Nature Will launches Natura-1, a 50-billion-parameter hand-control model trained on 1 million hours of human video
Physical AI lab Nature Will said in a press release on Oct 13 that it launched Natura-1, a 50-billion-parameter mixture-of-experts foundation model for dexterous robot hands, pre-trained on more than one million hours of egocentric human data and running a 100-millisecond control loop.
The lab, led by CEO Ning Ding, published no benchmark results, robot hardware or customers in the release. On parameter count the model is about 21 times the 2.4-billion-parameter REKI model MEKAWIRE briefed this week.
A lab few people had heard of claims one of the largest robot-control models announced this year. Nature Will said in a press release datelined San Francisco on Oct 13 that it launched Natura-1, a 50-billion-parameter mixture-of-experts model that controls dexterous robot hands and follows natural-language instructions zero-shot. The company said it was pre-trained on more than one million hours of egocentric human data, and trained jointly on language, images, video and action.
What the release says, and what it leaves out
Natura-1 runs a 100-millisecond control loop and is designed for long, continuous workflows, reasoning about spatial constraints, relationships between multiple objects and negative instructions, the company said. Chief executive Ning Ding said the lab believes general models must use dexterous hands, and called Natura-1 a first step toward models that scale from human experience. The release describes it as the lab's first foundation model capable of general-purpose dexterous manipulation on real hardware.
The release does not name the robot or hand used, give any success rate or benchmark, disclose how the million hours were collected or licensed, or mention funding or customers. Its headline calls Natura-1 the first general foundation model for the physical world, a claim that cannot be checked against published results.
How it compares in MEKAWIRE data
Scale is the one verifiable number. At 50 billion parameters Natura-1 is about 21 times the 2.4-billion-parameter REKI vision-language-action model from Kyutai and Inria that MEKAWIRE briefed on Oct 16 (50 / 2.4 = 20.8). Its one million hours of human data is 1,000 times the 1,000 hours of human touch data used in the Spark-0 preprint, though the two are different kinds of data. A 100 ms loop is slow for fine force control, which suggests a split where the model plans and a faster low-level controller executes; the release does not say.
MEKAWIRE view
Human egocentric video has become the cheapest route to robot training data at scale, and robot-data companies such as Mecka AI, which raised $60m on Oct 7, are building businesses on it. Natura-1 is a bet that sheer scale on that data yields general dexterity. Until Nature Will publishes evaluations on named hardware, it is a parameter count and a data claim rather than a demonstrated capability.
What to watch: a technical report or benchmark results, disclosure of the hand platform and data sources, and any funding announcement that would show who is backing a model of this size.
- Ning Ding (founder)
- Natura-1, 50-billion-parameter mixture-of-experts
- Nature Will press release, San Francisco, Oct 13, 2026
- 100 ms
- Benchmarks, robot hardware, data provenance, funding, customers
- More than 1 million hours of egocentric human data; joint language, image, video and action training
- About 21x the parameters of REKI (2.4B)
- Dexterous hand control, zero-shot instruction following, long-horizon workflows