Approaches for Exploiting Neural Networks for Semi-supervised Myoelectric Control of Robot Hands

Human-In-The-Loop (HITL) control strategies using surface electromyography (sEMG) face challenges with labeling in supervised learning. Unsupervised regression methods for sEMG signals have limitations in controlling multiple grasp motions. This paper presents two semi-supervised regression approaches using neural networks (NN) for sEMG-based robot hand control. The first approach uses soft-DTW divergence as a loss function for minimally supervised NN training. The second combines Non-Negative Matrix Factorization (NMF) with self-supervised NN regression. Offline tests show the soft-DTW NN performs similarly to a standard MSE-based NN, and the self-supervised regression outperforms traditional unsupervised methods, enabling multiple grasp actions.