Research On Method Of Deep Reinforcement Learning For Robot Motion Planning

Robot motion planning plays an important role in industrial manufacturing and people's lives,in the field of automated factory and the driverless car are inseparable from the support of robotics and related technologies.In recent years,robot motion planning has become one of the hot topics in robotics,automatic control,and artificial intelligence.However,the robot motion planning is mostly based on the accurate environment model,which cannot deal with the motion planning problem under uncertain conditions.The main research of this thesis is robot motion planning problem under uncertain conditions,using deep reinforcement learning method to solve this problem.Introducing the knowledge of motion planning,analyzing the problems of traditional motion planning methods.This thesis introduces the theory of reinforcement learning and deep learning,analyzing the problems of reinforcement learning application in robot motion planning,and combining deep learning with reinforcement learning to solve the high dimension problem of traditional reinforcement learning.This thesis analyzes the problems in the application of deep reinforcement learning in robot motion planning,namely reward function model,exploration strategy and neural network structure,then gives corresponding solutions.The main contents of the thesis are as follows:1.This thesis studies the reward function model in reinforcement learning.Analyzing the effects of the reward function model and the problem of existing,for more rapid training system,this thesis uses a reward function model based on system convergence speed,then does contrast experiments between different reward function models.2.This thesis studies the balance between exploitation and exploration in reinforcement learning.This thesis analysis problem of ?-greedy strategy.To ensure the action of small value will be executed during the exploration of the new environment,this thesis uses an N?-greedy strategy to deal with the balance between exploitation and exploration,then does contrast experiments between ?-greedy strategy and N?-greedy strategy.3.This thesis analyzes the effects of neural network structure during in the motion planning.For design a neural network structure suitable for current motion planning task,this thesis designs the neural network of different structures and carries out correlation experiments.4.This thesis implements a detailed algorithm based on deep reinforcement learning,develops a robot motion planning system based on deep reinforcement learning.This thesis carries out the motion planning experiment in the environment with obstacles and no obstacles,verifies the feasibility of the method to solve the problem of robot motion planning under uncertain conditions.