This project aims to build a deep learning pipeline for hand pose estimation from surface electromyography (sEMG) signals recorded by a smart wristband equipped with muscle activity sensors. We used the emg2pose dataset, which includes data from 193 users, 370 hours, 16-channel sEMG signals at 2 kHz (Salter, Warren, Schlager, et al. 2024). This publicly available dataset is found in the GitHub repository: https://github.com/facebookresearch/emg2pose.

We focused on the core deployment plan, generalization to new users/poses, sensor placements, and trajectory quality. We established a baseline LTSM model and added small, well-ablated improvement through an spatiotemporal learning approach. This project is packaged as a reproducible PyTorch pipeline that can be run in Google Colab. Additionally, we included deployment by publishing our trained model checkpoints and inference code to Hugging Face.