SAFE defines multitask failure detection: one detector that must work on tasks and environments it never saw in training. It reads a VLA’s internal features, which turn out to carry generic signals of success and failure, and outputs a single failure score, trained on both successful and failed rollouts. It is tested on OpenVLA, π0 and π0-FAST in simulation and on real robots, with conformal prediction setting the trade-off between accuracy and detection time.
Relevance
Generalist policies need generalist monitors, and SAFE is often cited for that claim. Read with FAIL-Detect and Hide-and-Seek for runtime monitoring in robust robotics.
Links
- arXiv · project page
- Subfield: Safety and verification, Perception