The models, tests, research and parts behind physical AI, from labs in the US, China and elsewhere.
| Section | US | China | Other | Total |
|---|---|---|---|---|
| Foundation models | 29 | 16 | 4 | 49 |
| World models | 9 | 9 | 1 | 19 |
| Benchmarks (by maintainer) | 10 | 9 | 5 | 24 |
| Datasets | 9 | 4 | 1 | 14 |
Builder, open weights and published scores. Flagged scores are the developer's own.
| Model | Developer | Released | Weights | What it claims | Source |
|---|---|---|---|---|---|
| Isaac GR00T N2World-action model | NVIDIAUS | Previewed 2026-03-16; not released | Closed | Preview of a world-action successor to the N1.x VLAs, based on NVIDIA's DreamZero research; slated for end of 2026.
No public checkpoint. The ranking claims are NVIDIA's own. |
nvidianews.nvidia.com read Oct 26, 2026 |
| Gemini 4 Argon (robotics use)Embodied-reasoning VLM | Google DeepMind / GoogleUS | 2026-10-13 | Closed | 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.
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. |
developers.googleblog.com read Oct 26, 2026 |
| Dyna-2.1World-action model | Dyna RoboticsUS | 2026-09-28 | Closed | Full-stack physical agent (100 Hz whole-body controller + ~5 Hz Dyna-2 policy + VLM workflow orchestrator) doing hour-long loco-dexterous laundry workflows.
Dyna's own arithmetic: 95% per-step success over about 79 steps leaves a 1.7% chance of finishing unaided. |
dyna.co read Oct 26, 2026 |
| Helix 2.5Dual-system VLA | Figure AIUS | 2026-09-17 | Closed | Claims zero-shot generalization to 30 unseen homes from human-video (Index) pretraining.
The 56% is measured against Figure's own policy trained from scratch (9%), not against another lab's model. |
figure.ai read Oct 26, 2026 |
| GEN-1.5VLA | Generalist AIUS | 2026-08-19 | Closed | Adds one-shot in-context task learning ('learn a new task in seconds, from a single example').
|
generalistai.com read Oct 26, 2026 |
| Skild AI S1 (powers Skild Brain)VLA | Skild AIUS | 2026-08-18 | Closed | Built from the ground up as an in-context learner: new tasks from one demonstration video without retraining.
|
skild.ai read Oct 26, 2026 |
| Dyna-2World-action model | Dyna RoboticsUS | 2026-08-15 | Closed | Claims a 1-million-hour scaling law for world-action models; one-step video sampling 10,203 ms -> 110 ms on one H100.
|
dyna.co read Oct 26, 2026 |
| Gemini Robotics 2 (VLA) + On-Device 2VLA | Google DeepMindUS | 2026-07-30 | Closed | 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.
The VLA and On-Device 2 are early access for partners. |
deepmind.google read Oct 26, 2026 |
| Gemini Robotics-ER 2Embodied-reasoning VLM | Google DeepMindUS | 2026-07-30 | Closed | Upgrade over ER 1.6 with safety-instruction and controlling-a-VLA evaluations.
Google gives no ERQA figure for ER 2. |
blog.google read Oct 26, 2026 |
| ACT-2VLA | Sunday RoboticsUS | 2026-07-17 | Closed | Claims 99.1% zero-shot laundry folding in unseen homes with a single checkpoint; socks/underwear out of scope.
A preview; the number of evaluation homes is not given. |
sunday.ai read Oct 26, 2026 |
| Model | Developer | Released | Weights | What it claims | Source |
|---|---|---|---|---|---|
| pi0.7VLA | Physical IntelligenceUS | 2026-04-16 | Closed | Steerable generalist prompted by language, metadata, control modality and visual subgoals; first signs of compositional generalization.
Weights are not stated as released. |
pi.website read Oct 26, 2026 |
| Gemini Robotics-ER 1.6Embodied-reasoning VLM | Google DeepMindUS | 2026-04-14 | Closed | Public ER model that adds agentic-vision instrument reading.
|
deepmind.google read Oct 26, 2026 |
| GEN-1VLA | Generalist AIUS | 2026-04-02 | Closed | Claims production-grade reliability (99%) on simple physical tasks from ~1 hour of per-task robot data.
Early-access partners only. |
generalistai.com read Oct 26, 2026 |
| Isaac GR00T N1.7VLA · 3B | NVIDIAUS | 2026-03-16 | Open (restricted license) | First GR00T release with commercial licensing; no task-success table on HF card.
Commercial license; the model card gives no task-success table. |
huggingface.co read Oct 26, 2026 |
| Helix 02Dual-system VLA · 10M (S0); S1/S2 not stated | Figure AIUS | 2026-01-27 | Closed | Adds System 0 whole-body controller (1 kHz) beneath S1 (200 Hz) and S2.
|
figure.ai read Oct 26, 2026 |
| RhoVLA · 5B | Microsoft ResearchUS | 2026 | Open | Microsoft's open-weight 5B bimanual VLA family with in-house embodiment-specific variants.
The project page names no license. |
microsoft.github.io read Oct 26, 2026 |
| pi0.6 / pi*0.6 (Recap)VLA · 5B VLM backbone | Physical IntelligenceUS | 2025-11-17 | Closed | First VLA from PI trained with RL from its own deployment experience (Recap: advantage-conditioned policies).
Weights are not stated as released. |
pi.website read Oct 26, 2026 |
| GEN-0VLA · 10B+ | Generalist AIUS | 2025-11-04 | Closed | Scaling-law claim: power-law link between pretraining data and post-training performance, phase transition near 7B. No published scores. |
generalistai.com read Oct 26, 2026 |
| Gemini Robotics 1.5 + ER 1.5VLA | Google DeepMindUS | 2025-09-25 | Closed | 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.
The VLA is partner-only; ER 1.5 is public via the Gemini API. |
deepmind.google read Oct 26, 2026 |
| Atlas LBMVLA | Boston Dynamics / Toyota Research InstituteUS | 2025-08-20 | Closed | Single LBM controls hands and feet of Atlas through packing/sorting with recovery from disruptions; company says new skills need no new code. No published scores. Press page only; no model card or paper. |
Snippet-only pressroom.toyota.com read Oct 26, 2026 |
| TRI Large Behavior ModelVLA | Toyota Research InstituteUS | 2025-07-07 | Closed | Large controlled study finding performance predictably increases with pretraining scale and diversity. No published scores. |
arxiv.org read Oct 26, 2026 |
| Dyna-1 (Dynamism v1)VLA | Dyna RoboticsUS | 2025-06-24 | Closed | First Dyna foundation model: continuous 24-hour autonomous napkin folding.
Earlier press reported 700+ napkins at 99%; the primary page says 850+ at 99.4%. |
dyna.co read Oct 26, 2026 |
| Redwood AIVLA · 160M | 1X TechnologiesUS | 2025-06-10 | Closed | Small 160M model running at about 5 Hz on NEO's onboard GPU. No published scores. Weights are not stated as released. |
1x.tech read Oct 26, 2026 |
| SmolVLAVLA | Hugging FaceUS/France | 2025-06-02 | Open | Compact open VLA with an asynchronous inference stack decoupling action prediction from execution. No published scores. |
arxiv.org read Oct 26, 2026 |
| pi0.5VLA | Physical IntelligenceUS | 2025-04-22 | Open | First end-to-end learned system shown doing long-horizon kitchen/bedroom cleaning in entirely new homes via heterogeneous co-training. No published scores. |
arxiv.org read Oct 26, 2026 |
| OpenVLA-OFTVLA | Stanford UniversityUS | 2025-02-27 | Open | Fine-tuning recipe (parallel decoding, action chunking, continuous actions) lifting OpenVLA LIBERO from 76.5% to 97.1% with 26x higher throughput.
|
arxiv.org read Oct 26, 2026 |
| HelixDual-system VLA · 7B (S2) + 80M (S1) | Figure AIUS | 2025-02-20 | Closed | First humanoid VLA with a 7-9 Hz VLM (S2) driving a 200 Hz visuomotor policy (S1).
|
figure.ai read Oct 26, 2026 |
| Covariant RFM-1VLA | Covariant (licensed to Amazon, 2024)US | 2024-03 (RFM-1) | Closed | Warehouse-picking foundation model; Amazon took a non-exclusive license to RFM-1 in August 2024 and hired the founders. No published scores. No later Covariant model was found. |
Snippet-only aboutamazon.com read Oct 26, 2026 |
| Isaac GR00T N1.5VLA · 3B | NVIDIAUS | Not disclosed | Open (restricted license) | Medium-sized cross-embodiment model; license is non-commercial (superseded by N1.7 Open Model License). No published scores. Non-commercial license; superseded by N1.7. |
huggingface.co read Oct 26, 2026 |
| Model | Developer | Released | Weights | What it claims | Source |
|---|---|---|---|---|---|
| ACE-Brain-0.6Embodied-reasoning VLM · 9B (fine-tune of Qwen3.5-9B) | ACE Robotics (Daxiao Robotics)China | 2026-10-19 | Closed | SenseTime co-founder Xiaogang Wang's company; 9B spatial-grounding VLM claiming to beat GPT-6 Astra on grounding.
