A Mobile Robot Control Method In Complex Scenes Based On Deep Reinforcement Learning

MarketsandMarkets,a world-renowned market research organization,reports that the global mobile robot market is about $14 billion in 2018 and is expected to reach $50 billion by 2023.Among them,mobile service robots and logistics warehousing mobile robots are becoming more and more important.And intelligent mobile robots are getting more and more attention and favor.In the research and application of mobile robots,autonomous obstacle avoidance and navigation technology are the most basic and important issues.They are also the focus of mobile robot research,and have great significance in mobile robot theory research and engineering practice.Based on the special project of Youth Research and Innovation Program of Beijing University of Posts and Telecommunications,this paper deeply studies the mobile robot control problem in the complex environment with dynamic/static obstacles,and deeply studies the autonomous obstacle avoidance and navigation control technology of mobile robot based on deep reinforcement learning(DRL)and its improved method.It aims to improve the control accuracy of mobile robots.The main research contents are as follows:1.A dynamic obstacle avoidance and autonomous navigation strategy for mobile robots based on DRL is proposed.The state space representation strategy of mobile robot kinematics based on the Actor-Critic framework is designed.The representation method of the continuous motion space of mobile robot is constructed according to the distribution of dynamic/static obstacles around the mobile robot.Finally,the mobile is established.Robotic obstacle avoidance and navigation models.2.Aiming at the shortcomings of the generalization ability of the native Actor-Critic framework,a mobile robot obstacle avoidance and navigation model based on the improved Actor-Critic framework is proposed.Firstly,the long-short-time memory unit is introduced and added to the model-aware end,so that the mobile robot has a continuous sequence of memory functions,which improves its reasoning and cognitive ability to some extent.Secondly,the improved discrete motion space and layered training of the mobile robot decision-making end are designed to improve the network training speed and learning effect.Finally,the simulation experiment of the improved mobile robot obstacle avoidance and navigation strategy model has improved the performance.3.A mobile robot test platform consisting of mobile robots,solid-high controllers,inertial navigation sensors,laser radar and other related equipment was built.Aiming at the different scene requirements of AGV mobile robots,mobile robot obstacle avoidance and autonomous navigation experiments were carried out in static scenes and dynamic scenes.Experiments show that the obstacle avoidance and navigation strategies based on improved DRL can effectively solve mobile robots'autonomous obstacle avoidance and navigation task requirements for different scenarios.