First-person data of people doing real work.

We record operators restocking shelves, coiling cables and putting away gym equipment from their own point of view, with stereo video, depth, 1 kHz head pose and hand motion, and deliver it ready for robot-learning, VLA and world-model training.

Download the free sample Ask for the full dataset

One real episode, segment by segment

Restocking canned chips · 103.6 s · 33 captioned segments

2.2–4.3 sPick up the black can of chips from the carton

Reach or pick up
Place
Adjust or align
Pause or look
Start / end gesture

What every episode contains

Each recording is one person completing one task, captured on a head-mounted Pico headset. All streams share one time base and come with per-episode calibration, so you can project depth into the video or the video into the world without guessing.

Stereo video: Left and right RGB at 1920 × 1456, 30 fps, undistorted, with rectified intrinsics.
Depth: Per-frame depth from the headset's depth sensor, plus pose-aligned point clouds.
Head pose: 6-DoF headset pose at about 1 kHz.
Hands: MANO hand reconstruction and 3D hand keypoints in the full delivery.
Action labels: Contiguous, time-stamped segments describing what the operator is doing, with skill tags in the full delivery.
Calibration: Camera intrinsics and distortion, camera-to-IMU extrinsics, stereo baseline and a shared microsecond time base.
Format: MCAP with Foxglove schemas, plain NumPy and JSON files, and an export path to LeRobot.

10,000+ hoursin the full collection

1,203 episodesin the public sample

21 tasksretail shelving and facility work

~1 kHzhead-pose sampling rate

Start free, then scale up

The public sample lets you check the format and quality on your own pipeline before you commit. The sample is licensed for non-commercial use; training commercial models needs a commercial license.

Feature	Basic	Advanced	Full delivery
Access	Hugging Face, instant approval	Hugging Face, reviewed	By agreement
Scale	1,203 episodes	1,203 episodes	10,000+ hours, or a custom set
Video, depth, head pose	Included	Included	Included
Point clouds and MCAP	—	Included	Included
Action captions	—	Included	Segments with skill tags
MANO hands	—	—	Included
License	CC BY-NC 4.0	CC BY-NC 4.0	Research or commercial license

For research groups

If you are working toward a paper on motion prediction, world models, hand–object interaction or robot learning, we can provide a larger subset at no cost, in exchange for an acknowledgement of OpenElephant in the publication.

Research licenses cover non-commercial use within your group, allow you to train models and publish results, and do not allow redistribution of the raw data.

Custom collection

Need a specific environment, object set or task list? We plan and capture it for you:

You describe the scenes and the behaviours your model has to learn.
We turn that into a task list for you to review.
We capture at scale, check quality, and deliver in the format you train on.

Talk to us about the data

Tell us what you are training and roughly how many hours you need. We are based in Beijing (UTC+8).

jiuchen@openelephant.ai