Research On Robot Obstacle Avoidance Based On Improved Dueling Network

Robot obstacle avoidance is a branch of motion planning.Motion planning is to find a human-constrained trajectory for individuals in a given environment.The constraints depend on the tasks people expect to accomplish,such as making the robotic arms complete a grasping action in the simplest way,or let the robot find a collision-free path in an obstacle environment and so on.With the rapid development of industrial technology and the upsurge of research on intelligent products such as unmanned aerial vehicles and driverless cars in today's society,the problem of motion planning has gradually been paid attention by researchers and industry.At present,the mature methods of motion planning have achieved good results,but most of them need manual teaching or manual control.They have high cost,poor flexibility and are difficult to cope with changes.However,with the development of deep learning algorithm in recent years,the reinforcement learning algorithm which originally stayed at the theoretical stage has made considerable progress.Deep learning is a method of processing complex data by using multi-layer network.The deep reinforcement learning algorithm combined with deep learning has achieved excellent results in GO game and man-machine confrontation.Deep reinforcement learning describes the process of interaction between the individual and the environment,that is,the learning process from ignoring the environment to adapting to the environment and completing the corresponding tasks.In this paper,the existing deep reinforcement learning method is applied to the problem of obstacle avoidance of robots.Aiming at the shortcomings of the existing algorithm,an improved algorithm is proposed.The autonomous learning ability of the algorithm is used to train the obstacle avoidance characteristics of the simulated robot,and finally to try on the real robot.Firstly,this paper introduces the research background and current situation of motion planning and reinforcement learning algorithm,then introduces the application of traditional reinforcement learning algorithm in obstacle avoidance problem,then introduces the deep reinforcement learning algorithm and its improved algorithm combined with the neural network in deep learning.Finally,according to the advantages and disadvantages of the existing algorithm,a deep reinforcement learning algorithm based on improved dueling network is proposed and applied to real robots.The main contents of this paper are as follows:1.This paper studies the application of traditional reinforcement learning algorithm in obstacle avoidance.Firstly,the basic concepts of traditional reinforcement learning are introduced,then the basic principles of reinforcement learning are introduced.Finally,the performance of traditional reinforcement learning algorithm is observed through simulation experiments and its advantages and disadvantages are analyzed.2.This paper studies the obstacle avoidance method based on deep reinforcement learning.It combines neural network with reinforcement learning idea for the first time,and improves the ability of the reinforcement learning algorithm to adapt to complex environments through the excellent data processing ability of the neural network.This paper first introduces the basic concepts of neural networks and network optimization methods,then introduces the classical deep reinforcement learning algorithm DQN and its improved algorithms: Double DQN and Dueling DQN.Finally,through experiments,compare the performance differences between deep reinforcement learning algorithm and traditional reinforcement learning algorithm on obstacle avoidance of robot.3.An improved algorithm is proposed based on the existing deep reinforcement learning algorithm.Firstly,paper describe the advantages and disadvantages of the existing deep enhancement learning algorithms: there is no evaluation of the environmental state value and the target network is not trainable in DQN algorithm.Then introduces its own improved algorithm,which is based on the idea of dueling network and Actor-Critic.Finally,the trained network model is loaded into the real mobile robot,and the reasons for the difference between the simulation effect and the real effect are analyzed.