Calibrating a modular robotic joint using neural network approach

In this study a neural network has been proposed to calibrate a joint module of two rotary degrees of freedom. A feedforward neural network has been trained to predict errors in the joint angles using a fast backpropagation learning rule and then implemented in the control system to correct the errors. To improve calibration effectiveness, a calibration scheme using two neural networks has been suggested where the second network is trained by learning the residual errors of the first trained network. Satisfying accuracy of the neural network calibration has been verified by simulations. It has been found in this study case that the network using a sinusoid transfer function exhibited better converging performance