RoboCoach: World Models as Active Coaches for Compositional Robot Skills
RoboCoach uses a world model to target robot skill demos, lifting Franka success from 13.3% to 75%.
RoboCoach is a world-model coaching framework that uses imagined failures to choose which robot subtask demonstrations to collect and which skill experts to update. Its Route-Imagine-Diagnose-Improve loop runs reusable experts inside the action-conditioned COACHWORLD model and records the first subtask that fails. With 150 additional demonstrations, success rose from 13.3% to 75.0% on Franka and from 40.0% to 83.8% on AgileX, and coached experts reached 35.0% on four held-out compositions versus 0% for a uniformly updated shared policy. Imagined and real success correlated at rho 0.840 across 22 task-policy pairs.
- RIDI loop finds the first failing subtask inside COACHWORLD.
- 150 extra demos raise Franka success from 13.3% to 75.0%.
- AgileX success rises from 40.0% to 83.8% under the same budget.
- Held-out compositions reach 35.0% versus 0% for a shared policy.
- Imagined and deployed success correlate at rho 0.840 over 22 pairs.
- AI models
- COACHWORLD
From the original article
1 min read at huggingface.co
Long-horizon robot manipulation reuses skills across many task compositions, but improving these compositions with additional end-to-end demonstrations is costly. A practical self-improving system must decide both what to teach next and where to apply that supervision. We present ROBOCOACH, a world-model-guided coaching framework that uses imagined failures to guide demonstration requests and expert updates.
By Jiajun Liu, Yifan Chen, Yichao Liu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.