Hide-and-Seek trains a failure detector for vision-language-action (VLA) policies from one success or failure label per trajectory, with no per-step annotation. Contrastive losses between and within trajectories push the detector to find the actions where things start to go wrong, so the failure score rises at a sensible moment instead of being smeared over the whole run. It is tested on LIBERO, VLABench and a real robot with OpenVLA, π0 and π0.5, with alarm thresholds set by conformal prediction.
Relevance
Most VLA monitors only get told whether the episode failed. If a detector can find the onset from that weak label, it is a cheap route to runtime monitoring for robust robotics. Read it next to SAFE, which works on the same multitask setting.
Links
- arXiv
- Subfield: Safety and verification, Perception