Path Planning of UGV via Deep Reinforcement Learning with Observation Enhancement

Unmanned Ground Vehicles (UGVs) have emerged as a cost-effective unmanned vehicle and have found extensive applications in various fields, and traditional UGV path planning algorithms rely on a comprehensive understanding of the environment and accurate UGV modeling, leading to limited generalization capabilities. In this paper, we propose a UGV path planning algorithm which based on reinforcement learning method to enable autonomous navigation in partially observation environment. We first introduce the Duel Deep Q-Network algorithm to model UGV path planning tasks, including designing the space of action and state, and reward function. Then, to address the challenge of inefficient UGV path planning in complex environments due to incomplete perception, we incorporate a recurrent prediction module into the algorithm. This module enhances the memory of temporal perception information of the UGV, effectively complementing the current state information and supporting efficient UGV path planning in complex environments. To demonstrate the efficacy of the presented algorithm, we design a typical UGV path planning scenario. Comparative analysis with baseline algorithms demonstrates that our method ensures higher path planning capability for UGV in complex environments.