Scholarlink

<p>For unmanned surface vehicles (USVs), how to find an effective, feasible path that substantially improves mission success rates and time efficiency in dynamic marine environments is a critical issue. To address the path planning problem for USVs using deep reinforcement learning (DRL) in dynamic ocean environments, an improved algorithm based on Deep Q-Networks (DQN) is proposed, which is called Fast Guided Deep Q-Network Algorithm (FG-DQN). This algorithm combines DQN with the artificial potential field (APF) method and uses the A* algorithm to initialize a guiding path in a global static environment and to provide prior knowledge for the USVs. Additionally, the configuration of the reward function using APF and the guiding path effectively reduces the frequency of random movements during the early exploration phase of the DQN algorithm, which accelerates convergence, improves the computational efficiency of path planning, and increases path safety. Finally, the performance of the presented algorithm is validated through experiments in a 2D environment. Compared with traditional reinforcement learning methods such as Q-learning and Sarsa, as well as the original DQN algorithm, FG-DQN is more effective for USV path planning.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Path Planning for Unmanned Surface Vehicles in Dynamic Environments Based on Artificial Potential Field and Global Guided Reinforcement Learning

  • 期刊论文

  • 发表日期：

    2025年5月16日

查看全文
  • Shanqiang Li,
  • Chaoxi Li

摘要

For unmanned surface vehicles (USVs), how to find an effective, feasible path that substantially improves mission success rates and time efficiency in dynamic marine environments is a critical issue. To address the path planning problem for USVs using deep reinforcement learning (DRL) in dynamic ocean environments, an improved algorithm based on Deep Q-Networks (DQN) is proposed, which is called Fast Guided Deep Q-Network Algorithm (FG-DQN). This algorithm combines DQN with the artificial potential field (APF) method and uses the A* algorithm to initialize a guiding path in a global static environment and to provide prior knowledge for the USVs. Additionally, the configuration of the reward function using APF and the guiding path effectively reduces the frequency of random movements during the early exploration phase of the DQN algorithm, which accelerates convergence, improves the computational efficiency of path planning, and increases path safety. Finally, the performance of the presented algorithm is validated through experiments in a 2D environment. Compared with traditional reinforcement learning methods such as Q-learning and Sarsa, as well as the original DQN algorithm, FG-DQN is more effective for USV path planning.