Robot Path Planning Method Based on Deep Reinforcement Learning

摘要

Aiming at the excessive dependence of traditional path planning methods on map information and the lack of self-learning and self-adaptive capabilities, a path planning method for robot based on deep reinforcement learning is proposed. The paper takes lidar data of Gazebo simulation environments built on the ROS platform as input. Learn direct action control from environment information through end-to-end learning, adopt neural network to fit value-based non-model time difference Q learning algorithm, reasonably design environment models, and the number of state spaces, the optimal decision strategy is learned by maximizing robot and dynamic environment interaction of cumulative reward. Simulation results show the method can meet the requirements of intelligent perception and decision-making by only relying on some map information.