer": "NVIDIA", "country": "US", "country_label": "US", "released": "2026-02-06", "type": "video world model", "role": "simulator", "role_detail": "data generation / policy evaluation / planning", "weights": "open", "summary": "World model that learns dexterous controls from 44,000 hours of egocentric human video, distilled to run in real time (10.81 frames per second).", "caveat": "The abstract names no benchmark scores.", "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2602.06949", "captured_at": "2026-10-26", "quote": "a foundation world model that learns diverse interactions and dexterous controls from 44k hours of egocentric human videos" } ] }, { "id": "hy-world-2-0", "name": "HY-World 2.0", "developer": "Tencent Hunyuan", "country": "China", "country_label": "China", "released": "2026-04-15", "type": "interactive/generative environment", "role": "simulator", "role_detail": "simulation assets for planning / evaluation", "weights": "open", "summary": "Open multimodal 3D world model that turns text, images or video into persistent 3D scenes exported as meshes, splats and point clouds for simulation pipelines.", "caveat": "Tencent compares it to World Labs' Marble; vendor claim. Robotics is one use case, not a tested benchmark.", "status": "snippet-only", "sources": [ { "url": "https://rits.shanghai.nyu.edu/ai/tencent-open-sources-hy-world-2-0-a-multi-modal-3d-world-model/", "captured_at": "2026-10-26", "quote": "Tencent Open-Sources HY-World 2.0: A Multi-Modal 3D World Model" } ] }, { "id": "genie-envisioner-2-0", "name": "Genie Envisioner 2.0", "developer": "AgiBot (Zhiyuan)", "country": "China", "country_label": "China", "released": "2026-04", "type": "world-action model", "role": "simulator", "role_detail": "planning / evaluation / data generation", "weights": "closed", "summary": "Closed-loop robot video world simulator for training and evaluating robots.", "caveat": "A reprint reports first place in a CVPR 2026 WorldArena track; not confirmed against the official board.", "status": "snippet-only", "sources": [ { "url": "https://gitcode.csdn.net/6a2964dd10ee7a33f27a9022.html", "captured_at": "2026-10-26", "quote": "在近期公布的CVPR 2026 WorldArena世界模型赛道评测中,智元(Agibot)自研的Genie Envisioner 2.0(GE 2.0)获得总分第一。" } ] }, { "id": "gaia-3", "name": "GAIA-3", "developer": "Wayve", "country": "Other", "country_label": "UK", "released": "2025-12-02", "type": "driving world model", "role": "simulator", "role_detail": "evaluation (closed-loop AV validation; robotics-adjacent)", "weights": "closed", "summary": "15B-parameter generative world model for autonomous-driving validation: generates and scores counterfactual scenarios offline.", "caveat": "Not a robot model; included as an adjacent evaluation world model.", "status": "snippet-only", "sources": [ { "url": "https://www.automotiveworld.com/news/wayve-unveils-gaia-3-for-autonomous-driving-tests/", "captured_at": "2026-10-26", "quote": "wayve unveils gaia 3 for autonomous driving tests" } ] }, { "id": "qwen-robotworld", "name": "Qwen-RobotWorld", "developer": "Alibaba Tongyi Lab (Qwen)", "country": "China", "country_label": "China", "released": "2026-06", "type": "video world model", "role": "simulator", "role_detail": "data generation / planning / policy evaluation", "weights": "closed", "summary": "Language-driven video world model that forecasts scene changes after an action; trained on 8.6 million video-text pairs across 20+ robot platforms (company figures).", "caveat": "Enterprise pilot only; not broadly available.", "status": "snippet-only", "sources": [ { "url": "https://forklog.com/en/news/alibaba-unveils-qwen-robot-suite-for-navigation-manipulation-and-world-modeling", "captured_at": "2026-10-26", "quote": "Alibaba unveils Qwen robot suite for navigation, manipulation and world modeling" } ] }, { "id": "marble", "name": "Marble", "developer": "World Labs", "country": "US", "country_label": "US", "released": "2025-11", "type": "interactive/generative environment", "role": "simulator", "role_detail": "simulation / data generation (3D environments)", "weights": "api", "summary": "Builds 3D environments from images, video or text; World Labs' first commercial product. AMD agreed to acquire World Labs (reported at about $8.2B).", "caveat": "No robotics benchmark found; deal details from secondary reports.", "status": "snippet-only", "sources": [ { "url": "https://next.ink/258807/amd-rachete-world-labs-et-ses-modeles-de-monde-pour-82-milliards-de-dollars/", "captured_at": "2026-10-26", "quote": "AMD rachète World Labs et ses « modèles de monde » pour 8,2 milliards de dollars" } ] }, { "id": "wall-b", "name": "WALL-B", "developer": "X Square Robot", "country": "China", "country_label": "China", "released": "2026-04-21", "type": "world-action model", "role": "policy", "role_detail": "policy pretraining / direct control", "weights": "partly open", "summary": "Embodied model that trains vision, language, action, touch and physical prediction in one network. Company demo at WRC 2026: 1,816 parcels per hour at over 98% accuracy.", "caveat": "Company-reported demo; WALL-B weight status not confirmed.", "status": "snippet-only", "sources": [ { "url": "https://macaubusiness.com/x-square-robot-shows-the-full-embodied-ai-loop-at-wrc-2026-from-foundation-model-to-real-world-deployment/", "captured_at": "2026-10-26", "quote": "In April 2026, the company introduced WALL-B, an embodied AI foundation model built on its World Unified Model architecture." } ] }, { "id": "dreamzero", "name": "DreamZero", "developer": "NVIDIA GEAR", "country": "US", "country_label": "US", "released": "2026-02-17", "type": "world-action model", "role": "policy", "role_detail": "direct control", "weights": "open", "summary": "14B world-action model on a video-diffusion backbone; NVIDIA reports over 2x better generalisation to new tasks and environments than leading VLAs and about 7 Hz closed-loop control. Basis for the GR00T N2 preview.", "caveat": "Generalisation figure is the developer's own. Listed as open source on RoboArena, where it sits second (1735).", "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2602.15922", "captured_at": "2026-10-26", "quote": "over 2x improvement in generalization to new tasks and environments compared to state-of-the-art VLAs in real robot experiments" } ] }, { "id": "lingbot-va-2-0", "name": "LingBot-VA 2.0", "developer": "Robbyant (Ant Group)", "country": "China", "country_label": "China", "released": "2026-07-09", "type": "world-action model", "role": "policy", "role_detail": "direct control", "weights": "unknown", "summary": "Billed as the first video-action world model pretrained from scratch for embodiment rather than adapted from a video generator. Version 1 reported 92.0% on RoboTwin 2.0 and 98.5% on LIBERO.", "caveat": "Scores are the developer's own (v1 paper, arXiv 2601.21998).", "status": "verified", "sources": [ { "url": "https://www.webull.com/news/15201670514402304", "captured_at": "2026-10-26", "quote": "the industry's first embodied-native video-action world model" } ] } ], "benchmarks": [ { "id": "erqa", "name": "ERQA (Embodied Reasoning QA)", "maintainer": "Google DeepMind", "country": "US", "country_label": "US", "introduced": "2025-03", "kind": "Embodied reasoning", "measures": "400 multiple-choice visual questions on spatial, trajectory, action, state-estimation, pointing, multi-view and task reasoning; scored as accuracy.", "metric": "accuracy (%)", "evaluator": "vendor-run", "as_of": "2026-10-13", "leaders": [ { "model": "Gemini 4 Argon", "developer": "Google DeepMind", "country": "US", "country_label": "US", "value": "87%", "as_of": "2026-10-13", "vendor_reported": true } ], "caveat": "The only published top score is Google's own model on Google's own benchmark. Other models' scores appear only in an image chart, so no comparison is shown. There is no official leaderboard.", "status": "verified", "sources": [ { "url": "https://developers.googleblog.com/gemini-4-argon-for-robotics/", "captured_at": "2026-10-26", "quote": "You can see this leap in capability on the Embodied Reasoning (ERQA) benchmark, which tests a model's ability to understand and solve physical, spatial problems in 3D environments. Using pure reasoning, Argon scores 87%, outperforming every other frontier model." }, { "url": "https://arxiv.org/abs/2503.20020", "captured_at": "2026-10-26" }, { "url": "https://arxiv.org/html/2503.20020", "captured_at": "2026-10-26" }, { "url": "https://github.com/embodiedreasoning/ERQA", "captured_at": "2026-10-26" } ] }, { "id": "roboarena", "name": "RoboArena", "maintainer": "Atreya et al. (distributed evaluator network)", "country": "US", "country_label": "US", "introduced": "2025-06", "kind": "Real robot", "measures": "Crowd-sourced, double-blind pairwise real-robot comparisons of generalist policies on the DROID platform; evaluators choose their own tasks and scenes; policies get an Elo-style score.", "metric": "Elo-style score from pairwise preferences (std and num_evals reported)", "evaluator": "third-party", "as_of": "2026-10-26", "leaders": [ { "model": "Spirit v1.6 (listed as j2-vla)", "developer": "Spirit AI (inferred from display name)", "country": "China", "country_label": "China", "value": "1788 (±106, 25 evals)", "as_of": "2026-10-26", "vendor_reported": false }, { "model": "DreamZero", "developer": "NVIDIA", "country": "US", "country_label": "US", "value": "1735 (±43, 190 evals)", "as_of": "2026-10-26", "vendor_reported": false }, { "model": "Apricot", "developer": null, "country": "Unverified", "country_label": "Unverified", "value": "1661 (±61, 63 evals)", "as_of": "2026-10-26", "vendor_reported": false }, { "model": "pi0.5-DROID", "developer": "Physical Intelligence", "country": "US", "country_label": "US", "value": "1608 (±31, 745 evals)", "as_of": "2026-10-26", "vendor_reported": false } ], "caveat": "Double-blind pairwise real-robot comparisons. The top two are not statistically separated (wide uncertainty on few evaluations). The developer of the top entry is inferred from its display name. Elo-style scores are not comparable with success rates.", "status": "verified", "sources": [ { "url": "https://robo-arena.github.io/", "captured_at": "2026-10-26", "quote": "{\"num_evals\":25,\"open_source\":true,\"policy\":\"j2-vla\",\"score\":1788,...,\"std\":106.0} (site JS: j2-vla:\"Spirit v1.6\")" }, { "url": "https://arxiv.org/abs/2506.18123", "captured_at": "2026-10-26" }, { "url": "https://arxiv.org/abs/2602.15922", "captured_at": "2026-10-26" } ] }, { "id": "robodojo-live", "name": "RoboDojo (live leaderboard)", "maintainer": "AI MMLab Club (non-profit) with academic partners", "country": "Unverified", "country_label": "Unverified", "introduced": "2026-07", "kind": "Sim and real", "measures": "Unified sim-and-real benchmark for generalist manipulation policies: 42 simulation tasks across five capability dimensions (generalization, memory, precision, long-horizon, open-vocabulary instruction following) and 18 real-world tasks; each leaderboard cell is score / success rate.", "metric": "score / success rate (%) averaged over the five dimensions", "evaluator": "self-submitted", "as_of": "2026-10-26", "leaders": [ { "model": "GO-3 Nano Preview (agent track)", "developer": "AgiBot", "country": "China", "country_label": "China", "value": "61.90 / 56.42%", "as_of": "2026-10-26", "vendor_reported": true }, { "model": "PyRUA-Lean (agent track)", "developer": "PKU DAGroup", "country": "China", "country_label": "China", "value": "55.26 / 50.35%", "as_of": "2026-10-24", "vendor_reported": true }, { "model": "HexaShell-Preview (agent track)", "developer": "HexaFuture", "country": "China", "country_label": "China", "value": "53.18 / 46.85%", "as_of": "2026-10-20", "vendor_reported": true }, { "model": "VPP2 (best non-agent entry)", "developer": "Robotera", "country": "China", "country_label": "China", "value": "39.26 / 32.26%", "as_of": "2026-10-08", "vendor_reported": true }, { "model": "GPT-6-Astra (agent track)", "developer": "OpenAI (run by the RoboDojo team)", "country": "US", "country_label": "US", "value": "28.97 / 22.48%", "as_of": "2026-09-16", "vendor_reported": false }, { "model": "InternW0-Delta (real-world board)", "developer": "Shanghai AI Lab", "country": "China", "country_label": "China", "value": "38.9 / 31.7%", "as_of": "2026-10-21", "vendor_reported": true } ], "caveat": "Score / success rate, simulation average unless marked real-world. Entries are contributed by their developers. Agent-track entries (agent plus world-action-model stacks) are not like-for-like with single-policy VLAs. Entrants are overwhelmingly Chinese labs. Superseded the paper table below within four months.", "status": "verified", "sources": [ { "url": "https://robodojo-benchmark.com/", "captured_at": "2026-10-26", "quote": "Add GO-3 Nano Preview (61.90 / 56.42%, contributed by AGIBOT) to the sim leaderboard (agent track)." }, { "url": "https://arxiv.org/abs/2607.04434", "captured_at": "2026-10-26" }, { "url": "https://arxiv.org/html/2607.04434", "captured_at": "2026-10-26" } ] }, { "id": "robodojo-paper", "name": "RoboDojo (paper table)", "maintainer": "Tianxing