Online Unsupervised Arm Posture Adaptation for sEMG-based Gesture Recognition on a Parallel Ultra-Low-Power Microcontroller

Surface electromyography (sEMG) is a State-of-the-Art (SoA) data source for natural and dexterous control in human-machine interaction for industrial, commercial, and rehabilitation use cases. Despite non-invasiveness and versatility, a major challenge for sEMG-based control is the inherent presence of many signal variability factors, which hamper the generalization of automated learning models. In this work, we propose an unsupervised adaptation technique for sEMG classification and apply it to arm posture variability. The approach relies on aligning the Principal Components (PCs) of new data with the PCs of the training set. No classifier retraining is required, and the PCs are estimated online, consuming one sample at a time without storing any data. We validate our method on the UniBo-INAIL dataset, showing that it recovers 37% to 51% of the inter-posture accuracy drop. We deploy our solution on GAP9, a parallel ultra-low-power microcontroller, obtaining a latency within 3.57 ms and an energy consumption within 0.125 mJ per update step. These values satisfy the constraints for real-time operation on embedded devices. Our solution is unsupervised and thus suitable for real incremental learning conditions where ground truth is not available.