On Hierarchical Path Planning Based on Deep Reinforcement Learning in Off- Road Environments

The uneven terrains and complex obstacles in off-road environments present significant challenges to the stability and safety of vehicle path planning. In this paper, an improved D* Lite algorithm is employed to generate global waypoints for path planning. The dynamic window approach (DWA) is used for local path planning guided by these waypoints. The selection of DWA parameters is formulated as a Markov decision process (MDP), and a deep reinforcement learning (DRL) algorithm with a novel reward function that takes elevation maps, vehicle pose, goal information, and obstacle information as inputs is proposed to achieve the optimization of the parameters for off-road path planning scenarios. To validate the effectiveness of our algorithm, we conduct experiments in 3D simulation environments with different terrains and obstacles. Simulation results demonstrate that the proposed algorithm achieves a higher success rate and lower average elevation gradient compared to prior path planning methods in uneven terrains.