Where the data
that comes in stays.
A demonstration has to be stored somewhere, and when it leaves it has to leave with its rights. Aeffer Hub is a repository for uploading physical demonstration data, versioning it, previewing it in 3D in the browser and sharing it with a consent scope attached. If models and general data have Hugging Face, data that comes out of the human body needs this.
Upload it, check it, share it with rights attached
All six are at the design stage. Two of them — the viewer and the schema — are already built and you can check them now.
- HB-01 · REPOIn design
Repositories and versions
One dataset is one repository. It has a provider/name address, and every time sessions are added a new version accumulates. You can trace which model was trained on which version.
- HB-02 · VIEWERAvailable now
3D preview in the browser
Play a demonstration back as a 3D point cloud before downloading. Sensor placement, hand close-ups and millimetre-scale measurement are all visible in the browser.
Open the viewer - HB-03 · SCHEMAAvailable now
Schema validation on upload
We accept data only after checking that it matches the aeffer-v1 format and that every frame records a pose_source grade. It exports to LeRobot v3.
Schema reference - HB-04 · RIGHTSIn design
Rights metadata required
Data without a consent scope, a commercial-training flag, third-party provision terms and withdrawal terms does not go up. If a demonstrator withdraws, that session drops out of later versions.
- HB-05 · ACCESSIn design
Three levels of visibility
Public, organisation-internal, private. For physical data the default is private. Public can only be chosen where the demonstrator's consent allows publication.
- HB-06 · COMPUTEIn design
Training without export
Technical copy protection does not work on training data. So alongside this we are preparing a way to run training inside the hub without the data leaving it.
This is how one dataset looks
Below is an example of a screen still being designed. It is not real data, and the names and figures are placeholders.
meta/info.jsonschema version · fps · joint definitionsmeta/tasks.parquettask instructionsrights/consent.jsonconsent scope · withdrawal terms (required)data/chunk-000/episode_000000.parquetjoint SE(3) · aperture · pose_sourcevideos/ego/episode_000000.mp4egocentric videoderived/isolation folder for estimated values
$ aeffer login$ aeffer push ./session --to example-lab/kitchen-dishes-ko \ --rights consent.json✓ Schema validated · aeffer-v1✓ Rights metadata · consent scope checked✓ pose_source grade · recorded per frame✓ Uploaded · version v1 · private
Nothing goes up without a rights file. That single line is the biggest difference from a general-purpose hub.
Compatible, not competing
The standard repository and format for robot data are already Hugging Face and LeRobot. We make sure our data exports to that standard, and add on top of it what physical data needs.
| Item | General-purpose hub | Aeffer Hub |
|---|---|---|
| Data covered | Models, text, images, robotics — every field | Only demonstration data from the human body |
| Rights information | Licence tag, optional | Consent scope and withdrawal terms required; upload refused without them |
| Preview | Mostly tabular row views | 3D motion playback · sensor placement · mm measurement |
| Quality marking | Relies on the provider's description | Per-frame pose_source grade (measured · estimated · interpolated · synthetic) |
| Default visibility | Public | Private (physical data) |
| Withdrawal | Deletion at repository level | Withdrawal per demonstrator → session excluded from the next version |
| Format | LeRobot v3 and others | aeffer-v1, exports to LeRobot v3 |
Three groups using the same repository
- AFFERENT · UPLOADING
Collectors and studios
They do not know where or how to store what they have captured so that it becomes sellable.
- Automatic format checks on upload
- Consent form templates and rights file generation
- A record of where their data has been used
- SYNAPSE · BETWEEN
University labs
They want to share data gathered for research, but likeness rights and consent scope get in the way.
- Private and organisation-level sharing
- Running shared evaluation sets
- Training without export, which protects trade secrets
- EFFERENT · RECEIVING
Robotics companies and model teams
Most of their time goes on finding data that is confirmed usable for commercial training.
- Search only data with confirmed rights
- Check quality in 3D before downloading
- Pin a version to reproduce a training run
The order we open in
We have not put dates on these. We will set them once the first pilot capture is done and there is real data.
- H0 · INFORMATION
Design and announcementNow
This page. We publish the viewer and the schema reference first, and take feedback.
- H1 · INTERNAL
Internal loading
We upload the data from the first pilot ourselves, and refine the upload, validation and versioning flow against real data.
- H2 · INVITE
Invite-only beta
We invite a few labs and capture studios that hold data. Private repositories only.
- H3 · PUBLIC
Public
We open public repositories, search and discussion. After enough data worth publishing has accumulated.
We will tell you first
when it opens
The invite-only beta starts with the labs and capture studios that need somewhere to put their data. Tell us what data you hold and how much of it, and we will factor it into the repository design.