Humanoid Robots Can Now Learn Actions by Watching Humans

jiqizhixin · x · 2026-07-17

This paper introduces Human-as-Humanoid: enabling humanoid robots to learn actions directly from synchronized first-person and third-person human videos, bypassing traditional teleoperation for data collection.

Methodologically, they first aligned the design of a 60-DoF humanoid robot with human anatomy. Then, using phased inverse kinematics, they translated human motions into robot action sequences. The authors claim this data collection method is 4.8–7.2 times faster than teleoperation.

Crucially, the paper demonstrates that policies trained solely on these converted human videos can successfully generalize to real robots, achieving a "zero-shot" learning effect from human demonstrations.

Original post →

More from Embodied

Embodied channel →