Nature Will, a physical artificial intelligence lab with a stated base in San Francisco, announced on October 13, 2026, the launch of Natura-1, which it describes as its first “general foundation model for the physical world.” It is a 50-billion-parameter mixture-of-experts model, pretrained on more than one million hours of first-person video recorded by humans, that takes language and images as input and drives 20-degree-of-freedom dexterous robot hands within a 100-millisecond control loop.
The company was founded in April 2026 by Ning Ding, who entered Tsinghua University’s Department of Computer Science and Technology for his PhD in 2018 and later stayed on as faculty, according to a post by ZhenFund, which has backed the company since its founding. “We believe general models that can handle every task must exist, and that those models must utilize dexterous hands,” Ding said in the announcement distributed by PR Newswire.
The lab’s central argument is that the physical world was shaped by and for the human hand: tools were made to fit our hands, and those tools were used to make everything else. That is why, Nature Will argues, no machine can be truly general without dexterity. The problem is that controlling hands with many degrees of freedom is extremely hard: contact dynamics leave little room for error and the action space grows exponentially with the number of joints. Most prior work has focused on simple grippers or systems built for a single task.
Natura-1’s bet is to learn from human experience rather than rely only on robot data. According to the company’s technical blog, robot data collection is limited by robot capability: a robot can only gather data where it already works. Humans, by contrast, operate in virtually every environment. The training corpus totals one million hours, with more than 61.7 million episodes, 317,842 tasks and 14,286 scenes, combining controlled recordings with first-person internet videos.
To turn that video into useful signals, Nature Will built its own infrastructure: a data store of tens of petabytes, a pipeline that removes duplicates and filters anomalies, reconstruction of 3D hand and body poses from monocular and stereo video, and hierarchical annotations describing everything from an activity’s overall goal down to each step. The company says its system processes 350,000 hours of video per week and that its motion labeling cuts error by 41% compared with open-source alternatives on an internal benchmark.
After pretraining, the model is adapted to a physical robot with a limited amount of robot manipulation data. Video and action are treated as two projections of the same physical state: Natura-1 uses past video as memory and learns to predict the future as part of its world model, letting it draw supervision from video, which is far more abundant than action labels. The company adds that it trains the model so that differences in the instruction yield different predictions, making language a signal that cannot easily be ignored.
In the published tests, Natura-1 runs on dual-arm robots with dexterous hands. Nature Will built its own benchmark with twelve dimensions grouped into dexterous control, instruction following and long-horizon execution. Because there is no public dexterous-hand foundation model for a direct comparison, it picked two references: π0.5, a well-known vision-language-action model, and GPT-6 Astra used as a code-based policy. According to the company, neither scored on its dexterous-hand platform.
For a somewhat fairer comparison, Nature Will evaluated both on a gripper platform, ARX AC One, where they scored 13.5 and 16.7, respectively. Natura-1, using 20-degree-of-freedom hands, reached 53.4 in its base version and 59.1 in its agentic version. These figures come from a benchmark designed by the company itself under non-identical conditions, so they should be read as indicative rather than as an independent comparison.
The demos include long tasks after specific tuning — making wontons, applying a screen protector to a phone or preparing a bowl of oats with sliced banana — and zero-shot instruction-following tests: choosing a bottle by ruling out six colors, closing a drawer by pushing it with the fingertips or tracking a lemon a person moves around. The company notes that six of the long tasks use additional tuning and that the rest share the same weights.
Nature Will also describes behaviors it did not set out to measure, such as sliding a card to the edge of a box so it can be grasped from below, using a rod to pull an out-of-reach basket closer or pressing the spacebar on a laptop, an object that never appeared in its robot data. The company presents them as an early sign of more general physical intelligence, an interpretation that will require external validation and testing in less controlled settings.
The announcement comes amid a race in so-called physical AI. In recent weeks, large technology companies and startups have launched tools and funding rounds for industrial and humanoid robots, and the industry is debating whether advances in language models can carry over to the physical world. Nature Will’s approach stands out for prioritizing dexterous hands over grippers and for betting on large-scale human video as the main source of experience.
The limitations are clear, and the company acknowledges them. Progress will depend on more capable and stable hands and sensors, broader pretraining data and models that need less adaptation for each robot. No customers, pricing, commercial availability or access for outside researchers have been announced, and the results have not been peer reviewed. The detailed write-up of the method does, however, let the community scrutinize the claims.
With Natura-1, Nature Will steps onto the international stage six months after its founding with an ambitious thesis: that human-level fine manipulation does not have to be programmed task by task, but can emerge from a foundation model as the human experience it is trained on grows. If that thesis holds outside the lab, robot hands could move closer to everyday tasks that remain beyond automation today.
