Nature Will Debuts Natura-1, a 50B Model for Robot Hands

Nature Will says its new Natura-1 is a 50-billion-parameter mixture-of-experts foundation model for dexterous hand control, pre-trained on over 1 million hours of egocentric human video and running on a 100-millisecond control loop. Robotics & Automation News covered the November 9 launch.

The model was trained jointly on language, images, video, and action. The company's claim is that it follows complex natural-language instructions zero-shot while "reasoning through spatial constraints, multi-object relationships, and negative conditions."

"We believe general models that can handle every task must exist, and that those models must utilize dexterous hands," CEO Ning Ding said. He framed Natura-1 as "our first step towards models that scale from human experience."

The headline benchmark numbers are first-party. On Nature Will's own dexterous-hands group, Natura-1 Agentic scored 59.1% and Natura-1 Base scored 53.4%; the two models the company tested against, π0.5 and GPT-6-Astra, "produced no measurable scores after testing." On a separate gripper group, GPT-6-Astra posted 16.7% and π0.5 posted 13.5%.

The report does not identify the partner hardware used in demos, the specific sources of the million-plus hours of egocentric human video, or any release date for Natura-1.