Research And Implementation Of Obstacle Avoidance Method For Mobile Robot Based On Deep Reinforcement Learning

With the development of intelligent technology,mobile robots have been widely used in service transportation,warehousing and transportation,security inspection and other work,and the scenes they face have become increasingly complex,which puts forward higher requirements for mobile robot obstacle avoidance technology.Traditional obstacle avoidance algorithms need to establish an accurate environment model,which limits the flexibility and adaptability of obstacle avoidance algorithms.Deep Reinforcement Learning(DRL)learns obstacle avoidance strategies by interacting with the environment,showing high adaptability to the environment,and has been successfully applied to simple environments,showing satisfactory performance and ease of use,but there are still problems of poor obstacle avoidance performance in crowded environments and difficulty in applying to long-distance tasks.In view of the above problems,this paper conducts the following research:(1)In order to improve the success rate of DRL obstacle avoidance in a crowded environment with multiple obstacles,a DRL obstacle avoidance method based on feature fusion is proposed.Action prediction using shallow features extracted from high-resolution lidar readings to improve obstacle avoidance accuracy.At the same time,to solve the learning difficulties caused by the lack of semantic information of shallow features,the deep features rich in semantic information and shallow features are fused together,and the attention mechanism is used in the fusion process to enable the network to learn important information according to the characteristics of the current environment.Features are enhanced.Experiments show that the feature fusion scheme increases the success rate of obstacle avoidance by 8.4% compared with existing algorithms on maps with crowded environments.(2)In order to solve the problem that the DRL obstacle avoidance algorithm is easy to fall into the local minimum because it cannot obtain global information in longdistance tasks,this paper combines the DRL obstacle avoidance algorithm with the global path planning and aims at the existing reinforcement learning probability The roadmap algorithm proposes a series of improvements.A candidate edge screening method that considers obstacles in the map is designed,which improves the abstraction of the road map to the environment map;reduces the distance between the waypoints in the process of obstacle avoidance execution through the transition of relay points;The original undirected graph structure is changed to a directed graph structure to solve the impact on the road map caused by the different success rates of mutual obstacle avoidance tasks between two points.Experiments show that compared with the algorithm before improvement,the success rate of long-distance tasks is increased by3.6%.(3)Combined with positioning,mapping and other algorithms,the above algorithms are implemented in software,and the implemented algorithms are deployed on the mobile robot platform.And to further improve the success rate of the task in the real environment,a behavior recovery method is designed,so that the mobile robot can recover the action through attitude adjustment and path re-planning when the action is limited.Finally,the algorithm is evaluated in a real indoor environment with a size of 42 m × 84 m,and the test results show that the algorithm can better deal with the structural scene obstacle avoidance problem,and fully illustrates the practicability of the method proposed in this paper.