Vision-language navigation (VLN) agents struggle in real-world indoor fires because smoke, heat, and low light degrade visual input, breaking the link between language instructions and what the agent can perceive. According to a new arXiv preprint, this degradation impairs three core capabilities: grounding instructions to visual cues, predicting waypoints, and exploring over long horizons. The authors argue that simply restoring images is a natural but insufficient response.
The proposed method, called FireNav, introduces decision-oriented reconstruction evidence routing. Instead of treating image restoration as a separate step, FireNav routes reconstruction efforts toward the evidence most relevant to the navigation decision at hand. This way, the agent prioritizes visual information that directly supports instruction grounding, waypoint selection, and sustained exploration in degraded scenes.
The preprint is early-stage and does not yet report experimental results, so the effectiveness of FireNav remains unverified. However, it highlights a growing research direction: adapting VLN systems to extreme, safety-critical environments rather than assuming clean, well-lit indoor settings.