Hide-and-Seek in Trajectories: Discovering Failure Signals for VLA Runtime Monitoring

Researchers introduced Hide-and-Seek, a runtime monitoring framework for detecting failures in vision-language-action robot policies during execution.

It uses inter-trajectory and intra-trajectory contrastive learning to find failure-indicative actions without step-level annotations.

In tests on LIBERO, VLABench, and a real robot with OpenVLA, π_0, and π_0.5, it outperformed prior methods and generalized to both seen and unseen tasks.

The work could make embodied AI safer and more dependable by enabling reliable failure detection while robots carry out language-driven tasks.