Improved DQN-based Robot Path Planning Algorithm for Mobile Robots

To resolve the challenge of sluggish convergence and suboptimal path quality in the deep Q-learning algorithm for path planning, this study investigates a robot path planning algorithm based on the Deep Q-Network (DQN) algorithm. Initially, The conventional artificial potential field approach is improved through eliminating the repulsive force and incorporating the map size within the gravitational field. A gravitational threshold is introduced to balance the effects of obstacle avoidance and goal attraction. To circumvent local minima, the improved APF is integrated with a stochastic exploration strategy, thereby accelerating convergence. Subsequently, the reward mechanism is refined by incorporating a distance-based reward between the intelligent agent and the target point. Finally, redundant nodes are eliminated by assessing the spatial relationship between paths and obstacles, and path smoothing is achieved using Bezier curves to enhance path quality. Simulation experiments conducted in a 10×10 grid environment using MATLAB demonstrate that the enhanced DQN algorithm reduces path length by 26% and increases convergence speed by 64%.