Effect of shrinking embodiment gap via more human-like robot hardware on transfer and zero-shot execution

Ascertain whether reductions in the human–robot embodiment gap achieved by deploying robots with more human-like kinematics and dexterity enable stronger human-to-robot transfer and potentially zero-shot execution on novel tasks for the EgoScale policy, and characterize the conditions under which such transfer is realized.

Background

EgoScale demonstrates cross-embodiment transfer from human-pretrained representations to robots with substantially different hands and kinematics, aided by a small amount of aligned mid-training data. Despite these successes, an embodiment gap persists between human motion and robot actuation.

The authors point to an open direction: as robot hardware becomes more human-like, the embodiment gap may shrink, potentially enabling stronger transfer and even zero-shot execution. Establishing and quantifying this effect remains an unresolved question.