Research On Collision Avoidance And Navigation Strategy Of Mobile Robot Based On Deep Reinforcement Learning

Mobile robots play a wide range of roles in many areas of modern society,and autonomous collision avoidance and navigation technology is the primary key problem for mobile robots to study and is the basis and prerequisite for other functions of mobile robots.The classical approach to autonomous collision avoidance navigation requires the creation of a map,followed by path planning and trajectory tracking,combining technologies from multiple disciplines to accomplish the entire task of collision avoidance navigation.The premise of this approach is to build a complete a priori map for the robot,and when the scene changes,or in unfamiliar environments without a priori maps,traditional methods will not be able to complete the navigation task;in addition,the direct integration of multiple disciplinary techniques will continue to amplify the error in each link,which will affect the navigation effect.The deep reinforcement learning-based autonomous navigation technology does not rely on a priori maps and directly outputs the line and angular velocities of the robot through sensors to obtain environmental information,which is an instantaneous end-to-end navigation strategy that makes it play a unique role in the field of autonomous collision avoidance for mobile robots.Based on deep reinforcement learning algorithm,this paper focuses on the research of autonomous collision avoidance navigation strategy for mobile robots,and completes the algorithm design of end-to-end collision avoidance navigation task with no a priori map,relying only on sensor information.The key issues in the navigation task,including autonomous collision avoidance,path planning,and speed control,are studied and optimized to ensure and improve the safety and efficiency of the robot’s collision avoidance navigation.In addition,based on the simulation experiments,the above algorithms are experimentally verified in a real environment,which proves the feasibility and effectiveness of the algorithms.The main research contents and innovative results of this paper are as follows:(1)Based on the basic framework of classical deep reinforcement learning algorithms,including DDPG,PPO,and SAC,we design learning networks,reward functions,and complete the autonomous collision avoidance navigation task of mobile robots in unfamiliar environments by training in a simulation environment.(2)To address the problems of long training times for deep reinforcement learning algorithms and the lack of robot speed continuity and fragmentation of navigation trajectories during end-to-end navigation,While decomposing the navigation task and optimizing the network framework,a spline interpolation module and a local path planning module are introduced into the whole navigation framework to reasonably assign subtasks.Through this method,the mobile robot can plan the optimal route during autonomous collision avoidance navigation,reduce the navigation time,and improve the navigation efficiency while optimizing the smoothness and continuity of speed to achieve a better navigation effect.(3)The migration and experimental validation of the algorithm in a real environment are completed.The algorithm is designed to avoid collision and escape in the real environment,and the retraining system is designed based on the simple sparse map to further optimize the navigation effect in the specific environment.A Scout-mini mobile robot platform was built,and the algorithm was successfully transplanted to the robot to complete the navigation task in the office corridor environment and the lobby environment of the science and technology building to verify the feasibility and effectiveness of the algorithm.