Export is how your data leaves the platform. It turns a frozen dataset version into a dataset you can download, check and load into your own training code. LeRobot v3 is the default, and you can pick another format when your training code expects one. Because the version never changes, the same export settings always describe the same data.
Start an export
Open the export dialog
Open the version from Versions and choose Export at the top right. Pick a Format; the default is LeRobot v3, with camera videos included and encoded as H.264. The formats are described below.
Check what goes in
The top of the dialog shows What this export contains: how many episodes from how many recordings, how many AI-proposed episodes nobody reviewed, and any privacy findings. Choose which Splits to include and whether to Include labels, then choose Start.
Wait for it to finish
The export appears under Exports on the version page, first as Queued, then Running, then Succeeded. If you start an export with exactly the same settings as one that already exists, the existing one is reused instead of built again.
Export formats
Most teams stay with LeRobot v3: it is the format the platform works in, and it loads straight into LeRobot and Hugging Face tools. When your training stack expects something else, choose it in Format. Robot projects offer all six formats; human projects offer the first four.
| Format | Good for | Projects |
|---|---|---|
| LeRobot v3 (default) | LeRobot and Hugging Face tools. Camera videos as H.264. | Robot and human |
| LeRobot v2.1 | Training code that still uses LeRobot v2.1. | Robot and human |
| Original files | The files exactly as they were uploaded. | Robot and human |
| MP4 and JSON | MP4 video plus JSON metadata, for your own tools. | Robot and human |
| RLDS | OpenVLA and Octo. Register the dataset in their loader config. | Robot |
| robomimic HDF5 | robomimic. Choose Include camera images to add the frames. | Robot |
The rest of this page describes a LeRobot v3 pack; the other formats come with the same file list and SHA-256 checksums.
What the pack contains
A finished export is one export.zip: a complete LeRobot v3 dataset with its data tables, metadata and camera videos. With labels included, each episode also carries its labels and per-episode measurements in an embodiflow folder under meta. Next to the zip, artifact.json describes the pack: the episodes, frames, features, video settings and, for each recording, its review and privacy state.
Each row under Exports shows the format, status, creation time, Load check and size. The load check opens the pack with the LeRobot loader; only Passed means it loaded. The line under the format lists what is inside, for example camera videos and labels.
Download and verify
Choose Download on a succeeded export to get the pack. To check that nothing changed on the way, open the Files tab: it lists every file with its path, size and SHA-256, so you can compare them with your own checksum.
After you unzip the pack, the official LeRobot library reads the folder directly:
from lerobot.datasets.lerobot_dataset import LeRobotDataset and then LeRobotDataset("/path/to/unzipped").
Use it with Hugging Face tools
Above the exports list, Use with Hugging Face tools gives you three ready-to-copy snippets for this version: one for the EmbodiFlow SDK, one for LeRobot and one for hf download. The LeRobot and hf download snippets point HF_ENDPOINT at your team's Hugging Face–compatible address and ask for an access token in HF_TOKEN; replace <access-token> with your own. Choose Copy next to a snippet to take it.
Privacy before export
No written consent is needed for any project. What matters is anonymization: recordings of people must have passed anonymization before their data can be exported. If any have not, the export is refused with a message that names how many recordings need anonymization; run the anonymize step on them and check the result. Recordings with audio that cannot be redacted yet are refused too.
Privacy findings that do not block, such as a possible face left after blur, are listed in What this export contains and are never shown as passed. The same goes for unreviewed AI-proposed episodes: they are included, and the dialog says how many.