Neuro-accuracy compensator for industrial robots

Abstract

A nominal analytic kinematic model augmented by a neural network (NN) accuracy compensator has been used to determine accurately the relationship between robot world space co-ordinates and joint transducer readings. In contrast to model-based calibration approaches which have been used in an attempt to model and identify the specific error source, the NN-based calibration provides a generic model of robot accuracy which accounts for various errors, with the error source information being represented in the distributed network weight connections. A novel network architecture, based on Pi-sigma neural networks, has been designed so that it has sufficient approximations capability, which is equivalent to the higher-order polynomials, to approximate the relationship between the accuracy compensations (both in the world space and in joint space) and robot configurations, while maintaining an efficient network leaning ability. The authors' results for a full-pose calibration of a six DOF (degree of freedom) Puma 560 Robot have shown that the neural network approach can achieve better accuracy compared with the kinematic model-based calibration. The forward neuro-accuracy compensation (compensated in the world space) demonstrates a decrease in the average position and orientation error from 4.35 mm and 2.55 degree to 0.24 mm and 0.44 degree, in the range from 0.90 mm and 0.41 degree to 0.15 mm and 0.37 degree respectively in the calibrated area. In addition, error compensation is much more efficient than conventional numerical iterative compensation algorithms, suggesting that the neuro-accuracy compensator can be implemented on-line.< <ETX xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">&gt;</ETX>