Ways to tell, while a robot runs, that it is failing or has been fed something wrong. Most of these are on my to-read list, so the table comes from their abstracts.
| Detector | What it watches | Training data | Alarm | Reported |
|---|---|---|---|---|
| SAFE | a VLA’s internal features | successful and failed rollouts | conformal threshold | works on unseen tasks with OpenVLA, π0 and π0-FAST, in simulation and on real robots |
| FAIL-Detect | scores distilled from the policy’s inputs and outputs | no failure data | conformal threshold | flow-based density scores were the most consistent, and flagged failures sooner than baselines |
| Hide-and-Seek | contrastive signals within and between trajectories | one success or failure label per run | conformal threshold | finds the step where things start to go wrong, on LIBERO, VLABench and a real robot |
| VLM judges (FailBench) | camera video of the attempt | none (general VLMs) or fine-tuned | the model’s verdict | best of 13 reaches 0.77 balanced accuracy, and they lean towards “success” |
| Detect and Suppress | one sparse-autoencoder feature inside a VLA | patched and clean inputs | linear probe | suppressing the feature only when the probe fires restores success under intermittent attacks |
| C-3PO | its own pose estimate against the input points and the shape prior | self-training | two certificates | certified outputs are always correct; it abstains on the rest |
| SOTER | the robot’s state against a safe set | none (reachability) | switch to a safe controller | the advanced controller ran more than 96% of 104 simulated hours |
Notes
- Detectors that read the policy itself (SAFE, FAIL-Detect, Hide-and-Seek) have settled on conformal thresholds. Conformal guarantees need data that looks like the calibration data, which drift and attacks both break.
- Outside judges are cheap and lean towards saying yes. No Free Checker makes the general point: the cheaper and earlier a verdict, the less it tells you.
- Only C-3PO and SOTER say anything about what to do next, and both do it by refusing or falling back. See failure recovery.
Related: Safety and verification, runtime monitoring.