EgoSteer-Egocentric

Egocentric hand-motion labels (wrist pose, fingertips, camera pose; EgoSteer 74-d state/action layout, LeRobot v3.0) estimated by the EgoSmith pipeline from public egocentric video datasets. This repository contains no images or video. Each sub-folder is a complete labels-only LeRobot dataset; every frame names the original recording and frame number it was estimated from, so the pixels can be re-attached from the source dataset obtained under its own terms.

Datasets

folder	source	episodes	frames	source licence	labels licence
taco/	TACO	1,977	304,054	CC-BY-SA-4.0	CC-BY-SA-4.0
oakink2/	OakInk2	887	169,881	CC-BY-SA-4.0	CC-BY-SA-4.0
egodex/	EgoDex	147,588	37,061,953	CC-BY-NC-ND-4.0	CC-BY-NC-4.0
holoassist/	HoloAssist	11,426	1,148,545	CDLA-Permissive-2.0	CC-BY-4.0
epic_kitchens/	EPIC-KITCHENS	108,077	4,936,621	CC-BY-NC-4.0	CC-BY-NC-4.0
ego4d/	Ego4D	584,042	13,770,599	custom	CC-BY-NC-4.0
egocentric_10k/	Egocentric-10K	194,225	47,860,714	Apache-2.0	CC-BY-4.0
egocentric_100k/	Egocentric-100K	496,357	237,603,974	Apache-2.0	CC-BY-4.0

Each folder's README.md, LICENSES/<source>/LICENSE.txt and NOTICE.txt carry that dataset's full terms, attribution and provenance; the licence of the labels in a folder is the one in its last column. Copyright and attribution lines:

• taco: TACO: Benchmarking Generalizable Bimanual Tool-ACtion-Object Understanding (Liu et al., CVPR 2024) — see source URL
• oakink2: OakInk2 (Zhan et al., CVPR 2024) — see source URL
• egodex: Copyright (C) 2025 Apple Inc. All Rights Reserved.
• holoassist: HoloAssist (Wang et al., ICCV 2023), Microsoft — see source URL
• epic_kitchens: EPIC-KITCHENS-100 (Damen et al.), University of Bristol — see source URL
• egocentric_10k: Egocentric-10K, Copyright Build AI (builddotai), released under the Apache License 2.0 — see source URL
• egocentric_100k: Egocentric-100K, Copyright Build AI (builddotai), released under the Apache License 2.0 — see source URL

Required citations:

• ego4d:

  Egocentric Live 4D Perception (Ego4D) Database: A large-scale first-person video database, supporting research in multi-modal machine perception for daily life activity. Ego4D Consortium 2020. https://sites.google.com/view/ego4d/home
  

Loading one dataset

from huggingface_hub import snapshot_download
from lerobot.datasets.lerobot_dataset import LeRobotDataset

root = snapshot_download("EgoSteer/EgoSteer-Egocentric", repo_type="dataset", allow_patterns=["taco/**"])
ds = LeRobotDataset(repo_id="local/taco", root=f"{root}/taco")

The first load materialises an Arrow copy of data/ in the Hugging Face datasets cache (about 700 bytes per frame; egocentric_100k needs ~170 GB): point HF_DATASETS_CACHE at a disk with room or use StreamingLeRobotDataset.

Re-attaching the video

Per frame: source_frame_index (frame number counted in decode order from the start of the media; never convert it from time) and source_frame_observed. Per episode: source_media (path relative to the source dataset root), source_media_fps, source_media_frame_start/end, source_media_undistort, calibration/source_camera. rehydrate.py in this repository does it for one folder (needs numpy, pyarrow, opencv-python and the ffmpeg CLI): it lists the source files a folder needs, then fetches each indexed frame from your copy of the source dataset, applies the undistortion / resize, encodes h264 and writes a complete LeRobot dataset with observation.images.head:

python rehydrate.py --labels_root taco --list_media > needed.txt        # which source files to download
python rehydrate.py --labels_root taco --media_root path/to/TACO --output_dir taco_with_video --workers 16

The result converts to training shards with EgoSteer's lerobot_to_wds.py (head-only form: 116-d lowdim, cameras = ["head"]). The source datasets need registration or a signed agreement, so they cannot be fetched automatically; the Egocentric-10K / 100K clips are on the Hugging Face Hub and only the listed tars are needed, e.g. python -c "from huggingface_hub import snapshot_download; snapshot_download('builddotai/Egocentric-100K', repo_type='dataset', local_dir='path/to/Egocentric-100K', allow_patterns=open('needed.txt').read().split())". Ego4D frames refer to the videos undistorted by VITRA's data/preprocessing/undistort_video.py (https://github.com/microsoft/VITRA): run it on the listed Ego4D videos first.

Frame K of a video with ffmpeg (decode-order frame number, no time conversion):

ffmpeg -i <video> -vf "select='eq(n\,K)'" -fps_mode passthrough -frames:v 1 frame.png

Where the original frames come from

folder	source media	source_media looks like	frame numbering / transform
taco/	TACO release, folder Egocentric_RGB_Videos/	Egocentric_RGB_Videos/(brush, brush, bowl)/20230919_036/color.mp4	30 fps video; frame k = k-th decoded frame. A few episodes end a handful of frames after the video: those frames were cut.
oakink2/	OakInk2 release, data/<sequence>.tar (or the unpacked folder)	data/scene_01__A001++seq__…/104422070969 (the egocentric camera's image folder)	an image folder, not a video: source_frame_index = position in the name-sorted file list (000001.png, 000005.png, …, one file per entry of the annotation's frame_id_list). The tar can be read in place.
egodex/	Apple EgoDex raw release	part2/basic_pick_place/820.mp4	30 fps video; frame k = k-th decoded frame.
holoassist/	HoloAssist raw release, video/<recording>/Export_py/Video_pitchshift.mp4	R078-21July-SmallPrinter/Export_py/Video_pitchshift.mp4	variable frame rate (27.5–30 fps): count frames from the start of the file in decode order, never convert source_frame_index to a time. The same original frame can be referenced by two consecutive dataset frames.
epic_kitchens/	EPIC-KITCHENS-100 videos	P01/videos/P01_01.MP4	native 50 / 59.94 fps video; source_frame_index counts native frames (the episodes were cut by VITRA-1M from these videos).
ego4d/	Ego4D videos as prepared by VITRA-1M (undistorted with VITRA's per-video intrinsics)	<video_uid>.mp4	30 fps; frame numbers refer to VITRA's undistorted copy of the Ego4D video, whose frames correspond one-to-one to the original. Obtain Ego4D under its agreement and apply VITRA's undistortion, or use VITRA-1M's released videos.
egocentric_10k/	Hugging Face builddotai/Egocentric-10K	factory_001/workers/worker_001/factory_001_worker_001_0003.mp4 (a member of the worker's *_part*.tar)	30 fps; the fisheye frame must be undistorted with calibration/source_camera (cv2.fisheye.undistortImage(img, K, D, Knew=K), source_media_undistort = opencv_fisheye_knew_k) and then resized to calibration/head_image_size.
egocentric_100k/	Hugging Face builddotai/Egocentric-100K	factory021/worker005/part0004.tar/factory_021_worker_005_0169.mp4 (<tar>/<member>: the video inside that tar)	30 fps; same fisheye undistortion as Egocentric-10K, then resize to calibration/head_image_size.

In every case the original frame is finally resized (whole frame, no crop) to calibration/head_image_size; the labels' calibration/head_intrinsics are given at that size.

See each folder's README for the format details and the format document of the EgoSmith release.