A neural network approach to the robot inverse calibration problem

In this paper, methods for robot inverse calibration are described. The position and orientation in space of the end effector (pose) errors of a six degrees of freedom (DOF) Puma robot are measured, using a precision co-ordinate measuring machine (CMM), at discrete locations distributed in the calibration volume. The corresponding joint corrections are obtained to compensate the pose error using a nonlinear optimisation method. This nonlinear optimisation method is more accurate and robust than the, widely-used, iterative Newton-Raphson method. However, the computation necessary for nonlinear optimisation is prohibitively expensive for online joint compensation. Therefore, an artificial neural network (NN) based approach has been developed to achieve constant-time solutions. A simple feedforward network architecture with a higher-order approximation capability is designed to ensure efficient and accurate network learning, where the training patterns are based on the results of the nonlinear optimisation method. Both simulation and experimental results are presented to show the effectiveness of the approach