Chen et al.", "country": "Unverified", "country_label": "Unverified", "introduced": "2026-07", "kind": "Simulation", "measures": "30 policies evaluated through XPolicyLab on the simulation suite; 3 seeds x 50 trials per task; mean score / success rate across five capability dimensions.", "metric": "score / success rate (%)", "evaluator": "authors", "as_of": "2026-07-03", "leaders": [ { "model": "Hy-Embodied-0.5-VLA", "developer": "Tencent Hunyuan", "country": "China", "country_label": "China", "value": "13.07 / 8.80%", "as_of": "2026-07-03", "vendor_reported": false }, { "model": "Spatial Forcing", "developer": "Academic (Li et al.)", "country": "Unverified", "country_label": "Unverified", "value": "12.38 / 8.04%", "as_of": "2026-07-03", "vendor_reported": false }, { "model": "pi0.5", "developer": "Physical Intelligence", "country": "US", "country_label": "US", "value": "11.41 / 6.91%", "as_of": "2026-07-03", "vendor_reported": false } ], "caveat": "Frozen table from the benchmark paper: 30 policies, 3 seeds x 50 trials per task, all run by the authors. Shown separately from the live board because the two are not the same test conditions. A human teleoperator scores about 88% on the first dimension.", "status": "verified", "sources": [ { "url": "https://arxiv.org/html/2607.04434", "captured_at": "2026-10-26", "quote": "Hy-Embodied-0.5-VLA (Zhang et al., 2026a) 11.77 / 8.39% 13.81 / 8.00% 25.74 / 14.92% 13.37 / 12.11% 0.65 / 0.58% 13.07 / 8.80%" } ] }, { "id": "libero-pro", "name": "LIBERO-PRO", "maintainer": "Zhou et al., Huazhong University of Science and Technology", "country": "China", "country_label": "China", "introduced": "2025-10", "kind": "Simulation", "measures": "LIBERO extended with perturbations along four dimensions (manipulated objects, initial states, task instructions, environments) to separate generalization from memorization; success rate.", "metric": "success rate", "evaluator": "authors", "as_of": "2025-10", "leaders": [ { "model": "pi0.5 (best of four tested)", "developer": "Physical Intelligence", "country": "US", "country_label": "US", "value": "0.97 original; 0.0 under task perturbation; 0.4 under position perturbation", "as_of": "2025-10", "vendor_reported": false } ], "caveat": "A stress test, not a leaderboard. Every model tested scores 0.0 when the task instruction is perturbed, but pi0.5 keeps 0.4 / 0.2 / 0.1 / 0.2 across the four suites when object positions are perturbed, so \"collapses to zero\" is not universal. Scores are 0-1 fractions; the unit is not stated.", "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2510.03827", "captured_at": "2026-10-26", "quote": "although existing models achieve over 90% accuracy under the standard LIBERO evaluation, their performance collapses to 0.0% under our generalized setting." }, { "url": "https://arxiv.org/html/2510.03827", "captured_at": "2026-10-26" }, { "url": "https://github.com/Zxy-MLlab/LIBERO-PRO", "captured_at": "2026-10-26" } ] }, { "id": "libero-plus", "name": "LIBERO-Plus", "maintainer": "Fei et al.", "country": "China", "country_label": "China", "introduced": "2025-10", "kind": "Simulation", "measures": "Robustness of VLA models on LIBERO under controlled perturbations across seven dimensions: objects layout, camera viewpoints, robot initial states, language instructions, light conditions, background textures and sensor noise.", "metric": "success rate (%) per perturbation dimension and Total", "evaluator": "authors", "as_of": "2025-10", "leaders": [ { "model": "OpenVLA-OFT+ (authors' own fine-tune)", "developer": "LIBERO-Plus authors", "country": "China", "country_label": "China", "value": "79.6 total", "as_of": "2025-10", "vendor_reported": true }, { "model": "OpenVLA-OFT", "developer": "Stanford", "country": "US", "country_label": "US", "value": "69.6 total", "as_of": "2025-10", "vendor_reported": false }, { "model": "pi0-FAST", "developer": "Physical Intelligence", "country": "US", "country_label": "US", "value": "61.6 total", "as_of": "2025-10", "vendor_reported": false } ], "caveat": "The top score is the authors' own model trained on LIBERO-Plus data; plain OpenVLA-OFT scores 69.6. Drops \"below 30%\" happen for some model and perturbation pairs, not all. The authors find models tend to ignore the language instruction.", "status": "verified", "sources": [ { "url": "https://github.com/sylvestf/LIBERO-plus", "captured_at": "2026-10-26", "quote": "models exhibit extreme sensitivity to perturbation factors, including camera viewpoints and robot initial states, with performance dropping from 95% to below 30% under modest perturbations." }, { "url": "https://arxiv.org/abs/2510.13626", "captured_at": "2026-10-26" }, { "url": "https://arxiv.org/html/2510.13626", "captured_at": "2026-10-26" } ] }, { "id": "simplerenv-audit", "name": "SimplerEnv and the 2026 validity audit", "maintainer": "SimplerEnv: Li, Su, Xiao et al.; audit: Jiang et al. (TTIC, Argonne, UChicago)", "country": "US", "country_label": "US", "introduced": "2024-05", "kind": "Simulation", "measures": "SimplerEnv: simulated real-to-sim evaluation of manipulation policies on Google Robot and WidowX+Bridge setups, validated by sim-vs-real correlation. Audit (arXiv 2606.04233, 2026-06): four diagnostics (shortcut solvability, statistical significance, creeping overfitting, data-source dependence) applied to LIBERO, CALVIN, SimplerEnv, RoboCasa, RoboTwin 2.0.", "metric": "success rate; audit reports share of SOTA claims that are provably significant", "evaluator": "authors", "as_of": "2026-06", "leaders": [], "caveat": "No SimplerEnv leaderboard exists. The audit found that only 19.8% (LIBERO) and 19.7% (SimplerEnv) of published state-of-the-art claims are provably significant, against 53.3% (RoboCasa) and 73.7% (RoboTwin 2.0). A tiny probe with no language encoder scores near reported state of the art on LIBERO.", "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2606.04233", "captured_at": "2026-10-26", "quote": "On LIBERO and SimplerEnv, only 19.8% and 19.7% of SOTA claims, respectively, are provably significant." }, { "url": "https://arxiv.org/html/2606.04233", "captured_at": "2026-10-26" }, { "url": "https://arxiv.org/abs/2405.05941", "captured_at": "2026-10-26" }, { "url": "https://github.com/simpler-env/SimplerEnv", "captured_at": "2026-10-26" } ] }, { "id": "behavior-1k", "name": "BEHAVIOR-1K Challenge (2025)", "maintainer": "Stanford Vision and Learning Lab", "country": "US", "country_label": "US", "introduced": "2025-10", "kind": "Simulation", "measures": "Full-length household mobile-manipulation tasks in BEHAVIOR-1K/OmniGibson; ranked by Q-score, the fraction of a task's goal conditions satisfied with partial credit; also reports full-task success rate. 2026 edition: 100 tasks, 7 scenes (4 new), 20,000 teleop demos (1,950 hours), baselines pi0.5 and GR00T N1.7.", "metric": "held-out Q-score (primary); full task success rate", "evaluator": "third-party", "as_of": "2025-11-17", "leaders": [ { "model": "Robot Learning Collective", "developer": "Independent", "country": "Unverified", "country_label": "Unverified", "value": "Q 0.2599; full-task success 0.1240", "as_of": "2025-11-14", "vendor_reported": false }, { "model": "Comet", "developer": "NVIDIA Research", "country": "US", "country_label": "US", "value": "Q 0.2514; full-task success 0.1140", "as_of": "2025-11-17", "vendor_reported": false }, { "model": "SimpleAI Robot", "developer": "Beijing Simple AI Technology", "country": "China", "country_label": "China", "value": "Q 0.1591; full-task success 0.1080", "as_of": "2025-11-17", "vendor_reported": false }, { "model": "The North Star", "developer": "Huawei", "country": "China", "country_label": "China", "value": "Q 0.1204; full-task success 0.0760", "as_of": "2025-11-16", "vendor_reported": false } ], "caveat": "Ranked by held-out Q-score, which gives partial credit for goal conditions met; full-task success is much lower. Held-out entries are verified by the BEHAVIOR team. A 2026 edition closed submissions on 2026-10-16; its results were not available.", "status": "verified", "sources": [ { "url": "https://behavior.stanford.edu/challenge/archive/2025/leaderboard.html", "captured_at": "2026-10-26", "quote": "1Robot Learning CollectiveIndependentStandardcodereport0.11200.12400.26050.259920251114" }, { "url": "https://behavior.stanford.edu/challenge/index.html", "captured_at": "2026-10-26" }, { "url": "https://huggingface.co/spaces/behavior-1k/2026-challenge-leaderboard", "captured_at": "2026-10-26" } ] }, { "id": "x2real", "name": "X2Real", "maintainer": "X Square Robot", "country": "China", "country_label": "China", "introduced": "2026-09", "kind": "Sim and real", "measures": "Evolvable simulation benchmark on NVIDIA Isaac Lab-Arena for faithfully evaluating real-world performance of manipulation policies: 44 hierarchical long-horizon tasks across 10 capability dimensions, rule-based DAG subtask scoring (success rate and progress score), domain randomization and disjoint train/eval pipelines.", "metric": "success rate (%) and progress score (0-100), hierarchically macro-averaged", "evaluator": "vendor-run", "as_of": "2026-10-13", "leaders": [ { "model": "wall-oss-preview", "developer": "X Square Robot", "country": "China", "country_label": "China", "value": "53.9% in-distribution; 34.3% out-of-distribution", "as_of": "2026-10-13", "vendor_reported": true }, { "model": "DreamZero", "developer": "NVIDIA", "country": "US", "country_label": "US", "value": "4.85% in-distribution", "as_of": "2026-10-13", "vendor_reported": false } ], "caveat": "Run by the vendor whose own model tops it; all four policies were post-trained by the authors on the same demonstrations from the vendor's arm. The headline 0.84 sim-to-real correlation is for progress score; for success rate it is 0.74.", "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2609.27449", "captured_at": "2026-10-26", "quote": "Taking all tasks into account, our simulation environment achieves 0.74 linear correlation coefficient in success rate and 0.84 in progress score." }, { "url": "https://arxiv.org/html/2609.27449", "captured_at": "2026-10-26" } ] }, { "id": "robocurve-rc", "name": "RoboDojo-RC Tier 1 (Robocurve)", "maintainer": "Robocurve PBC", "country": "US", "country_label": "US", "introduced": "2026-09", "kind": "Real robot", "measures": "Real-world scores of frontier LLM agents and VLAs on bimanual tasks adapted from RoboDojo, 120 task attempts per configuration, caps of 40 model calls for LLMs and 3,600 control steps for VLAs; reports mean score, cost per attempt and time per attempt.", "metric": "mean score (%) over 120 attempts", "evaluator": "third-party", "as_of": "2026-10-26", "leaders": [ { "model": "Dream-1", "developer": "Dream Labs", "country": "Unverified", "country_label": "Unverified", "value": "59.3%", "as_of": "2026-10-19", "vendor_reported": false }, { "model": "Gemini 4 Argon", "developer": "Google DeepMind", "country": "US", "country_label": "US", "value": "48.2%", "as_of": "2026-10-13", "vendor_reported": false }, { "model": "GPT-6 Astra", "developer": "OpenAI", "country": "US", "country_label": "US", "value": "42.5%", "as_of": "2026-09-03", "vendor_reported": false }, { "model": "GPT-6.1 Sol", "developer": "OpenAI", "country": "US", "country_label": "US", "value": "41.7%", "as_of": "2026-09-29", "vendor_reported": false }, { "model": "Claude Opus 5.5", "developer": "Anthropic", "country": "US", "country_label": "US", "value": "38.2%", "as_of": "2026-09-22", "vendor_reported": false }, { "model": "MolmoAct2", "developer": "Ai2", "country": "US", "country_label": "US", "value": "18.9%", "as_of": "2026-05-04", "vendor_reported": false }, { "model": "GR00T N1.7", "developer": "NVIDIA", "country": "US", "country_label": "US", "value": "9.5%", "as_of": "2026-04-17", "vendor_reported": false } ], "caveat": "Independent evaluator running models on its own bimanual arms, 120 attempts per configuration, so single-digit gaps are not significant. No Chinese model is on the board, so it cannot support a US-China comparison. Dates are model release dates.", "status": "verified", "sources": [ { "url": "https://robocurve.org/", "captured_at": "2026-10-26", "quote": "Dream-1 59.3% $0.15 per attempt (hosted inference)" }, { "url": "https://robocurve.org/all-runs/", "captured_at": "2026-10-26" }, { "url": "https://robocurve.org/rci/", "captured_at": "2026-10-26" }, { "url": "https://robocurve.org/blog/seed-raise/", "captured_at": "2026-10-26" } ] }, { "id": "roboharm", "name": "RoboHarm v1", "maintainer": "Robocurve PBC", "country": "US", "country_label": "US", "introduced": "2026-09", "kind": "Safety", "measures": "Whether frontier robot policies refuse unsafe instructions: five hazardous tasks (stab a baby doll, heat a can of compressed air, put a screwdriver in a toaster, drop a power bank in water, mix bleach and ammonia), one fixed wording each, 20 trials per policy per instruction, 300 trials total; human reviewers label each trial into five outcomes.", "metric": "share of trials refused (safety) vs completed", "evaluator": "third-party", "as_of": "2026-09-18", "leaders": [ { "model": "Claude Fable 5.1", "developer": "Anthropic", "country": "US", "country_label": "US", "value": "Refused 20/100; carried out 34/100", "as_of": "2026-09-18", "vendor_reported": false }, { "model": "GPT-6 Astra", "developer": "OpenAI", "country": "US", "country_label": "US", "value": "Refused 2/100 (3 for any reason); carried out 60/100", "as_of": "2026-09-18", "vendor_reported": false }, { "model": "MolmoAct2", "developer": "Ai2", "country": "US", "country_label": "US", "value": "Refused 0/100; carried out 6/100", "as_of": "2026-09-18", "vendor_reported": false } ], "caveat": "Measures the share of hazardous instructions refused versus carried out: five tasks, 20 trials each. A widely quoted \"97%\" is the share GPT-6 Astra attempted, not completed. MolmoAct2's low completion reflects capability, not safety. The live board has since moved. No Chinese model has been tested.", "status": "verified", "sources": [ { "url": "https://robocurve.org/roboharm/", "captured_at": "2026-10-26", "quote": "Pooled trials by outcome, safest on the left. Fable refused 20 of 100, Astra 2, MolmoAct2 none." }, { "url": "https://robocurve.org/", "captured_at": "2026-10-26" }, { "url": "https://www.analyticsinsight.net/news/gpt-6-astra-faces-physical-ai-safety-test-as-model-attempts-97-hazardous-instructions", "captured_at": "2026-10-26" } ] }, { "id": "aa-robotics", "name": "AA-Robotics v0.1", "maintainer": "Artificial Analysis", "country": "US", "country_label": "US", "introduced": "2026-10", "kind": "Real robot", "measures": "Frontier language models in coding-agent harnesses directly control a 6-axis robot arm (Hiwonder NexArm) via an HTTP API with four cameras; eight tabletop tasks; a human judge scores the end state out of 100, partial credit allowed; zero-shot.", "metric": "mean score per task (0-100)", "evaluator": "third-party", "as_of": "2026-10-14", "leaders": [ { "model": "GPT-6 Astra (Codex harness)", "developer": "OpenAI", "country": "US", "country_label": "US", "value": "69.4", "as_of": "2026-10-14", "vendor_reported": false }, { "model": "GPT-6.1 Sol", "developer": "OpenAI", "country": "US", "country_label": "US", "value": "64.6", "as_of": "2026-10-14", "vendor_reported": false }, { "model": "Gemini 4 Argon", "developer": "Google DeepMind", "country": "US", "country_label": "US", "value": "55.3", "as_of": "2026-10-14", "vendor_reported": false }, { "model": "Claude Opus 5.5", "developer": "Anthropic", "country": "US", "country_label": "US", "value": "48.8", "as_of": "2026-10-14", "vendor_reported": false }, { "model": "Claude Fable 5.1", "developer": "Anthropic", "country": "US", "country_label": "US", "value": "46.3", "as_of": "2026-10-14", "vendor_reported": false } ], "caveat": "Research preview: language models in coding-agent harnesses drive a 6-axis arm; a human judge scores eight tasks out of 100. One trial per cell. Only US models tested. Models take about half an hour per task against a human's 20 seconds.", "status": "verified", "sources": [ { "url": "https://artificialanalysis.ai/articles/aa-robotics", "captured_at": "2026-10-26", "quote": "GPT-6 Astra places first in AA-Robotics v0.1 with a score of 69, followed by GPT-6.1 Sol, Gemini 4 Argon, Claude Opus 5.5 and Claude Fable 5.1." } ] }, { "id": "worldarena-2", "name": "WorldArena 2.0", "maintainer": "Tsinghua University-led team", "country": "China", "country_label": "China", "introduced": "2026-05", "kind": "World-model eval", "measures": "Embodied world-model benchmark extended along modality (vision to visuotactile), functionality (policy evaluation and planning to interactive RL environments for policy optimization) and platform (simulators to real robots, multiple embodiments).", "metric": "perceptual-quality metrics (16 metrics, 6 dimensions) and downstream task success when training policies inside the world model", "evaluator": "authors", "as_of": "2026-10-26", "leaders": [ { "model": "WoVR (Click Bell task)", "developer": "Jiang et al.", "country": "Unverified", "country_label": "Unverified", "value": "75.0% vs 43.75% baseline", "as_of": "2026-10-26", "vendor_reported": false }, { "model": "Ctrl-World (Adjust Bottle task)", "developer": "Guo et al.", "country": "Unverified", "country_label": "Unverified", "value": "70.7% vs 55.1% baseline", "as_of": "2026-10-26", "vendor_reported": false }, { "model": "Cosmos-Predict-2.5 (Adjust Bottle task)", "developer": "NVIDIA", "country": "US", "country_label": "US", "value": "63.5%", "as_of": "2026-10-26", "vendor_reported": false } ], "caveat": "Tests world models as places to train policies. There is no composite score; leaders differ by task. The authors note a gap remains versus training in the native simulator.", "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2605.17912", "captured_at": "2026-10-26", "quote": "WoVR (Jiang et al., 2026) leads the learned environments on Click Bell, whereas Ctrl-World (Guo et al., 2025) leads on Adjust Bottle, reaching 75.00% and 70.70% under proxy rewards compared to 43.75% and 55.08% for SFT." }, { "url": "https://arxiv.org/html/2605.17912", "captured_at": "2026-10-26" } ] }, { "id": "nist-humanoid-baseline", "name": "NIST Humanoid Robot Baseline Performance Benchmark", "maintainer": "NIST Intelligent Systems Division", "country": "US", "country_label": "US", "introduced": "2026-04", "kind": "Standard (proposed)", "measures": "Proposed low-footprint set of locomotion and manipulation tasks derived mostly from previously standardized NIST test methods, for the minimum capabilities expected of commercial humanoids: basic mobility and manipulation/dexterity, coordinated loco-manipulation, whole-body awareness and control in confined-space manipulation, and minimal reasoning, task and scene understanding.", "metric": "quantifiable performance metrics from NIST test methods (not specified on page); results aggregated under data sharing agreements", "evaluator": "none", "as_of": "2026-05-15", "leaders": [], "caveat": "A proposal seeking participants: no results, leaderboard or timeline yet. NIST says the DARPA Robotics Challenge (2013-2014) was the last rigorous head-to-head humanoid measurement.", "status": "verified", "sources": [ { "url": "https://www.nist.gov/el/intelligent-systems-division-73500/humanoid-robot-baseline-performance-benchmark", "captured_at": "2026-10-26", "quote": "Test results will be collected under pre-approved data sharing agreements between NIST and the robot manufacturers in order to protect IP and any attribution if needed." } ] }, { "id": "world-humanoid-robot-games-2026", "name": "2nd World Humanoid Robot Games (Beijing)", "maintainer": "Beijing organizers", "country": "China", "country_label": "China", "introduced": "2026-08", "kind": "Competition", "measures": "Track-and-field races (100m, 400m, 1,500m), jumps, plus competitions across nine real-world application scenarios (household tasks, industrial operations, emergency rescue, e.g. a 'Smart Charging Service' challenge); 666 teams from 16 countries and more than 2,000 robots.", "metric": "race time; scenario task completion", "evaluator": "organizer", "as_of": "2026-08-26", "leaders": [ { "model": "Tianzhuo Team (1,500 m)", "developer": "Tianzhuo Team", "country": "China", "country_label": "China", "value": "2:21.64", "as_of": "2026-08-23", "vendor_reported": false }, { "model": "Tianjiao Team (100 m, large group)", "developer": "Tianjiao Team", "country": "China", "country_label": "China", "value": "8.64 s", "as_of": "2026-08", "vendor_reported": false }, { "model": "Tiangong Omni (400 m, small group)", "developer": "Beijing Innovation Center of Humanoid Robotics", "country": "China", "country_label": "China", "value": "45.66 s", "as_of": "2026-08", "vendor_reported": false } ], "caveat": "Source is a Beijing government release. 666 teams from 16 countries took part; the winners listed are Chinese teams. The release compares the 1,500 m time to a human record it does not name.", "status": "verified", "sources": [ { "url": "https://english.beijing.gov.cn/beijinginfo/sci/latesttrends/202608/t20260828_4840947.html", "captured_at": "2026-10-26", "quote": "In the 1,500m event, Tianzhuo Team claimed the title with a time of 2 minutes 21.64 seconds, a significant improvement on the previous Games' winning time of 6 minutes 34.40 seconds and faster than the 3-minute-26-second human record." }, { "url": "https://english.beijing.gov.cn/beijinginfo/sci/latesttrends/202608/t20260825_4836357.html", "captured_at": "2026-10-26" } ] }, { "id": "etown-half-marathon-2026", "name": "Beijing E-Town Humanoid Half-Marathon", "maintainer": "Beijing E-Town organizers", "country": "China", "country_label": "China", "introduced": "2026-04", "kind": "Competition", "measures": "Humanoid robots racing the 21.0975 km half-marathon course alongside human runners (over 100 teams), more than ten terrain types, fully autonomous and remote-controlled classes ranked together but timed separately (remote times multiplied by 1.2).", "metric": "finish time", "evaluator": "organizer", "as_of": "2026-04-19", "leaders": [ { "model": "Shandian (Team Monkey King)", "developer": "Team Monkey King", "country": "China", "country_label": "China", "value": "50:26", "as_of": "2026-04-19", "vendor_reported": false } ], "caveat": "Secondary coverage attributes the winning robot to Honor; that attribution is unverified. The release does not say whether it ran autonomously or remote-controlled (remote times were multiplied by 1.2).", "status": "verified", "sources": [ { "url": "https://english.beijing.gov.cn/latest/news/202604/t20260419_4593508.html", "captured_at": "2026-10-26", "quote": "The robot \"Shandian\" (闪电) from Team Monkey King (齐天大圣队) clocked a time of 50 minutes and 26 seconds and was declared the winner." } ] }, { "id": "robotwin-2", "name": "RoboTwin 2.0", "maintainer": "RoboTwin team", "country": "China", "country_label": "China", "introduced": "2025-06", "kind": "Simulation", "measures": "Dual-arm manipulation, 50 tasks across five robot embodiments with structured domain randomization (clutter, lighting, background, tabletop height, language); leaderboard score = (clean-to-clean + clean-to-random)/2 average success, 100 trials per task on Aloha-AgileX.", "metric": "average success rate (%) of clean2clean and clean2random", "evaluator": "self-submitted", "as_of": "2026-10-25", "leaders": [ { "model": "GO-3 Nano Preview (co-train)", "developer": "AgiBot", "country": "China", "country_label": "China", "value": "85.01%", "as_of": "2026-10-25", "vendor_reported": true }, { "model": "PrismWAM", "developer": "Tsinghua & NVIDIA", "country": "China", "country_label": "China/US", "value": "83.67%", "as_of": "2026-10-23", "vendor_reported": true }, { "model": "CloudRobo-WM", "developer": "CloudRobo Team", "country": "Unverified", "country_label": "Unverified", "value": "81.95%", "as_of": "2026", "vendor_reported": true } ], "caveat": "Average of clean and randomized success over 50 dual-arm tasks. Entries are contributor-submitted; \"co-train\" entries train on both settings, so they are not comparable with clean-only training.", "status": "verified", "sources": [ { "url": "https://robotwin-platform.github.io/leaderboard", "captured_at": "2026-10-26", "quote": "Add GO-3 Nano Preview co-train results (88.54 / 81.48%)." }, { "url": "https://arxiv.org/abs/2506.18088", "captured_at": "2026-10-26" } ] }, { "id": "robocasa365", "name": "RoboCasa365", "maintainer": "Nasiriany et al. (UT Austin, NVIDIA)", "country": "US", "country_label": "US", "introduced": "2026-03", "kind": "Simulation", "measures": "Simulation benchmark for household mobile manipulation: 365 everyday tasks across 2,500 kitchen environments, 600+ hours of human demos and 1,600+ hours of synthetic demos; supports multi-task, foundation-model training and lifelong learning evaluations.", "metric": "task success rate", "evaluator": "none", "as_of": "2026-03", "leaders": [], "caveat": "No leader in the paper. Scores seen on aggregator sites were not verified. The 2026 audit found 53.3% of RoboCasa state-of-the-art claims provably significant, better than LIBERO or SimplerEnv.", "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2603.04356", "captured_at": "2026-10-26", "quote": "RoboCasa365 introduces 365 everyday tasks across 2,500 diverse kitchen environments, with over 600 hours of human demonstration data and over 1600 hours of synthetically generated demonstration data" } ] }, { "id": "embodiedbench", "name": "EmbodiedBench", "maintainer": "Yang et al.", "country": "Unverified", "country_label": "Unverified", "introduced": "2025-02", "kind": "Embodied reasoning", "measures": "1,128 testing tasks across four environments for vision-driven embodied agents (high-level household to low-level navigation/manipulation), six capability subsets; 24 MLLMs evaluated.", "metric": "average task success (%)", "evaluator": "authors", "as_of": "2025-06-05", "leaders": [ { "model": "GPT-4o", "developer": "OpenAI", "country": "US", "country_label": "US", "value": "28.9% average", "as_of": "2025-06-05", "vendor_reported": false } ], "caveat": "The paper's own 2025 result; the leader is a 2025 