Comparison Of Weakly Supervised Regression Of sEMG Signals With State-Of-the-Art Unsupervised Methods For Robot Hand Control: A Pilot Study

The development of control strategies for multi-articulated robotic hands remains a key challenge in robotics. A promising approach leverages machine learning applied to surface electromyography (sEMG) for detecting human grasping intentions and controlling robotic devices. Most current methods rely on supervised learning to regress sEMG signals into robotic hand commands. However, unsupervised approaches are increasingly favored as they bypass the need for labor-intensive or imprecise instant-by-instant labeling of sEMG signals. This study explores three sEMG regression methods: unsupervised Nonnegative Matrix Factorization (NMF) and autoencoder (AE), and a recently developed novel weakly supervised technique leveraging the soft-Dynamic Time Warping (soft-DTW) divergence. Unlike standard unsupervised methods, soft-DTW supports nonlinear fitting, crucial for capturing the complex dynamics of human grasping intentions. By smoothing the traditional DTW metric, soft-DTW measures similarity between sequences while handling temporal misalignments, making it particularly effective for weakly supervised regression tasks avoiding for the necessity of instant-by-instant labeling of sEMG data. Preliminary experiments indicate that the soft-DTW-based approach surpasses NMF and AE in regressing grasp closure profiles for power, ulnar, and tripodal motions, offering a significant step forward in intuitive robotic hand control.