Shengshu's Motus2 Robot World Model Closes the Loop: Act, Predict, Self-Evaluate
量子位 · wechat · 2026-09-12
Shengshu (生数科技) released Motus2, a robot world model unifying three skills in one model — action (WAM), prediction (AC-WM), and evaluation (VM) — forming a closed self-improvement loop: generate action → predict outcome → evaluate → update policy. The original Motus beat Pi-0.5 by 35+ points absolute success rate across 50 tasks.
Key techniques:
- Action-first information flow forces the model to emit actions before seeing results, preventing future-frame leakage that sinks many video+action-control approaches
- Best-of-N planning + model-based RL: at inference, candidate actions are imagined and scored; in training, scores become policy-update signals. Real-robot success rose from 65% (base) to 75% (combined)
- Lightweight tactile expert reuses the backbone's intermediate features, lifting paper-tearing/cup tasks from 60% to 72.5%
- Memory: caching recent real observations reaches 57.5% on hidden-object tasks, far above compressed-history variants
- 130k-hour egocentric human data pipeline (monocular → binocular+hand → robot adaptation): robot-domain mid-training lifts 5-task success from 51% to 84%
Deployed on 22-DoF SharpaWave and 20-DoF WUJI Hand2 dexterous hands for bulb-screwing, page-turning and multi-finger manipulation. Under Shengshu's L1-L5 world-model roadmap, Motus2 reaches L3 (acting in the world) and begins touching L4 recursive self-improvement; long-horizon open-world learning remains open.
Related event: Shengshu Unveils Motus2, a Self-Evolving Robot World Model(2 posts)→
More from Embodied
- VJ interface for ReCamera built with Astra in 10 minutes reacts to people and music — Stefania_druga · 2026-09-12
- visionOS 27 Quick View brings fully animated 3D models into your space — Scobleizer · 2026-09-12
- Robotics economics: pick-and-place worth $1/box in factories, six-figure contracts in data centers — paigeinsf · 2026-09-12
- GPT-6 Astra Rebuilds Office Scan in Blender, Drops Humanoid Gym Into Newton Sim — liuziwei7 · 2026-09-12
- Building a physical Codex Pet for OpenAI Dev Day — gabrielchua · 2026-09-12
- Skild AI founder explains why robotics data needs four sources, each flawed — deepakpathak · 2026-09-12