No Free Checker surveys about 150 verifiers for robot policies: success detectors, reward models, runtime monitors, safety filters and temporal-logic specifications. It compares them on availability (how cheap, early and frequent a verdict is) and credibility (how much a high score tells you about the task), and finds that across human, formal, learned and model-intrinsic verifiers, credibility falls as availability rises. It ends with nine metrics meant to make claims about a verifier checkable.
Relevance
A map to read before picking a monitor for robust robotics. The trade-off it names is the one every runtime monitor faces in the field: the cheap early alarms are the ones you can least trust.
Links
- arXiv · code
- Subfield: Safety and verification, Robot learning