Processing of sEMG signals for online motion of a single robot joint through GMM modelization

This paper evaluates the use of Gaussian Mixture Model (GMM) trained through Electromyography (EMG) signals to online estimate the bending angle of a single human joint. The parameters involved in the evaluation are the number of Gaussian components, the channel used for model, the feature extraction method, and the size of the training set. The feature extraction is performed through Wavelet Transform by investigating several kind of configuration. Two set of experimental data are collected to validate the proposed framework from 6 different healthy subjects. Trained GMMs are validated by comparing the joint angle estimated through Gaussian Mixture Regression (GMR) with the one measured on new unseen data. The goodness of the estimated date are evaluated by means of Normalized Mean Square Error (NMSE), while the time performances of the retrieval system are measured at each phase in order to analyze possible critical situations. Achieved results show that our framework is able to obtain high performances in both accuracy and computation time. The whole procedure is tested on a real humanoid robot by remapping the human motion to the robotic platform in order to verify the proper execution of the original movement.