Company-reported via secondary press; no primary page or weights. |
Snippet-only technode.com read Oct 26, 2026 |
| Natura-1VLA · 50B mixture-of-experts | Nature WillChina | 2026-10-13 | Unknown | Claims first 50B-parameter model for 20-DoF dexterous hands, pretrained on over a million hours of human egocentric data, 100 ms control loop.
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. |
naturewill.ai read Oct 26, 2026 |
| Xiaomi-Robotics-1VLA · 5B (per repo name) | Xiaomi RoboticsChina | 2026-08-05 | Open | Apache-2.0 open 5B VLA on a Qwen3-VL-4B-Instruct backbone pre-trained on 100K hours of UMI data. No published scores. Xiaomi says it ranked first on RoboCasa365 and RoboDojo; the model card gives no numbers, so none is shown. |
huggingface.co read Oct 26, 2026 |
| Qwen-RobotManipVLA | Alibaba (Qwen)China | 2026-06-16 | Closed | Part of Qwen's first robot suite (RobotNav, RobotManip, RobotWorld).
Enterprise pilot only; no weights release found. |
Snippet-only forklog.com read Oct 26, 2026 |
| Genie Operator-2 (GO-2)Dual-system VLA | AgiBotChina | 2026-04-17 | Closed | Action chain-of-thought plus asynchronous dual-system architecture; core work accepted to CVPR 2026 and ACL 2026.
Weights availability is not stated. |
agibot.com read Oct 26, 2026 |
| HY-Embodied-0.5Embodied-reasoning VLM · 2B and 32B | Tencent Hunyuan / Robotics XChina | 2026-04-09 | Partly open | Only the 2B edge model's weights are stated as open; 32B claimed comparable to Gemini 3.0 Pro.
Only the 2B model's weights are open. ERQA scores across labs use different model sizes and setups, so they are not directly comparable. |
ai.gitcode.com read Oct 26, 2026 |
| Xiaomi-Robotics-0VLA | Xiaomi RoboticsChina | 2026-02-13 | Open | Predecessor to Xiaomi-Robotics-1; real-time execution on consumer GPU claimed. No published scores. Press names LIBERO, CALVIN and SimplerEnv; the abstract gives no numbers. |
arxiv.org read Oct 26, 2026 |
| RDT2VLA | Tsinghua University (TSAIL)China | 2026-02-03 | Open | Scales UMI handheld-gripper data to 10,000+ hours for zero-shot cross-embodiment transfer. No published scores. The arXiv page does not itself state that weights are released; the open license comes from secondary reporting. |
arxiv.org read Oct 26, 2026 |
| UnifoLM-VLA-0VLA | Unitree RoboticsChina | 2026-01-29 | Open (restricted license) | Open humanoid-manipulation VLA with a LIBERO fine-tune; 12 manipulation categories from a single policy per HF page. No published scores. 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). |
github.com read Oct 26, 2026 |
| Spirit v1.5VLA | Spirit AI (Qianxun Intelligence)China | 2026-01-14 | Open | Model that topped the RoboChallenge real-robot leaderboard in January 2026, per company statements to state media.
A third-party catalog says Spirit was later removed from the RoboChallenge leaderboard after a methodology change. This could not be confirmed. |
Snippet-only news.cgtn.com read Oct 26, 2026 |
| Model | Developer | Released | Weights | What it claims | Source |
|---|---|---|---|---|---|
| Galaxea G0PlusDual-system VLA · 3B (G0Tiny: 250M) | Galaxea AIChina | 2026-01-04 | Partly open | Open-weight 3B VLA released under a non-commercial G0 PLUS Community License (post-2026-01-04 commits). No published scores. Newer releases use a non-commercial license. |
raw.githubusercontent.com read Oct 26, 2026 |
| Thinker-4BEmbodied-reasoning VLM · 4B | UBTech RoboticsChina | 2025-12 | Open (restricted license) | Open 4B VLM on Qwen3-VL-4B-Instruct; license is CC BY-NC-SA 4.0 (non-commercial). No published scores. The model card claims records on seven benchmarks but names none. Non-commercial license. |
huggingface.co read Oct 26, 2026 |
| WALL-OSS / WALL-OSS-0.5VLA | X Square RobotChina | 2025-09-15 (0.5: 2026-05) | Open | Open VLA family; 0.5 positioned as deployment-ready; Apache-2.0 per repo. No published scores. A press figure versus pi0.5 is not shown because it could not be traced to a primary page. |
github.com read Oct 26, 2026 |
| GR-3VLA | ByteDance SeedChina | 2025-07-21 | Closed | Large VLA co-trained on VL data, VR human trajectories, and robot data on ByteDance's own ByteMini. No published scores. |
arxiv.org read Oct 26, 2026 |
| AgiBot GO-1 (Genie Operator-1)VLA | AgiBot (Zhiyuan)China | 2025-03-09 | Open | Generalist policy using latent action representations, released with the large AgiBot World dataset.
|
arxiv.org read Oct 26, 2026 |
| GraspVLA / AstraBrainVLA | GalbotChina | 2025 (GraspVLA) | Partly open | GraspVLA is the only Galbot model with released weights; AstraBrain is described as an end-to-end brain/cerebellum model. No published scores. Release dates conflict across sources. GraspVLA weights are non-commercial (CC BY-NC 4.0). |
Snippet-only roboticscenter.ai read Oct 26, 2026 |
| Model | Developer | Released | Weights | What it claims | Source |
|---|---|---|---|---|---|
| Neuron AIVLA | NEURA RoboticsGermany | 2026-10-14 | Closed | Perception-Language-Action model fusing depth, tactile, force-torque and sound with memory; runs on-robot; 'time-to-skill' named as key metric. No published scores. No benchmarks or weights statement. |
implicator.ai read Oct 26, 2026 |
| FLUX 3 ActionWorld-action model · 7B | Black Forest LabsGermany | 2026-09-23 | Open (restricted license) | Open-weight 7B joint video+action flow-matching model; policies for DROID and low-cost SO-101.
The page gives 42.2% in its text and 42.92% in its table. Weights carry a custom license. |
bfl.ai read Oct 26, 2026 |
| FLUX-mimicWorld-action model | mimic robotics (with Black Forest Labs)Switzerland | 2026-07-23 | Closed | Video-action model that attaches an action decoder to a generative video model; 30 minutes of robot data to fine-tune.
Release date differs across sources (July 23 vs August 13, 2026). |
Snippet-only businesswire.com read Oct 26, 2026 |
| RLDX-1VLA | RLWRLDKorea | 2026-05-05 | Unknown | Dexterous-manipulation policy built on a Multi-Stream Action Transformer, tested on a high-DoF humanoid.
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. |
arxiv.org read Oct 26, 2026 |
Models that learn how the physical world behaves from video. Some simulate worlds to train or test a robot; others drive the robot directly.
