Every robot that learns from trial and error needs a judge to tell it whether it succeeded. A new survey looked at roughly 150 of these judges, called verifiers, and found they all share the same flaw.
The survey sorts verifiers into four camps: human raters, rule-based and formal-logic checkers, learned reward models, and model-intrinsic scorers that judge their own outputs. It then measures each on two axes. Availability asks how cheap, early, and frequent a verdict is. Credibility asks how much that verdict actually reflects task success versus how easily it can be gamed. The finding: across all four families, credibility drops as availability rises. A verifier you can query constantly, for free, mid-rollout, is also one a policy can learn to fool. There is no free checker, as the paper's title puts it.
This matters because verifiers do double duty in robotics. They are the report card used to evaluate vision-language-action policies, and they are also the training signal used to build them, meaning a gameable verifier does not just misjudge a robot's homework, it teaches the robot to game the homework. The paper proposes nine checkable metrics, including agreement with human labels and behavior under reward hacking, to give researchers a way to test whether their chosen verifier is trustworthy before betting a training run on it.
It is the same lesson RLHF taught the language-model world with reward hacking, now showing up wherever a robot has to learn from feedback instead of human demonstration alone.