Nature Will, a lab working on physical AI, has launched Natura-1, a foundation model for controlling dexterous robot hands and following instructions zero-shot, the company announced.

The company says Natura-1 is a 50-billion-parameter mixture-of-experts model trained on language, images, video and action, and pre-trained on more than 1 million hours of egocentric human data. It runs a 100-millisecond control loop that links language and vision to high-degree-of-freedom hand movements.

According to Nature Will, the model can run long, continuous workflows on its own and follow complex natural-language instructions involving spatial constraints, relationships between several objects and negative conditions. It is the company's first physical foundation model that can do general-purpose dexterous manipulation on real hardware.

Ning Ding, chief executive of Nature Will, said Natura-1 is “our first step towards models that scale from human experience.”

Source: Robotics & Automation News

Background

Nature Will

AI lab

A physical AI lab that works on general foundation models for robots.

Description from Robotics & Automation News

Natura-1

AI model

Nature Will's foundation model for dexterous robot hand control and zero-shot instruction following.

Description from Robotics & Automation News

Ning Ding

Chief executive of Nature Will

Chief executive officer of Nature Will.

Description from Robotics & Automation News