The model learns physics from video and is used to train the robot's policy or to choose its actions directly.
| Model | Developer | Released | Weights | What it claims | Source |
|---|---|---|---|---|---|
| VPP2world-action model | RoboteraChina | 2026-10-09 | Open | 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. Developer-contributed leaderboard entry. Code at github.com/roboterax/video-prediction-policy-2. |
technode.global, robodojo-benchmark.com read Oct 26, 2026 |
| GE-Act 2.0world-action model | AgiBot (Zhiyuan)China | 2026-09-04 | Closed | 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. Developer's own paper; weight release not confirmed. |
arxiv.org read Oct 26, 2026 |
| Praxis-1world-action model | RunwayUS | 2026-09 | Closed | 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. Figures are vendor-reported. |
runway.com read Oct 26, 2026 |
| Daimon-TWMtactile world model | Daimon RoboticsChina | 2026-08 | Closed | Tactile world model for force, friction, deformation and slip, aimed at planning and real-time control. Shown at IROS 2026. Launch date disputed in trade press; performance claims are company-reported. |
Snippet-only robotics247.com read Oct 26, 2026 |
| LingBot-VA 2.0world-action model | Robbyant (Ant Group)China | 2026-07-09 | Unknown | 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. Scores are the developer's own (v1 paper, arXiv 2601.21998). |
webull.com read Oct 26, 2026 |
| MotuBrainworld-action model | ShengShu TechnologyChina | 2026-04-30 | Closed | 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. Paper result from the developer; weight release not confirmed. |
arxiv.org read Oct 26, 2026 |
| WALL-Bworld-action model | X Square RobotChina | 2026-04-21 | Partly open | 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. Company-reported demo; WALL-B weight status not confirmed. |
Snippet-only macaubusiness.com read Oct 26, 2026 |
| DreamZeroworld-action model | NVIDIA GEARUS | 2026-02-17 | Open | 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. Generalisation figure is the developer's own. Listed as open source on RoboArena, where it sits second (1735). |
arxiv.org read Oct 26, 2026 |
| 1X World Modelvideo world model | 1X TechnologiesUS | 2026-01-12 | Closed | 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. No success rates or benchmark numbers published. |
aap.com.au read Oct 26, 2026 |
| General Intuition (MIRA)interactive/generative environment | General IntuitionUS | 2025-10-16 | Closed | 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. Robotics is a secondary application; the robot claim is company-reported. |
Snippet-only techcrunch.com read Oct 26, 2026 |
The model generates worlds, data or test scenarios; the robot's policy is trained or scored inside it.
| Model | Developer | Released | Weights | What it claims | Source |
|---|---|---|---|---|---|
| Oasis 3interactive/generative environment | DecartUS | 2026-06 | API only | API-accessible interactive world model that generates action-responsive simulation for robot learning, starting with autonomous vehicles. Latency and speed claims are vendor-reported. |
Snippet-only wandb.ai read Oct 26, 2026 |
| Qwen-RobotWorldvideo world model | Alibaba Tongyi Lab (Qwen)China | 2026-06 | Closed | 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). Enterprise pilot only; not broadly available. |
Snippet-only forklog.com read Oct 26, 2026 |
| Genesis World 1.0simulation platform | Genesis AIUS | 2026-05 | Open | Open-source robotics simulator with its own renderer (Nyx) and GPU compute layer (Quadrants). Company claims one hour of real-world testing equals 100 days in simulation; not independently checked. |
Snippet-only app.dealroom.co read Oct 26, 2026 |
| HY-World 2.0interactive/generative environment | Tencent HunyuanChina | 2026-04-15 | Open | 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. Tencent compares it to World Labs' Marble; vendor claim. Robotics is one use case, not a tested benchmark. |
Snippet-only rits.shanghai.nyu.edu read Oct 26, 2026 |
| Genie Envisioner 2.0world-action model | AgiBot (Zhiyuan)China | 2026-04 | Closed | Closed-loop robot video world simulator for training and evaluating robots. A reprint reports first place in a CVPR 2026 WorldArena track; not confirmed against the official board. |
Snippet-only gitcode.csdn.net read Oct 26, 2026 |
| DreamDojovideo world model | NVIDIAUS | 2026-02-06 | Open | 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). The abstract names no benchmark scores. |
arxiv.org read Oct 26, 2026 |
| GAIA-3driving world model | WayveUK | 2025-12-02 | Closed | 15B-parameter generative world model for autonomous-driving validation: generates and scores counterfactual scenarios offline. Not a robot model; included as an adjacent evaluation world model. |
Snippet-only automotiveworld.com read Oct 26, 2026 |
| Marbleinteractive/generative environment | World LabsUS | 2025-11 | API only | Builds 3D environments from images, video or text; World Labs' first commercial product. AMD agreed to acquire World Labs (reported at about $8.2B). No robotics benchmark found; deal details from secondary reports. |
Snippet-only next.ink read Oct 26, 2026 |
Released for synthetic data and evaluation as well as policy pretraining.
| Model | Developer | Released | Weights | What it claims | Source |
|---|---|---|---|---|---|
| Cosmos 3video world model | NVIDIAUS | 2026-06-01 | Open | 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. NVIDIA says it ranks first among open models on four leaderboards but gives no numbers. |
investor.nvidia.com read Oct 26, 2026 |
Leader on each test as of the date shown. An audit of LIBERO and SimplerEnv found only about one in five claimed top scores were statistically significant, so read close scores as ties.
Google DeepMind · US
400 multiple-choice visual questions on spatial, trajectory, action, state-estimation, pointing, multi-view and task reasoning; scored as accuracy.
Metric: accuracy (%)
| Model | Developer | Country | Score | As of |
|---|---|---|---|---|
| Gemini 4 Argon | Google DeepMind | US | 87% Vendor-reported | Oct 13, 2026 |
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.
developers.googleblog.com, arxiv.org, arxiv.org 2, github.com read Oct 26, 2026
Atreya et al. (distributed evaluator network) · US
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)
| Model | Developer | Country | Score | As of |
|---|---|---|---|---|
| Spirit v1.6 (listed as j2-vla) | Spirit AI (inferred from display name) | China | 1788 (±106, 25 evals) | Oct 26, 2026 |
| DreamZero | NVIDIA | US | 1735 (±43, 190 evals) | Oct 26, 2026 |
| Apricot | Not disclosed | Unverified | 1661 (±61, 63 evals) | Oct 26, 2026 |
| pi0.5-DROID | Physical Intelligence | US | 1608 (±31, 745 evals) | Oct 26, 2026 |
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.
robo-arena.github.io, arxiv.org, arxiv.org 2 read Oct 26, 2026
AI MMLab Club (non-profit) with academic partners · Unverified
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
| Model | Developer | Country | Score | As of |
|---|---|---|---|---|
| GO-3 Nano Preview (agent track) | AgiBot | China | 61.90 / 56.42% Vendor-reported | Oct 26, 2026 |
| PyRUA-Lean (agent track) | PKU DAGroup | China | 55.26 / 50.35% Vendor-reported | Oct 24, 2026 |
| HexaShell-Preview (agent track) | HexaFuture | China | 53.18 / 46.85% Vendor-reported | Oct 20, 2026 |
| VPP2 (best non-agent entry) | Robotera | China | 39.26 / 32.26% Vendor-reported | Oct 8, 2026 |
| GPT-6-Astra (agent track) | OpenAI (run by the RoboDojo team) | US | 28.97 / 22.48% | Sep 16, 2026 |
| InternW0-Delta (real-world board) | Shanghai AI Lab | China | 38.9 / 31.7% Vendor-reported | Oct 21, 2026 |
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.
robodojo-benchmark.com, arxiv.org, arxiv.org 2 read Oct 26, 2026
Tianxing Chen et al. · Unverified
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 (%)
| Model | Developer | Country | Score | As of |
|---|---|---|---|---|
| Hy-Embodied-0.5-VLA | Tencent Hunyuan | China | 13.07 / 8.80% | Jul 3, 2026 |
| Spatial Forcing | Academic (Li et al.) | Unverified | 12.38 / 8.04% | Jul 3, 2026 |
| pi0.5 | Physical Intelligence | US | 11.41 / 6.91% | Jul 3, 2026 |
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.
arxiv.org read Oct 26, 2026
Zhou et al., Huazhong University of Science and Technology · China
LIBERO extended with perturbations along four dimensions (manipulated objects, initial states, task instructions, environments) to separate generalization from memorization; success rate.
