Evaluating Generalization Capability of Bio-inspired Models for a Myoelectric Control: A Pilot Study

The rapid growing interest in the field of wearable robots opens the challenge for the development of intuitive and natural control strategies for establishing an effective human-robot interaction. The myoelectric control could be a valid solution for achieving this goal, since it is a strategy based on decoding the human motor intentions from surface electromyographic signals (sEMG) and mapping them into the control output. In this work we propose a bio-inspired myocontrol approach able to generalize the hand force estimation in the central point of a rectangle shaped workspace, after being trained and interpolated using data acquired on the vertexes only. We compared performance of the proposed approach (featuring factorization and clustering techniques for building muscles patterns valid in the whole workspace) versus the ones obtained using the classical muscle synergies extraction strategy and without the use of muscle synergies. Obtained results show that all the three tested approaches were able to generalize. Moreover, with the proposed approach, we were able to closely estimate the muscle patterns in the testing point, suggesting the possibility of a better generalization capability (compared to other approaches) when increasing the size of the workspace.