Recent advancements in artificial intelligence (AI) have had a significant impact on various sectors, especially maritime navigation. Since human behavior plays a major role in maritime accidents, there is a growing need for autonomous solutions to enhance safety. Additionally, the maritime industry faces a shortage of skilled captains, further necessitating the development of autonomous systems. Consequently, this study focuses on applying deep reinforcement learning (DRL) techniques to develop Maritime Autonomous Surface Ships (MASS), particularly addressing ship collision avoidance in compliance with the International Regulations for Preventing Collisions at Sea (COLREGs). We introduce a collision avoidance system within a custom Unity simulation environment, designed to handle static obstacles and dynamic encounter scenarios. Our approach utilizes Proximal Policy Optimization (PPO) to train an autonomous agent through a comprehensive state space and continuous rudder control. A detailed reward structure guides the agent towards safe and efficient navigation. The effectiveness of this methodology is validated through experiments in path-following, overtaking, head-on, and crossing give-way scenarios. Results show that the PPO algorithm enables COLREG-compliant maneuvers and successful navigation in these scenarios. This research advances autonomous maritime navigation, demonstrating DRL’s potential to enhance safety and efficiency in maritime operations.
Autonomous Shipping in Complex Situations
Lect. Notes on Data Eng. and Comms.Technol.
International Conference on P2P, Parallel, Grid, Cloud and Internet Computing ; 2024 ; San Benedetto del Tronto, Italy November 13, 2024 - November 15, 2024
Advances on P2P, Parallel, Grid, Cloud and Internet Computing ; Kapitel : 30 ; 327-335
17.11.2024
9 pages
Aufsatz/Kapitel (Buch)
Elektronische Ressource
Englisch
Human-Centred Autonomous Shipping
GWLB - Gottfried Wilhelm Leibniz Bibliothek | 2024
|Human-Centred Autonomous Shipping
GWLB - Gottfried Wilhelm Leibniz Bibliothek | 2024
|