Metric: success rate
| Model | Developer | Country | Score | As of |
|---|---|---|---|---|
| pi0.5 (best of four tested) | Physical Intelligence | US | 0.97 original; 0.0 under task perturbation; 0.4 under position perturbation | Oct 2025 |
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.
arxiv.org, arxiv.org 2, github.com read Oct 26, 2026
Fei et al. · China
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
| Model | Developer | Country | Score | As of |
|---|---|---|---|---|
| OpenVLA-OFT+ (authors' own fine-tune) | LIBERO-Plus authors | China | 79.6 total Vendor-reported | Oct 2025 |
| OpenVLA-OFT | Stanford | US | 69.6 total | Oct 2025 |
| pi0-FAST | Physical Intelligence | US | 61.6 total | Oct 2025 |
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.
github.com, arxiv.org, arxiv.org 2 read Oct 26, 2026
SimplerEnv: Li, Su, Xiao et al.; audit: Jiang et al. (TTIC, Argonne, UChicago) · US
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
No ranked results shown.
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.
arxiv.org, arxiv.org 2, arxiv.org 3, github.com read Oct 26, 2026
Stanford Vision and Learning Lab · US
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
| Model | Developer | Country | Score | As of |
|---|---|---|---|---|
| Robot Learning Collective | Independent | Unverified | Q 0.2599; full-task success 0.1240 | Nov 14, 2025 |
| Comet | NVIDIA Research | US | Q 0.2514; full-task success 0.1140 | Nov 17, 2025 |
| SimpleAI Robot | Beijing Simple AI Technology | China | Q 0.1591; full-task success 0.1080 | Nov 17, 2025 |
| The North Star | Huawei | China | Q 0.1204; full-task success 0.0760 | Nov 16, 2025 |
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.
behavior.stanford.edu, behavior.stanford.edu 2, huggingface.co read Oct 26, 2026
X Square Robot · China
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
| Model | Developer | Country | Score | As of |
|---|---|---|---|---|
| wall-oss-preview | X Square Robot | China | 53.9% in-distribution; 34.3% out-of-distribution Vendor-reported | Oct 13, 2026 |
| DreamZero | NVIDIA | US | 4.85% in-distribution | Oct 13, 2026 |
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.
arxiv.org, arxiv.org 2 read Oct 26, 2026
Robocurve PBC · US
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
| Model | Developer | Country | Score | As of |
|---|---|---|---|---|
| Dream-1 | Dream Labs | Unverified | 59.3% | Oct 19, 2026 |
| Gemini 4 Argon | Google DeepMind | US | 48.2% | Oct 13, 2026 |
| GPT-6 Astra | OpenAI | US | 42.5% | Sep 3, 2026 |
| GPT-6.1 Sol | OpenAI | US | 41.7% | Sep 29, 2026 |
| Claude Opus 5.5 | Anthropic | US | 38.2% | Sep 22, 2026 |
| MolmoAct2 | Ai2 | US | 18.9% | May 4, 2026 |
| GR00T N1.7 | NVIDIA | US | 9.5% | Apr 17, 2026 |
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.
robocurve.org, robocurve.org 2, robocurve.org 3, robocurve.org 4 read Oct 26, 2026
Robocurve PBC · US
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
| Model | Developer | Country | Score | As of |
|---|---|---|---|---|
| Claude Fable 5.1 | Anthropic | US | Refused 20/100; carried out 34/100 | Sep 18, 2026 |
| GPT-6 Astra | OpenAI | US | Refused 2/100 (3 for any reason); carried out 60/100 | Sep 18, 2026 |
| MolmoAct2 | Ai2 | US | Refused 0/100; carried out 6/100 | Sep 18, 2026 |
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.
robocurve.org, robocurve.org 2, analyticsinsight.net read Oct 26, 2026
Artificial Analysis · US
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)
| Model | Developer | Country | Score | As of |
|---|---|---|---|---|
| GPT-6 Astra (Codex harness) | OpenAI | US | 69.4 | Oct 14, 2026 |
| GPT-6.1 Sol | OpenAI | US | 64.6 | Oct 14, 2026 |
| Gemini 4 Argon | Google DeepMind | US | 55.3 | Oct 14, 2026 |
| Claude Opus 5.5 | Anthropic | US | 48.8 | Oct 14, 2026 |
| Claude Fable 5.1 | Anthropic | US | 46.3 | Oct 14, 2026 |
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.
artificialanalysis.ai read Oct 26, 2026
Tsinghua University-led team · China
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
| Model | Developer | Country | Score | As of |
|---|---|---|---|---|
| WoVR (Click Bell task) | Jiang et al. | Unverified | 75.0% vs 43.75% baseline | Oct 26, 2026 |
| Ctrl-World (Adjust Bottle task) | Guo et al. | Unverified | 70.7% vs 55.1% baseline | Oct 26, 2026 |
| Cosmos-Predict-2.5 (Adjust Bottle task) | NVIDIA | US | 63.5% | Oct 26, 2026 |
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.
arxiv.org, arxiv.org 2 read Oct 26, 2026
NIST Intelligent Systems Division · US
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
No ranked results shown.
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.
nist.gov read Oct 26, 2026
Beijing organizers · China
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
| Model | Developer | Country | Score | As of |
|---|---|---|---|---|
| Tianzhuo Team (1,500 m) | Tianzhuo Team | China | 2:21.64 | Aug 23, 2026 |
| Tianjiao Team (100 m, large group) | Tianjiao Team | China | 8.64 s | Aug 2026 |
| Tiangong Omni (400 m, small group) | Beijing Innovation Center of Humanoid Robotics | China | 45.66 s | Aug 2026 |
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.
english.beijing.gov.cn, english.beijing.gov.cn 2 read Oct 26, 2026
Beijing E-Town organizers · China
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
| Model | Developer | Country | Score | As of |
|---|---|---|---|---|
| Shandian (Team Monkey King) | Team Monkey King | China | 50:26 | Apr 19, 2026 |
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).
english.beijing.gov.cn read Oct 26, 2026
RoboTwin team · China
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
| Model | Developer | Country | Score | As of |
|---|---|---|---|---|
| GO-3 Nano Preview (co-train) | AgiBot | China | 85.01% Vendor-reported | Oct 25, 2026 |
| PrismWAM | Tsinghua & NVIDIA | China/US | 83.67% Vendor-reported | Oct 23, 2026 |
| CloudRobo-WM | CloudRobo Team | Unverified | 81.95% Vendor-reported | 2026 |
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.
robotwin-platform.github.io, arxiv.org read Oct 26, 2026
Nasiriany et al. (UT Austin, NVIDIA) · US
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
No ranked results shown.
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.
arxiv.org read Oct 26, 2026
Yang et al. · Unverified
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 (%)
| Model | Developer | Country | Score | As of |
|---|---|---|---|---|
| GPT-4o | OpenAI | US | 28.9% average | Jun 5, 2025 |
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.
arxiv.org read Oct 26, 2026
Zhang et al. · China
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
No ranked results shown.
No leader in the paper. Numbers on aggregator sites come from a different harness and are not shown.
arxiv.org read Oct 26, 2026
Open X-Embodiment collaboration (21 institutions) · US-led international
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
No ranked results shown.
A dataset and cross-robot evaluation rather than a ranked board; included as the cross-embodiment data foundation.
arxiv.org read Oct 26, 2026
Mees, Hermann, Rosete-Beas, Burgard (University of Freiburg) · Germany
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)
No ranked results shown.
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.
arxiv.org, arxiv.org 2 read Oct 26, 2026
Hao Su lab (UCSD / Hillbot) · US
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
No ranked results shown.
Primarily a fast GPU simulator and task library; no single capability leaderboard.
arxiv.org read Oct 26, 2026
AgiBot (with Dexmal) · China
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
| Model | Developer | Country | Score | As of |
|---|---|---|---|---|
| PrismBot | vivo | China | Reasoning-to-Action champion | Jun 5, 2026 |
Organizer press release; ranks without scores. 526 teams from 27 countries. World Model track standings were not retrieved.
agibot.com read Oct 26, 2026
Dated evidence by research direction, then datasets and compute.