model and no 2026 board was found. Models do well at high-level planning and poorly at low-level manipulation.", "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2502.09560", "captured_at": "2026-10-26", "quote": "MLLMs excel at high-level tasks but struggle with low-level manipulation, with the best model, GPT-4o, scoring only 28.9% on average." } ] }, { "id": "vlabench", "name": "VLABench", "maintainer": "Zhang et al.", "country": "China", "country_label": "China", "introduced": "2024-12", "kind": "Simulation", "measures": "Language-conditioned manipulation: 100 task categories with strong randomization, 2,000+ objects, world-knowledge/common-sense transfer and long-horizon reasoning, evaluating both action policies and language-model capabilities.", "metric": "success / progress score", "evaluator": "none", "as_of": "2024-12", "leaders": [], "caveat": "No leader in the paper. Numbers on aggregator sites come from a different harness and are not shown.", "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2412.18194", "captured_at": "2026-10-26", "quote": "VLABench provides 100 carefully designed categories of tasks, with strong randomization in each category of task and a total of 2000+ objects." } ] }, { "id": "open-x-embodiment", "name": "Open X-Embodiment (RT-X)", "maintainer": "Open X-Embodiment collaboration (21 institutions)", "country": "Other", "country_label": "US-led international", "introduced": "2023-10", "kind": "Real robot", "measures": "Dataset of 22 robots, 527 skills (160,266 tasks) from 21 institutions, and RT-X models evaluated for positive transfer across robots; a dataset/standard rather than a ranked leaderboard.", "metric": "success rate vs single-robot baselines", "evaluator": "none", "as_of": "2023-10", "leaders": [], "caveat": "A dataset and cross-robot evaluation rather than a ranked board; included as the cross-embodiment data foundation.", "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2310.08864", "captured_at": "2026-10-26", "quote": "We assemble a dataset from 22 different robots collected through a collaboration between 21 institutions, demonstrating 527 skills (160266 tasks)." } ] }, { "id": "calvin", "name": "CALVIN", "maintainer": "Mees, Hermann, Rosete-Beas, Burgard (University of Freiburg)", "country": "Other", "country_label": "Germany", "introduced": "2021-12", "kind": "Simulation", "measures": "Open-source simulated benchmark for language-conditioned long-horizon manipulation: chains of language-specified tasks, zero-shot to novel instructions and environments; headline metric average tasks completed in a row (of 5).", "metric": "average tasks completed (of 5)", "evaluator": "none", "as_of": "2021-12", "leaders": [], "caveat": "Older and widely used. The 2026 audit found that moving blocks within the training range drops the average from 4.17 to 3.14 tasks completed.", "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2112.03227", "captured_at": "2026-10-26", "quote": "On CALVIN, resampling block poses within the training range lowers average tasks completed out of five (ATC) from 4.17 to 3.14." }, { "url": "https://arxiv.org/abs/2606.04233", "captured_at": "2026-10-26" } ] }, { "id": "maniskill3", "name": "ManiSkill3", "maintainer": "Hao Su lab (UCSD / Hillbot)", "country": "US", "country_label": "US", "introduced": "2024-10", "kind": "Simulation", "measures": "GPU-parallelized robotics simulator and task suite (12 domains incl. mobile manipulation, humanoids, dexterous hands); simulation with rendering 10-1000x faster than other platforms, 30,000+ FPS.", "metric": "simulation throughput; per-task success", "evaluator": "none", "as_of": "2024-10", "leaders": [], "caveat": "Primarily a fast GPU simulator and task library; no single capability leaderboard.", "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2410.00425", "captured_at": "2026-10-26", "quote": "We introduce and open source ManiSkill3, the fastest state-visual GPU parallelized robotics simulator with contact-rich physics targeting generalizable manipulation." } ] }, { "id": "agibot-world-challenge-2026", "name": "AgiBot World Challenge 2026 (ICRA)", "maintainer": "AgiBot (with Dexmal)", "country": "China", "country_label": "China", "introduced": "2026-06", "kind": "Competition", "measures": "Two tracks, Reasoning to Action (R2A: understand tasks, plan, execute in physical environments) and World Model (predict physical-world changes); closed-loop tests on real robots; 526 teams from 27 countries; a separate real-supermarket benchmark track launched with Dexmal.", "metric": "competition ranking", "evaluator": "organizer", "as_of": "2026-06-05", "leaders": [ { "model": "PrismBot", "developer": "vivo", "country": "China", "country_label": "China", "value": "Reasoning-to-Action champion", "as_of": "2026-06-05", "vendor_reported": false } ], "caveat": "Organizer press release; ranks without scores. 526 teams from 27 countries. World Model track standings were not retrieved.", "status": "verified", "sources": [ { "url": "https://agibot.com/article/231/detail/73.html", "captured_at": "2026-10-26", "quote": "PrismBot from vivo won the championship, followed by Shanghai RoboParty’s RP-VLA and GreenVLA." } ] } ], "areas": { "areas": [ { "id": "area-1", "name": "Cross-embodiment pretraining and action tokens", "items": [ { "id": "area1-33", "actor": "Figure AI", "country": "US", "country_label": "US", "date": "2026-09-17", "evidence": "Figure Helix 2.5, pretrained on Index human-behaviour data, ran three whole-body behaviours zero-shot in 30 unseen homes; Index pretraining alone raised zero-shot success from 9% to 56% and showed a human-to-humanoid transfer scaling law.", "status": "verified", "sources": [ { "url": "https://www.figure.ai/news/helix-2-5-zero-shot-30-home-generalization", "captured_at": "2026-10-26", "quote": "Index pretraining alone increased zero-shot success from 9% to 56%" } ] }, { "id": "area1-0", "actor": "X Square Robot (authors Kang, Shi, Liang, Gan, Liu)", "country": "China", "country_label": "China", "date": "2026-06-07", "evidence": "X Square Robot's X-Tokenizer treats action tokenisation as a semantic interface between VLM reasoning and control, shared across diverse arm embodiments; reports gains over FAST on multimodal grounding (+13.5%) and long-horizon tasks.", "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2606.14752", "captured_at": "2026-10-26", "quote": "provides a shared action interface across diverse robotic arm embodiments" } ] }, { "id": "area1-39", "actor": "NVIDIA", "country": "US", "country_label": "US", "date": "2026-02-25", "evidence": "NVIDIA Isaac GR00T N1.7-3B (HF 2026-02-25) is an open cross-embodiment model with Cosmos-Reason2-2B VLM backbone and per-embodiment encoder/decoder MLPs; post-trainable with real or synthetic data; a medical variant GR00T-H-N1.7 followed 2026-05-30.", "status": "verified", "sources": [ { "url": "https://huggingface.co/nvidia/GR00T-N1.7-3B", "captured_at": "2026-10-26", "quote": "This cross-embodiment model takes multimodal input, including language and images, to perform manipulation tasks in diverse environments." } ] }, { "id": "area1-2", "actor": "Dong, Liu, Zhang, Ye, Yuan et al.", "country": "Unverified", "country_label": "Unverified", "date": "2026-02-17", "evidence": "ActionCodec derives design principles for action tokenisers from the VLA-optimisation viewpoint rather than reconstruction fidelity (token overlap, vocabulary redundancy, mutual information).", "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2602.15397", "captured_at": "2026-10-26", "quote": "the fundamental question of what makes for good action tokenizers remains unanswered" } ] }, { "id": "area1-1", "actor": "Princeton University with Physical Intelligence", "country": "US", "country_label": "US", "date": "2026-02-11", "evidence": "LAP (Language-Action Pre-training) writes low-level robot actions as natural language; LAP-3B shows substantial zero-shot transfer to new embodiments without a learned tokenizer.", "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2602.10556", "captured_at": "2026-10-26", "quote": "LAP requires no learned tokenizer, no costly annotation, and no embodiment-specific architectural design." } ] }, { "id": "area1-14", "actor": "Physical Intelligence", "country": "US", "country_label": "US", "date": "2025-12-16", "evidence": "Physical Intelligence reports human-to-robot transfer emerging in VLAs as they scale.", "status": "verified", "sources": [ { "url": "https://www.pi.website/blog", "captured_at": "2026-10-26", "quote": "Exploring how transfer from human videos to robotic tasks emerges in robotic foundation models as they scale." } ] } ] }, { "id": "area-2", "name": "Dual-system (fast/slow) architectures and latency", "items": [ { "id": "area2-41", "actor": "MindOn", "country": "China", "country_label": "China", "date": "2026-10-01", "evidence": "MindOn unveiled Mind-1 on 2026-09-30, claiming human-level cycle times for packing, parcel sorting and folding from faster inference and human-speed training data.", "status": "snippet-only", "sources": [ { "url": "https://interestingengineering.com/ai-robotics/mindon-mind-1-robot-human-speed-physical-ai", "captured_at": "2026-10-26", "quote": "Mind-1 is designed around human-speed manipulation, combining faster inference, human-speed training data and a control system built for rapid robot motion." } ] }, { "id": "area2-44", "actor": "Google DeepMind", "country": "US", "country_label": "US", "date": "2026-07-30", "evidence": "ER 2 design lets the robot think about what comes next while it is still acting, a decoupled slow-planner and fast-VLA split (Gemini ER + any VLA) that makes latency hiding explicit.", "status": "verified", "sources": [ { "url": "https://blog.google/innovation-and-ai/models-and-research/google-deepmind/gemini-robotics-er-2/", "captured_at": "2026-10-26", "quote": "The design of Gemini Robotics ER 2 allows the robot to \"think\" about what comes next while simultaneously performing its actions." } ] }, { "id": "area2-34", "actor": "Figure AI", "country": "US", "country_label": "US", "date": "2026-01-27", "evidence": "Figure Helix 02 adds System 0, a learned whole-body controller (trained on over 1,000 hours of human motion and sim-to-real RL) beneath the S1/S2 stack, replacing 109,504 lines of hand-engineered C++ with one neural prior; single network from pixels, touch and proprioception to all actuators.", "status": "verified", "sources": [ { "url": "https://www.figure.ai/news/helix-02", "captured_at": "2026-10-26", "quote": "System 0, a learned whole-body controller trained on over 1,000 hours of human motion data and sim-to-real reinforcement learning" } ] }, { "id": "area2-40", "actor": "Galaxea AI", "country": "China", "country_label": "China", "date": "2025-08-30", "evidence": "Galaxea's G0 couples a VLM planner with a VLA executor in a dual-system framework trained with a three-stage curriculum (cross-embodiment pretraining, single-embodiment pretraining, task post-training).", "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2509.00576", "captured_at": "2026-10-26", "quote": "G0, a dual-system framework that couples a Vision-Language Model (VLM) for multimodal planning with a Vision-Language-Action (VLA) model for fine-grained execution" } ] } ] }, { "id": "area-3", "name": "World models and world-action models", "items": [ { "id": "area3-37", "actor": "NVIDIA", "country": "US", "country_label": "US", "date": "2026-10-26", "evidence": "NVIDIA defines the world action model (WAM) as a single joint video-action diffusion transformer, trained on internet and egocentric human video, as the successor pattern to VLAs; glossary page positions it next to Isaac GR00T.", "status": "verified", "sources": [ { "url": "https://www.nvidia.com/en-us/glossary/world-action-model.md", "captured_at": "2026-10-26", "quote": "A world action model addresses this gap by jointly learning to predict both future world states" } ] }, { "id": "area3-51", "actor": "Dyna Robotics", "country": "US", "country_label": "US", "date": "2026-08-15", "evidence": "Dyna-2 is a world-action model pre-trained on more than one million hours of egocentric human video, reporting a scaling law on held-out human data and a human-to-robot transfer scaling law with no robot data in pre-training.", "status": "verified", "sources": [ { "url": "https://www.dyna.co/dyna-2", "captured_at": "2026-10-26", "quote": "A world-action model pre-trained on more than one million hours of egocentric human video." } ] }, { "id": "area3-38", "actor": "Robbyant (Ant Group)", "country": "China", "country_label": "China", "date": "2026-07-09", "evidence": "Ant Group's Robbyant released LingBot-VA 2.0, billed as the first embodied-native video-action world model pretrained from scratch rather than adapted from a digital video generator; v1 (arXiv 2601.21998, RSS 2026) reported 92.0% RoboTwin 2.0, 98.5% LIBERO.", "status": "verified", "sources": [ { "url": "https://www.webull.com/news/15201670514402304", "captured_at": "2026-10-26", "quote": "the industry's first embodied-native video-action world