Detaching phenomena in the learning control of manipulation of rigid objects

Abstract In this paper the problem of the dexterous manipulation of rigid objects, in presence of incipient sliding or detaching phenomena, and using hybrid learning control techniques, is discussed. Such phenomena have not been largely considered in the literature so far. The objective is to show that, even in presence of such events, that could violate the constraints specific of the manipulation task, the learning procedure is effective; in particular it is demonstrated that for sufficiently smooth object's motions and adequately planned (internal) contact forces, after a possibly large, but limited number of trials the learning procedure guarantees that no further conditions of sliding/detaching occur. From that trials on the classical learning algorithms assure the convergence to zero of task errors.