Deep Reinforcement Learning for Ship Collision Avoidance and Path Tracking

Abstract This research focuses on collision avoidance between ships in the maritime industry, addressing human decision errors as a significant contributor to ship-to-ship collision. Emphasizing the crucial role of automation in maritime operations, the study employs advanced algorithms, specifically Deep Q Network, to enhance dynamic collision avoidance capabilities. By automating decision-making processes, the research aims to significantly reduce collision occurrences between ships, highlighting these algorithms as promising within the realm of deep reinforcement learning. Utilizing a three-degree-of-freedom dynamic model and Krisco field ship as a benchmark hull, rigorous numerical simulations validate the proposed model’s accuracy. The reinforcement learning agent, trained on this dynamic model, strives to optimize collision avoidance and waypoint tracking, demonstrated through numerical simulations and model experiments with a scaled version for a comprehensive evaluation of its maritime safety efficacy.