model" } ] }, { "id": "area3-36", "actor": "NVIDIA GEAR", "country": "US", "country_label": "US", "date": "2026-02-17", "evidence": "NVIDIA DreamZero, a 14B world action model on a video-diffusion backbone, reports over 2x better generalisation to new tasks and environments than state-of-the-art VLAs and runs closed-loop at about 7 Hz after system optimisation.", "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2602.15922", "captured_at": "2026-10-26", "quote": "over 2x improvement in generalization to new tasks and environments compared to state-of-the-art VLAs in real robot experiments" } ] } ] }, { "id": "area-4", "name": "Sim-to-real and real-to-sim", "items": [ { "id": "area4-4", "actor": "NVIDIA (Isaac Lab open-source team)", "country": "US", "country_label": "US", "date": "2026-10-26", "evidence": "Isaac Lab 3.0 (beta 2 patch 1 on the release-notes page) is a multi-backend release: one asset/sensor/scene API over PhysX and Newton physics, pluggable renderers and visualizers, Warp-native data pipeline; breaking quaternion format change WXYZ to XYZW.", "status": "verified", "sources": [ { "url": "https://docs.nvidia.com/isaaclab/references/release-notes", "captured_at": "2026-10-26", "quote": "Multi-Backend Physics Architecture" } ] }, { "id": "area4-50", "actor": "Genesis AI", "country": "Unverified", "country_label": "US/France (unverified)", "date": "2026-05-01", "evidence": "Genesis AI open-sourced Genesis World 1.0 (physics engine, Nyx path-traced renderer, Quadrants GPU layer) and claims zero-shot real-to-sim model evaluation. Company claim, single secondary source.", "status": "snippet-only", "sources": [ { "url": "https://app.dealroom.co/news/note/genesis-ai-open-sources-world-1-0-simulation-platform-for-robotics", "captured_at": "2026-10-26", "quote": "Genesis AI open-sources Genesis World 1.0 simulation platform for robotics" } ] }, { "id": "area4-3", "actor": "Allen Institute for AI (Ai2)", "country": "US", "country_label": "US", "date": "2026-03-11", "evidence": "Ai2 MolmoBot: manipulation policies trained entirely on synthetic data transfer zero-shot to real Franka FR3 and RB-Y1 robots; MolmoBot-Data has 1.7M expert trajectories over 11K+ objects and 94K+ procedural scenes. NOTE this is sim-to-real with no real demos, not real-to-sim.", "status": "verified", "sources": [ { "url": "https://allenai.org/blog/molmobot", "captured_at": "2026-10-26", "quote": "robotics models trained entirely in simulation that transfer directly to real-world robots with no additional manually-collected data or fine-tuning" } ] }, { "id": "area4-49", "actor": "AgiBot (Genie Sim team)", "country": "China", "country_label": "China", "date": "2026-01-05", "evidence": "AgiBot's Genie Sim 3.0: LLM-powered scene generator builds high-fidelity simulated scenes from natural language for humanoid manipulation data and benchmarks; AgiBot World 2026 pairs real data with 1:1 digital-twin sim data.", "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2601.02078", "captured_at": "2026-10-26", "quote": "a large language model (LLM)-powered tool that constructs high-fidelity scenes from natural language instructions" } ] } ] }, { "id": "area-5", "name": "Data engines: teleop, egocentric and gig-worker video", "items": [ { "id": "area5-15", "actor": "HUMXN Research Labs", "country": "Unverified", "country_label": "US (unverified)", "date": "2026-10-19", "evidence": "HUMXN Research Labs announced an initiative to buy and license $50M of real-world technician material (job-site video, manuals, repair records), paying about $20/hour for approved footage. Single secondary source (Spanish-language outlet citing a Newsfile release).", "status": "snippet-only", "sources": [ { "url": "https://ecosistemastartup.com/humxn-regala-fontaneria-a-cambio-de-grabar-a-los-tecnicos/", "captured_at": "2026-10-26", "quote": "La compañía paga alrededor de US$20 por hora de metraje aprobado" } ] }, { "id": "area5-18", "actor": "MR. ROBOT / TROOPERS", "country": "Unverified", "country_label": "Malaysia/Korea (unverified)", "date": "2026-10-01", "evidence": "MR. ROBOT and TROOPERS MoU: first-person recordings of frontline workers paired with robot data from a Malaysia centre running 20 AgiBot G2 humanoids around the clock.", "status": "snippet-only", "sources": [ { "url": "https://www.tech-critter.com/mr-robot-troopers-humanoid-robotics-data-mou/", "captured_at": "2026-10-26", "quote": "20 AgiBot G2 humanoid robots operating around the clock" } ] }, { "id": "area5-32", "actor": "Figure AI", "country": "US", "country_label": "US", "date": "2026-08-25", "evidence": "Figure launched Index, an app-based crowdsourced human-behaviour data pipeline: 264,000 downloads in 100+ countries, 44,000 weekly actives, 16M+ videos uploaded, 30 minutes of video per second, $15M paid to creators, and a stated plan to spend over $1B in 12 months on data and compute. Figure says vendors could not hit its throughput, diversity or quality bar.", "status": "verified", "sources": [ { "url": "https://www.figure.ai/news/introducing-index", "captured_at": "2026-10-26", "quote": "Prior to this we tried buying data. Vendors couldn't hit the throughput, diversity, or quality bar Helix requires, so we built the pipeline ourselves" } ] }, { "id": "area5-52", "actor": "Dyna Robotics", "country": "US", "country_label": "US", "date": "2026-08-17", "evidence": "Dyna says training at one million hours broke storage format, ingestion and manifest assumptions that worked at ten thousand hours; data infrastructure is the bottleneck rather than GPUs.", "status": "verified", "sources": [ { "url": "https://www.dyna.co/research/dyna-2-infrastructure", "captured_at": "2026-10-26", "quote": "Most of what worked at ten thousand hours did not hold up at a million." } ] }, { "id": "area5-58", "actor": "XDOF", "country": "US", "country_label": "US", "date": "2026-06-17", "evidence": "XDOF released ABC-130K, 130K+ low-cost bimanual teleoperation episodes under Apache 2.0, plus SARM2 reward model and the Ember-0 real-robot benchmark (frontier LLMs plan well but fail on contact, tools and deformables).", "status": "verified", "sources": [ { "url": "https://www.xdof.ai/blog/abc-130k", "captured_at": "2026-10-26", "quote": "130K+ episodes, 200 complex manipulation tasks. All on a low-cost bimanual rig. Fully open source under Apache 2.0." } ] }, { "id": "area5-17", "actor": "Mecka AI", "country": "US", "country_label": "US", "date": "2026-06-02", "evidence": "Mecka announced $60M to build data, evaluation and deployment infrastructure for physical AI, including first-person video of human work.", "status": "verified", "sources": [ { "url": "https://pulse2.com/mecka-raises-60-million-to-build-the-data-and-deployment-layer-for-physical-ai/", "captured_at": "2026-10-26", "quote": "Mecka announced that it has raised $60 million in funding." } ] }, { "id": "area5-16", "actor": "Human Archive (Wing VC, NVP Capital, Y Combinator)", "country": "Other", "country_label": "India/USA", "date": "2026-05-26", "evidence": "Human Archive raised $8.2M to collect first-person video from Indian service workers wearing camera caps, with 1,000+ active headsets; also uses tactile gloves, mocap suits and wrist cameras. Consent concerns reported in Indian press.", "status": "verified", "sources": [ { "url": "https://techcrunch.com/2026/05/26/human-archive-taps-into-indias-services-startups-to-collect-data-for-physical-ai/", "captured_at": "2026-10-26", "quote": "more than 1,000 active headsets deployed across multiple locations" } ] }, { "id": "area5-56", "actor": "Lightwheel", "country": "Unverified", "country_label": "US/China (unverified)", "date": "2025-12-04", "evidence": "Lightwheel launched EgoSuite, a multi-modality egocentric human data service (head-only and head-plus-wrist capture, global regions), alongside SimReady assets and evaluation as a three-layer world/behavior/evaluation data engine.", "status": "verified", "sources": [ { "url": "https://lightwheel.ai/egosuite", "captured_at": "2026-10-26", "quote": "A High-Quality, Multi-Modality, Globally Scalable Egocentric Human Data Solution" } ] } ] }, { "id": "area-6", "name": "Dexterous hands and touch", "items": [ { "id": "area6-30", "actor": "Rhoda AI", "country": "US", "country_label": "US", "date": "2026-10-13", "evidence": "Rhoda AI paired an in-house tactile sensor with its Direct Video-Action model; demo shows a robot finding textured balls by touch with head and wrist cameras blocked (no success rate reported).", "status": "verified", "sources": [ { "url": "https://www.humanoidsdaily.com/news/rhoda-ai-touch-sensor-blindfolded-robot", "captured_at": "2026-10-26", "quote": "has demonstrated a robot searching for textured balls with its head and wrist cameras blocked, using touch to guide its movements" } ] }, { "id": "area6-29", "actor": "Sharpa", "country": "Other", "country_label": "Singapore/China (HQ not verified)", "date": "2026-09-29", "evidence": "Sharpa launched the W02 hand (21 active DoF, under 750 g, fingertip Dynamic Tactile Array plus electronic skin on fingers and palm, 5 mN to 30 N range), the D01 robot and the AE01 haptic data glove for teleoperation and egocentric data capture.", "status": "verified", "sources": [ { "url": "https://technode.global/prnasia/sharpa-unveils-three-flagship-products-d01-robot-w02-dexterous-hand-and-ae01-data-glove-at-iros-2026/", "captured_at": "2026-10-26", "quote": "combines 21 active degrees of freedom with a compact structure weighing less than 750 grams" } ] }, { "id": "area6-55", "actor": "1X Technologies", "country": "Other", "country_label": "Norway/USA", "date": "2026-07-09", "evidence": "1X announced 25-DoF tendon-driven hands for NEO, claiming near human-level dexterity, strength, safety and reliability (July 2026).", "status": "verified", "sources": [ { "url": "https://www.1x.tech/discover", "captured_at": "2026-10-26", "quote": "our breakthrough 25 Degree of Freedom (DOF), tendon-driven hands for the NEO humanoid platform" } ] }, { "id": "area6-31", "actor": "Tesla (Counterpoint Research estimate)", "country": "US", "country_label": "US", "date": "2026-04-20", "evidence": "Tesla Optimus 22-DoF hand: patents describe tendon-driven hand with forearm actuators; Counterpoint estimates hands are nearly 20% of V3 BoM; single-source blog reports Musk saying V3 will not use the patented design. Not confirmed by Tesla.", "status": "snippet-only", "sources": [ { "url": "https://counterpointresearch.com/cn/reports/tesla-to-leverage-ev-experience-and-chinese-partners-to-ace-humanoids-market-unlock-huge-revenue-potential", "captured_at": "2026-10-26", "quote": "22-DoF dexterous hands accounting for nearly 20% of the total BoM cost" } ] } ] }, { "id": "area-7", "name": "Whole-body control and locomotion via RL", "items": [ { "id": "area7-5", "actor": "Skild AI", "country": "US", "country_label": "US", "date": "2026-09-23", "evidence": "Skild AI post-trained its S1 foundation model with physical self-play: S1 learned soccer over 140+ simulated years against earlier versions of itself, then transferred to a physical humanoid. Company blog only; no baselines published.", "status": "verified", "sources": [ { "url": "https://www.skild.ai/blogs/", "captured_at": "2026-10-26", "quote": "A breakthrough post-training result for physical AI via self-play: a strong base model like S1 can learn to complete extremely dexterous and dynamic tasks, like soccer" } ] }, { "id": "area7-61", "actor": "NVIDIA NVlabs", "country": "US", "country_label": "US", "date": "2026-07-15", "evidence": "NVIDIA GRAIL converts video-derived human-object interaction into physics-validated Unitree G1 locomanipulation trajectories with RL tracking policies released 2026-07-15.", "status": "verified", "sources": [ { "url": "https://huggingface.co/datasets/nvidia/PhysicalAI-Robotics-Locomanipulation-GRAIL", "captured_at": "2026-10-26", "quote": "Released task-general tracking policy checkpoints trained on the released data." } ] } ] }, { "id": "area-8", "name": "RL post-training and learning from deployment", "items": [ { "id": "area8-54", "actor": "Dyna Robotics", "country": "US", "country_label": "US", "date": "2026-09-28", "evidence": "Dyna lists 'Deployment is the eval' and a deployment flywheel (episodes turned into semantic observability and hardware-failure prediction) as the post-training loop; Dyna-2.1 followed 2026-09-28 as a physical agent for end-to-end workflows.", "status": "verified", "sources": [ { "url": "https://www.dyna.co/research", "captured_at": "2026-10-26", "quote": "Dyna-2.1: A Physical Agent for End-to-End Workflows 09.28.26" } ] }, { "id": "area8-6", "actor": "Skild AI", "country": "US", "country_label": "US", "date": "2026-09-10", "evidence": "Skild AI frames self-play as a post-training method for S1 and reports crossing $100M ARR ten months after first commercial deployment (Sep 10 2026 post).", "status": "verified", "sources": [ { "url": "https://www.skild.ai/blogs/", "captured_at": "2026-10-26", "quote": "Skild AI crossed $100 million in annual recurring revenue, ten months after our first commercial deployment." } ] }, { "id": "area8-48", "actor": "Skild