Vision-language navigation is moving into photorealistic, physically simulated scenes, according to a new arXiv paper. The leading zero-shot agents in this setting place a vision-language model in the control loop, using it to choose every move and when to stop.

The paper, titled "Structure over Reasoning: Unified Zero-Shot Navigation with a Small Vision-Language Model," suggests that relying on structural cues may be more effective than heavy reasoning. It specifically advocates for a small vision-language model, noting that robots favor small models in practice.

The abstract is brief, so details on the method and experiments are not yet available. But the central claim is clear: for zero-shot navigation, structure can outperform reasoning while keeping the model small enough for real robots.