A GNN-based Decentralized Path Planning for Agricultural Robot Team: Work in Progress

• Published in: International Conference on Control, Automation and Systems (Online) pp. 1480 - 1483
• Main Authors: Jo, Yuseung, Son, Hyoung Il
• Format: Conference Proceeding
• Language: English
• Published: ICROS 29.10.2024
• Subjects: Adaptation models; Agricultural Robot; Agricultural robots; Computational modeling; Decentralized; Graph Neural Network; Graph neural networks; Heuristic algorithms; Multi-UGV; Path planning; Prediction algorithms; Robot sensing systems; Scalability; Testing
• ISSN: 2642-3901
• DOI: 10.23919/ICCAS63016.2024.10773055

In a wide, dense, and unstructured agricultural environment, the deployment of autonomous mobile robots is an attractive option. In such the environment, large robot systems are subject to physical limitations such as communication distance and sensor measurements. This limitation is solved effectively with distributed path planning and coordination. A graph neural networks (GNNs) are an effective approach for efficient communication of multi-robot systems. In this paper, we propose a GNN-based decentralized path planning framework for agricultural robot team. The proposed model used a graph neural network for responsiveness to dynamic environmental changes, scalability, and efficient local information exchange among the adjacent agents. A graph neural network takes as input the observable features (e.g., states, subgoal, obstacle) of each agent for a partial observation scenario. As the action policy to predict the behavior of the agents, the model trained the tradition optimal multi-agent pathfinding algorithm, conflict-based search algorithm. Through the simulation-based validation, the model was confirmed to have performance comparable to existing expert algorithms, responsiveness to dynamic environments, and scalability.