A decentralized path planning model based on deep reinforcement learning

• Published in: Computers & electrical engineering Vol. 117; p. 109276
• Main Authors: Guo, Dong, Ji, Shouwen, Yao, Yanke, Chen, Cheng
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01-07-2024
• Subjects: Automated guided vehicles; Decentralized control; Deep reinforcement learning; Dynamic couple method; Path planning; Deep reinforcement learning; Path planning; Dynamic couple method; Decentralized control; Automated guided vehicles
• ISSN: 0045-7906; 1879-0755
• DOI: 10.1016/j.compeleceng.2024.109276

The escalating usage of automated guided vehicles (AGVs) on a grand scale accentuates the scalability limitations inherent in traditional, centrally-controlled path planning algorithms. Addressing this challenge, we introduce a pioneering decentralized path planning model, DPPDRL, built upon the foundations of deep reinforcement learning. Drawing inspiration from human path planning and dynamic coupling techniques, DPPDRL imbues AGVs with the autonomy to independently dictate and dynamically adapt their paths in real time. Notably, DPPDRL champions the adoption of diverse policies, thereby facilitating efficient collision avoidance and exploration of potential routes. Simulation-based evidence substantiates critical findings, including remarkable improvements in training speed and scalability, heightened adaptability to environmental variations and agent number changes, and applicability across any number of AGVs, provided a reasonable task distribution. However, DPPDRL's performance becomes limited when tasks are densely concentrated. Addressing this issue through task distribution and hierarchical reinforcement learning is a promising direction for future research. These findings underscore the transformative potential of DPPDRL in managing extensive, dynamic AGV operations.