AGV Path Planning Based On Deep Reinforcement Learning

The transformation from Made in China to Intelligent Manufacturing in China has led to the development of smart logistics systems and smart warehousing systems.Automated Guided Vehicles(AGVs),as an important conveyor belt for flexible manufacturing systems and smart warehousing systems,have assumed the task of transportation.In discrete manufacturing enterprises,the material as the main operation object is the center of overall operation and management,and the material logistics process is also a key link that affects the overall production efficiency.Automated and intelligent path planning can greatly improve the efficiency of logistics,thereby reducing enterprise production costs and improving enterprise production efficiency.Therefore,the research on AGV path planning technology is a necessary support for the intelligent transformation of discrete manufacturing flexible production lines and storage systems into modern industries.At present,the main navigation method of AGV is electromagnetic,visual and other navigation methods by arranging the route in advance.Therefore,the flexibility of AGV is greatly restricted.In order to improve the flexibility and adaptive ability of AGV navigation in complex and unknown environments,this paper mainly studies the AGV path planning method based on deep reinforcement learning(Deep Reinforcement Learning,DRL),which can realize the multi-modal sensor information of the AGV.End-to-end path planning for intelligent decision-making based on intelligent perception and action selection.The main research contents of this paper are as follows:(1)Describe the Markov Decision Process(MDP)of the AGV path planning problem and build an intelligent warehouse simulation environment.The MDP description of the AGV path planning problem is carried out from the state space,action space and reward function,and ROS and Gazebo are used to build an experimental platform for the intelligent storage simulation environment;(2)Designed an AGV path planning method based on Dueling DDQN-PER,including a neural network structure for multi-modal sensor information processing and an end-to-end path planning method process that implements environmental information to action mapping.This method introduces a global path planning pre-training strategy,Thereby improving the speed of the AGV’s path planning training in a complex environment;(3)This paper conducts three groups of AGV path planning experiments from simple to complex.The first group of experiments compares and analyzes the performance of the improved DQN method.The second group of experiments compares and analyzes the path planning performance based on the DRL method and RRT.The third group of experiments verifies the DRL-based path planning performance.The adaptive capability of the AGV path planning method in a complex environment.