Figure AI US Sep 17, 2026
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.
figure.ai read Oct 26, 2026
X Square Robot (authors Kang, Shi, Liang, Gan, Liu) China Jun 7, 2026
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.
arxiv.org read Oct 26, 2026
NVIDIA US Feb 25, 2026
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.
huggingface.co read Oct 26, 2026
Dong, Liu, Zhang, Ye, Yuan et al. Unverified Feb 17, 2026
ActionCodec derives design principles for action tokenisers from the VLA-optimisation viewpoint rather than reconstruction fidelity (token overlap, vocabulary redundancy, mutual information).
arxiv.org read Oct 26, 2026
Princeton University with Physical Intelligence US Feb 11, 2026
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.
arxiv.org read Oct 26, 2026
Physical Intelligence US Dec 16, 2025
Physical Intelligence reports human-to-robot transfer emerging in VLAs as they scale.
pi.website read Oct 26, 2026
MindOn China Oct 1, 2026
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.
Snippet-only interestingengineering.com read Oct 26, 2026
Google DeepMind US Jul 30, 2026
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.
blog.google read Oct 26, 2026
Figure AI US Jan 27, 2026
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.
figure.ai read Oct 26, 2026
Galaxea AI China Aug 30, 2025
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).
arxiv.org read Oct 26, 2026
NVIDIA US Oct 26, 2026
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.
nvidia.com read Oct 26, 2026
Dyna Robotics US Aug 15, 2026
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.
dyna.co read Oct 26, 2026
Robbyant (Ant Group) China Jul 9, 2026
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.
webull.com read Oct 26, 2026
NVIDIA GEAR US Feb 17, 2026
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.
arxiv.org read Oct 26, 2026
NVIDIA (Isaac Lab open-source team) US Oct 26, 2026
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.
docs.nvidia.com read Oct 26, 2026
Genesis AI US/France (unverified) May 1, 2026
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.
Snippet-only app.dealroom.co read Oct 26, 2026
Allen Institute for AI (Ai2) US Mar 11, 2026
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.
allenai.org read Oct 26, 2026
AgiBot (Genie Sim team) China Jan 5, 2026
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.
arxiv.org read Oct 26, 2026
HUMXN Research Labs US (unverified) Oct 19, 2026
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).
Snippet-only ecosistemastartup.com read Oct 26, 2026
MR. ROBOT / TROOPERS Malaysia/Korea (unverified) Oct 1, 2026
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.
Snippet-only tech-critter.com read Oct 26, 2026
Figure AI US Aug 25, 2026
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.
figure.ai read Oct 26, 2026
Dyna Robotics US Aug 17, 2026
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.
dyna.co read Oct 26, 2026
XDOF US Jun 17, 2026
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).
xdof.ai read Oct 26, 2026
Mecka AI US Jun 2, 2026
Mecka announced $60M to build data, evaluation and deployment infrastructure for physical AI, including first-person video of human work.
pulse2.com read Oct 26, 2026
Human Archive (Wing VC, NVP Capital, Y Combinator) India/USA May 26, 2026
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.
techcrunch.com read Oct 26, 2026
Lightwheel US/China (unverified) Dec 4, 2025
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.
lightwheel.ai read Oct 26, 2026
Rhoda AI US Oct 13, 2026
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).
humanoidsdaily.com read Oct 26, 2026
Sharpa Singapore/China (HQ not verified) Sep 29, 2026
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.
technode.global read Oct 26, 2026
1X Technologies Norway/USA Jul 9, 2026
1X announced 25-DoF tendon-driven hands for NEO, claiming near human-level dexterity, strength, safety and reliability (July 2026).
1x.tech read Oct 26, 2026
Tesla (Counterpoint Research estimate) US Apr 20, 2026
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.
Snippet-only counterpointresearch.com read Oct 26, 2026
Skild AI US Sep 23, 2026
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.
skild.ai read Oct 26, 2026
NVIDIA NVlabs US Jul 15, 2026
NVIDIA GRAIL converts video-derived human-object interaction into physics-validated Unitree G1 locomanipulation trajectories with RL tracking policies released 2026-07-15.
huggingface.co read Oct 26, 2026
Dyna Robotics US Sep 28, 2026
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.
dyna.co read Oct 26, 2026
Skild AI US Sep 10, 2026
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).
skild.ai read Oct 26, 2026
Skild AI US Aug 18, 2026
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.
skild.ai read Oct 26, 2026
Physical Intelligence US Mar 19, 2026
PI extracts an 'RL Token' from VLA models for fast online RL on precise tasks with only a few hours of data (March 2026).
pi.website read Oct 26, 2026
Physical Intelligence US Nov 17, 2025
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.
pi.website read Oct 26, 2026
XDOF US Oct 23, 2026
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.
xdof.ai read Oct 26, 2026
Robocurve US Sep 18, 2026
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.
popularai.org, robocurve.org read Oct 26, 2026
Yin, Liu, Yang, Li, Xue et al. Unverified Jun 17, 2026
ROBOSHACKLES builds a human-injury-prevention safety dataset for embodied foundation models from real DROID observations with hazard-aware image editing and rollout synthesis.
arxiv.org read Oct 26, 2026
Zhang, Zhang, Fan, Shen, Cai et al. Unverified Apr 24, 2026
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.
arxiv.org read Oct 26, 2026
Pollen Robotics / Hugging Face France/USA Aug 31, 2026
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.
huggingface.co read Oct 26, 2026
Hugging Face LeRobot US/France May 21, 2026
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).
techtimes.com read Oct 26, 2026
Menlo Research US May 7, 2026
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).
opensourceforu.com read Oct 26, 2026
Dyna Robotics US Sep 29, 2026
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.
dyna.co read Oct 26, 2026
Google DeepMind US Jul 30, 2026
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.
blog.google read Oct 26, 2026
Stanford (lab attribution unclear) US Jul 1, 2026
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.
Snippet-only aiweekly.co read Oct 26, 2026
Kim, Pak, Long, Tian, Atanasov et al. Unverified Jun 11, 2026
SERF conditions a VLA on an online spatiotemporal feature map of the environment and robot body to improve long-horizon mobile manipulation.
arxiv.org read Oct 26, 2026
Qian, He, Shi, Xiao, Jiang et al. Unverified Apr 15, 2026
ESCAPE builds a depth-free persistent 3D episodic spatial memory coupled with an adaptive execution policy for long-horizon mobile manipulation.
arxiv.org read Oct 26, 2026
Physical Intelligence US Mar 3, 2026
PI Multi-Scale Embodied Memory (MEM) gives VLAs long- and short-term memory for tasks longer than ten minutes.
pi.website read Oct 26, 2026
Nebius with NVIDIA Netherlands/USA Oct 21, 2026
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).
nebius.com read Oct 26, 2026
Physical Intelligence US Oct 13, 2026
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.
pi.website read Oct 26, 2026
NVIDIA US Sep 22, 2026
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.
blogs.nvidia.com read Oct 26, 2026
StarVLA Community China (Hong Kong) Apr 6, 2026
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.
arxiv.org read Oct 26, 2026
KAIST (Hwangbo lab) Korea Sep 24, 2026
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.
news.kaist.ac.kr, natureasia.com read Oct 26, 2026
Dyna Robotics US Aug 27, 2026
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.
dyna.co read Oct 26, 2026
Google DeepMind US Jul 30, 2026
Gemini Robotics ER 2 introduces multi-robot collaboration so different robots share a workspace and complete workflows a single robot could not.
