A new training system teaches robots by making them fail inside a simulation first, then using those imagined mistakes to decide exactly what to teach next.
Researchers built RoboCoach, a framework that pairs reusable robot skill modules with a shared world model called CoachWorld. The system runs each skill module in simulation, flags the first subtask where it breaks down, and uses that record to pick which real-world demonstrations to collect and which module to retrain. Tested across two simulation suites and two real robot platforms, Franka and AgileX arms, simulated failure rates tracked real-world failure rates closely, with a correlation of 0.840 across 22 task-policy pairs. Adding just 150 targeted demonstrations chosen this way raised Franka's success rate from 13.3% to 75.0% and AgileX's from 40.0% to 83.8%.
Most attempts to improve long-horizon robot manipulation just throw more demonstrations at the whole task, which is expensive and often wasted on skills that already work fine. By using a world model to predict where a robot will actually trip up, RoboCoach aims that costly human demonstration time at the real bottleneck instead of spreading it evenly. The coached modules also carried over to four new task combinations, averaging 35.0% success versus 0% for a baseline updated without that targeted guidance.
The results come from controlled comparisons with two robot arms in a lab, not a crowded warehouse floor, and a 0.840 correlation in simulation is not the same as reliability in the wild.