Research And Implementation Of Robotic Arm Grab Control System Based On Deep Reinforcement Learning And Embedded Technology

With the wider and wider application of robots in real life,the intelligent development of robots has also played a very important role in people’s lives.In the process of controlling the manipulator to grasp objects,the traditional method mostly teaches the manipulator to grasp objects at a fixed position.Such manipulator itself has no ability to perceive the external environment.Therefore,in the process of carrying out the grasping task,it is particularly vulnerable to the influence of external uncertainties,thus affecting the effect of grasping.If the position of the object changes,It is very likely that the manipulator can not accurately complete the grasping action.In the practical application of robots,facing the complex working environment,it is very important that robots can adapt to the changes of the environment at any time.However,the actual environmental state is dynamic and uncertain,which brings great obstacles to the practical application of robots.Deep Reinforcement Learning(DRL)can improve the adaptability of manipulators to unknown environments to a large extent because of its high tolerance to environment.Aiming at the problem of traditional manipulator grabbing control,this paper introduces a depth camera,uses the feature extraction ability of convolution neural network to pre-detect the target in the picture,takes the status picture with the location information of the pre-detection frame as the input of the grabbing control network,and uses the optimal strategy of Brahman in deep reinforcement learning to select the pre-detection frame,and then uses the strong depth to select the pre-detection frame.The Q-value network in the learning process calculates and processes each pixel in the selected pre-detection frame.The output of the network is set as the capture success rate of each pixel,and the maximum value is selected as the position coordinate of the manipulator.Through the design of the overall architecture of the algorithm,the training efficiency of the grasping control algorithm is improved,and the evaluation mechanism of the grasping results is designed to update and learn the whole grasping system in reverse.The realization of the control system algorithm is to configure TensorFlow environment under Ubuntu operating system,write the whole algorithm using Python language,and simulate a real manipulator grasping system using V-REP software based on ROS system.The algorithm is transplanted into the embedded system by combining hardware and software to verify its operability in the hardware,and an easy-to-operate interface is designed.Through the interface,the algorithm can be successfully used to control the manipulator grasping operation,which has the characteristics of intuitive and easy-to-operate.Finally,the grab control system proposed in this paper is compared with the traditional grab control system,which proves the rationality and effectiveness of the deep reinforcement learning control method.