The labels in this dataset are free for research and other non-commercial use (CC BY-NC 4.0). For commercial use, write to hello@arete-robotics.com. Access is approved automatically.

Egocentric-10K Actions

648 robot-learning episodes from 30 real factories, labelled automatically from first-person video, with factory-level splits and a hand-forecasting benchmark.

Egocentric-10K by Build AI is 10,000 hours of head-mounted video from real factories, released as raw video. This dataset turns a slice of it into training data. Every frame has two-hand 3D keypoints and a wrist-and-grip action. Every episode has an instruction, sub-steps, the objects touched and an action verb. Everything is stored in LeRobot v3 format.

This is release 2.0. The free 41-minute sample (Apache 2.0) has the same features, without splits, verbs or the benchmark.


Episodes	648 clips, 2.15 hours, 116,276 frames
Coverage	30 factories, one worker each, 56 source videos, 6–24 clips per worker
Clip length	3–15 s, median 13 s
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, objects, verb
Splits	train 484 / val 48 / test 116 episodes, by factory
Labelling	Fully automatic; not checked by people (see Limitations)

What is new in 2.0

Three times the data. 648 episodes instead of 211, with up to 12 clips from each of a worker's videos instead of 8.
Factory-level splits in meta/splits.json. A whole factory goes to one split, so no workplace, worker or bench appears in both training and test. The rule is sha1(factory) % 10: 0 is test, 1 is val, the rest is train. It does not depend on which factories a release contains, so later releases keep the same assignment.
Action verbs. Each episode has one verb, mostly from a fixed list: place 240, assemble 153, insert 132, inspect 31, sort 26, solder 15, fold 11, screw 9, and 13 rarer ones. The counts are in meta/quality.json.
A quality report in meta/quality.json. It gives, for each episode, how often one or both hands were found and how much the wrist estimate jitters.
A forecasting benchmark with baselines (below).

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 "solder electronic components to circuit board"

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.

import json
from huggingface_hub import hf_hub_download
from lerobot.datasets.lerobot_dataset import LeRobotDataset

repo = "arete-robotics/egocentric-10k-actions"
splits = json.load(open(hf_hub_download(repo, "meta/splits.json", repo_type="dataset")))
train = LeRobotDataset(repo, episodes=splits["episodes"]["train"])
frame = train[0]
print(frame["task"], frame["observation.hands.kp3d"].shape, frame["action"].shape)

File in `meta/`	Content
`captions.json`	Instruction, sub-steps, objects and verb for each episode
`sources.json`	Factory, worker, video and start time in Egocentric-10K for each episode
`splits.json`	The split rule, the factories in each split and the episode lists
`quality.json`	Per-episode detection rates and wrist jitter, plus dataset totals
`benchmark.json`	Baseline results for the forecasting benchmark

scripts/bench.py is the benchmark script.

Quality

Measure	Value
Frames with at least one hand found	86.7%
Frames with both hands found	42.4%
Wrist jitter, median over episodes	2.6 cm (90th percentile 4.8 cm)

Wrist jitter is the median size of the frame-to-frame second difference of the 3D wrist position. Real hand motion contributes very little to it at 15 fps, so it is mostly estimation noise, and most of that noise is in depth. Use it to filter episodes: for example, keep those under 3 cm.

Benchmark: where will the hand be in one second?

Given one second (15 frames) of a wrist's 3D path, predict the next second (15 frames). Windows start every 5 frames, for each hand wherever it is found in all 30 frames: 8,006 for training and 2,111 for testing. Errors are in centimetres on the test factories, which no method sees during training.

Method	ADE (cm) ↓	FDE (cm) ↓
Static: the hand stays where it is	6.37	8.69
Constant velocity	13.62	23.30
Ridge regression on the past path	6.37	8.55

The simple methods barely beat standing still. Constant velocity does worst because it extrapolates the depth noise. A good model will probably need to look at the image as well as the path: what the hand is reaching for, and what step of the task comes next. scripts/bench.py in this repository reproduces these numbers, and its windows() function gives the exact training and test windows for your own model. We would be glad to list better results here.

huggingface-cli download arete-robotics/egocentric-10k-actions --repo-type dataset --local-dir ego10k-actions
python ego10k-actions/scripts/bench.py ego10k-actions --out results.json

How it was made

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

Sampling: one worker from each of 30 factories, chosen with a fixed random seed. The first 4 minutes of up to two of that worker's videos are used (56 videos in all).
Hands: MediaPipe Hand Landmarker (Apache 2.0) runs on every frame.
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.
Episodes: stretches with hands in view (at least 70% of frames), 3–15 seconds long, up to 12 per video and spread through it.
Language: Qwen3-VL-8B-Instruct (Apache 2.0) looks at 4 frames per episode and writes an instruction, sub-steps, objects and a verb. Episodes it judged idle (11 of 659) were dropped.

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, hands very close to the camera 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. We checked the 12 preview clips by eye. One instruction was wrong: "insert wire into component" for a worker using a power drill. Three were correct but vague, such as "place small white object on black surface". The model was asked to choose a verb from a fixed list, but it sometimes used its own (solder, drill, apply).
Few factories per split. The test split has 5 factories and the val split 2, so results can depend on which factories those are.
Workers are visible. Hands, arms, clothing and sometimes other workers appear in the video, as in the original dataset. Do not try to identify them.

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. A commercial licence for these labels, or a larger slice of Egocentric-10K, is also available.

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

Licence and attribution

Video: from Egocentric-10K by Build AI, licensed under Apache 2.0. Clips were cut, scaled to 640×360 and re-encoded by Arete Robotics.
Labels (everything in data/ apart from the video, and everything in meta/): © Arete Robotics, licensed under CC BY-NC 4.0. Commercial licences on request.

If you use this dataset, please cite both:

@dataset{arete2026egocentric10kactions,
  author = {Arete Robotics},
  title = {Egocentric-10K Actions},
  version = {2.0},
  year = {2026},
  publisher = {Hugging Face Datasets},
  url = {https://huggingface.co/datasets/arete-robotics/egocentric-10k-actions}
}

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

א — Arete Robotics

Downloads last month: 71

Total file size:

1.92 GB