Path Planning for Autonomous Vehicles in Park Environments Using Probabilistic Roadmap Method and Improved DQN Algorithm

The increasing complexity of autonomous driving in park environments necessitates advanced path planning methods to ensure safety, efficiency and comfort. This paper addresses this challenge by proposing a novel path planning method that combines the Probabilistic Roadmap Method (PRM) and Deep Reinforcement Learning (DRL). The study develops a comprehensive risk assessment model, integrating collision risk, path length cost and comfort cost into a multilevel risk cost framework to enable quantitative risk assessment in park environments. Additionally, the paper constructs connectivity, length, comfort and security risk assessment matrices based on the risk model to evaluate the risk costs among different sampling nodes. An improved algorithm, KMDQN, is introduced, leveraging the Kullback–Leibler Divergence (KLD) to enhance policy optimization and balance exploration–exploitation tradeoffs. The KMDQN algorithm also incorporates a KMEANS-based empirical replay mechanism to break sample similarity and temporal correlation, improving sample utilization. Experimental validation in a simulated park environment shows that the KMDQN algorithm effectively plans shorter, safer and more efficient paths compared to traditional A* and RRT algorithms, as well as the standard DQN algorithm. The KMDQN algorithm maintains stable training, avoids local optima and optimizes exploration strategies, demonstrating its potential to significantly improve the safety and efficiency of autonomous driving in complex park environments.