blog.google read Oct 26, 2026
| Dataset | Country | Released | Size | Human video (h) | License | Source |
|---|---|---|---|---|---|---|
| AgiBot World 2026AGIBOT | China | 2026 (staged; Theme 3 on 2026-09-01) | All real-world data on the AgiBot G2 across commercial and home spaces, with simulated twins in Genie Sim; total hours not stated. | — | CC-BY-NC-SA-4.0 | huggingface.co read Oct 26, 2026 |
| AgiBot World 2026, Theme 3 (reinforcement learning)AGIBOT | China | 2026-09-01 | 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 | — | Not disclosed | agibot.com read Oct 26, 2026 |
| MolmoBot-DataAllen Institute for AI | US | 2026-03-11 | 1.7M expert trajectories, 11K+ objects, 94K+ procedural environments, 8 task types. | — | Open (exact license not checked) | Snippet-only allenai.org read Oct 26, 2026 |
| NVIDIA Flora (video-labeling models)NVIDIA GEAR lab | US | 2026-10-20 | Labels 200,000 hours of human and robot video per day on 512 GPUs; outputs captions and reward signals. | — | Non-commercial research (reports conflict) | research.nvidia.com read Oct 26, 2026 |
| Egocentric-10KBuild AI | US | 2025-11-09 | 10,000 hours, 1.08B frames, 16.4 TB of head-mounted factory-worker video from 85 factories. | 10,000 | Apache-2.0 (gated, HF conditions) | huggingface.co read Oct 26, 2026 |
| Galaxea Open-World DatasetGalaxea AI | China | 2025-08-30 | About 500 hours, 50 scenes, 150+ tasks, single embodiment. | — | CC-BY-NC-SA-4.0 (gated) | arxiv.org read Oct 26, 2026 |
| RoboMINDBeijing Humanoid Robot Innovation Center and BAAI | China | 2025-01-02 | Over 1 TB, multi-embodiment (AgileX, Tien Kung, Franka and others). | — | Apache-2.0 | huggingface.co read Oct 26, 2026 |
| Eidon Tracker POV (+ IMU)Eidon AI (released on shutdown) | US | 2026-09-11 | 9.05 TB, about 1,274 hours of first-person household video; IMU variant adds arm tracking. | 1,274 | CC-BY-4.0 (from HF API tags) | huggingface.co read Oct 26, 2026 |
| LeRobot community datasetsHugging Face LeRobot | US/France | Count as of 2026-05-21 | more than 58,000 community datasets, up from 1,145 at end of 2024; robotics is the largest dataset category on the Hub | — | Mixed per dataset | techtimes.com read Oct 26, 2026 |
| DROID 1.0.1DROID collaboration (LeRobot copy) | US | 2025-07-16 | 95,658 episodes, 27,630,375 frames, 49,630 tasks (info.json) | — | Apache-2.0 | huggingface.co read Oct 26, 2026 |
| Open X-EmbodimentOpen X-Embodiment collaboration | US/international | 2023-10 | 1M-10M episodes from 22 robots. | — | CC-BY-4.0 (mirror) | Snippet-only huggingface.co read Oct 26, 2026 |
| ABC-130KXDOF | US | 2026-06-15 | 130,000+ bimanual teleoperation episodes, 195 tasks on low-cost YAM stations. | — | Apache-2.0 | xdof.ai read Oct 26, 2026 |
| Open-H-Embodiment (surgical)NVIDIA with 30+ institutions (incl. China, Hong Kong, UAE) | USA (hosted); contributors from China, HK, UAE, Europe | 2026-02-06 | Paired kinematics and video across surgical and ultrasound robots; size not stated. | — | CC-BY-4.0 | huggingface.co read Oct 26, 2026 |
| GRAIL locomanipulation (Unitree G1)NVIDIA NVlabs | US | 2026-04-28 | Physics-validated human-object interaction trajectories (pickup, stairs, slopes, sitting); counts not stated. | — | CC-BY-NC-4.0 | huggingface.co read Oct 26, 2026 |
| Chip or platform | Country | Spec | Used by | Date | Source |
|---|---|---|---|---|---|
| Jetson AGX Thor / T5000NVIDIA | US | up to 2070 FP4 TFLOPS; 128GB memory; 40-130W; 7.5x performance and 3.5x energy efficiency of AGX Orin | humanoid and robot makers per NVIDIA (1X, Agile Robots, Boston Dynamics, Amazon Robotics, FANUC) | Oct 26, 2026 | nvidia.com read Oct 26, 2026 |
| NVIDIA modules in Unitree robotsNVIDIA chips inside Unitree robots | China (robot) / US (chip) | Thor 2070 FP4 TFLOPS, 128GB, 40-130W; Orin NX 100 TOPS; AGX Orin 275 TOPS | Unitree H2, Unitree G1 (back expansion dock) | Oct 26, 2026 | Snippet-only scan.co.uk read Oct 26, 2026 |
| GR00T reference humanoid (Unitree H2 + Jetson Thor)NVIDIA with Unitree | US / China | Jetson Thor Blackwell compute in Unitree H2 frame; ships later in 2026 to labs | Unitree H2; early buyers Stanford, ETH Zurich, Ai2, UC San Diego | Oct 1, 2026 | Snippet-only aiweekly.co read Oct 26, 2026 |
| SiMa.ai Modalix and 2028 roadmapSiMa.ai | US | Series C $150M at a $1.45B valuation; new hardware in H1 2028 at 1,000 dense TOPS. | targets drones, humanoid robots, ADAS, cockpits; customers Bosch, Emerson, Micron, Synopsys, TRUMPF | Sep 28, 2026 | thenextweb.com read Oct 26, 2026 |
| Qualcomm to acquire PickNik (MoveIt)Qualcomm | US | software acquisition; MoveIt 1 and 2 remain open source; price not disclosed; subject to closing conditions | Dragonwing robotics platforms (Arduino VENTUNO Q first per press reports) | Sep 23, 2026 | therobotreport.com read Oct 26, 2026 |
| Figure and Nscale: up to 100,000 Vera Rubin GPUsFigure AI / Nscale / NVIDIA | US / UK | up to 100,000 Vera Rubin GPUs; Figure also states over $1B in 12 months on data and compute | Figure Helix training (offboard, not onboard compute) | Sep 3, 2026 | figure.ai read Oct 26, 2026 |
| Jetson T3000 (Thor family)NVIDIA | US | 865 FP4 TFLOPS; 32GB LPDDR5X; 273GB/s; 8-core Neoverse Arm CPU; 25 GbE; roughly half size and power of T5000 | Thor family platform users named by NVIDIA: 1X, Agile Robots, Amazon Robotics, Boston Dynamics, FANUC, Hitachi, Techman Robot | Jul 15, 2026 | blogs.nvidia.com read Oct 26, 2026 |
| Jetson T2000 (Thor family)NVIDIA | US | 400 FP4 TFLOPS; 16GB memory; module availability Q1 2027 (T3000 and T2000) | target: visual AI agents, AMRs, industrial manipulators | Jul 15, 2026 | blogs.nvidia.com read Oct 26, 2026 |
| Jetson AGX Orin memory optimisationNVIDIA | US | software memory optimisation cut usage by up to 15GB, moving from AGX Orin 64GB to 32GB module | UBTech, Agile Robots, Connect Tech (per NVIDIA); UBTech is a Chinese humanoid maker running Jetson | Jul 15, 2026 | blogs.nvidia.com read Oct 26, 2026 |
| Qualcomm Dragonwing IQ10Qualcomm | US | Up to 700 TOPS; 18-core Oryon CPU; 64 GB in-package memory; early access June 2026. | Targets AMRs and full-size humanoids; no customers named | Jun 1, 2026 | hothardware.com read Oct 26, 2026 |
| Tesla AI5Tesla | US | Taped out April 2026; Musk says one AI5 gives 5x the useful compute of two AI4. Power figures unconfirmed. | Optimus and Tesla supercomputer clusters first | Apr 15, 2026 | notateslaapp.com read Oct 26, 2026 |
| D-Robotics RDK S100 / S600D-Robotics (spun off from Horizon Robotics, May 2024) | China | CPU + BPU + MCU on one chip (S100, launched 2025-06-11); Series B1 $120M reported March 2026. | 20+ embodied-AI partners (unnamed) | Mar 16, 2026 | kr-asia.com read Oct 26, 2026 |
| Rockchip RK3588 / RK3568Rockchip | China | RK3588 8nm, up to 6 TOPS NPU; RK3568 22nm, 1 TOPS; RK3688 (Armv9.3, 16-32 TOPS reported) not confirmed shipping | UBTech, Unitree, Deep Robotics (Rockchip humanoid/robot customers per 36Kr) | Oct 25, 2025 | eu.36kr.com read Oct 26, 2026 |
| Jetson Thor at Chinese humanoid makersNVIDIA / Unitree, AgiBot, Galbot, EngineAI, UBTech | US / China | No primary source found that Jetson Thor itself is export-restricted to China; secondary reports say Chinese humanoid makers were early adopters. | Unitree, AgiBot, Galbot, EngineAI, UBTech (secondary reporting) | Aug 25, 2025 | Snippet-only scmp.com read Oct 26, 2026 |
Humanoid build cost, three analyst splits of the bill of materials, and teardown facts.
| Estimate by | Scope | Estimate | Date | Source |
|---|---|---|---|---|
| Goldman Sachs | Humanoid manufacturing cost, low end to state of the art | $30k-$150k (2024), down from $50k-$250k (2023) | Feb 2024 | goldmansachs.com read Oct 26, 2026 |
| Morgan Stanley | Tesla Optimus Gen2 bill of materials, excluding software | $50k-$60k per unit | Feb 6, 2025 | web.archive.org read Oct 26, 2026 |
| BofA Global Research | Humanoid built in China | $35,000 (2025), falling below $17,000 by 2030; pilot-stage builds $90k-$100k | Mar 12, 2026 | institute.bankofamerica.com read Oct 26, 2026 |
| McKinsey | Typical humanoid bill of materials | $30k-$150k now; under $20k is the widely cited long-term target | Apr 14, 2026 | mckinsey.com read Oct 26, 2026 |
Each firm splits the bill of materials into different categories and measures a different robot or year, so the shares are shown as published and not combined.
