Diffusion route planner: Data-driven modeling of human ship navigation that implicitly balances safety, efficiency, and rule compliance under complex geographical constraints

This paper presents a diffusion route planner (DRP), a data-driven framework for modeling human ship navigation from AIS–radar-integrated data. The DRP employs diffusion policy (DP) trained on 181 days of real-world navigation trajectories collected from Kobe University's training ship Fukae-Maru. We first evaluated the DRP in 63 basic one-on-one scenarios and found that it consistently adhered to COLREGs Rules 13–17 without intruding into safety buffer regions. We then validated its performance in six real-sea scenarios featuring multi-ship encounters and complex geographical constraints, demonstrating efficient routing with sufficient safety margins. Compared to existing data-driven approaches, the DRP handled more complicated and realistic situations due to the offline training using AIS–radar-fused data. Compared to optimization-based approaches, the DRP showed potential to account for even greater complexity, such as remaining COLREGs rules and local traffic regulations. These findings underscore the importance of real-world data acquisition and are expected to accelerate future research on data-driven autonomous maritime navigation. • A diffusion-based route planner was proposed to model human ship navigation. • The method was trained on 181 days of AIS–radar-integrated real-sea shipping data. • COLREGs Rules 13–17 were satisfied in all basic one-on-one encounter scenarios. • Safe and efficient routes were generated in geographically constrained scenarios. • The approach reproduced human-like behavior without explicit rule encoding.