Research on real-time collision avoidance and path planning of USVs in multi-obstacle ships environment

• Published in: Ocean engineering Vol. 295; p. 116890
• Main Authors: Xu, Xinli, Cao, Yunlong, Cai, Peng, Zhang, Weidong, Chen, Hongtian
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.03.2024
• Subjects: Collision avoidance; Compound experience replay mechanism; Deep deterministic policy gradient; Sum-tree data extraction; Unmanned surface vessels; Compound experience replay mechanism; Deep deterministic policy gradient; Unmanned surface vessels; Collision avoidance; Sum-tree data extraction
• ISSN: 0029-8018; 1873-5258
• DOI: 10.1016/j.oceaneng.2024.116890

Collision avoidance is the key to ensure the safe navigation of unmanned surface vessels (USVs). This paper proposes a collision avoidance and path planning algorithm (CAPP) algorithm for real-time collision avoidance and path planning of USV in the marine environment with multiple obstacle ships. Firstly, a mathematical model of both the encounter and action situations is introduced to address the problem of unifying the quantitative standard for encounter situations. Secondly, a deep deterministic policy gradient algorithm, suitable for continuous action space, is improved. A reward function model integrating international regulations for preventing collisions at sea (COLREGs) and driving habits is proposed to solve the problems of stability, economy, and comfort in the process of sailing for USVs. Thirdly, a compound experience replay mechanism is proposed to increase the training efficiency of the network model by selecting the experience with high priority. Additionally, a sum-tree data extraction method is introduced to improve sampling efficiency. Finally, test results on the simulation and training platform are used to verify the functionality and effectiveness of the proposed CAPP algorithm. •A mathematical model of both the encounter and action situations is established.•A reward function model integrating COLREGs and driving habits is proposed.•A compound experience replay mechanism is proposed to increase the training efficiency.•A sum-tree data extraction method is introduced to improve sampling efficiency.