Tesla Optimus Gen2 teardown, by part type · Feb 6, 2025
Counts a torque or force sensor in every joint as "sensors", which is why sensing is far larger here than in McKinsey's split. By body region the 28 actuators are about 47% (upper body) and 48% (lower body) of the total.
web.archive.org, web.archive.org 2, web.archive.org 3 read Oct 26, 2026
Generic humanoid, by domain · Apr 14, 2026
These five domains are 85-90% of unit cost; the rest is cooling and wiring harnesses.
mckinsey.com read Oct 26, 2026
Humanoid built in China, 2030 projection · Mar 12, 2026
A projection for 2030, not today's mix.
institute.bankofamerica.com read Oct 26, 2026
| Claim | Value | Robot or scope | Country | Source |
|---|---|---|---|---|
| Actuators 8 | ||||
| Morgan Stanley: Optimus Gen2 uses 50 DoF driven by 28 actuators (14 linear, 14 rotary); humanoids in development have 16-60 DoF. Tesla's newest hand (likely Gen3) is 22 DoF vs 27 in the human hand. | 28 actuators (14 linear + 14 rotary); 50 DoF; hand 22 DoF | Tesla Optimus Gen2 | US | web.archive.org read Oct 26, 2026 |
| Morgan Stanley teardown bill per rotary actuator: frameless torque motor + torque/force sensor + harmonic reducer + cross-roller bearing (+ angular-contact bearings and 2 encoders); linear actuators: frameless torque motor + 1D force sensor + ball screw (upper body/forearm) or planetary roller screw (thigh, calf) + 4-point contact bearing + encoder. Report notes ball and planetary roller screws are both in use because of limited supply/cost of roller screws. | rotary: harmonic; linear: ball screw upper body, roller screw legs | Tesla Optimus Gen2 | — | web.archive.org read Oct 26, 2026 |
| A press relay of a BofA report says its model robot uses 16 rotary actuators (harmonic reducers) and 14 linear actuators (planetary roller screws), with a 6-DoF hand. BofA's own 2026 page says only "dozens of actuators". | 16 rotary + 14 linear = 30; hand 6 DoF | BofA model humanoid | — | outlookbusiness.com read Oct 26, 2026 |
| McKinsey: OEMs coordinate 20 to 40 primary structural actuators per humanoid at kilohertz control rates with fail-operational safety logic; most use Nvidia Jetson with Elmo, Novanta (Celera Motion) or Synapticon servo controllers. | 20-40 primary actuators | humanoid, generic | — | mckinsey.com read Oct 26, 2026 |
| Figure states Figure 03 is vertically integrated across actuators, batteries, sensors, structures and electronics, all designed in-house, while buying individual components from volume-capable partners. Figure 03 actuators are claimed 2x faster with better torque density; no per-robot actuator count is published on the page. | actuators, battery, sensors, structure, electronics in-house | Figure 03 | US | Vendor-reported figure.ai read Oct 26, 2026 |
| Morgan Stanley's teardown shows planetary roller screws only in the thigh and calf (8 of 14 linear actuators) and ball screws elsewhere. No Tesla statement says every linear joint uses roller screws. | Roller screws in thigh and calf only (teardown) | Tesla Optimus | US | Snippet-only en.wikipedia.org read Oct 26, 2026 |
| 36Kr reports a $685M Tesla actuator order with Sanhua Intelligent Controls and says Chinese makers hold 70% of Optimus supply-chain share. Tesla has not confirmed either figure. | $685M order; 70% share (unconfirmed) | Tesla Optimus | China | Snippet-only eu.36kr.com read Oct 26, 2026 |
| Figure pairs 100,000 robots with 3,000,000 actuators over four years: about 30 actuators per robot (a derived ratio, not a stated spec). | 3,000,000 / 100,000 = 30 actuators per robot (derived) | Figure 02/03 (BotQ plan) | US | Vendor-reported figure.ai read Oct 26, 2026 |
| Motors and magnets 3 | ||||
| Morgan Stanley: humanoids use either frameless torque motors (lower technical barriers) or coreless motors (greater technical barriers); reducers are either harmonic or planetary. | frameless vs coreless; harmonic vs planetary | humanoids generally | — | web.archive.org read Oct 26, 2026 |
| McKinsey: China controls ~69% of rare-earth mining and 90% of magnet processing and refining capacity; Musk publicly noted magnet supply constraints affected Optimus production. | 69% mining; 90% magnet processing | global NdFeB for actuators | China | mckinsey.com read Oct 26, 2026 |
| humanoid.guide relays TechBuzz China component tracker: Chinese coreless cup motors from Aoyi Technology at about 100 yuan (~$14) each versus 4,000+ yuan (~$560) for imported German equivalents (Maxon, Faulhaber); roughly a 40x gap for hand-class motors. | ~CNY 100 vs CNY 4,000+ | coreless hand motors | China | humanoid.guide read Oct 26, 2026 |
| Reducers 6 | ||||
| McKinsey flags harmonic/strain-wave drives as concentrated among Harmonic Drive, Nabtesco and emerging Chinese players such as Leaderdrive; precision-bound, capital-intensive, hard to scale. | bottleneck cluster 1 | humanoid supply chain | Japan | mckinsey.com read Oct 26, 2026 |
| humanoid.guide relays PICEA Motion: a high-end strain-wave (harmonic) reducer costs $800-$1,500 per unit. A trade estimate, not an audited price. | $800-$1,500 per high-end harmonic reducer | humanoid joints | — | humanoid.guide read Oct 26, 2026 |
| humanoid.guide, relaying a TechBuzz China tracker, puts a typical full-size humanoid at about 14-15 harmonic plus 12-15 planetary reducers. | 14-15 harmonic + 12-15 planetary | full-size humanoid | — | humanoid.guide read Oct 26, 2026 |
| humanoid.guide (citing J.P. Morgan): Leaderdrive holds 30-40% of China's harmonic reducer market; 2025 revenue up 47% to 570.7 million yuan and net profit more than doubled to 124.4 million yuan. | 30-40% of China harmonic market | China harmonic reducers | China | humanoid.guide read Oct 26, 2026 |
| Morgan Stanley's Exhibit 50: 'Humanoids mainly use harmonic and planetary reducers, but RV reducers could be an alternative.' Cycloidal/RV is not the mainstream humanoid choice in this taxonomy. | harmonic + planetary mainstream; RV alternative | humanoids generally | — | web.archive.org read Oct 26, 2026 |
| Cone Drive (US) manufactures harmonic gearing in Traverse City, Michigan, with single-stage ratios up to 160:1 (RLC model); the page gives no torque ratings or robotics volume. A US-made harmonic source exists but its humanoid volume is undisclosed. | up to 160:1; Traverse City MI | harmonic gearing | US | Vendor-reported conedrive.com read Oct 26, 2026 |
| Screws 3 | ||||
| McKinsey rates planetary roller screws an even more acute bottleneck than strain-wave gears: narrow supplier base (SKF named, plus select high-precision Asian suppliers), long lead times, few substitutes. | acute bottleneck | humanoid supply chain | Sweden | mckinsey.com read Oct 26, 2026 |
| 36Kr, relaying Huaxin Securities, estimates the planetary roller screws in Optimus cost about 2,000 yuan (~$280) each. | ~CNY 2,000 (~$280) per screw | Tesla Optimus | China | Snippet-only eu.36kr.com read Oct 26, 2026 |
| 36Kr, relaying Huaxi Securities: the top four overseas roller-screw suppliers held about 78% of the market in 2022; Chinese makers Nanjing Machine Tool Works and Jining Boter about 8% each. | overseas top-4 ~78% (2022); Beite 2.6M sets/yr planned | planetary roller screws | China | Snippet-only eu.36kr.com read Oct 26, 2026 |
| Hands and touch 8 | ||||
