Active Maritime Copilot for Ambiguity and Risk-Driven Decisions

Cross-modal conflicts in maritime navigation—where a vessel’s verbal communication contradicts its physical maneuvers (e.g., promising to give way while maintaining speed) pose severe risks to safety. Current autonomous systems often process sensor data and linguistic inputs in isolation, failing to detect such discrepancies. We present a Multimodal Agentic Framework that serves as a “Watchful Copilot,” using Retrieval-Augmented Generation (RAG) to cross-reference navigational dialogue with real-time kinematic data. To manage uncertainty, a Risk-Prioritized Interface employs progressive disclosure, escalating from a “Green” (Verified) state to a “Yellow” (Ambiguous) state, where the agent visualizes supporting evidence and requests human supervision for clarification. Preliminary validation in a 2D simulation benchmark (N=13) provides initial evidence that this human-in-the-loop workflow may support reduced cognitive load and appropriate trust calibration in high-ambiguity scenarios, warranting further investigation.