Peking University's Open-HumanWAM: open recipe scales human experience to 88.65% on RoboTwin

CyberRobooo · x · 2026-11-19

Peking University's DA Group released Open-HumanWAM, a fully open recipe (paper, code, models, datasets) for embodied foundation models that learn from human experience: egocentric video + language + hand-motion pretraining, then robot post-training, scaling from 100 to 50,000 hours of human data. Results: RoboTwin zero-shot 13.94% (rank 1 open-source, no task-specific post-training), 88.65% post-trained (rank 1 in reported comparisons), and real-robot zero-shot improving from 20.0% to 30.5% with human-robot hybrid data. Deployed on Astribot S1, it demos multi-stage household tasks including a 14-stage desk organization sequence.

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

Original post →

More from Embodied

Embodied channel →