A path planning strategy unified with a COLREGS collision avoidance function based on deep reinforcement learning and artificial potential field

• Published in: Applied ocean research 2021-08, Vol.113, p.102759, Article 102759
• Main Authors: Li, Lingyu, Wu, Defeng, Huang, Youqiang, Yuan, Zhi-Ming
• Format: Article
• Language: English
• Subjects: Algorithms; Artificial potential field; Autopilots; Avoidance behaviour; Collision avoidance; Collisions; COLREGS collision avoidance; Deep learning; Deep reinforcement learning; International regulations; Navigation; Path planning; Ships
• ISSN: 0141-1187; 1879-1549
• DOI: 10.1016/j.apor.2021.102759

•A DRL method is designed to handle COLREGS collision avoidance path planning, which can ensure that each action of the USV is the optimal solution in the current state.•Simulated real-time sensor information is chosen as the input data of the DQN, which is used to simulate the practical navigation of the USVs.•The APF algorithm is utilized to improve the action space and reward function of the DQN to solve the sparse reward conundrum. Improving the autopilot capability of ships is particularly important to ensure the safety of maritime navigation.The unmanned surface vessel (USV) with autopilot capability is a development trend of the ship of the future. The objective of this paper is to investigate the path planning problem of USVs in uncertain environments, and a path planning strategy unified with a collision avoidance function based on deep reinforcement learning (DRL) is proposed. A Deep Q-learning network (DQN) is used to continuously interact with the visually simulated environment to obtain experience data, so that the agent learns the best action strategies in the visual simulated environment. To solve the collision avoidance problems that may occur during USV navigation, the location of the obstacle ship is divided into four collision avoidance zones according to the International Regulations for Preventing Collisions at Sea (COLREGS). To obtain an improved DRL algorithm, the artificial potential field (APF) algorithm is utilized to improve the action space and reward function of the DQN algorithm. A simulation experiments is utilized to test the effects of our method in various situations. It is also shown that the enhanced DRL can effectively realize autonomous collision avoidance path planning.