Robot Navigation Algorithm Based On Reinforcement Learning In Unknown Environment

The research on mobile robot can not get around the navigation control all the time.At present,there are many excellent algorithms applied to robot navigation,and many achievements have been achieved,but most of these algorithms rely on environmental maps or are limited in theory.Therefore,the purpose of this paper is to design an end-to-end navigation model which does not depend on the environment map,and to improve the practical application ability of the robot.The main research contents are as follows:First of all,this paper summarizes the significance and background of the research,briefly introduces the principles and limitations of several traditional navigation algorithms,focuses on the principle and application status of reinforcement learning navigation,analyzes the problems of reinforcement learning in the field of navigation according to different algorithm schemes,and designs an end-to-end navigation model based on deep reinforcement learning by using traditional DQN algorithm.In the experiment,the effects of different state parameters and network models on the navigation effect are compared,which provides a model reference for the follow-up model improvement.Secondly,in order to solve the problem that it is difficult to design the decay rate of greedy-random exploration strategy in reinforcement learning,an adaptive exploration method based on old and new strategies is proposed,and the proportion of exploration is determined by using different degrees of action between new and old strategies.So as to make exploration and utilization more balanced.Thirdly,in order to solve the problem of slow convergence of reinforcement learning using discrete reward function,the reward function is optimized.According to the purpose of navigation algorithm,the reward function is designed as a continuous reward function which includes two parts: accomplishing the goal and avoiding obstacles.After performing each action,the corresponding different reward value is obtained.The improved reward function enables the robot to make better use of environmental information.Finally,in order to overcome the collision of the robot under the reinforcement learning navigation model,the mode of action selection is improved,and a collision-free end-to-end navigation model with fuzzy decision is designed.The improved algorithm model combines the efficient obstacle avoidance ability of fuzzy decision-making and the learning ability of reinforcement learning in unknown environment.