CORALL: A COLREGs-Guided Risk-Aware LLM for Decision-Making in Maritime Autonomous Surface Ships

The paper introduces a novel Large Language Models (LLMs)-based solution for real-time COLREGs-based decision-making in collision encounters. The COLREGs, published by the International Maritime Organisation (IMO), have underpinned maritime collision avoidance regulations in open seas to ensure vessel safety. As COLREGs were originally defined for human seafarers, their translation into computer language for autonomous operations has received significant attention in the literature; however, a universal solution remains unavailable. A key requirement is explainable decision-making for autonomous operations or captain assistance. This study leverages the power of LLMs to develop such a decision-making algorithm. The proposed COLREGs-guided Risk-Aware LLM (CORALL) algorithm integrates a supervisory decision-making module with an execution layer for path tracking and collision avoidance under external disturbances. The resulting framework enables COLREGs-compliant, risk-informed LLM-based navigation, with online performance demonstrated on a high-speed ship model. The tailored LLM processes navigation outputs and risk indices, identifies the COLREGs encounter type, and generates decisions with accompanying explanations. The proposed solution is tested on all the 22 Imazu's benchmark problems, evaluated using the Imazu benchmark scenarios, a widely accepted standard in maritime collision avoidance research comprising 22 progressively complex encounter situations. To complement the study, the proposed algorithm is verified using a developed Hardware-in-the-Loop (HIL) test rig, with end-to-end execution times measured to demonstrate the feasibility of real-time implementation on low-cost processors. This risk-aware AI solution, thought to be the first of its kind, opens a door to future research on advanced decision-making algorithms in safe maritime transport and similar operational environments.