Motus2: a robot world model that grades its own actions
Shengshu Technology, the Chinese world-model lab, shipped the rare robotics release where the loop actually closed: one model that acts, imagines, and scores itself — then learns from the grade.
Shengshu released Motus2, a "self-evolving" general world model for dexterous manipulation that puts three capabilities — acting, predicting, and evaluating — inside one shared-parameter video-action model, according to the project page and the paper on arXiv. A policy proposes candidate action chunks, an action-conditioned simulator predicts the visual consequences of each, and a value model ranks them before one is executed; the robot then re-observes reality and plans again. Shengshu frames the cycle — generate action, predict outcome, evaluate, update strategy — as a first step toward recursive self-improvement in embodied AI. The trick that makes it honest is timing: Motus2's mid-training uses an action-first information flow so the model can't peek at future frames when choosing the current action, the shortcut that quietly wrecked earlier "video generation plus action control" recipes. Failed trajectories, normally filtered out as noise, are kept to teach dynamics and value instead.
The ablations are concrete. On phone-placement and multi-finger tasks, the base policy averaged 65% success; adding Best-of-N planning alone lifted it to 67.5%, model-based reinforcement learning alone to 72.5%, and both together to 75%. A separate line: a lightweight tactile expert reuses the backbone's cached features to refine each short action sub-chunk just before execution — keeping touch cheap enough to run at hand speed — lifting cup-pickup and paper-tearing tasks from 60% to 72.5%. On memory, caching full observation history beat compressing it on hide-the-block and history-cued button tasks (57.5% average), though Shengshu presents that as a compromise, not a solution.
The data story may matter most: roughly 130,000 hours of first-person human video underpins pretraining, but human data alone yielded 51% on five real-robot tasks while adding robot mid-training jumped it to 84%. Scaling stereo human data from 2,000 to 20,000 hours kept cutting validation error with no plateau yet. Motus2 runs on high-DoF hands — 22-DoF Sharpa Wave, 20-DoF WUJI Hand 2 — doing bulb-screwing and page-turning.
This is the "predict the world, then act in it" idea we unpacked in AI 101 — What is a world model? finally wired into a learning loop — and Shengshu is explicit that touch transfer across embodiments and long-horizon reliability remain open.
What to watch: whether the self-evaluation loop holds up on tasks measured in minutes rather than seconds.
Is a robot that grades its own attempts the roadmap to general manipulation, or a very expensive way to learn to tie its own shoelaces? Tell us in the comments.
Sources: QbitAI · Motus2 project page · arXiv:2608.30237
