Data outlier tests and elimination based on sensitivity analysis in robot periodic motion and its application to robot parameter identification

In this paper, eliminating data outliers, we identify the minimum set of robot dynamics parameters suitable for the control system. While the least-squares method tests for outliers based on the output error, it is desirable to use the parameter error from parameter identification point of view. Therefore, we use the stochastic parameter identification method which is previously proposed. Since this method shapes the auto-covariance of the parameter errors based on the sensitivity of the control system, the parameter errors can be considered to follow a multidimensional normal distribution represented by an ellipsoid. In order to test based on Mahalanobis distance, we derive the range of rejection and propose an eliminating method. In addition, though the conventional stochastic parameter identification requires prior data thinning to prevent handling extremely large orthogonal complementary space matrices and generalized inverse, it is necessary to use all data to perform appropriate tests. Therefore, we also derive a speed-up algorithm that takes advantage of the sparsity of matrices. By applying the proposed method to the weaving motion of the planar three-link manipulator which corresponds to a welding robot, and designing the control system, we verify that the designed controller reduces the position error of the end-effector through experiments.