Robot arm is an important part of the robot to perform operational tasks,but different companies have a variety of robot arm structure, respectively, have different control equations, so there is no unified control method. In the 1990s,the neural network was used in solving the inverse kinematics of the manipulator,and the pose as the input of the neural network and the joint angle as the output of the network. The object of the study from the two joints, the three joints to the present mainstream six joint manipulator. The input and output of the neural network are also changing. The input can be the target pose, the current pose and the difference between the target pose and the current pose or the image of the target object acquired by the camera. The output can be the joint angle,the difference between the target joint angle and the current joint angle or the torque of the motor. But the obvious disadvantage is that the neural network can only output one solution, and there are multiple solutions to reach the target position or grasp objects, and the effect of use neural networks to simulate this non-functional relationship is greatly influenced by the training set.Sergey Levine et al use the camera’s image and motor commands as input to the neural network, and the grasp success probability as the output of the neural network. For a certain input, the output is deterministic and only. This input and output play the advantage of the neural network simulate function relationship. But when it find solutions, it randomly generate a batch of motor command input neural network to obtain the grasp success probability, and the use the CEM algorithm to find the motor command which have the highest success probability. This Method don’t play the advantage of the back propagation of the neural network to find the derivative of the motor command and optimize the motor command same as optimize the weight of the neural network. What’s more, this method need to pre-train a lot of training data, do not have hot start characteristics.In this paper, an iterative learning algorithm with memory unit is proposed.The algorithm is driven by the target pose, and the derivative of the square sum of the difference between the target position and the current position to the current joint angle is find through the back propagation of the neural network and through gradient descent with linear search algorithm to find the optimal joint angle, executing the motor command to find the joint angle and get the real position. If meet the error requirements then to the end and if not meet the requirements, the just practical data will be added to the memory unit and training the neural network, continue to find the joint angle. Through these two seamlessly connected and circular optimization process which contain optimizing the joint angle and optimizing the weight of neural network to achieve the purpose of iterative learning and hot start. Through using a limited memory unit, the neural network can use less data to achieve convergence, and do not have to remember a large number of training data. For the task which has different error requirements, using the methods proposed in this paper training neural network has greater adaptability.For the can’t reachable positions that were judged by the neural network, in this paper, a tentative learning strategy algorithm is proposed, which can 100%achieve any target position with arbitrary precision without limiting the number of attempts.The proposed algorithm is a reverse application of the function relationship simulated by the neural network. Through constructing a benign circular ecology,the neural network can learning by itself. The method has universal applicability. |