sEMG-based Regression of Hand Kinematics with Temporal Convolutional Networks on a Low-Power Edge Microcontroller

Human-Machine Interfaces based on gesture control are a very active field of research, aiming to enable natural interaction with objects. Nowadays, one of the most promising State-of-the-Art (SoA) methodology for robotic hand control relies on the surface electromyographic (sEMG) signal, a non-invasive approach that can provide accurate and intuitive control when coupled with decoding algorithms based on Deep Learning (DL). However, the vast majority of the approaches so far have focused on sEMG classification, producing control systems that limit gestures to a predefined set of positions. In contrast, sEMG regression is still a new field, providing a more natural and complete control method that returns the complete hand kinematics. This work proposes a regression framework based on TEMPONet, a SoA Temporal Convolutional Network (TCN) for sEMG decoding, which we further optimize for deployment. We test our approach on the NinaPro DB8 dataset, targeting the estimation of 5 continuous degrees of freedom for 12 subjects (10 able-bodied and 2 trans-radial amputees) performing a set of 9 contralateral movements. Our model achieves a Mean Absolute Error of 6:89 degrees, which is 0:15 degrees better than the SoA. Our TCN reaches this accuracy with a memory footprint of only 70:9kB, thanks to int8 quantization. This is remarkable since high-accuracy SoA neural networks for sEMG can reach sizes up to tens of MB, if deployment-oriented reductions like quantization or pruning are not applied. We deploy our model on the GAP8 edge microcontroller, obtaining 4:76ms execution latency and an energy cost per inference of 0:243mJ, showing that our solution is suitable for implementation on resource-constrained devices for real-time control.