| Morgan Stanley Exhibit 48 teardown: Optimus Gen2 hands ~$9.5k (17.2% of BOM) using 12 coreless motors, 12 planetary reducers, 12 encoders, 2 6D force sensors; legs ~$7.3k thigh + $7.3k calf (13.2% each); waist and pelvis ~$7.8k (14.2%); head ~$2.1k (3.8%); battery ~$0.3k (0.5%). | hands $9.5k = 17.2% of BOM | Tesla Optimus Gen2 | — | web.archive.org read Oct 26, 2026 |
| Tesla's Optimus program lead (Q3 2026 call) called the hand the hardest hardware area, more complicated than the locomotion actuators, with bespoke gears still in development. | hand = hardest remaining hardware; bespoke gears in development | Tesla Optimus V3 (pre-launch) | US | Vendor-reported singjupost.com read Oct 26, 2026 |
| Figure developed its own first-generation fingertip tactile sensor after judging market options not durable enough; each fingertip detects forces as small as three grams. Each Figure 03 hand also has an embedded palm camera. | 3 g sensitivity per fingertip; palm camera per hand | Figure 03 | US | Vendor-reported figure.ai read Oct 26, 2026 |
| Two Tesla patents (filed October 2024) describe a tendon-driven hand with actuators in the forearm, 4 DoF per finger plus 2 at the wrist. These are patent filings, not the shipping hand. | 4 DoF/finger + 2 wrist; tendons; actuators in forearm | Tesla Optimus V3 (patent) | US | teslarati.com read Oct 26, 2026 |
| Tesla has not published how many actuators drive the Optimus hand. Third-party counts conflict: Morgan Stanley lists 12 coreless motors per Gen2 hand, and other accounts range from 6 to 50. None comes from Tesla. | Not disclosed | Tesla Optimus Gen2/Gen3 hands | US | en.wikipedia.org read Oct 26, 2026 |
| Meta FAIR announced Digit 360 fingertip tactile sensor: forces as small as 1 millinewton, over 18 sensing features, over 8 million taxels; GelSight Inc to manufacture and distribute it (US supplier). Digit Plexus is the open hardware-software platform. | 1 mN; 8M+ taxels; 18+ sensing features | Meta Digit 360 / GelSight | US | Vendor-reported ai.meta.com read Oct 26, 2026 |
| Shadow Robot Dexterous Hand: 24 DoF driven by 20 motors (tendon-driven), 4.3 kg, two tactile fingertips as standard (up to five); price on request. The research-grade end of the range. | 24 DoF; 20 motors; 4.3 kg | Shadow Dexterous Hand | UK | Vendor-reported shadowrobot.com read Oct 26, 2026 |
| Linkerbot (China) press release: shipped its 10,000th dexterous hand; Linker Hand line spans 6 to 42 DoF (flagship L30 with +/-0.2 mm repeatability); claims over 80% global share of the high-DoF hand segment (company-stated, no independent audit); $150M raised since April 2025. | 10,000th hand shipped; 6-42 DoF; >80% high-DoF share (self-reported) | Linkerbot Linker Hand | China | Vendor-reported prnewswire.com read Oct 26, 2026 |
| Sensors 6 | ||||
| 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. | limited vendors; calibration-bound | humanoid supply chain | US | mckinsey.com read Oct 26, 2026 |
| humanoid.guide relays X.PIN: a six-axis force-torque sensor still costs 5,000 to 30,000 yuan (about $700-$4,200). | CNY 5,000-30,000 (~$700-$4,200) | 6-axis F/T | — | humanoid.guide read Oct 26, 2026 |
| 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. | 128 ch; 189x360 deg; 40 m; 265 g | Hesai JT128 | China | Vendor-reported hesaitech.com read Oct 26, 2026 |
| 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. | AksIM-2 20-bit; Orbis 14-bit; RM08 12-bit | humanoid joint encoders | Slovenia | Vendor-reported rls.si read Oct 26, 2026 |
| 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. | 23-bit internal; 3x3x0.75 mm | humanoid joint encoder IC | China | Vendor-reported aap.com.au read Oct 26, 2026 |
| 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. | 0.05 deg; 1 kHz | Galbot | China | Vendor-reported autonews.gasgoo.com read Oct 26, 2026 |
| Compute 1 | ||||
| Musk (Q3 2026 call): Tesla plans to put its in-house AI5 chip in Optimus around April 2027, about 10x the compute of AI4. | AI5 in Optimus ~Apr 2027; ~10x AI4 compute | Tesla Optimus | US | Vendor-reported singjupost.com read Oct 26, 2026 |
| Power and thermal 4 | ||||
| Morgan Stanley's teardown lists a 2.3 kWh, 52 V chest battery pack at about $300 (~0.5% of the bill of materials). | 2.3 kWh, 52 V, ~$300 | Tesla Optimus Gen2 | — | web.archive.org read Oct 26, 2026 |
| Figure 03's battery is designed and built in-house: 2.3 kWh, five hours at peak performance, structural enclosure, UN38.3 certified. | 2.3 kWh; 5 h peak run time | Figure 03 | US | Vendor-reported figure.ai read Oct 26, 2026 |
| UBTECH Walker S2 swaps its own battery in about three minutes (dual-battery design) and carries 15 kg. | autonomous battery swap in 3 minutes; 15 kg payload | UBTECH Walker S2 | China | Vendor-reported ubtrobot.com read Oct 26, 2026 |
| 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. | joint heat is the endurance limit (qualitative) | humanoid joint modules | China | amdt.hkust-gz.edu.cn read Oct 26, 2026 |
| Wiring 2 | ||||
| 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. | bore ~1/3 of OD; XT30 + CAN on G1; igus bus cable tested >15M torsion cycles | humanoid joints | — | humanoid.guide read Oct 26, 2026 |
| 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. | 4 conductors per joint port; 6 CAN branches | Menlo Asimov; Unitree G1 | US | humanoid.guide read Oct 26, 2026 |
| Assembly and whole robots 10 | ||||
| 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%). | brain ~4%; battery+structure ~0.8% | Tesla Optimus Gen2 | US | web.archive.org read Oct 26, 2026 |
| 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. | G1 ~$16k; Digit ~$250k | Unitree G1; Agility Digit | China | web.archive.org read Oct 26, 2026 |
| 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. | Early production Oct 2026; deliveries ~end-2027 | Tesla Optimus | US | Vendor-reported singjupost.com read Oct 26, 2026 |
| 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. | 12,000/yr line capacity; 100,000 over 4 years | Figure BotQ | US | Vendor-reported figure.ai read Oct 26, 2026 |
| 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. | 23 DoF; 35 kg; 90 N.m knee; ~2 h; $13.5K | Unitree G1 | China | Vendor-reported unitree.com read Oct 26, 2026 |
| 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. | 31 DoF; 70 kg; 360 N.m; 0.972 kWh; $29,900 | Unitree H2 | China | Vendor-reported unitree.com read Oct 26, 2026 |
| 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. | $4,900-$5,900; 20-26 DoF | Unitree R1 | China | Vendor-reported unitree.com read Oct 26, 2026 |
| 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. | 56 DoF; 90 kg; 4 h; self-swapping battery | Boston Dynamics Atlas (product) | US | Vendor-reported bostondynamics.com read Oct 26, 2026 |
| 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. | 30 kg class; tendon drive; hand 22 DoF each; 4 h | 1X NEO | US | Vendor-reported 1x.tech read Oct 26, 2026 |
| 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. | -10 kg Gen2; PEEK structural parts | Tesla Optimus Gen2 | — | web.archive.org read Oct 26, 2026 |
Entries come from a research sweep read on Oct 26, 2026. Missing values show as "Not disclosed". Country columns describe where a lab or company is based; they are not a score.