Realization and Performance Characterization of a Myoelectric Control System for Robotic Hands Based on Kernel Ridge Regression

In the field of human-robot interaction, the research is still far to find a solution to a stable control for hand prosthesis. In particular, one of the most promising methodology is represented by the use of electromyographic signals(EMG) of the muscles as a interface between the human and the artificial limb. The EMG is already used to control robotic systems that present a little number of degrees of freedom (d.o.f.), but for more complex controls able to regulate 6 or more hand's degrees of freedom several problems persist. The nonstationarity of the EMG and the nonlinear relation related to the hand configuration, became the main problem that we have to manage. The reason is that the EMG signals change over time under the influence of various factors. One of the state of the art approaches to address this problem is the use of incremental Ridge Regression with Random Fourier Features (iRRRFF) for the myocontrol algorithm, the iRRRFF is a machine learning algorithm for nonlinear mapping that is also able to update the model with new data. This enables the possibility of a continuous adaptation to the changes in the signals. In this work we implement this control in a Matlab/Simulink environment and we create a standard procedure for its use. We use an acquisition scheme that permits to acquire 8 EMG signals of the muscles around the forearm to test the performance of this type of myocontrol and, especially, we focus on the influence of carrying out different training protocols. This was tested involving two type of subject: one expert user that already used a similar EMG system in the past and a naïve user.