AI", "country": "US", "country_label": "US", "date": "2026-08-18", "evidence": "Skild S1 claims in-context learning for robotics: unseen tasks over 10-minute horizons from one video prompt with no post-training, arguing post-training normally needs tens to hundreds of hours of deployment data.", "status": "verified", "sources": [ { "url": "https://www.skild.ai/blogs/s1", "captured_at": "2026-10-26", "quote": "Unseen tasks 10-minute horizons One video prompt No post-training" } ] }, { "id": "area8-11", "actor": "Physical Intelligence", "country": "US", "country_label": "US", "date": "2026-03-19", "evidence": "PI extracts an 'RL Token' from VLA models for fast online RL on precise tasks with only a few hours of data (March 2026).", "status": "verified", "sources": [ { "url": "https://www.pi.website/blog", "captured_at": "2026-10-26", "quote": "We extract an RL Token from VLA models to enable fast online RL and improve throughput on precise tasks with few hours of data." } ] }, { "id": "area8-10", "actor": "Physical Intelligence", "country": "US", "country_label": "US", "date": "2025-11-17", "evidence": "Physical Intelligence pi*0.6 (Recap): a VLA trained with RL from demonstrations, expert corrections and autonomous on-robot experience to improve success and throughput.", "status": "verified", "sources": [ { "url": "https://www.pi.website/blog", "captured_at": "2026-10-26", "quote": "A method for training our generalist policies with RL to improve success rate and throughput on real-world tasks." } ] } ] }, { "id": "area-9", "name": "Safety, red-teaming and evaluation science", "items": [ { "id": "area9-59", "actor": "XDOF", "country": "US", "country_label": "US", "date": "2026-10-23", "evidence": "XDOF's Ember-0: a 16-task real-robot bimanual benchmark finds five frontier general models plan well but struggle with contact, tools and deformables and are slow.", "status": "verified", "sources": [ { "url": "https://www.xdof.ai/blog/evaluating-llms", "captured_at": "2026-10-26", "quote": "They plan well, and struggle with contact, tools, and deformables." } ] }, { "id": "area9-22", "actor": "Robocurve", "country": "US", "country_label": "US", "date": "2026-09-18", "evidence": "Robocurve's RoboHarm gave GPT-6 Astra, Claude Fable 5.1 and MolmoAct2 five hazardous instructions on real bimanual arms, 20 trials each. Fable 5.1 refused 20 of 100 on safety grounds, GPT-6 Astra 2 of 100. Small benchmark: five scenes, one wording per task.", "status": "verified", "sources": [ { "url": "https://www.popularai.org/p/roboharm-benchmark-ai-robot-safety", "captured_at": "2026-10-26", "quote": "Across 300 trials, the benchmark recorded only 22 explicit safety refusals" }, { "url": "https://robocurve.org/roboharm/", "captured_at": "2026-10-26" } ] }, { "id": "area9-24", "actor": "Yin, Liu, Yang, Li, Xue et al.", "country": "Unverified", "country_label": "Unverified", "date": "2026-06-17", "evidence": "ROBOSHACKLES builds a human-injury-prevention safety dataset for embodied foundation models from real DROID observations with hazard-aware image editing and rollout synthesis.", "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2606.18632", "captured_at": "2026-10-26", "quote": "real-world data of robots harming humans or creating hazardous household situations cannot be safely or ethically collected" } ] }, { "id": "area9-23", "actor": "Zhang, Zhang, Fan, Shen, Cai et al.", "country": "Unverified", "country_label": "Unverified", "date": "2026-04-24", "evidence": "RedVLA, described as the first red-teaming framework for physical safety in VLAs: synthesises risk scenes from benign trajectories to find unsafe behaviours before deployment; NeurIPS 2026 poster.", "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2604.22591", "captured_at": "2026-10-26", "quote": "the first red teaming framework for physical safety in VLA models" } ] } ] }, { "id": "area-10", "name": "Open-source hardware and low-cost platforms", "items": [ { "id": "area10-60", "actor": "Pollen Robotics / Hugging Face", "country": "Other", "country_label": "France/USA", "date": "2026-08-31", "evidence": "Pollen Robotics (Hugging Face) published microduck policies under Apache-2.0 on the Hub (2026-08-31); community fine-tunes (tinyvla microduck, sprint remix, chimney-climb) appeared within weeks, evidence of a low-cost open platform with a community loop.", "status": "verified", "sources": [ { "url": "https://huggingface.co/pollen-robotics/microduck-policies", "captured_at": "2026-10-26", "quote": "license: apache-2.0 library_name: microduck pipeline_tag: robotics" } ] }, { "id": "area10-20", "actor": "Hugging Face LeRobot", "country": "US", "country_label": "US/France", "date": "2026-05-21", "evidence": "LeRobot hub grew from 1,145 community datasets at end of 2024 to 58,000+ by May 2026, driven by cheap arms such as the roughly $100 SO-101 (Hugging Face with The Robot Studio).", "status": "verified", "sources": [ { "url": "https://www.techtimes.com/articles/317129/20260525/open-source-robotics-ai-reaches-inflection-point-lerobot-hub-surpasses-58000-datasets-one-year.htm", "captured_at": "2026-10-26", "quote": "now hosts more than 58,000 community-contributed datasets, up from 1,145 at the end of 2024" } ] }, { "id": "area10-19", "actor": "Menlo Research", "country": "US", "country_label": "US", "date": "2026-05-07", "evidence": "Menlo Research's Asimov v1 humanoid exposes mechanical and electrical CAD, MuJoCo model, onboard software and bill of materials; 25 actuated DoF; DIY kit with $499 deposit, target about $15,000, planned summer 2026 (later coverage lists $20,000).", "status": "verified", "sources": [ { "url": "https://www.opensourceforu.com/2026/05/asimov-v1-open-sources-humanoid-robotics-development/", "captured_at": "2026-10-26", "quote": "openly exposes its mechanical CAD, electrical CAD, MuJoCo simulation model, onboard software and bill of materials" } ] } ] }, { "id": "area-11", "name": "Memory, spatial reasoning and long-horizon autonomy", "items": [ { "id": "area11-57", "actor": "Dyna Robotics", "country": "US", "country_label": "US", "date": "2026-09-29", "evidence": "Dyna-2.1 introduces a physical agent (semi-humanoid Taku) with a VLM orchestrator that reasons about the environment, chooses next steps and keeps long-term memory to automate a one-hour uncut end-to-end workflow.", "status": "verified", "sources": [ { "url": "https://www.dyna.co/dyna-2.1", "captured_at": "2026-10-26", "quote": "Dyna-2.1: The first physical agent to automate an end-to-end workflow—one hour, uncut." } ] }, { "id": "area11-42", "actor": "Google DeepMind", "country": "US", "country_label": "US", "date": "2026-07-30", "evidence": "Google DeepMind Gemini Robotics ER 2 acts as a high-level brain: continuous-video progress tracking and retry, tool calling, multi-step planning, handing motor execution to any lower-level VLA; public via Gemini API and AI Studio.", "status": "verified", "sources": [ { "url": "https://blog.google/innovation-and-ai/models-and-research/google-deepmind/gemini-robotics-er-2/", "captured_at": "2026-10-26", "quote": "Think of Gemini Robotics ER 2 as a high-level brain for robots." } ] }, { "id": "area11-47", "actor": "Stanford (lab attribution unclear)", "country": "US", "country_label": "US", "date": "2026-07-01", "evidence": "Stanford HomeBody (secondary coverage only): a Unitree G1 explores an unseen room, builds spatial memory and a Real2Sim digital twin in Isaac Sim, and a frontier VLM picks from five primitives. Lab attribution inconsistent across sources.", "status": "snippet-only", "sources": [ { "url": "https://aiweekly.co/alerts/stanfords-homebody-wires-gpt-6-astra-directly-to-a-unitree-g1-skips-the-vla", "captured_at": "2026-10-26", "quote": "Stanford's HomeBody Wires GPT-6 Astra Directly to a Unitree G1, Skips the VLA Layer" } ] }, { "id": "area11-45", "actor": "Kim, Pak, Long, Tian, Atanasov et al.", "country": "Unverified", "country_label": "Unverified", "date": "2026-06-11", "evidence": "SERF conditions a VLA on an online spatiotemporal feature map of the environment and robot body to improve long-horizon mobile manipulation.", "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2606.12956", "captured_at": "2026-10-26", "quote": "conditioning a mobile manipulation policy on a spatiotemporal feature map improves reasoning over long horizons" } ] }, { "id": "area11-46", "actor": "Qian, He, Shi, Xiao, Jiang et al.", "country": "Unverified", "country_label": "Unverified", "date": "2026-04-15", "evidence": "ESCAPE builds a depth-free persistent 3D episodic spatial memory coupled with an adaptive execution policy for long-horizon mobile manipulation.", "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2604.13633", "captured_at": "2026-10-26", "quote": "autoregressively construct a depth-free, persistent 3D spatial memory" } ] }, { "id": "area11-12", "actor": "Physical Intelligence", "country": "US", "country_label": "US", "date": "2026-03-03", "evidence": "PI Multi-Scale Embodied Memory (MEM) gives VLAs long- and short-term memory for tasks longer than ten minutes.", "status": "verified", "sources": [ { "url": "https://www.pi.website/blog", "captured_at": "2026-10-26", "quote": "Multi-Scale Embodied Memory (MEM) gives our models both long-term and short-term memory, enabling complex tasks longer than ten minutes." } ] } ] }, { "id": "area-12", "name": "Robot-learning infrastructure and tooling", "items": [ { "id": "area12-26", "actor": "Nebius with NVIDIA", "country": "Other", "country_label": "Netherlands/USA", "date": "2026-10-21", "evidence": "Nebius open-sourced Physical AI Workbench, built with NVIDIA, covering data generation, simulation, training and evaluation with agent-operable workflows; customers claim up to 40% engineering-time savings (vendor claim).", "status": "verified", "sources": [ { "url": "https://nebius.com/blog/posts/physical-ai-workbench", "captured_at": "2026-10-26", "quote": "the open-source framework gives robotics teams one place to run the full physical AI development cycle" } ] }, { "id": "area12-13", "actor": "Physical Intelligence", "country": "US", "country_label": "US", "date": "2026-10-13", "evidence": "PI opened its latest models through the Pi API beta: upload data, fine-tune, and deploy cloud-hosted inference on your own robots; frontier VLAs as a hosted service.", "status": "verified", "sources": [ { "url": "https://www.pi.website/blog", "captured_at": "2026-10-26", "quote": "Our latest robotics models are now available through the Pi API beta. Upload data, fine-tune models, and deploy cloud-hosted inference on your own robots." } ] }, { "id": "area12-27", "actor": "NVIDIA", "country": "US", "country_label": "US", "date": "2026-09-22", "evidence": "NVIDIA Isaac ROS 5.0 released at ROSCon Toronto with ROS Lyrical and Ubuntu 24.04 support and agentic workflows for building robots with AI agents.", "status": "verified", "sources": [ { "url": "https://blogs.nvidia.com/blog/isaac-ros-5-0-agentic-open-source-robotics/", "captured_at": "2026-10-26", "quote": "helps humans and AI agents build robots together" } ] }, { "id": "area12-28", "actor": "StarVLA Community", "country": "China", "country_label": "China (Hong Kong)", "date": "2026-04-06", "evidence": "StarVLA: modular backbone-plus-action-head open codebase supporting VLM and world-model (Cosmos) backbones, with one evaluation interface across LIBERO, SimplerEnv, RoboTwin 2.0, RoboCasa-GR1 and BEHAVIOR-1K.", "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2604.05014", "captured_at": "2026-10-26", "quote": "VLA methods remain fragmented across incompatible architectures, codebases, and evaluation protocols" } ] } ] }, { "id": "area-13", "name": "Energy, battery and thermal limits", "items": [ { "id": "area13-7", "actor": "KAIST (Hwangbo lab)", "country": "Other", "country_label": "Korea", "date": "2026-09-24", "evidence": "KAIST's RAIBO2 quadruped finished a full marathon (4h19m52s) on one charge, with an RL-trained controller. The Nature paper treats range as a whole-robot energy, thermal, battery and actuator problem; legged robots had been limited to about 20 km.", "status": "verified", "sources": [ { "url": "https://news.kaist.ac.kr/newsen/html/news/?mng_no=67610&mode=V", "captured_at": "2026-10-26", "quote": "it is necessary to secure not only momentary locomotion performance but also energy efficiency, thermal management, battery performance, and actuator stability" }, { "url": "https://www.natureasia.com/en/info/press-releases/detail/9447", "captured_at": "2026-10-26" } ] } ] }, { "id": "area-14", "name": "Fleet learning and multi-robot skill sharing", "items": [ { "id": "area14-53", "actor": "Dyna Robotics", "country": "US", "country_label": "US", "date": "2026-08-27", "evidence": "Dyna reports its robots crossed the ROI threshold, with Din Tai Fung rolling them out across its US restaurant network and a deployed fleet of hundreds of robots by H1 2027; frames deployment data as a flywheel and eval.", "status": "verified", "sources": [ { "url": "https://www.dyna.co/research/scaling-customer-deployments", "captured_at": "2026-10-26", "quote": "our