Can a robot become too good for its own monitor? That is the question posed by a new arXiv preprint, which looks at what happens when a fixed vision-language model (VLM) is used to judge whether a robot has completed a task while the robot's policy is continuously improved. The authors set up a controlled tray-placement experiment with 20 seeds, where the VLM's success judgments are evaluated against a policy that gets better over time.

The central issue is coverage shift. As the policy improves, it visits states and actions that differ from those the VLM was originally trained on. The monitor is then operating outside its comfort zone, and failures that occur in these new regions become harder to detect. In other words, the very improvements that make the policy more capable can also make it less transparent to a fixed observer.

The finding has practical implications for robot learning pipelines. If success monitors are not updated or retrained alongside the policy, they may silently degrade, giving false confidence in system performance. The authors argue for monitoring systems that can adapt to the changing distribution of policy behavior.