Path Planning for Multi-AGV Systems Based on Globally Guided Reinforcement Learning Approach

The path planning for multi-AGV systems is particularly challenging in dynamic environments, where the AGVs must effectively reach their target locations while avoiding collisions with dynamic obstacles and other AGVs. Traditional approaches often rely on a re-planning strategy, triggering the planning algorithm whenever an AGV encounters a conflict. However, this method frequently results in unnecessary detours and deadlock issues. To overcome these challenges, this paper introduces a hierarchical reinforcement learning-based approach to multi-AG V path planning. This method establishes a globally guided reinforcement learning framework (G2-rMAPPO) by integrating global guided paths with the rMAPPO algorithm. Initially, a global path planner is employed to derive an optimal guided path for resolving static conflicts, while a rMAPPO-based local reinforcement learning planner is utilized to navigate dynamic conflicts. Additionally, a dense reward function is designed to effectively address the multi-AGV path planning problem. The feasibility and effectiveness of this approach are validated through experimental simulations.