quietbit/egocentric-10k-actions-sample

Egocentric-10K Actions (sample)

Robot-learning labels for real factory work, generated automatically from first-person video.

Egocentric-10K by Build AI is 10,000 hours of head-mounted video from real factories, but it comes as raw video with camera intrinsics only. This sample shows what the same footage looks like once it is turned into training data: every frame has two-hand 3D keypoints, wrist poses with grip state that can stand in for robot actions, and a language instruction for each clip. Everything is stored in LeRobot v3 format.

Episodes	211 clips, 41 minutes, from 28 workers in 28 factories
Frame rate	15 fps
Video	640×360, original fisheye view (OpenCV fisheye intrinsics in every frame)
Labels	2D and 3D hand keypoints (21 per hand), hand presence, wrist pose, pinch distance, actions, instruction, sub-steps
Labelling	Fully automatic; not checked by people (see Limitations)

Features

Key	Shape	Content
observation.images.ego	360×640×3 video	Head-mounted camera
observation.intrinsics	8	fx, fy, cx, cy, k1–k4 for the 640×360 image (OpenCV fisheye model)
observation.hands.kp2d	126	[left, right] × 21 joints × (u, v, score), pixels
observation.hands.kp3d	126	[left, right] × 21 joints × (x, y, z), metres, camera frame (x right, y down, z forward)
observation.hands.valid	2	1 if the left / right hand was found in this frame
observation.hands.ee	20	Per hand: wrist xyz, palm rotation (6D: first two columns), thumb–index pinch distance
action	20	observation.hands.ee of the next frame: a wrist-and-gripper stand-in for robot actions
task	text	One instruction per episode, for example "place metal pins into machine"

Joint order follows MediaPipe: wrist, then thumb, index, middle, ring and little finger from base to tip. All values are zero when a hand is not found.

from lerobot.datasets.lerobot_dataset import LeRobotDataset

ds = LeRobotDataset("quietbit/egocentric-10k-actions-sample")
frame = ds[0]
print(frame["task"], frame["observation.hands.kp3d"].shape, frame["action"].shape)

Sub-steps and touched objects for each episode are in meta/captions.json.

How it was made

With the open-source tool ego2lerobot (Apache 2.0). You can run the same pipeline on your own video.

1. Hands: MediaPipe Hand Landmarker (Apache 2.0) runs on every frame.
2. 3D: MediaPipe's metric hand shape is placed in the camera frame by least squares against the keypoints after removing the fisheye distortion, using the calibrated intrinsics.
3. Episodes: stretches with hands in view (at least 70% of frames), 3–15 seconds long, up to 8 per video and spread through it; one video per worker, one worker per factory.
4. Language: Qwen3-VL-8B-Instruct (Apache 2.0) looks at 4 frames per episode and writes an instruction, sub-steps and objects. Episodes it judged idle (13 of 224) were dropped.

meta/sources.json gives the factory, worker, video and start time in Egocentric-10K for every episode.

Limitations

• Depth is estimated, not measured. The distance to each hand comes from the average size of a human hand. Expect errors of several centimetres, and more for unusually small or large hands.
• Fast or blurred hands are missed. Gloved hands and hands at the edge of the fisheye view are detected less reliably.
• Left and right are sometimes swapped. This is most likely when only one hand is visible.
• Captions are machine-written. They can be generic ("assemble or inspect component"), or describe the action wrongly. In our spot checks, one of the 12 preview clips had a wrong instruction.

Need this for your own video?

We label first-person video for robot learning: hand tracks, actions, language, segmentation and LeRobot export. We work with public datasets or your own footage, and orders can be small. If you need depth, we can also record new demonstrations with iPhone LiDAR.

Contact chizkiyahuohayon@gmail.com. Send us a few minutes of video and we will label them for free so you can judge the quality.

Attribution

Video from Egocentric-10K by Build AI, licensed under Apache 2.0. Clips were cut, scaled to 640×360 and re-encoded, and all labels were added by Quietbit Labs. If you use this dataset, please also cite the original:

@dataset{buildaiegocentric10k2025,
  author = {Build AI},
  title = {Egocentric-10k},
  year = {2025},
  publisher = {Hugging Face Datasets},
  url = {https://huggingface.co/datasets/builddotai/Egocentric-10K}
}

א — Quietbit Labs