sEMG-Based Minimally Supervised Regression Using Soft-DTW Neural Networks for Robot Hand Grasping Control

One of the major challenges in robotics consists in developing successful control strategies for robotic grasping devices. In this scenario, one of the most interesting approaches regards the exploitation of surface electromyography(sEMG). In this work, we propose a novel sEMG-based minimally supervised regression approach capable of performing nonlinear fitting without the necessity for point-by-point training data labelling. The proposed method exploits a differentiable version of the Dynamic Time Warping (DTW) similarity - referred to as soft-DTW divergence - as loss function for a flexible neural network architecture. This is a different paradigm with respect to state-of-the-art approaches in which sEMG-based control of robot hands is mainly realized using supervised or unsupervised machine learning based regression. An experimental session was carried out involving 10 healthy subjects in an offline experiment for systematic and statistical evaluations, and an online experiment for the evaluation of the control of a robot hand. The reported results demonstrate that the proposed soft-DTW neural network can be trained by means of a labelling that does not require to be temporally aligned with the sEMG training dataset, while reporting performances comparable with a standard mean square error(MSE)-based neural network. Also, the subjects were able to successfully control a robot hand for grasping motions and tasks with error levels comparable to state-of-the-art regression approaches.