Incremental Learning of Muscle Synergies: From Calibration to Interaction

Abstract

In the previous chapter it has been shown how sEMG gathered from only two loci of muscular activity with opposite mechanical actions can be used to control the synergy-inspired robotic hand described in Chap. 8. Here, the problem of simpli-fying the control of a multi-DoF, multi-DoA mechatronic system—more specifically a prosthetic hand—is tackled from the opposite perspective, i.e. by leveraging the information contained in the sEMG gathered from multiple sources of activity. Nat-ural, reliable and precise control of a dexterous hand prosthesis is a key ingredient to the restoration of a missing hand's functions, to the best extent allowed for by the current technology. However, this kind of control, based upon machine learning applied to synergistic muscle activation patterns, is still not reliable enough to be used in the clinics. In this chapter we propose to use incremental machine learn-ing to improve the stability and reliability of natural prosthetic control. Incremental learning enforces a true, endless adaptation of the prosthesis to the subject, the envi-ronment, the objects to be manipulated; and it allows for the adaptation of the subject to the prosthesis in the course of time, leading to the exploitation of reciprocal learn-ing. If proven successful in the large, this idea will prepare the shift from prostheses, which need to be calibrated, to prostheses that interact with human beings.