Machine learning-based local collision avoidance for maritime navigation

• Main Author: Zou, Yixuan
• Other Authors: Jiang Xudong
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2024
• Subjects: Engineering; Maritime navigation; Machine learning; Collision avoidance
• Online Access: https://hdl.handle.net/10356/177278
• Institution: Nanyang Technological University

This report investigates the application of Deep Reinforcement Learning (DRL) in collision avoidance in maritime navigation. In this study, the transfer of DRL methods from land to sea environment was studied, with a focus on the ability of this extensive range of techniques to generalize. Our thorough analysis classifies significant DRL studies according to important variables such as evaluating potential dangers, selecting algorithms, adhering to International Regulations for Preventing Collisions at Sea (COLREGs), and utilizing Automatic Identification System (AIS) data. In this study, we also examine the differences between DRL approaches that are implemented on-policy and off-policy and evaluate their appropriateness for various navigation scenarios. As a former study for the model training in sea environment, agent was trained in various ground environment using TD3 algorithm during the research. By capitalizing on the sophisticated dynamics of the simulation environment and utilizing the strengths of DRL, this integrated approach guarantees a solid foundation for further research of autonomous vessel navigation.