open access

https://hdl.handle.net/2115/100453

Quantitative representations of COLREG-compliant behaviour from large-scale AIS trajectory data

英語

Autonomous ships

2026-07-01

Proceedings of the 10th International Symposium on Reliability Engineering and Risk Management (ISRERM 2026)

The safe integration of autonomous ships into existing maritime traffic requires reliable models of navigational behaviour that conform to the International Regulations for Preventing Collisions at Sea (COLREGs). In this presentation we give a brief introduction to how machine-learning methods can be used to extract quantitative representations of COLREGcompliant behaviour from large-scale AIS trajectory data. We first indicate how to establish encounter-detection and situation-classification procedures that identifies head-on, crossing, and overtaking scenarios, along with give-way and stand-on roles. Combining trajectory embedding and sequence modelling, this can be used to identify patterns characteristic of human navigators in different COLREG situations. Supervised learning, imitation learning, and inverse reinforcement learning may then be applied to infer decision models that map encounter geometry to heading and speed adjustments. This way one can capture implicit navigational preferences based on real traffic. Finally, we outline how these learned behaviour policies can serve as priors or constraint mechanisms in autonomous-ship decision systems, enabling behaviour that is both predictable to human mariners and consistent with maritime collision-avoidance rules.

会議発表論文

Version of Record