UHAS: A Unified Action Space for Cross-Embodiment Robot Manipulation

DJiafei · x · 2026-08-11

Current robot manipulation policies are typically tied to specific robotic hand embodiments, limiting cross-platform transfer. This paper proposes the Unified Hand Action Space (UHAS), a sphere-based action representation that models hand movements as geometric deformations of a canonical sphere. A Cascade Inverse Kinematics (CIK) algorithm then maps this shared representation to embodiment-specific joint configurations.

Using reinforcement learning, the researchers trained dexterous manipulation policies directly within UHAS for in-hand cube reorientation. Evaluated across multiple embodiments—including the Allegro, LEAP, Shadow Hand, and MANO Human Hand—the method demonstrated stable control, zero-shot transfer to unseen hands, rapid finetuning, and successful real-world deployment.

Original post →

More from Embodied

Embodied channel →