boootaooo/forceband-emg2force

ForceBand EMG2Force dataset

Wrist sEMG + IMU paired with per-finger fingertip force and synchronized RGB video, recorded with the ForceBand wrist band — the pretraining data of the EMG2Force model in "ForceBand: Learning Forceful Manipulation with sEMG" (arXiv:2606.26093 · project page · code).

• 8.61 h of cleaned-up EMG + IMU + force at 1 kHz (8.39 h valid), 53 recordings, Oct 2025 – Apr 2026
• 6.37 h of it with RGB video (40 recordings; video only for recordings before March 2026)
• ~7.7 GB (0.55 GB time series + 7.2 GB video)

Files

segments/<segment_id>/timeseries.hdf5              EMG / IMU / force, masks, trial table (1 kHz)
segments/<segment_id>/rgb_<camera>.mp4              RGB video, one per camera (if any)
segments/<segment_id>/rgb_<camera>_timestamps.npy   capture time of each frame
segments/<segment_id>/meta.json                     durations, provenance, video info
manifest.csv · recording_log.csv · summary.json     per-segment index, recording times, totals

A segment is one continuous recording, including the rest periods between trials. segment_id is the local start time of the recording.

Format

timeseries.hdf5, group forceband/timeseries/:

dataset	shape	meaning
timestamps	(N,)	seconds from segment start (1 kHz)
emg	(N, 8)	8-channel sEMG, filtered
imu	(N, 10)	quaternion xyzw, gyro (rad/s), accel (m/s²)
force	(N, 5)	thumb … pinky force, sensor units (÷ 18 ≈ N)
valid / imu_valid	(N,)	real EMG + force / IMU data present
in_trial, task_index	(N,)	trial mask and task id (−1 between trials)

forceband/trials (K, 4) lists task_index, trial_in_task, start_s, end_s.

Video: H.264 MP4 at native resolution (848×480 camera; 640×480 cam_ego + cam_gt in Dec 2025). Frame i was captured at rgb_<camera>_timestamps.npy[i] on the same clock as timestamps — use these times to align, since the container's 30 fps is nominal.

Quick start

from huggingface_hub import snapshot_download
import h5py, numpy as np

root = snapshot_download("boootaooo/forceband-emg2force", repo_type="dataset")
# time series only: snapshot_download(..., allow_patterns=["*.hdf5", "*.json", "*.csv", "*.md"])
seg = f"{root}/segments/2025-11-04-15-13-41"
with h5py.File(f"{seg}/timeseries.hdf5") as f:
    g = f["forceband/timeseries"]
    t, emg, imu, force = (g[k][()] for k in ("timestamps", "emg", "imu", "force"))
frame_t = np.load(f"{seg}/rgb_camera_timestamps.npy")    # align frames with np.searchsorted(t, frame_t)

Notes

• EMG: OpenBCI timestamps evened out, then 58–62 Hz bandstop, 5–50 Hz bandpass and 50/60 Hz noise removal — the preprocessing used for training. All streams are resampled to 1 kHz.
• Some recordings had force-sensor drift; their force carries a baseline correction (force_tared).
• The EMG contains occasional large spikes (filter edge transients, electrode artifacts) not flagged by valid.
• Two recordings have no IMU (imu is NaN). Video shows the participants, including faces.

License

Apache-2.0, the same as the ForceBand code (see LICENSE).

Citation

@article{he2026forceband,
  title   = {ForceBand: Learning Forceful Manipulation with sEMG},
  author  = {He, Botao and Wang, Zhi and Kuang, Linna and Ghosh, Ishaan and Malik, Jitendra and Fermuller, Cornelia and Wu, Tingfan and Mao, Jiayuan and Liu, Ruoshi and Qi, Haozhi and Aloimonos, Yiannis},
  journal = {arXiv preprint arXiv:2606.26093},
  year    = {2026}
}