Robot Arm Dynamics Control Based on Deep Learning and Physical Simulation

General robot arm dynamics is calculated by building robot arm physical model and calculating Newton-Euler formula. In the process of dynamic modeling, the accuracy of robot dynamic control will be affected by many factors, such as estimation of dynamic parameters, simplification of robot model, inaccurate parameters of friction model. In this paper, we proposal a method of calculating robot dynamics using deep learning networks and physical simulation. The deep learning network implement an end-to-end method to calculate robot dynamics. It means that the robot arm modeling process is not required. The robot arm physical simulation based on physical engine is carried out, and the deep learning network online training system is constructed by simultaneously communicating with the physical simulation through Socket communication. The experimental data shows that the maximum error of the joint torque calculation is 15.8% and 94.7% of the joint torque error is less than 10%. The calculation results of this method are approximately the same as those traditional dynamic method. However, the work presented in this paper is a first step towards combining robot dynamics and deep learning. It focuses on using a new method to solve the traditional problem.