Learning Velocity Kinematics: Experimental Comparison of On-line Regression Algorithms

The increasing complexity of tasks addressed by humanoid robotics requires accurate mechanical models which are difficult to obtain in practice. One approach is to let the robot learn its own models. In [16], two algorithms were compared on learning the velocity kinematics model of iCub: XCSF and LWPR. This comparison was based on simulated data. In this paper, we propose to extend this study to data recorded from the iCub robot. We analyze the behavior of these algorithms in presence of large noise in real conditions of use. We also add the study of a third algorithm, iRFRLS. After a detailed study on the tuning of the three algorithms, we show that the results obtained in [16] are still valid on real data: XCSF converges more slowly, but to a lower error than LWPR. However, we show that iRFRLS outperforms these two algorithms.