Session: 07-10-02: Mobile Robots and Unmanned Ground Vehicles

Paper Number: 113113

113113 - A Decentralized Multi-Agent Path Planning Approach Based on Imitation Learning and Global Static Feature Extraction 

In the era of Industry 4.0, mobile robots are widely used in smart material transportation, assembly and patrolling, and efficient path planning methods are critical to ensure these robots work properly. Traditional multi-agent path planning usually adopts a centralized approach, in which a central planning unit generates paths for all agents based on global information. Recent studies explore to develop decentralized methods in which each agent makes its own movement decisions based on local information. Compared to centralized methods, decentralized methods can reduce the computational burden of the central planning unit and show greater scalability as the trained models for small-scale maps can be smoothly transferred to the environments of large-scale maps. However, previous decentralized methods often suffer from limited successful deployment as they cannot fully exploit the information from global static environment. Moreover, these methods heavily rely on offline learning, which requires large amounts of training data (usually more than tens of thousands of cases), thus their computational efficiency is instable in complex environments.

 

To bridge these gaps, we propose a decentralized multi-agent path planning approach based on imitation learning with global static feature extraction. The overall structure of this approach mainly includes three layers: information extraction, information aggregation and action output. The information extraction layer of this approach uses a convolutional neural network to extract features from local field-of-view observations and expert paths planned under a global static map. Then, the information aggregation layer uses a graph attention network to aggregate feature information from other agents within a certain range. Finally, the action output layer converts the feature information into movement instructions for agents. In addition, we develop a fixing mechanism that can provide guidance for agents to escape from local traps by utilizing the expert paths planned under the global static map. In the training of the proposed approach, a supervised contrastive learning method is leveraged to reduce the use of large amount of expert training data without sacrificing the performance of the approach.

 

We examine the validity of the proposed approach in a simulated gridded environment with different map sizes, obstacle densities and agent numbers. Experimental results show that our approach outperforms other decentralized path planning methods in success rate and generalizability. Moreover, our approach is computationally efficient and scalable, making it a feasible solution for real-world applications. Our study contributes to the development of practical decentralized multi-agent path planning methods for mobile robots in smart manufacturing.

Presenting Author: Bohan Feng Shanghai Jiao Tong University

Presenting Author Biography: Bohan Feng is currently studying at University of Michigan-Shanghai Jiao Tong University Joint Institute, Shanghai Jiao Tong University for a PhD degree in Control Science and Engineering. His research interests include multi-agent system and machine learning for intelligent manufacturing.

Authors:

Bohan Feng Shanghai Jiao Tong University
Youyi Bi Shanghai Jiao Tong University
Mian Li Shanghai Jiao Tong University
Liyong Lin Contemporary Amperex Technology Co., Limited