Semantic Hazard Detection for Maritime Autonomy with Vision-Language Models

This paper addresses a critical challenge in maritime autonomy: handling out-of-distribution situations that require semantic understanding. It proposes a novel approach using vision-language models (VLMs) to detect hazards and trigger safe fallback maneuvers, aligning with the requirements of the IMO MASS Code. The focus on a fast-slow anomaly pipeline and human-overridable fallback maneuvers is particularly important for ensuring safety during the alert-to-takeover gap. The paper's evaluation, including latency measurements, alignment with human consensus, and real-world field runs, provides strong evidence for the practicality and effectiveness of the proposed approach.