Research on Mobile Robot Path Planning Based on Improved Distributional Reinforcement Learning

Abstract

Path planning plays a crucial role in mobile robot navigation. An effective technique in path planning can enhance the operational efficiency of mobile robots. In this paper, we introduce an improved distributional reinforcement learning algorithm, PQR-DQN, for path planning problems in known environments. The algorithm combines the prioritized experience replay technique with the QR-DQN algorithm to improve the utilization of important samples. Additionally, it utilizes a residual network structure to reduce the estimation error in value distribution. To tackle the issue of sparse rewards in path planning tasks, we present a dynamic composite reward function. This function integrates heuristic distance, azimuth, and a turn penalty function, thereby offering abundant reward signals for the robot. During the training phase, PQR-DQN exhibits faster convergence speed and superior ability to learn effective strategies in comparison to alternative RL algorithms. Compared to traditional path planning algorithms, the PQR-DQN algorithm can effectively reduce the number of turning points, lower the angle of turns, and shorten the path length.