Experimental comparison of parameter estimation methods in adaptive robot control

In the literature on adaptive robot control a large variety of parameter estimation methods have been proposed, ranging from tracking-error-driven gradient methods to combined tracking- and prediction-error-driven least-squares type adaptation methods. This paper presents experimental data from a comparative study between these adaptation methods, performed on a two-degrees-of-freedom robot manipulator. Our results show that the prediction error concept is sensitive to unavoidable model uncertainties. We also demonstrate empirically the fast convergence properties of least-squares adaptation relative to gradient approaches. However, in view of the noise sensitivity of the least-squares method, the marginal performance benefits, and the computational burden, we (cautiously) conclude that the tracking-error driven gradient method is preferred for parameter adaptation in robotic applications.