General foundation model for physical AI? Nature Will bets on 1 million hours of video

General foundation model for physical AI

There's a new entrant in the race to build a general foundation model for physical AI, or at least for hands. Nature Will, a physical AI lab with an office in San Francisco, launched Natura-1 today: a a 50-billion-parameter mixture-of-experts model for dexterous hand control, trained jointly on language, images, video and action. The headline number is the data: over 1 million hours of egocentric human video, footage shot from a person's point of view while they use their hands.

The goal is zero-shot instruction following: tell the robot what you want in ordinary language, and it just does it, including with objects and in settings that it hasn't seen before.

"Our goal is to allow ordinary language to inform a growing range of physical capability in robotics, creating models that can follow different requests, handle unfamiliar objects, and find a way forward when interactions change," Nature Will CEO Ning Ding said in a statement.

CTO Haozhan Li puts the focus on hands bluntly.

"We believe dexterous hands are central to the long-term future of embodied intelligence," he said. "Hands already work with all the tools humans have developed, and are closely aligned with actions recorded in human experience."

Human video and physical AI

That logic is why human video is suddenly the hottest data source in robotics.

There's essentially no internet of robot data, but there's a vast and growing supply of humans using their hands. Figure is paying people to film themselves doing chores through Index, which has passed 16 million uploaded videos and $15 million paid to contributors. Figure is now backing that up with an initial $3.5 billion compute commitment to train Helix. Genesis AI, which raised a $105 million seed round, pairs a 20-degree-of-freedom hand with a sensor glove and video from head-mounted cameras to train its GENE-26.5 model.

Natura-1 is the purest version of that bet so far. No glove, no teleoperation fleet mentioned, no custom hand. Just a very large pile of first-person video and a very large model.

Of course, not everyone thinks video is enough.

When I talked to 1X CEO Bernt Børnich about the company's new world model lab, he made the case that video is missing the most important part: what happens when you act. Video won't tell a robot how hard to squeeze a glass so it neither slips nor shatters.

"You observe your state, you take an action, you see the result — kind of like the scientific method," Børnich said. "That's not available for the data out there right now, because you just have the results. It's just video."

Boston Dynamics goes further. In explaining its new Atlas hands, BD argued that even sensor gloves capture proxy signals rather than the actual action signals needed to control a dexterous robot hand, and that reinforcement learning in simulation is essential to cracking manipulation.

So there's a real split in the industry. One camp says scale up human data and the skills will transfer. The other says force, contact and consequence have to come from robots, real or simulated.

Natura-1 lands firmly in the first camp.

Speed and more details

There's also a speed question. Natura-1 runs a 100-millisecond control loop, which means 10 decisions a second. That's quick for a 50-billion-parameter model, but not next to Figure's Helix, where the System 1 layer sends control updates to the hands 200 times a second. 

Dexterous work, of course, lives and dies on reacting to slips and contact.

That gap isn't necessarily fatal. Researchers from MIT, Caltech, Nvidia, UC San Diego and Tsinghua recently showed with VLASH that robots can plan their next actions while still executing current ones, completing some tasks 1.5 to 2X faster. A slower model that thinks ahead can still move smoothly. We just don't know yet if that's how Natura-1 works.

We also need more data on success rates, benchmarks, and head-to-head comparisons with Helix, π0 or GENE-26.5, plus more info on availability: open weights, licensing, or an API.

Those matter, because the industry's biggest gap isn't demos. It's deployment. As Kyber Labs co-founder Tyler Habowski told me earlier this year: "There are literally zero robot hands deployed right now doing routine work."

In independent GPT-6 tests, a general model hit 19 of 20 on simple block-in-bowl tasks but just 2 of 20 on precision insertions. Precision is where hands earn their keep, and it's exactly where the video-only bet will be tested hardest.

I'm meeting with Nature Will's CEO this week, and will update this with what I learn.

For now, Natura-1 is a big model, a big dataset, and a big claim. The proof will be a robot hand doing real work.

Sources

• Emailed Nature Will press release, Natura-1 announcement (Oct. 13, 2026)
• How Figure committed $3.5 billion for AI compute — Forbes
• 1X launches humanoid robot world model lab — Forbes
• These robot hands can literally make you an egg scramble breakfast — Forbes
• Boston Dynamics' new Atlas hands — HumanoidMegaHub
• Unitree's founder tested GPT-6 on robots — HumanoidMegaHub
• New AI technique helps robots complete tasks twice as fast — Live Science