Host: The Japan Society of Naval Architects and Ocean Engineers
Name : 2025 Annual Autumn Meeting
Number : 41
Location : Himeji Culture and Convention Center Operation Consortium
Date : November 17, 2025 - November 18, 2025
Pages 9-11
Maritime autonomous surface ships (MASS) require route planning that ensures safety and compliance with the International Regulations for Preventing Collisions at Sea (COLREGs) under complex multi-ship and geographical conditions. However, designing explicit objective functions for multi-objective optimization remains challenging due to trade-offs among safety, efficiency, and rule compliance. This paper presents three data-driven approaches: inverse reinforcement learning (IRL), imitation learning (IL), and diffusion policy (DP), which learn directly from expert demonstrations without predefined objectives. Using simulator data from professional captains and real-world navigation data from the training ship Fukae-Maru, IRL inferred COLREGs-consistent reward functions, IL reproduced human-like collision-avoidance maneuvers, and DP generated feasible trajectories in congested waterways. The results demonstrate that learning-based planners can achieve COLREGs-compliant and human-acceptable navigation behaviors. The proposed framework provides a scalable foundation for extending autonomous navigation to cover remaining COLREGs rules and diverse local maritime regulations.
