The Application Research Of Robot Control Based On Deep Learning And Deep Reinforcement Learning

In recent years,with the development of artificial intelligence technology,intelligent robot control system is a hot field in robotic research.As the main branches of artificial intelligence technology,deep learning and deep reinforcement learning have achieved remarkable results in theory and application.Therefore,under the framework of deep learning and deep reinforcement learning algorithm,this paper studies the application of deep learning and deep reinforcement learning in robot control.The main works are as follows:(1)Aiming at the problem of robot positioning control,a robot positioning control strategy for unknown objects based on Convolutional Neural Network is proposed,which combines deep learning with robot positioning control.Firstly,the model of multi-objective recognition and detection network based on Convolutional Neural Network is used to identify and detect all objects on the robot operating platform,so as to obtain the category information of all unknown objects on the robot operating platform.Secondly,the user randomly selects the object to be positioned online according to the multi-objective recognition and detection results,and then in the whole process of robot positioning control,the robot positioning control system uses the network to detect the object selected by the user from many objects and calculate the current image feature,so as to realize the intelligent perception for the unknown object.Finally,according to the image feature error,the vision control law is designed by sliding mode control concept to drive the robot’s claw movement,thus completing the robot positioning control task for unknown object.Five sets of experiments of positioning control for unknown objects have been completed on MOTOMAN-SV3X industrial robot in complex natural scenes.The experimental results show that the proposed positioning control strategy can accomplish the robot positioning control task for unknown objects without obtaining the prior model information of the operating object,and has high positioning accuracy.(2)Aiming at the problem of robot grasping control,combining deep reinforcement learning with robot grasping control,a robot grasping control policy for the object based on HER-DDPG is studied.Based on this policy,a robot grasping control policy for the object based on LS-HERDDPG is proposed.Firstly,the Hindsight Experience Replay(HER)and Deep Deterministic Policy Gradient method are introduced into the robot grasping control policy,a robot grasping control policy for the object based on HER-DDPG is established.Secondly,the determination method of Learning Sample(LS)is introduced into the robot grasping control policy based on HER-DDPG,a robot grasping control policy for the object based on LS-HER-DDPG is established.Then,using the learning samples acquired by interactive exploration between the robot and environment,the robot grasping control policy for the object is continuously updated through the policy learning iterative process.Finally,according to the optimal robot grasping control policy,the robot grasping control task for the object is completed.The validity of the robot grasping control policy based on HER-DDPG and LS-HER-DDPG is verified by experiments on OpenAI Fetch simulation platform.The experimental results show that the learning curves of the robot grasping control policy based on HER-DDPG and LS-HER-DDPG increase gradually with the increase of epochs.At the same time,the robot grasping control policy based on LS-HER-DDPG has higher learning rate than the robot grasping control policy based on HER-DDPG,which shows that the robot grasping control policy based on LS-HER-DDPG improves the efficiency of policy learning on the basis of the robot grasping control policy based on HER-DDPG.