Robot Calibration combining Kinematic Model and Neural Network for enhanced Positioning and Orientation Accuracy

摘要

Traditionally, the calibration of robots is pursued either using model-based or model-free methods. Only a few attempts to combine both approaches were reported, particularly the combination of geometric calibration and artificial neural network (ANN). The latter was mostly used to compensate the positioning error, however. This paper introduces an ANN for compensation of residual positioning as well as orientation error. Moreover, the ANN compensation can be applied with or without prior geometric calibration. An automatic measurement procedure was developed and nearly 14 000 robot poses were measured using a laser tracker. Five-fold cross validation on the training data was applied to find the best parameters of the ANN. These tests indicate that better accuracy is achievable by combining geometric calibration and ANN. Applying this combination on the test data reduced the maximum/average position error to 6.28 %/4.26 % and the maximum/average orientation error to 7.41 %/3.34 % of the original values (obtained without calibration).