The paper tackles reward design in embodied reinforcement learning, where vision-language models (VLMs) are seen as a promising tool. Existing methods typically train VLMs as reward regressors, but these demand extensive annotations and are primarily designed for finetuning already pretrained policies, limiting their broader use.

The authors propose an alternative: reframe VLMs as success detectors, guided by exploration-augmented key-observation gating. This conceptual shift could sidestep the need for dense reward annotations and extend applicability beyond finetuning scenarios.

The abstract does not include experimental results, so the method's effectiveness is not yet established. Still, moving from regressor to success detector is a notable direction for reducing the burden of reward specification in embodied RL.