Peking University's Open-HumanWAM trains humanoid robots on 50,000 hours of human video

CyberRobooo · x · 2026-11-19

Peking University's DAGroup released Open-HumanWAM, a World-Action Model trained on 50,000 hours of egocentric human video to learn hand movements, object interactions, and everyday manipulation, then adapted to robots with robot-collected data. Deployed on the Astribot S1, it handled multi-stage household tasks including a 14-stage desk organization sequence. In real-robot tests across 20 tasks, adding 45,938 hours of human data raised average zero-shot success from 20% to 30.5%.

Related event: Peking University open-sources Open-HumanWAM for household robots(2 posts)→

Original post →

More from Embodied

Embodied channel →