End-to-End Servo Control Of Robot With Deep Convolution Neural Network And Evolution Strategy

The traditional robotic grasping method is robust to those repetitive and fixed tasks,but it is difficult to apply to intelligent control scenarios that require robot to have self-learning capabilities.The rapid development of deep learning and deep reinforcement learning technology provides a new idea for intelligent control of robot.In recent years,domestic and foreign scholars have done a lot of research on the robot intelligent control technology of learning-based.However,end-to-end robot servo control of learning-based and visual feedback is still a challenging task.Based on the deep convolution neural network and covariance matrix adaptation evolutionary strategies,end-to-end servo control of robotic grasping under monocular vision is realized in this papper.Under the control of the servo mechanism,the robot can continuously adjust the posture through the feedback of the monocular image,and finally reach a better grasping position.Firstly,the mapping problem in the control task is analyzed.Based on residual unit,a deep convolution neural network is designed to map monocular images,motor commands to the probability of grasp success.The time complexity and space complexity of prediction model are also analyzed.The trained prediction model can provide heuristic feedback signal for robot posture adjustment.Secondly,in order to train the prediction model,A robot experiment platform based on vision is built and experiment scheme of data acquisition is designed,and over 40000 grasp attemps were collected in more than three months,result in 180000 training samples.By using the weighted cross-entropy loss and multiple performance evaluation indicators,the training problem of prediction model in the case of serious imbalance between positive and negative samples is solved.At the same time,the trained network is visualized based on GBP algorithm.Thirdly,the process of searching optimal motor command is modeled as a black box optimization problem with constraints,and then a search algorithm for optimal motor command is proposed based on CMA-ES and the trained network.By introducing a decision rule,the decision problem of whether the robot performs the optimal motor command to adjust its posture or grasp in current position is solved,and finally the continuous servo control of robotic grasping is realized.Finally,a robot grasping system is established based on the proposed algorithm,and an experimental evaluation scheme is designed.The experimental results show that the system can achieve more than 70% success rate for the objects in training set.