Scholarlink

<p>In logistics factories, Automated Guided Vehicles (AGVs) are often tasked with transporting materials between different locations. In addition to achieving optimal global path planning, they must also navigate path planning in complex environments with dynamic obstacles. To address the poor generalization performance of current single-agent path planning algorithms in complex scenarios with dynamic obstacles, this paper proposes an AGV global path planning model based on Deep Q-Network (DQN) and a distributed training framework. Firstly, the convergence efficiency of the algorithm is improved by optimizing the environmental state settings, refining the reward function, and enhancing the experience replay extraction mechanism. Secondly, a state-dynamic network model inspirated by ResNet, called DQNRN, is proposed to further enhance the convergence speed and of the model. Finally, a distributed training framework UpgradedDQN_Multi is introduced to improve AGV decision-making in complex environments. Through simulations in environments of varying scales and comparative experiments with D_Star Lite, the original DQN, DDQN, DuelingDQN, and an improved Rapidly-Exploring Random Tree (RRT*) algorithm, the method’s path planning capabilities in complex scenarios with dynamic obstacles are validated. UpgradedDQN_Multi achieves an average reduction of 13.963% in search steps relative to the RRT* algorithm, indicating its ability to explore the solution space more effectively. Furthermore, on some larger-scale maps, the UpgradedDQN_Multi algorithm outperforms the D*star algorithm, achieving a lower average number of search steps.</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

AGV Path Planning with Dynamic Obstacles Based on Deep Q-Network and Distributed Training

  • 期刊论文

  • 发表日期：

    2025年2月25日

查看全文
  • Tingbo Xie,
  • Xifan Yao,
  • Zhenhong Jiang,
  • Junting Meng

摘要

In logistics factories, Automated Guided Vehicles (AGVs) are often tasked with transporting materials between different locations. In addition to achieving optimal global path planning, they must also navigate path planning in complex environments with dynamic obstacles. To address the poor generalization performance of current single-agent path planning algorithms in complex scenarios with dynamic obstacles, this paper proposes an AGV global path planning model based on Deep Q-Network (DQN) and a distributed training framework. Firstly, the convergence efficiency of the algorithm is improved by optimizing the environmental state settings, refining the reward function, and enhancing the experience replay extraction mechanism. Secondly, a state-dynamic network model inspirated by ResNet, called DQNRN, is proposed to further enhance the convergence speed and of the model. Finally, a distributed training framework UpgradedDQN_Multi is introduced to improve AGV decision-making in complex environments. Through simulations in environments of varying scales and comparative experiments with D_Star Lite, the original DQN, DDQN, DuelingDQN, and an improved Rapidly-Exploring Random Tree (RRT*) algorithm, the method’s path planning capabilities in complex scenarios with dynamic obstacles are validated. UpgradedDQN_Multi achieves an average reduction of 13.963% in search steps relative to the RRT* algorithm, indicating its ability to explore the solution space more effectively. Furthermore, on some larger-scale maps, the UpgradedDQN_Multi algorithm outperforms the D*star algorithm, achieving a lower average number of search steps.