Predict the Contact That Changes the Decision

Abstract

Robot world models may predict plausible futures while ranking near-identical recovery actions incorrectly. We propose to train a small contact-event critic on action branches replayed from the same simulator state. The critic will rank recovery actions by the likelihood of task-relevant events such as clearing a jam or losing a grasp. The study will measure whether paired supervision improves downstream decisions at fixed model capacity and rollout cost, including under novel object geometry.

Keywords