An End-to-End Path Planner Combining Potential Field Method With Deep Reinforcement Learning

This article presents an end-to-end path planning and motion control method based on deep reinforcement learning (DRL), aimed at enhancing the autonomous navigation capabilities of wheeled robots in mapless environments. By physically modeling the robot and optimizing LiDAR inputs, our method precisely processes the relative distances between obstacles and the robot, thereby improving obstacle avoidance precision and adaptability. The adopted end-to-end approach directly outputs control commands, simplifying the process from path planning to execution and reducing execution errors. Incorporating the concept of the potential field method, we propose the potential field-soft actor-critic (PF-SAC) reinforcement learning (RL) method and the network. Experimental results demonstrate that our method significantly outperforms the deep deterministic policy gradient (DDPG) method in terms of average rewards in dynamic environments, achieving a success rate of 95% after 500 training episodes. In varied environments and tasks, our method exhibits a 17.5% reduction in task completion time compared to the baseline soft actor-critic (SAC) method. Moreover, in real-world deployments, the robot effectively avoids obstacles and successfully reaches its destination, demonstrating the effectiveness and robustness of our method in practical applications.