Nature Will says its new Natura-1 model can control dexterous robot hands using instructions that combine vision and language. The company describes a 20-degree-of-freedom system designed for physical tasks, trained on about one million hours of first-person video.
Watch Desk analysis
What happened
In its announcement, Nature Will says Natura-1 is a 50-billion-parameter mixture-of-experts model that runs within a 100-millisecond control loop. The company says it can follow zero-shot instructions and carry out long-horizon workflows, including on deformable materials and in changing surroundings.
Nature Will says it evaluated the model on dual-arm hardware and that it can generalise to unseen tasks. The company also says the model can be adapted to specific robot hardware through post-training on limited data. Those are the company’s reported capabilities, not an independent performance assessment.
Why it matters
A robot hand has to do more than recognise an object: it must act, adjust and keep going when the world refuses to match the demo. Natura-1’s stated combination of vision, language and a fast control loop puts the focus on that practical challenge, rather than on another model that only produces text.
If the reported generalisation and adaptation hold up across different tasks and hardware, they could make it easier to move from a carefully staged demonstration to useful physical work. That is a meaningful ambition. It is also where impressive robotics claims meet the less forgiving business of real-world testing.
Our read
Natura-1 is a substantial robotics announcement, with a useful emphasis on dexterity and adapting to different hardware. The most important next evidence is not another impressive clip, but clear results across tasks, materials and environments. Until then, treat the capabilities as Nature Will’s claims, not a settled measure of what the system can reliably do.
What to watch
- How the model performs on tasks and hardware beyond the company’s demonstrations.
- What post-training data and effort are needed to adapt it to a new robot.
- Whether independent evaluations support the claimed generalisation and control speed.
Discussion spark: What would convince you that a robot-hand model generalises beyond demonstrations: independent tests across hardware, or reliable performance on a smaller set of real jobs?
Sources and evidence
- The Will to Act – Natura-1 – Nature Will (13 October 2026, 11:06 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.