Shengshu's Motus2 world model hits 84% success on dexterous robot tasks via self-evolving loop

生数科技 · wechat · 2026-09-12

Shengshu Technology released Motus2, a self-evolving general world model for dexterous manipulation that unifies policy, simulator, and evaluator in a single shared-parameter model, closing the loop of action generation, future prediction, outcome evaluation, and policy updates. A lightweight tactile expert module refines actions using contact feedback.

Motus2 is pretrained on 130k hours of egocentric human video, then adapted with 100+ hours of robot trajectories and human-robot alignment data. It averages 84% success across five real-robot tasks: human-video pretraining alone reaches 51%, and adding robot mid-training lifts it 33 points; model-based RL plus Best-of-N inference-time planning raises success from 65% to 75%. It has been deployed on 20+ DoF tactile hands including SharpaWave and WuJiHand2, with demos like lightbulb screwing, two-handed paper tearing, page flipping, and hidden-object search.

Original post →

More from Embodied

Embodied channel →