FAIL-Detect detects failures of imitation-learning policies without any failure data, treating the task as sequential out-of-distribution detection. It distils policy inputs and outputs into scalar scores that track epistemic uncertainty, and sets alarm thresholds with conformal prediction. Learned scores, especially from a flow-based density estimator, were the most consistent, and it flags failures more accurately and sooner than the baselines across several manipulation tasks.
Relevance
Background for the 2026 VLA detectors in this queue, several of which use the same conformal thresholding (SAFE, Hide-and-Seek). Read for runtime monitoring in robust robotics.
Links
- arXiv
- Subfield: Safety and verification, Robot learning