RoboScript hits 83.4% success on RoboTwin 2.0 with one demo per task, beating trained policies by 20+ points

H-EmbodVis · hf · 2026-11-11

HUST and ArcheOne AI propose World In-Context Learning (WICL), a new paradigm for agentic robot manipulation, implemented as RoboScript.

Key insight: what an agent learns in context is semantic (task steps, object roles), not visual appearance, so experience transfers across worlds
Method: a frozen model starts from a single demonstration, rehearses the task in simulation, writes feedback into its context rather than its weights, and freezes the result as a Task Memory for deployment on new scenes
Results: across all 50 RoboTwin 2.0 tasks with one demo per task and zero gradient updates, RoboScript reaches 83.4% mean success in clean and randomized settings, over 20 points above the strongest trained action policy
Generalizes across visual distributions, improves via reflection, scales with the foundation model, and transfers to a physical dual-arm robot via cross-embodiment memory adaptation

Code is open source.

More from Embodied