Research On Path Planning Based On Deep Reinforcement Learning

In recent years,the use of robots has become increasingly prevalent in both production and daily life.Path planning is a core issue in robotics,but traditional path planning algorithms have limited efficiency,adaptability,and autonomy in complex environments.This study aims to improve the efficiency and autonomy of indoor robot path planning by using deep reinforcement learning.The study includes the following main contributions:(1)An ESKF filtering method is innovatively applied to the optimization process of the Gmapping algorithm backend to address map drift caused by a single sensor during robot mapping.This method fuses pose information obtained by IMU as prior data and the pose obtained by laser radar as posterior estimate for data fusion.By reinforcing the robot’s pose correction in the mapping process,the positioning accuracy and map resolution are improved.Results demonstrate that the Gmapping algorithm with ESKF data fusion effectively solves the map drift problem,and obtains a high-precision map.(2)To improve the autonomy of mobile robot path planning,the DDPG(Deep Deterministic Policy Gradient)algorithm is used.This algorithm allows for direct actions in a continuous action space,avoiding grid division and simplifying the calculation process.An artificial potential field is introduced in the reward function of the DDPG algorithm to consider the robot’s pose in the reward function,accelerating the convergence speed of the algorithm during training and improving the efficiency of robot path planning.(3)A robot simulation platform in the ROS2 system and Gazebo environment is built to verify the superiority of the DDPG algorithm combined with the artificial potential field.Experimental data comparison shows that the DDPG algorithm with the artificial potential field has more advantages than the DQN algorithm during training.Moreover,comparisons of planning time,path smoothness,and planning path length show that the DDPG algorithm with the artificial potential field is not inferior to the DQN algorithm and the artificial potential field method in all aspects,and its advantage is more obvious in long-distance path planning.