WICL: robots learn manipulation in-context, 83.4% success on all 50 RoboTwin tasks
_akhaliq · x · 2026-11-12
The paper proposes World In-Context Learning (WICL), based on the observation that what an agent learns in context is semantic — task steps and object roles rather than visual appearance — so experience transfers across worlds. A frozen model starts from one demonstration, rehearses the task in a simulator, and writes world feedback into its context instead of its weights. The RoboScript infrastructure freezes learned knowledge as a Task Memory and deploys it with data isolation on real robots. On all 50 RoboTwin 2.0 tasks, RoboScript hits 83.4% mean success with one demo per task and no gradient updates, beating the strongest trained action policy by 20+ points; it generalizes across visual distributions, improves via reflection, and transfers to a physical dual-arm robot through cross-embodiment memory adaptation. Code is open.
Related event: RoboScript's World In-Context Learning Hits 83.4% Success with Single Demo(2 posts)→
More from Embodied
- World Labs and AMD enter robotics, Li Fei-Fei appears at CoRL — breadli428 · 2026-11-12
- SimEX: agents learn manipulation skills in imperfect simulation and transfer them to real robots — chris_j_paxton · 2026-11-12
- Feather Robotics' $33,000 semi-humanoid tried hands-on: 15 lb payload per arm — chris_j_paxton · 2026-11-12
- Lie bracket visualized: how second-order residue lets wheeled robots move without sliding — burny_tech · 2026-11-12
- UCLA's TriMI: a $800 tri-modal robotic hand interface for demos, teleop and execution — chris_j_paxton · 2026-11-12
- CoRL 2026: papers say teleop is optional, but demo-floor humanoids still rely on it — jonstephens85 · 2026-11-12