Vision-language-action models for embodied navigation typically rely on RGB images. This becomes a problem in "lights-out" industrial environments where there is no visible illumination. A new paper on arXiv introduces ADAPT to address this gap.

ADAPT stands for Audio-Depth Action Planning with Transition-Aware Rewards. Instead of using RGB, it leverages audio and depth information to guide navigation. The method also incorporates transition-aware rewards, which likely help the model make better decisions about state changes, though the abstract does not provide further detail.

The abstract is brief and does not include experimental results or specific performance numbers. It only states that the authors "study" the problem, leaving the full findings to the full paper. Still, the proposal highlights a practical limitation in current navigation models for industrial settings.