GNN-RL: A Graph Neural Network-Based Reinforcement Learning Algorithm for 3D Path Planning of AUVs

This is the authors' abstract. We don't add key points for this paper.

To enhance the intelligence of autonomous trajectory planning for Autonomous underwater vehicles (AUVs) in complex Ocean environments, we propose a reinforcement learning algorithm for 3D path planning of AUVs based on graph neural networks. Incorporating with Digital Bathymetric Elevation Model (DBEM) and factors such as sea flow, mines, and sonar as threat information, we simulate the planning scenarios for AUVs in military applications. Based on the Actor-Critic framework of reinforcement learning, this paper designs an online learning process using a Monte Carlo tree search-based policy value network. We employ dynamic directed graphs to describe the real-time state information of the AUV and design sparse and dense rewards. Experimental results show that the algorithm (GNN-RL) designed in this paper has good convergence and generalization. Compared with traditional algorithms, GNN-RL can output planning results more quickly while ensuring the safety of the AUV, meeting the needs of real-time route planning for AUVs.