Improved Kinematics Calibration of Industrial Robots by Neural Networks

The paper presents a preliminary study on the feasibility of a Neural Networks based methodology for the calibration of Industrial Manipulators to improve their accuracy. A Neural Network is used to predict the pose inaccuracy due to general sources of error in the robot (e.g. geometrical inaccuracy, load deflection, stiffness and backlash of the mechanical members, etc. . . ). The network is trained comparing the ideal model of the robot with measures of the actual poses reached by the robot. A back-propagation learning algorithm is applied. The Neural Network output can be used by the robot controller to compensate for the errors in the pose. The proposed calibration technique appears extremely simple. It does not need any information on the pose errors nature, but only the ideal robot kinematics and a set of experimental pose measures. Different schemes of calibration procedures are applied to a simulated SCARA robot and to a Stewart Platform and compared, in order to select the most suitable. Results of the simulations are presented and discussed.