EMGNet: An EMG Dataset for Neural Decoding

Surface electromyography (EMG) can be used to interact with and control robotic systems via intent recognition. However, most machine learning algorithms used to decode EMG signals have been trained on relatively small datasets with limited subjects, which can affect their widespread generalization across different users and activities. Motivated by these limitations, we developed EMGNet - a large-scale dataset to support research and development in EMG neural decoding, with an emphasis on human locomotion. EMGNet combines 7 open-source datasets with processed EMG signals for 132 healthy subjects from four different countries (152 GB total size). We include manually annotated EMG data for four muscles (i.e., tibialis anterior, medial gastrocnemius, rectus femoris, and biceps femoris) and six activities, including standing, level-ground walking, stair ascent and descent, and ramp ascent and descent. Each individual dataset included in EMGNet was modified to achieve a consistent data structure for ease of use and to establish a standardized pipeline for signal processing (e.g., filtering, normalization, and windowing) and for training machine learning algorithms. The mission of EMGNet is to provide a large-scale, open-source platform to support the development and comparison of next-generation EMG neural decoding algorithms for myoelectric control and robotics.