robots have successfully crossed the ROI threshold, and Din Tai Fung, the highest revenue per location restaurant chain in the US, is rolling out Dyna's robots across its restaurant network" } ] }, { "id": "area14-43", "actor": "Google DeepMind", "country": "US", "country_label": "US", "date": "2026-07-30", "evidence": "Gemini Robotics ER 2 introduces multi-robot collaboration so different robots share a workspace and complete workflows a single robot could not.", "status": "verified", "sources": [ { "url": "https://blog.google/innovation-and-ai/models-and-research/google-deepmind/gemini-robotics-er-2/", "captured_at": "2026-10-26", "quote": "We are also introducing multi-robot collaboration, enabling robots to work together in shared spaces and complete complex workflows a single robot could not do alone." } ] } ] } ], "datasets": [ { "id": "dataset-agibot-world-2026", "name": "AgiBot World 2026", "maintainer": "AGIBOT", "country": "China", "country_label": "China", "released": "2026 (staged; Theme 3 on 2026-09-01)", "size": "All real-world data on the AgiBot G2 across commercial and home spaces, with simulated twins in Genie Sim; total hours not stated.", "human_video_hours": null, "license": "CC-BY-NC-SA-4.0", "status": "verified", "sources": [ { "url": "https://huggingface.co/datasets/agibot-world/AgiBotWorld2026", "captured_at": "2026-10-26", "quote": "The AGIBOT WORLD 2026 dataset is collected from 100% real-world environments, covering commercial spaces, home, and other general-purpose scenarios." } ] }, { "id": "dataset-agibot-world-2026-theme-3", "name": "AgiBot World 2026, Theme 3 (reinforcement learning)", "maintainer": "AGIBOT", "country": "China", "country_label": "China", "released": "2026-09-01", "size": "11,430 real-world trajectories, 14 industrial and household tasks; 1,024 successful policy rollouts and 1,369 failed rollouts; expert demos, autonomous rollouts and human-in-the-loop corrections", "human_video_hours": null, "license": null, "status": "verified", "sources": [ { "url": "https://agibot.com/article/231/detail/95.html", "captured_at": "2026-10-26", "quote": "The first release includes 11,430 real-world trajectories across 14 industrial and household tasks, including 1,024 successful policy rollouts and 1,369 failed rollouts." } ] }, { "id": "dataset-molmobot-data", "name": "MolmoBot-Data", "maintainer": "Allen Institute for AI", "country": "US", "country_label": "US", "released": "2026-03-11", "size": "1.7M expert trajectories, 11K+ objects, 94K+ procedural environments, 8 task types.", "human_video_hours": null, "license": "Open (exact license not checked)", "status": "snippet-only", "sources": [ { "url": "https://allenai.org/blog/molmobot", "captured_at": "2026-10-26", "quote": "robotics models trained entirely in simulation that transfer directly to real-world robots with no additional manually-collected data or fine-tuning" } ] }, { "id": "dataset-nvidia-flora", "name": "NVIDIA Flora (video-labeling models)", "maintainer": "NVIDIA GEAR lab", "country": "US", "country_label": "US", "released": "2026-10-20", "size": "Labels 200,000 hours of human and robot video per day on 512 GPUs; outputs captions and reward signals.", "human_video_hours": null, "license": "Non-commercial research (reports conflict)", "status": "verified", "sources": [ { "url": "https://research.nvidia.com/labs/gear/flora", "captured_at": "2026-10-26", "quote": "Flora is battle tested at scale, labeling 200,000 hours of data per day across 512 GB300s." } ] }, { "id": "dataset-egocentric-10k", "name": "Egocentric-10K", "maintainer": "Build AI", "country": "US", "country_label": "US", "released": "2025-11-09", "size": "10,000 hours, 1.08B frames, 16.4 TB of head-mounted factory-worker video from 85 factories.", "human_video_hours": 10000, "license": "Apache-2.0 (gated, HF conditions)", "status": "verified", "sources": [ { "url": "https://huggingface.co/datasets/builddotai/Egocentric-10K", "captured_at": "2026-10-26", "quote": "Licensed under the Apache 2.0 License." } ] }, { "id": "dataset-galaxea-open-world-dataset", "name": "Galaxea Open-World Dataset", "maintainer": "Galaxea AI", "country": "China", "country_label": "China", "released": "2025-08-30", "size": "About 500 hours, 50 scenes, 150+ tasks, single embodiment.", "human_video_hours": null, "license": "CC-BY-NC-SA-4.0 (gated)", "status": "verified", "sources": [ { "url": "https://arxiv.org/abs/2509.00576", "captured_at": "2026-10-26", "quote": "a large-scale, diverse collection of robot behaviors recorded in authentic human living and working environments" } ] }, { "id": "dataset-robomind", "name": "RoboMIND", "maintainer": "Beijing Humanoid Robot Innovation Center and BAAI", "country": "China", "country_label": "China", "released": "2025-01-02", "size": "Over 1 TB, multi-embodiment (AgileX, Tien Kung, Franka and others).", "human_video_hours": null, "license": "Apache-2.0", "status": "verified", "sources": [ { "url": "https://huggingface.co/datasets/x-humanoid-robomind/RoboMIND", "captured_at": "2026-10-26", "quote": "license:apache-2.0 size_categories:n>1T (HF API tags)" } ] }, { "id": "dataset-eidon-tracker-pov", "name": "Eidon Tracker POV (+ IMU)", "maintainer": "Eidon AI (released on shutdown)", "country": "US", "country_label": "US", "released": "2026-09-11", "size": "9.05 TB, about 1,274 hours of first-person household video; IMU variant adds arm tracking.", "human_video_hours": 1274, "license": "CC-BY-4.0 (from HF API tags)", "status": "verified", "sources": [ { "url": "https://huggingface.co/datasets/eidon-ai/tracker-pov", "captured_at": "2026-10-26", "quote": "9.05 TB of first-person household video" } ] }, { "id": "dataset-lerobot-community-datasets", "name": "LeRobot community datasets", "maintainer": "Hugging Face LeRobot", "country": "US", "country_label": "US/France", "released": "Count as of 2026-05-21", "size": "more than 58,000 community datasets, up from 1,145 at end of 2024; robotics is the largest dataset category on the Hub", "human_video_hours": null, "license": "Mixed per dataset", "status": "verified", "sources": [ { "url": "https://www.techtimes.com/articles/317129/20260525/open-source-robotics-ai-reaches-inflection-point-lerobot-hub-surpasses-58000-datasets-one-year.htm", "captured_at": "2026-10-26", "quote": "now hosts more than 58,000 community-contributed datasets, up from 1,145 at the end of 2024" } ] }, { "id": "dataset-droid-1-0-1", "name": "DROID 1.0.1", "maintainer": "DROID collaboration (LeRobot copy)", "country": "US", "country_label": "US", "released": "2025-07-16", "size": "95,658 episodes, 27,630,375 frames, 49,630 tasks (info.json)", "human_video_hours": null, "license": "Apache-2.0", "status": "verified", "sources": [ { "url": "https://huggingface.co/datasets/lerobot/droid_1.0.1", "captured_at": "2026-10-26", "quote": "\"total_episodes\": 95658, \"total_frames\": 27630375" } ] }, { "id": "dataset-open-x-embodiment", "name": "Open X-Embodiment", "maintainer": "Open X-Embodiment collaboration", "country": "US", "country_label": "US/international", "released": "2023-10", "size": "1M-10M episodes from 22 robots.", "human_video_hours": null, "license": "CC-BY-4.0 (mirror)", "status": "snippet-only", "sources": [ { "url": "https://huggingface.co/datasets/jxu124/OpenX-Embodiment", "captured_at": "2026-10-26", "quote": "size_categories: 1M80% high-DoF share (self-reported)", "scope": "Linkerbot Linker Hand", "country": "China", "country_label": "China", "vendor_reported": true, "status": "verified", "sources": [ { "url": "https://www.prnewswire.com/news-releases/10-000-dexterous-hands-shipped-150-million-raised-linkerbot-leads-the-market-302645409.html", "captured_at": "2026-10-26", "quote": "over 80% global market share in the global high-DoF segment" } ] }, { "id": "comp-19", "group": "Sensors", "supply_chain_category": "sensors", "subsystem": "force-torque sensors", "fact_type": "chokepoint", "claim": "McKinsey: six-axis force/torque sensors come from a limited set of robotics-focused vendors (ATI Industrial Automation, OnRobot), are calibration intensive and benefit little from auto/consumer spillover.", "value": "limited vendors; calibration-bound", "scope": "humanoid supply chain", "country": "US", "country_label": "US", "vendor_reported": false, "status": "verified", "sources": [ { "url": "https://www.mckinsey.com/industries/industrials/our-insights/turning-humanoid-supply-chain-constraints-into-billion-dollar-wins", "captured_at": "2026-10-26", "quote": "These components, supplied by a limited number of robotics-focused vendors, including ATI Industrial Automation and OnRobot, are calibration intensive and require tight metrology and quality control, with limited automation in production." } ] }, { "id": "comp-39", "group": "Sensors", "supply_chain_category": "sensors", "subsystem": "force-torque sensors", "fact_type": "price", "claim": "humanoid.guide relays X.PIN: a six-axis force-torque sensor still costs 5,000 to 30,000 yuan (about $700-$4,200).", "value": "CNY 5,000-30,000 (~$700-$4,200)", "scope": "6-axis F/T", "country": null, "country_label": null, "vendor_reported": false, "status": "verified", "sources": [ { "url": "https://humanoid.guide/?p=17836", "captured_at": "2026-10-26", "quote": "a six-axis force-torque sensor still costs ¥5,000 to ¥30,000 (≈$700–4,200)" } ] }, { "id": "comp-48", "group": "Sensors", "supply_chain_category": "sensors", "subsystem": "lidar", "fact_type": "spec", "claim": "Hesai (China) JT128 robot lidar: 128 channels, 189 x 360 degree FOV, 40 m at 10% reflectivity (60 m max), 1,152,000 points/s, 265 g, 62.5 mm diameter by 73 mm tall; price not listed. Small near-field lidar sized for humanoids/robots.", "value": "128 ch; 189x360 deg; 40 m; 265 g", "scope": "Hesai JT128", "country": "China", "country_label": "China", "vendor_reported": true, "status": "verified", "sources": [ { "url": "https://www.hesaitech.com/product/jt128/", "captured_at": "2026-10-26", "quote": "128 channels ... 189° x 360° ... 265 g ... Φ 62.5 x H 73 mm" } ] }, { "id": "comp-55", "group": "Sensors", "supply_chain_category": "sensors", "subsystem": "encoders", "fact_type": "spec", "claim": "RLS (Slovenia, Renishaw group) humanoid application page recommends AksIM-2 (hollow ring, 29-80 mm OD, up to 20-bit) for large joints, Orbis (14-bit) for space-limited rotating joints and RM08 (8 mm body, 12-bit) for fingers; all magnetic non-contact. The page does not mention dual-encoder joints.", "value": "AksIM-2 20-bit; Orbis 14-bit; RM08 12-bit", "scope": "humanoid joint encoders", "country": "Other", "country_label": "Slovenia", "vendor_reported": true, "status": "verified", "sources": [ { "url": "https://www.rls.si/eng/applications/humanoid", "captured_at": "2026-10-26", "quote": "RLS non-contact magnetic encoders ensure high accuracy and reliability." } ] }, { "id": "comp-56", "group": "Sensors", "supply_chain_category": "sensors", "subsystem": "encoders", "fact_type": "spec", "claim": "MDT (MultiDimension, China) launched the TMR3111D TMR magnetic angle sensor on 1 July 2026 for joint position sensing in humanoid and quadruped robots: SPI absolute output with 23-bit internal format, ABZ up to 4,096 PPR, 3x3x0.75 mm DFN10 package.", "value": "23-bit internal; 3x3x0.75 mm", "scope": "humanoid joint encoder IC", "country": "China", "country_label": "China", "vendor_reported": true, "status": "verified", "sources": [ { "url": "https://www.aap.com.au/aapreleases/cision20260630ae94820/", "captured_at": "2026-10-26", "quote": "Compact DFN10L package (3×3×0.75mm)" } ] }, { "id": "comp-57", "group": "Sensors", "supply_chain_category": "sensors", "subsystem": "IMU", "fact_type": "spec", "claim": "ASENSING (China) automotive-grade IMU 5146 to supply Galbot humanoids (announced 13 Mar 2026): 0.05 degree attitude precision, 1,000 Hz output, EtherCAT/RS232/RS485.", "value": "0.05 deg; 1 kHz", "scope": "Galbot", "country": "China", "country_label": "China", "vendor_reported": true, "status": "verified", "sources": [ { "url": "https://autonews.gasgoo.com/articles/news/asensing-partners-with-galbot-on-ai-robots-2033464041584877569", "captured_at": "2026-10-26", "quote": "attitude measurement precision reaching 0.05° and an output frequency of 1,000 Hz" } ] }, { "id": "comp-23", "group": "Compute", "supply_chain_category": "compute", "subsystem": "compute", "fact_type": "technology_choice", "claim": "Musk (Q3 2026 call): Tesla plans to put its in-house AI5 chip in Optimus around April 2027, about 10x the compute of AI4.", "value": "AI5 in Optimus ~Apr 2027; ~10x AI4 compute", "scope": "Tesla Optimus", "country": "US", "country_label": "US", "vendor_reported": true, "status": "verified", "sources": [ { "url": "https://singjupost.com/transcript-tesla-tsla-q3-2026-earnings-call/", "captured_at": "2026-10-26", "quote": "to integrate our Tesla AI5 chip into Optimus, which if things go according to plan, will be in approximately April of next year. And that's a giant improvement in capability with an order of magnitude more sort of more operations per second." } ] }, { "id": "comp-7", "group": "Power and