Deep Reinforcement Learning Mobile Robot Navigation For Local Path Planning

In the process of robot intelligence,mobile robots are the core foundation of the intelligence process,a key step to realize robot intelligence,and a research hotspot in the development of robots in recent years.The key to using mobile robots in increasingly complex environments is that the autonomous navigation of mobile robots can avoid obstacles and crowds to reach designated safe locations.In this paper,aiming at the navigation obstacle avoidance of mobile robots,a deep reinforcement learning(DRL)method of local path planning is used for mobile robot navigation.The following are the main research work of this article:(1)DRL based on Soft Actor-Critic(SAC)algorithm solves local path planning problemsIn the navigation process of mobile robots,traditional path planning methods had problems,such as complex calculations,difficulty in learning to converge,poor generalization,and easy to fall into local optima.This paper proposes to use a more stable SAC algorithm to solve the local path planning problem.And in the path planning process,according to the data error problem caused by indoor environmental noises when the mobile robot uses sensors to collect environmental information,a Multi-layer Grid Gap(MGM)method is proposed to solve it.The experimental results show that the local path planning based of the SAC algorithm are more stable than the existing deep reinforcement learning algorithm,and the results of the algorithm based on the MGM have been greatly improved.(2)A local path planning method based on a Social SAC(SSAC)algorithm is proposedThe local path planning method based on DRL training tends to learn only to obtain reward learning mode in navigation.This paper proposes a SAC algorithm based on sociality,which uses a reward function to enable mobile robots to obtain social obstacle avoidance.And in training process,the simulation Actor with an improved social force model is used to solve the problem that the path planning strategy in the deep reinforcement learning training process is difficult to obtain the human-computer interaction data.The experimental results show that the SSAC algorithm proposed to this paper has been greatly improved on all aspects of local path planning.(3)Deploy the local path planning algorithm of this article in the physical robot navigation system for testingTo verify the migration ability and actual effect of the algorithm simulation environment to the physical environment,this paper designs a physical robot navigation system,and deploys the algorithm in this paper to the navigation system for testing.According to the needs of the test,two sets of real-world experiments are carried out in this article.The experimental results show that the local path planning model of deep reinforcement learning in this article can be well transferred to the physical robot,and the effect of obstacle avoidance is significant.