Lo6yu/egocentric_dataset

Egocentric RGB-D + EMG/IMU Daily Activity Dataset

This dataset contains first-person daily activity recordings with synchronized RGB-D video, wrist EMG/IMU signals, hand keypoints, object masks, hand-object contact annotations, per-finger force annotations, and semantic action segments.

Overview

The dataset is designed for egocentric embodied AI and robot learning in everyday household activities. RGB-D frames provide visual and geometric context, EMG/IMU streams capture hand activity, and contact-force annotations describe where fingers interact with objects and how force is distributed across the hand.

These modalities support research on manipulation understanding, imitation learning, action segmentation, contact-aware perception, multimodal behavior modeling, and robot training from human demonstrations.

Dataset Highlights

• First-person RGB-D recordings from daily household tasks.
• Synchronized wrist EMG/IMU streams aligned to the RGB frame timeline.
• Hand keypoints, object masks, hand-object contact annotations, and per-finger force annotations.
• Semantic action segments for task-level and subtask-level analysis.
• training_ready/ exports for model and robot learning workflows.

Packages

Package	Files	Size GB	Validation
washing-machine-laundry	315	1.283	pass
tidy-bedroom	1188	5.213	pass
sweep-and-mop-floor	1333	6.241	pass
tidy-living-room	929	3.765	pass
tidy-dining-room	1610	7.08	pass
hand-cream-application	257	0.502	pass
sink-hand-washing	151	0.321	pass
paper-towel-hand-wiping	159	0.353	pass

Structure

packages/<package_id>/
  source_stage_a/          # original lossless RGB-D and sensor export
  clean/                   # aligned timelines and corrected sensors
  gold/                    # episode and standard model exports
  analysis/contact_force_v2/
  analysis/semantic_subtasks/
  events/
  training_ready/          # consolidated model-ready files

Training

For model or robot training, start from packages/<package_id>/training_ready/. It contains frame indexes, corrected EMG/IMU streams, semantic segments, contact-force tables, and standard imitation-learning exports when available.

Loader Examples

The repository includes two lightweight Python examples under examples/.

python examples/quickstart_loader.py --dataset-root .
python examples/training_sample_loader.py --dataset-root . --package-id hand-cream-application

quickstart_loader.py reads viewer/train.csv, lists all packages, and checks the expected public directories. training_sample_loader.py shows how to locate RGB/depth timelines, corrected EMG/IMU streams, semantic action segments, contact-force tables, RLDS episodes, and robomimic exports for one package.

The quickstart script uses only the Python standard library. The training sample script can also preview Parquet tables when pandas and pyarrow are installed.

Citation

If you use this dataset, please cite the dataset repository and reference the package IDs used in your experiments.