KU-MUSCLE Dataset

This dataset was created to study the biomechanics of hand movement to accelerate the development of robust machine learning (ML) algorithms that perform well in real-world, noisy, and variable conditions. Four types of hand movements were recorded from 28 subjects in comfortable, natural environments to introduce realistic variability. Participants performed each movement for as long as they preferred, adding subject-specific variability. The dataset can be applied to human-computer interaction systems or robotic prosthetic arm research.

EMG sensors collected signals around at 100 Hz, attached to participants’ hands. Participants were instructed to relax for 5 seconds before performing each movement. The participants performed four distinct exercises, which are
1. Hand grip movement,
2. Ball squeezing,
3. Thumb movement, and
4. Palm open and close.
We intentionally captured subject-specific variability in movement duration to reflect real-world differences in hand biomechanics.
At the root level, the dataset is organized into two main folders: the raw data folder and the processed data folder.

The raw data folder contains 28 subfolders, each corresponding to a different subject. Within each subject folder, there are four subfolders representing the four types of hand movements. Each movement subfolder contains JSON files named emg_recording_{timestamp}.json, which store the raw EMG values under the key "values" as well as the recording start and end times under "start_time" and "end_time".

The processed data folder contains preprocessed data in “processed_data.pickle” file and splitting information in “split_information.pickle” file. The preprocessing was performed in two stages. First, all signals were resampled to a fixed 100 Hz sampling rate, and second, the duration of each movement was fixed to 53 seconds. The split_information.pickle file contains various data split configurations, allowing researchers to use consistent splits so that their results are directly comparable with others.

The source code of the preprocessing can be found at this link https://github.com/mj-raihan/KU-MUSCLE.