World In-Context Learning for Agentic Robot Manipulation
RoboScript reports 83.4% mean success across 50 robot manipulation tasks, using one demonstration per task and no gradient updates.
The system starts with a frozen model and rehearses each task in simulation, storing feedback in its context rather than updating its weights. It then freezes that knowledge as a Task Memory for deployment in new scenes. Across clean and randomized RoboTwin 2.0 settings, the authors report beating the strongest trained action policy by more than 20 percentage points. They also report transfer to a physical dual-arm robot through adaptation of the memory across robot embodiments.
HF Daily Papers' note
The system starts with a frozen model and rehearses each task in simulation, storing feedback in its context rather than updating its weights. It then freezes that knowledge as a Task Memory for deployment in new scenes. Across clean and randomized RoboTwin 2.0 settings, the authors report beating the strongest trained action policy by more than 20 percentage points. They also report transfer to a physical dual-arm robot through adaptation of the memory across robot embodiments.
HF Daily Papers' note
score 5