Agent Environment Perception And Control Decision Based On Deep Reinforcement Learning

Collision avoidance,as an important function of mobile robots,has great applicationvalue in real scenes,such as automatic driving,warehousing and logistics.In recent years,with the development of deep reinforcement learning,more and more researchers try to use deep reinforcement learning to solve the problem of collision avoidance of mobile robots.Based on the deep reinforcement learning collision avoidance strategy,the decentralized control method is used to endow each mobile robot with independent self-help collision avoidance ability,which solves the problems of localization failure,large calculation and network delay caused by the previous centralized control collision avoidance strategy.However,the collision avoidance method based on deep reinforcement learning needs to first train the collision avoidance strategy in the simulation environment,and then migrate to the real environment.However,there is a huge gap between the simulation environment and the real environment,such as the texture and color of the object,which makes it difficult to transfer the trained collision avoidance strategy to the real environment.In order to reduce the gap between the simulation environment and the real environment and realize the effective migration of collision avoidance strategy,researchers focus on the laser perceptron.Because the laser perceptron can provide accurate and simple one-dimensional distance information,and completely ignore the difference between the simulation environment and the real environment in texture and color.But the expensive laser not only can not be used in large-scale in the real scene,but also has low robustness to some irregular objects.In order to solve the problem brought by laser,a novel pure vision collision avoidance method is proposed in this thesis.Through the depth estimation algorithm to obtain the distance information of the surrounding objects,and combined with the semantic information obtained by the deimage segmentation algorithm to encode the RGB image collected,a pseudo laser data is obtained.This pseudo laser data not only has the advantages of simple and accurate data of laser sensor,but also contains rich semantic information.Compared with the traditional laser data,the proposed pseudo laser data greatly improves the collision avoidance performance of mobile robot to irregular objects.At the same time,in order to realize the effective migration of collision avoidance algorithm from simulation to reality,a new data enhancement method is proposed,that is,by adding noise to the precise laser datacollected in the simulation environment,it can better simulate the pseudo laser data predicted in the real environment.Influenced by perceptron,the training process of robot in simulation environment is a Partially Observable Markov Decision Process.In order to make mobile robot better observe the surrounding environment,attention mechanism and Recurrent Neural Network mechanism are introduced to make mobile robot better grasp the surrounding scene.Experiments show that the method proposed in this paper can effectively avoid collisions for some irregular objects,such as tables and chairs.In addition,the collision avoidance strategy can be effectively migrated to the real environment through data augmentation to achieve efficient collision avoidance for mobile robots in real scenes.