Unified Hand Action Space Enables Cross-Embodiment Robotic Control

YuXiang_IRVL · x · 2026-07-11

A new method called UHAS (Unified Hand Action Space) has been proposed to solve the action representation challenge when training robot foundation models across different morphologies (such as robotic hands with varying mechanical structures or finger counts, and even humans).

Instead of learning latent actions, this approach relies on geometric principles: it maps the deformation of a canonical sphere to the control commands of different robots. Experiments show that a single policy can simultaneously control four completely different robotic hands to perform dexterous manipulation tasks.

Related event: UHAS Unifies Action Space for Cross-Embodiment Robotic Hands(3 posts)→

Original post →

More from Embodied

Embodied channel →