World models coach robots to improve skills with fewer demos
RoboCoach: World Models as Active Coaches for Compositional Robot Skills
RoboticsArtificial Intelligence
Summary
Robots need to learn complex tasks made of smaller skills, but teaching them everything with real examples is expensive and slow. The authors created a system called RoboCoach that imagines where a robot might fail and asks for help only on those parts, making learning more efficient. RoboCoach uses a virtual model to simulate tasks, finds the first skill that fails, and focuses demonstrations there to improve the robot step-by-step. Tests on robot simulations and real robots show big improvements in success rates with much less teaching needed.
What this means in practice
- •For robotics engineers: Train modular robot controllers more efficiently by focusing demonstrations on specific failing subtasks identified by simulated models.
- •For automation integrators: Improve performance of industrial robots by using virtual coaching to reduce the amount of real-world retraining needed for new task sequences.
Authors
Jiajun Liu, Yifan Chen, Yichao Liu, Jiayi Zhang, Ruoqu Chen, Shaoxuan Xie, Guocai Yao, Mengdi Xu, Sen Cui, Changshui Zhang
Abstract
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. Its Route-Imagine-Diagnose-Improve (RIDI) loop executes reusable skill experts inside COACHWORLD, our shared action-conditioned world model, and uses a progress judge to record the first subtask that fails to complete. Aggregated records select which subtask demonstrations to acquire and which expert adapters to update. Across two simulation suites and two real-robot platforms, imagined and deployed success correlate over 22 task-policy pairs (rho = 0.840). Controlled comparisons show that our coaching method outperforms matched baselines under matched data budgets and update schedules. With only 150 additional subtask demonstrations, success rises from 13.3% to 75.0% on Franka and from 40.0% to 83.8% on AgileX. The coached experts also transfer to four held-out compositions, achieving an average success of 35.0%, compared with 0% for a shared-policy baseline updated with uniformly acquired demonstrations. Together, these results show that world models can serve as active coaches, turning imagined failures into targeted supervision for modular policy improvement. Project Page: https://robocoach-ai.github.io/