Simulative and Experimental Evaluation of a Soft-DTW Neural Network for sEMG-Based Robotic Grasping

In this paper, we present a neural network architecture for minimally supervised regression of surface electromyographic (sEMG) signals into control commands to drive robotic grasping devices. The proposed architecture overcomes one of the limitations of state-of-the-art supervised regression approaches, which require an instant by instant labelling of the training dataset. This is achieved by deploying a differentiable version of the Dynamic Time Warping (DTW) similarity measure as loss function of a feed-forward neural network. The effectiveness of this approach was assessed both with simulation and experimental studies. We first used a model of the sEMG generation process to test the feasibility of the method. Then, we evaluated the proposed approach in a two-step experimental session involving a group of 10 subjects: an offline experiment was conducted to investigate neural network performance with desynchronized labelling, whereas an online experiment was carried out to control both a simulated and a real robotic hand. The obtained results demonstrate that the presented method allows minimally supervised regression of sEMG signals, reporting performances comparable with standard supervised approaches. In this relation, we show that the proposed soft-DTW neural network enables successful myocontrol of robotic hands even in presence of substantial temporal misalignments between sEMG trainset and related labelling, while supervised regression totally loses its capabilities. This means that the presented approach allows a greatly simplified training procedure that can pave the way to an innovative myocontrol framework characterized by highly simplified training procedures for the user without performance degradation.