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.
More from Embodied
- Weekend robot fights at autonomous labs: train your bot by chatting with an AI coach — dee_hw · 2026-09-12
- Flying humanoid robot installs AC on 30th floor in China, remote operator replaces rope climbers — TansuYegen · 2026-09-12
- 11 Years of Humanoid Robotics: From DARPA Falls to 2,056-Robot Games — TinfoilTricorn · 2026-09-12
- PokeBot CTO predicts zero-shot generalization on 80% of household tasks within a year — suchenzang · 2026-09-12
- GPT-6 Astra rumored to use recurrent-depth latent reasoning, sparking AI safety alarms — DJiafei · 2026-09-12
- Robot drags a 50kg rice bag needlessly — how do we measure robot "stupidity"? — ducha_aiki · 2026-09-12