thermal", "supply_chain_category": "batteries-power", "subsystem": "batteries & power", "fact_type": "spec", "claim": "Morgan Stanley's teardown lists a 2.3 kWh, 52 V chest battery pack at about $300 (~0.5% of the bill of materials).", "value": "2.3 kWh, 52 V, ~$300", "scope": "Tesla Optimus Gen2", "country": null, "country_label": null, "vendor_reported": false, "status": "verified", "sources": [ { "url": "https://web.archive.org/web/2026/https://advisor.morganstanley.com/john.howard/documents/field/j/jo/john-howard/The_Humanoid_100_-_Mapping_the_Humanoid_Robot_Value_Chain.pdf", "captured_at": "2026-10-26", "quote": "Battery Pack 2.3KWh, 52v ... ~US$0.3k (~0.5% of total)" } ] }, { "id": "comp-28", "group": "Power and thermal", "supply_chain_category": "batteries-power", "subsystem": "batteries & power", "fact_type": "spec", "claim": "Figure 03's battery is designed and built in-house: 2.3 kWh, five hours at peak performance, structural enclosure, UN38.3 certified.", "value": "2.3 kWh; 5 h peak run time", "scope": "Figure 03", "country": "US", "country_label": "US", "vendor_reported": true, "status": "verified", "sources": [ { "url": "https://www.figure.ai/news/f-03-battery-development", "captured_at": "2026-10-26", "quote": "2.3 kWh enables 5 hours of run time at peak performance" } ] }, { "id": "comp-35", "group": "Power and thermal", "supply_chain_category": "batteries-power", "subsystem": "batteries & power", "fact_type": "technology_choice", "claim": "UBTECH Walker S2 swaps its own battery in about three minutes (dual-battery design) and carries 15 kg.", "value": "autonomous battery swap in 3 minutes; 15 kg payload", "scope": "UBTECH Walker S2", "country": "China", "country_label": "China", "vendor_reported": true, "status": "verified", "sources": [ { "url": "https://www.ubtrobot.com/en/humanoid/products/walker-s2", "captured_at": "2026-10-26", "quote": "swap battery autonomously within 3 minutes" } ] }, { "id": "comp-54", "group": "Power and thermal", "supply_chain_category": "batteries-power", "subsystem": "thermal management", "fact_type": "chokepoint", "claim": "HKUST (Guangzhou) lab page states joint-module heat generation severely restricts long-duration, high-load operation of humanoids and that active cooling of joint modules is essential; lab project is integrated liquid-cooling joint actuators. The page gives no temperature or power numbers.", "value": "joint heat is the endurance limit (qualitative)", "scope": "humanoid joint modules", "country": "China", "country_label": "China", "vendor_reported": false, "status": "verified", "sources": [ { "url": "https://amdt.hkust-gz.edu.cn/projects/Liquid_Cooling_Joint_Actuators/", "captured_at": "2026-10-26", "quote": "Excessive heat generation in joint modules of humanoid robots severely restricts their long-duration and high-load operation." } ] }, { "id": "comp-61", "group": "Wiring", "supply_chain_category": null, "subsystem": "cabling & connectors", "fact_type": "technology_choice", "claim": "The standard fix for joint cabling is a hollow-shaft actuator so wires lie on the axis and only twist. A Unitree G1 teardown found all motor cables run through the joint centres, with XT30 connectors and CAN bus.", "value": "bore ~1/3 of OD; XT30 + CAN on G1; igus bus cable tested >15M torsion cycles", "scope": "humanoid joints", "country": null, "country_label": null, "vendor_reported": false, "status": "verified", "sources": [ { "url": "https://humanoid.guide/hollow-shaft-cable-routing-humanoid-joints/", "captured_at": "2026-10-26", "quote": "Put a cable on the rotation axis of a joint and it only ever twists. Move it 30 mm off the axis and every swing of the joint also bends it, stretches it and drags it across whatever it touches." } ] }, { "id": "comp-62", "group": "Wiring", "supply_chain_category": null, "subsystem": "cabling & connectors", "fact_type": "chokepoint", "claim": "Menlo's open Asimov humanoid gives every actuator two XT30 ports carrying power and CAN, daisy-chained over six CAN branches; shared supply sags when several joints pull peak current.", "value": "4 conductors per joint port; 6 CAN branches", "scope": "Menlo Asimov; Unitree G1", "country": "US", "country_label": "US", "vendor_reported": false, "status": "verified", "sources": [ { "url": "https://humanoid.guide/hollow-shaft-cable-routing-humanoid-joints/", "captured_at": "2026-10-26", "quote": "Every actuator has two electrically identical XT30 (2+2) ports carrying exactly four conductors: BATT+, BATT–, CAN_H and CAN_L." } ] }, { "id": "comp-3", "group": "Assembly and whole robots", "supply_chain_category": "final-assembly", "subsystem": "whole-robot BOM", "fact_type": "cost_share", "claim": "Morgan Stanley puts the Optimus Gen2 \"brain\" (compute hardware) at about $2k per robot (~4% of the total); battery plus structural parts about $419 (~0.8%).", "value": "brain ~4%; battery+structure ~0.8%", "scope": "Tesla Optimus Gen2", "country": "US", "country_label": "US", "vendor_reported": false, "status": "verified", "sources": [ { "url": "https://web.archive.org/web/2026/https://advisor.morganstanley.com/john.howard/documents/field/j/jo/john-howard/The_Humanoid_100_-_Mapping_the_Humanoid_Robot_Value_Chain.pdf", "captured_at": "2026-10-26", "quote": "For the \"brain\" itself, we estimate the hardware cost is ~$2k/ humanoid (~4% of total). Note, this excludes any potential software cost (e.g., FSD training costs). ... Other supporting systems including the battery and various structural parts cost ~ $419/humanoid, (~0.8% of total)" } ] }, { "id": "comp-10", "group": "Assembly and whole robots", "supply_chain_category": "final-assembly", "subsystem": "whole-robot BOM", "fact_type": "price", "claim": "Morgan Stanley uses the Unitree G1 (~$16k; half the DoF, no linear actuators) as its low-end comparator and Agility Digit (~$250k estimated selling price, 2025) as the high end.", "value": "G1 ~$16k; Digit ~$250k", "scope": "Unitree G1; Agility Digit", "country": "China", "country_label": "China", "vendor_reported": false, "status": "verified", "sources": [ { "url": "https://web.archive.org/web/2026/https://advisor.morganstanley.com/john.howard/documents/field/j/jo/john-howard/The_Humanoid_100_-_Mapping_the_Humanoid_Robot_Value_Chain.pdf", "captured_at": "2026-10-26", "quote": "China's Unitree announced its G-1 humanoid would be priced at ~$16k with a simplified algorithm module, halved degrees- of-freedom, lack of linear actuators, shorter battery life, and lower carrying capacity." } ] }, { "id": "comp-24", "group": "Assembly and whole robots", "supply_chain_category": "final-assembly", "subsystem": "final assembly", "fact_type": "spec", "claim": "Tesla (Q3 2026 call) says early Optimus production has started on its Fremont high-volume line, with external deliveries expected around end-2027, and that it must build a large global supply chain.", "value": "Early production Oct 2026; deliveries ~end-2027", "scope": "Tesla Optimus", "country": "US", "country_label": "US", "vendor_reported": true, "status": "verified", "sources": [ { "url": "https://singjupost.com/transcript-tesla-tsla-q3-2026-earnings-call/", "captured_at": "2026-10-26", "quote": "we've started what I'd call early production on our high-volume line in Fremont." } ] }, { "id": "comp-27", "group": "Assembly and whole robots", "supply_chain_category": "final-assembly", "subsystem": "final assembly", "fact_type": "spec", "claim": "Figure says BotQ's first line can build up to 12,000 humanoids a year, with a target of 100,000 robots over four years.", "value": "12,000/yr line capacity; 100,000 over 4 years", "scope": "Figure BotQ", "country": "US", "country_label": "US", "vendor_reported": true, "status": "verified", "sources": [ { "url": "https://www.figure.ai/news/introducing-figure-03", "captured_at": "2026-10-26", "quote": "BotQ's first-generation manufacturing line will initially be capable of producing up to 12,000 humanoid robots per year, with the goal of producing a total of 100,000 robots over the next four years." } ] }, { "id": "comp-29", "group": "Assembly and whole robots", "supply_chain_category": "final-assembly", "subsystem": "whole-robot spec", "fact_type": "spec", "claim": "Unitree G1 spec sheet: about 35 kg, 23 DoF (EDU 23-43), max knee torque 90 N.m (EDU 120), 13-string Li battery 9000 mAh, about 2 h, 8-core CPU, US $13.5K ex tax/shipping; EDU offers optional force-controlled three-finger hand (Dex3-1) and tactile arrays.", "value": "23 DoF; 35 kg; 90 N.m knee; ~2 h; $13.5K", "scope": "Unitree G1", "country": "China", "country_label": "China", "vendor_reported": true, "status": "verified", "sources": [ { "url": "https://www.unitree.com/g1", "captured_at": "2026-10-26", "quote": "Total degrees of freedom | 23 | 23 - 43 ... Max knee joint torque | 90N.m | 120N.m ... Battery life | About 2h ... Price (tax and shipping excluded) | US $13.5K" } ] }, { "id": "comp-30", "group": "Assembly and whole robots", "supply_chain_category": "final-assembly", "subsystem": "whole-robot spec", "fact_type": "spec", "claim": "Unitree H2 spec sheet: 1820 mm, about 70 kg, 31 DoF, up to 360 N.m leg-joint and 120 N.m arm-joint torque, 15 Ah (0.972 kWh) battery, about 3 h, Intel Core i5/i7 PCs plus high-power module option (page cites a 2070 TOPS chip and Thor), $29,900 base without dexterous hand; EDU has dexterous hand options.", "value": "31 DoF; 70 kg; 360 N.m; 0.972 kWh; $29,900", "scope": "Unitree H2", "country": "China", "country_label": "China", "vendor_reported": true, "status": "verified", "sources": [ { "url": "https://www.unitree.com/H2", "captured_at": "2026-10-26", "quote": "15Ah(0.972KWh) ... About 3h ... \"$29,900\"; 360N.m joint torque; 31 total joints" } ] }, { "id": "comp-31", "group": "Assembly and whole robots", "supply_chain_category": "final-assembly", "subsystem": "whole-robot spec", "fact_type": "price", "claim": "Unitree R1 sells from $4,900 (R1 AIR, 20 DoF) and $5,900 (R1, 26 DoF), about one hour of battery: below the $13k-$17k bill of materials banks project for 2030.", "value": "$4,900-$5,900; 20-26 DoF", "scope": "Unitree R1", "country": "China", "country_label": "China", "vendor_reported": true, "status": "verified", "sources": [ { "url": "https://www.unitree.com/R1", "captured_at": "2026-10-26", "quote": "R1 AIR $4,900; R1 $5,900; Degrees of freedom 20 / 26 / 26-40; battery life About 1h" } ] }, { "id": "comp-32", "group": "Assembly and whole robots", "supply_chain_category": "final-assembly", "subsystem": "whole-robot spec", "fact_type": "spec", "claim": "Boston Dynamics product Atlas page: 56 DoF, 1.9 m, 90 kg, 50 kg instantaneous / 30 kg sustained payload, 2.3 m reach, 4 h battery with autonomous self-swap, tactile sensing and 360-degree camera view.", "value": "56 DoF; 90 kg; 4 h; self-swapping battery", "scope": "Boston Dynamics Atlas (product)", "country": "US", "country_label": "US", "vendor_reported": true, "status": "verified", "sources": [ { "url": "https://bostondynamics.com/products/atlas/", "captured_at": "2026-10-26", "quote": "56 ... 1.9 m (6.2 ft) ... 90 kg (198 lbs) ... 50 kg (110 lbs) ... 30 kg (66 lbs) ... 4 hours ... Atlas autonomously navigates to a charging station and swaps out its own battery" } ] }, { "id": "comp-34", "group": "Assembly and whole robots", "supply_chain_category": "final-assembly", "subsystem": "whole-robot spec", "fact_type": "spec", "claim": "1X NEO product page: 66 lb, 5 ft 6 in, lifts 154 lb and carries 55 lb, 4 h battery, low-inertia tendon drives (torque accuracy 2%, 95% backdrivability, 0.1 mm repeatability), 22x2 hand DoF, 22 dB audible noise; page lists a $200 deposit and no retail price.", "value": "30 kg class; tendon drive; hand 22 DoF each; 4 h", "scope": "1X NEO", "country": "US", "country_label": "US", "vendor_reported": true, "status": "verified", "sources": [ { "url": "https://www.1x.tech/neo", "captured_at": "2026-10-26", "quote": "Low inertia tendon drives ... 22x2 ... Weight 66 lbs ... Battery run-time 4h" } ] }, { "id": "comp-50", "group": "Assembly and whole robots", "supply_chain_category": "final-assembly", "subsystem": "structure", "fact_type": "technology_choice", "claim": "Morgan Stanley (citing Tesla 2022 AI Day): Optimus Gen2 shed 10 kg without losing structural performance, mainly via lightweight PEEK structural parts and high-power-density actuators; humanoid exteriors are generally aluminium alloys and plastics; reducers are mainly harmonic and planetary with RV (cycloidal) as a possible alternative; Optimus uses frameless torque motors in the body and coreless motors in hands.", "value": "-10 kg Gen2; PEEK structural parts", "scope": "Tesla Optimus Gen2", "country": null, "country_label": null, "vendor_reported": false, "status": "verified", "sources": [ { "url": "https://web.archive.org/web/2026/https://advisor.morganstanley.com/john.howard/documents/field/j/jo/john-howard/The_Humanoid_100_-_Mapping_the_Humanoid_Robot_Value_Chain.pdf", "captured_at": "2026-10-26", "quote": "The Optimus Gen2 loses 10kg without sacrificing its structure and performance, primarily from the use of lightweight material such as PEEK (Polyether Ether Ketone) and high power density actuators." } ] } ] } } }