A verifier for robot policies reads a candidate behavior and returns a score, serving both evaluation of vision‑language‑action policies and their training. Verifiers come in many forms, such as success detectors, reward models, runtime monitors, safety filters, and temporal‑logic specifications. This survey examines roughly 150 verifiers and compares them along two axes:
Availability measures how cheap a verdict is, how early it arrives during a rollout, and how densely it can be queried. Verdicts that are cheaper, earlier, and more frequent have higher availability.
Credibility measures how much a high score tells us about the actual task. Credibility drops when judgments become gameable or self‑serving.
We group verifiers by the source of judgment: human verifiers, rule‑based and formal verifiers, learned and pretrained verifiers, and model‑intrinsic verifiers. Across these four families we find that credibility declines as availability rises. Regardless of who supplies the judgment, there is no free checker.
We then examine how to validate a verifier itself and what a high score truly indicates. The literature reports three common measures: agreement with human labels, the performance of the policy it trains, and behavior under reward hacking.
Finally, we propose nine metrics that make a verifier’s claim checkable and provide coordinates for verifiers that still need to be built.
Review