Study On Robot Following And Obstacle Avoidance Under Dynamic Environments

With the widespread application of unmanned driving technology in many scenarios,the obstacle avoidance of unmanned vehicles has attracted more and more attention.The route planning technology based on map with complete information has been developed maturely,but it cannot fully meet the actual needs.In the real environment,we cannot rely on a priori map to obtain complete information,only part of the static obstacle information in the map.When unmanned vehicles are in a dynamic environment with dynamic obstacles(such as pedestrians,other vehicles,etc.),higher requirements are placed on many practical tasks,such as the ability to follow driving and dynamic obstacle avoidance.This paper conducts a systematic research on the following driving and obstacle avoidance of unmanned vehicles in a dynamic environment.The main work and innovations are as follows:First,aiming at the problem of following driving in a dynamic environment,the speed obstacle method is combined with the DWA algorithm to realize the collision avoidance of dynamic and static obstacles.At the same time,effective following driving is realized by comprehensively optimizing the objective function.Aiming at the existing speed obstacle method,this paper changes the way to handle obstacles,using rectangles to describe dynamic and static obstacles more generally,so as to get closer to the actual scene;this paper optimizes the control of distance between the vehicle and the target,combining the distance between the vehicle and the target and the speed of the target to determine the appropriate follow-up target point,so the trajectory of the vehicle and the reasonable follow-up distance are optimized;the DWA algorithm is introduced to evaluate the speed of the vehicle within a reasonable range,so as to choose the control command that has the best effect in avoiding obstacles and following the car.At the same time,this paper has completed the integration of the obstacle avoidance and car following algorithm on the ROS platform,and conducted a series of experiments in actual scenarios,and achieved good results.Second,in view of the problems of slow convergence speed and unreasonable planning trajectory in the obstacle avoidance of deep reinforcement learning in a dynamic environment,combined with traditional path planning algorithms,a deep reinforcement learning obstacle avoidance navigation method integrating global guidance training is proposed.Combining some known information in the environment,as well as the initial position of the vehicle and target position,the traditional algorithm is used to plan an initial path composed of a series of waypoints as a global guide for deep reinforcement learning;according to the current position of the vehicle,choose appropriate waypoint in the initial path as the current temporary target of the vehicle,and the vehicle will move towards the temporary target.During the movement of the vehicle,at every moment,the vehicle will respond to the vehicle’s motion information and the relative position of the waypoint.The movement of the vehicle is evaluated accordingly as a training sample for deep reinforcement learning;when a vehicle collides with an obstacle,the robot will be given different punishment according to the different classification of the obstacle and the relative position relationship between the vehicle and the obstacle.This method can effectively improve the convergence speed of deep reinforcement learning,and improve the navigation efficiency of the vehicle in the navigation process,and make the movement of the vehicle environmentally friendly,reducing the impact on the surrounding environment.