coarse-supervision

Hide-and-Seek is a framework that detects robot execution failures in VLA models by localizing failure-indicative actions through contrastive learning without step-level annotations, achieving state-of-the-art multi-task failure detection.