A Decentralized Multi-Agent Path Planning Approach Based on Imitation Learning and Global Static Feature Extraction

Abstract

Abstract Multi-agent path planning is a crucial problem in numerous industrial robotic implementations, ranging from smart material transportation, mobile patrolling to automated warehousing. In this paper, we introduce a decentralized multi-agent path planning approach based on imitation learning and global static feature extraction. Our approach employs a convolutional neural network in the information extraction layer to obtain features from local field-of-view observations and expert-planned paths under a global static map. The information aggregation layer then uses a graph attention network to combine feature information from selected neighboring agents. The request-reply based selective communication is also applied in the information aggregation layer to identify appropriate neighboring agents to be included in the graph attention network. Finally, the action output layer translates the aggregated feature information into actions for each agent. Additionally, we develop a strategy switching mechanism that adaptively utilizes expert-planned paths under a global static map to support agents to escape from local traps. The effectiveness of our proposed approach is evaluated in simulated grid environments with varying map sizes, obstacle densities, and numbers of agents. Experimental results demonstrate that our approach outperforms other decentralized path planning methods in success rate and generalizability. Furthermore, our approach is computationally efficient and scalable, making it suitable for real-world applications.