RoboCoach: World Models as Active Coaches for Compositional Robot Skills
RoboCoach uses imagined robot-task failures to decide which subtask demonstrations to request and which skill experts to update.
The paper’s RIDI loop runs reusable skill experts inside a shared action-conditioned world model, then flags the first subtask that fails. Those failure records drive targeted data collection and adapter updates. In tests across simulation and real robots, imagined success tracked deployed success across 22 task-policy pairs. With 150 added subtask demonstrations, reported success rose to 75.0% on Franka and 83.8% on AgileX. HF Daily Papers' note
The paper’s RIDI loop runs reusable skill experts inside a shared action-conditioned world model, then flags the first subtask that fails. Those failure records drive targeted data collection and adapter updates. In tests across simulation and real robots, imagined success tracked deployed success across 22 task-policy pairs. With 150 added subtask demonstrations, reported success rose to 75.0% on Franka and 83.8% on AgileX. HF Daily Papers' note
score 4