Robot models learn from executed action alternatives to predict failures
In plain words
Robot world models predict what may happen after a robot acts. The paper finds that several released models often predicted success even for actions that failed, and asks whether those models can be repaired without changing how they are built.
CureWM changes actions from successful demonstrations, runs the changed actions, and trains the model on both the failures and successes it observes. In the tested simulations and a small physical-robot study, predictions better separated success from failure in several settings, but not every task. Better predictions do not show that the robot became better at completing tasks.
The research
A robot world model can predict success even when an action fails, which makes its predictions unreliable for distinguishing actions that work from those that do not. Can a released model learn that distinction from new training examples while keeping its architecture and training objective?
CureWM trains released robot world models on executed action alternatives, helping their predictions separate success from failure.
CureWM asks whether a released robot world model can learn to distinguish successful actions from failing ones without changing its architecture or training objective.
The method changes actions in successful demonstrations, executes the alternatives to verify their outcomes, then trains on both successful and failing replays alongside the original data.
In simulation, predictions separated success from failure better on LIBERO-Goal than with official-data training alone, and also improved on two other suites. A small physical-robot evaluation showed fewer failure sequences predicted as success-like.
Together, these results support predictive repair in tested settings, while the uneven suite results and false alarms on successful actions limit the conclusion.
These results concern prediction quality in the tested simulations and small hardware evaluations. They do not establish broader transfer or improved robot control; transfer varied by task, and the physical tests used one robot and unequal training-data counts.