Reinforcement learning-driven heuristic path planning method for automated special vehicles in unstructured environment

Abstract

Aiming at improving the adaptability of global path planning method for the Automated Special Vehicles (ASVs) in a variety of unstructured environments, a reinforcement learning (RL)-driven heuristic path planning method is proposed. The introduction of traditional heuristic algorithm avoids inefficiency of RL in the early learning phase, and it provides a preliminary planning path to be adjusted by RL. Furthermore, a reward function is designed based on vehicle dynamics to generate a smooth, stable, and efficient path. The simulation environments are established based on real terrain data. The algorithm's performance is evaluated by testing various starting and ending points across different terrains. This paper also examines how obstacle distributions and ground conditions affect ASV path planning. Results demonstrate that the proposed method generates collision-free, efficient paths while maintaining excellent adaptability to diverse complex terrains.