Robot Foundation Models Scale to 1M Hours of Human Video

The million-hour threshold

For most of the past decade, the main limit on robot learning was data. Large language models could train on the text of the internet. Robots had to learn from slow, expensive teleoperation logs gathered one machine at a time. October 2026 suggests that limit is starting to ease, at least on paper. Several labs now say they have pretrained robot foundation models on roughly a million hours of first-person human video. Around these launches, headline figures of 90 percent task success have circulated. Those numbers deserve a much closer look than they are getting.

The clearest example is Nature Will, a physical AI lab with an office in San Francisco. It announced Natura-1, a 50-billion-parameter mixture-of-experts model for dexterous robot hands. The model was trained jointly on language, images, video and action, using more than a million hours of egocentric human footage, and it runs a 100-millisecond control loop.22 The company says the model can follow plain-language instructions zero-shot, including with objects it has never seen. It covers hands only, not whole-body control or locomotion.22

Dyna Robotics has reported what may be the more important scientific result. Co-founder Jason Ma described pretraining on a million hours of egocentric human video. The team scaled its data tenfold at each step, from 100 hours to 100,000, and the model's predictions improved every time.24 The surprising part is that the model never saw a robot during pretraining, yet it got better at predicting robot motion trajectories as human data grew. Ma presents this as a cross-embodiment scaling law.24

The 90 percent problem

The success-rate claims are where the coverage diverges most, and the gap tells a clear story. One analysis points out that many headline numbers, including the million-hour training claims and some 90 percent success rates, come from the companies themselves and have not been independently checked.23 The same piece notes that these results usually come from structured settings with fixed lighting, known object positions and no clutter. A score earned in a tidy room says little about a messy kitchen.23 It also puts the best models at roughly 70 to 80 percent in controlled tests, and finds that models trained only on successful demonstrations struggle to recover when a task goes wrong partway through.23

Independent numbers back up that caution. Dream Labs, a new startup founded by former Nvidia GR00T researchers, says its Dream-1 model achieves about 80 percent zero-shot pick-and-place success on bimanual gripper robots it was not trained on, with other tasks landing between 50 and 60 percent.21 When the company paid for an outside evaluation from Robocurve, the model scored a mean of 59.3 percent across 120 attempts on the RoboDojo-RC Tier 1 benchmark.29 Figure says its Helix 2.5 model, pretrained on human-behavior data, succeeded on 56 percent of zero-shot trials in 30 unseen homes, compared with 9 percent when trained from scratch.18 Meta's V-JEPA 2 was trained on more than a million hours of internet video plus 62 hours of unlabeled robot footage, and is reported to reach 80 percent zero-shot success on pick-and-place in unfamiliar labs.27

Dyna's own history may be the most useful warning for anyone reading the headline numbers. Ma said its first-generation vision-language-action model folded napkins with 99.4 percent success in the lab. After ten months at the restaurant chain Din Tai Fung, only 75 percent of its work met customer standards, and it ran at one-third of the required speed.24 That gap led Dyna to drop the architecture entirely.24 Our reading is that 90 percent is not where the field stands today. The independently checked range for unseen settings sits somewhere between the mid-50s and about 80 percent, and real deployment cuts that further.

Video beats language as the robot's backbone

The more lasting story may be about architecture. Several groups are moving from vision-language-action (VLA) models toward so-called world action models (WAMs), which treat robot control as a video-prediction problem. Dream Labs' founders argue that robot brains should be built on video models rather than language models. They published that approach as DreamZero while still at Nvidia.21 Dyna made the same switch after its restaurant deployment and rebuilt around video prediction because pretrained video models were already available off the shelf.24 A new arXiv paper proposes a single mathematical framework that covers both VLAs such as OpenVLA and world action models such as DreamZero. It reports no new success rates, but it shows that researchers now see the two camps as rival families worth comparing directly.1619

The human-video bet is not limited to these startups. Nvidia's openly licensed GR00T N1.7 adds 20,000 hours of EgoScale human video to its pretraining, and Nvidia credits that data for better generalization and language-following. DEEP Robotics pretrained its DR-VLA-2.0 navigation and manipulation model on 10,000 hours of curated egocentric video before adapting it with teleoperated robot clips.13 China's Light Origins says its Light-O1 model learns a transferable human action prior from internet-scale video, and that its data pipeline processes 200,000 hours of video a week.12 Figure is collecting similar data through Index, an app that pays people to film themselves doing chores.22

The reason is economic as much as scientific. Ma notes that a million hours of teleoperation data is very hard to collect, while human video needs neither robots nor trained operators.24 He also concedes that his scaling curve flattened between 100,000 hours and a million, mainly because much of the extra footage had no labels. Whether the trend continues into tens of millions of hours depends on whether enough high-quality, annotated supply exists.24 There is also a physics problem: video cannot show how heavy a door is or whether a cup holds water.24

Capital follows the data thesis

Investors are paying for this thesis directly. Dream Labs came out of stealth with a $90 million seed round led by B Capital.28 Alongside it, the company launched what it calls the first public API for a robot foundation model, priced at $3 an hour for the flagship model and 75 cents for a lightweight version.28 The model runs in the company's data centers rather than on the robot, and Dream Labs has not yet published latency figures.21 Mecka AI, which collects human-movement data for robot training, raised $60 million in a Series B led by Sequoia, with Nvidia and Microsoft's M12 participating.8 In China, Sudo AI, which builds robot foundation models, raised more than RMB 1 billion, roughly $149 million.10

Humanoid hardware money is moving even faster. Astribot raised RMB 1 billion at an RMB 15 billion valuation, its fourth round in six months.6 CASBOT announced a Series A++ round worth several hundred million yuan.5 One tracker counted 112 robotics deals in the single week of October 12 to 18, with $9.5 billion disclosed, up from $910 million the week before.9 The largest bets are being made on compute. Figure has raised about $1.9 billion in total and signed a deal with Nscale for up to 100,000 Nvidia Vera Rubin GPUs, an initial $3.5 billion commitment, to train Helix.22

Will the middle layer get squeezed?

Not everyone thinks robot foundation models will remain a separate business. General Robotics CEO Ashish Kapoor describes robot intelligence as three layers: frontier models at the bottom, robot foundation models in the middle, and job-specific post-training on top. He expects the bottom two layers to merge.22 Unitree founder Wang Xingxing said tests of GPT-6 on his robots were strong enough that he moved his forecast for an embodied-AI tipping point up to within one to two years.22 Independent trials temper that optimism. Robocurve found GPT-6 Astra succeeded on 19 of 20 simple block-in-bowl attempts but only 2 of 20 precision insertions.22

Our view is that the million-hour milestone is real and matters. Human video looks like a working way to scale robot pretraining, and the cross-embodiment evidence from Dyna, Nvidia and Dream Labs points the same way. The 90 percent framing, though, gets ahead of the evidence. Deployment numbers make the stakes concrete: global humanoid production passed 20,000 units in 2025, but only about 10 percent went into real work.23 Lab breakthroughs have not yet become reliable labor. Until independent benchmarks and long-running field deployments catch up with company press releases, the gap between the 99 percent lab result and the 75 percent restaurant result is the number to watch.