Neural Network Control for Dynamics of a 3DOF Robot Arm

Abstract

This paper discusses the use of neural networks as an alternative to replace PID plus computed torque control in a 3DOF arm robot system. The neural network model used is multilayer perceptron with 3 inputs, 3 outputs and 30 hidden layers activated with the scaled gradient descent function. The training results show an R squared value of around 0.564, but the root mean square error of end effector external coordinate trajectories shows performance that does not exceed the computed torque control. The neural network model uses generated I/O data, and divide it for training, validation and testing 70%, 15% and 15% percent respectively. However, the graphically response of the neural network control system is better than that of PID plus computed torque control for the dynamic robot system. Therefore, further research is needed to improve the performance of this model by selecting appropriate architecture, hidden layer number and activating function.