MotorSense

MotorSense: A Video-EMG Dataset of Motoric Representations for Action Understanding (NeurIPS 2026)
Eadom Dessalene, Michael Maynord, Amir-Hossein Shahidzadeh, Botao He, Yasamin Tabatabaee, Ruoshi Liu, Ruohan Gao, Paul Pu Liang, Yiannis Aloimonos

MotorSense pairs egocentric video with bimanual forearm surface EMG and IMU. Participants wear an 8-channel EMG armband with a 6-axis IMU on each forearm while doing unscripted kitchen and household activities, filmed by a head-mounted camera (Project Aria glasses or a head-mounted smartphone). Video and EMG share one clock, so every frame can be matched to the muscle activity of both arms.

Code, loaders, pretraining and fine-tuning: https://github.com/edess442/MotorSense

At a glance

Recordings	370
Video	70.4 h, 5.97 M JPEG frames
Signals	2 × (8-ch EMG + 3-axis accelerometer + 3-axis gyroscope) at 500 Hz
Action annotations	5,209 verb/noun segments on 85 recordings (57 verbs, 180 nouns, EPIC-KITCHENS-100 format)
Contact annotations	4,700 per-hand intervals on 57 recordings, with touched vertices on a 6,173-vertex hand
Size	≈ 0.9 TB (frames ≈ 840 GB)

Capture setup	Recordings	Hours	Frame rate	Frame size
Project Aria Gen 1 glasses	18	22.0	15 fps	704 × 704
Project Aria Gen 2 glasses	19	10.0	10 / 30 fps	1280 × 960 / 1008 × 756
Head-mounted smartphone (ultra-wide)	333	38.4	30 fps	896 × 504

Layout

recordings/<recording>/          e.g. P01_2 (Aria participants P01-P08), M_3 (Aria Gen 2), S_042 (smartphone)
    frames.tar                   uncompressed tar; top-level member frames/ (frame_000000.jpg ..., manifest.json)
    camera_poses.csv             per-frame timestamps on the EMG clock (+ 6-DoF device pose for Aria)
    hand_bboxes.csv              per-frame left/right hand boxes
    left.csv, right.csv          EMG1-8, Accel1-3, Gyro1-3 per arm, with unix_time_s
annotations/actions/             MotorSense_{train,validation,all}.csv, verb/noun classes, splits.json
annotations/contacts/            contact_segments.csv, split.json
metadata/recordings.csv          capture setup, frames, fps, duration, resolution per recording
checkpoints/                     pretrained video encoders (Hiera-B, ResNet-50)

The column-level description of every file is in docs/DATA.md.

Download

The repository's script downloads, filters and unpacks (it extracts each frames.tar and deletes it):

git clone https://github.com/edess442/MotorSense.git && cd MotorSense
python scripts/download_motorsense.py --only metadata                # annotations + metadata (a few MB)
python scripts/download_motorsense.py --recordings S_144             # one recording to start with
python scripts/download_motorsense.py --subset actions               # the 85 action-annotated recordings
python scripts/download_motorsense.py                                # everything (≈ 0.9 TB)

Or with the Hub CLI directly:

hf download edessa/motorsense --repo-type dataset --local-dir data/MotorSense --include "recordings/S_144/*"
tar -xf data/MotorSense/recordings/S_144/frames.tar -C data/MotorSense/recordings/S_144

Notes

• The armbands stream over Wi-Fi, so samples are occasionally lost; lost samples are gaps in unix_time_s, not filled values.
• Recording and annotator names are anonymised. Smartphone sessions are not attributed to participants.
• Frames are not face-blurred.

Citation

@inproceedings{dessalene2026motorsense,
  title     = {{MotorSense}: A Video-{EMG} Dataset of Motoric Representations for Action Understanding},
  author    = {Dessalene, Eadom and Maynord, Michael and Shahidzadeh, Amir-Hossein and He, Botao and Tabatabaee, Yasamin and Liu, Ruoshi and Gao, Ruohan and Liang, Paul Pu and Aloimonos, Yiannis},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026}
}