Open data collections
A collection is a paid recording job. Someone needs demonstration data for a specific robot arm and a specific task, puts up a budget, and anyone with a matching setup can record part of it and get paid by the hour.
- Paid by the hour
- You keep your own dataset
- Requester stays anonymous
Every collection is anonymous. We never publish who commissioned it - no company name, no account, no country.
What a collection is
The marketplace sells finished datasets. A collection turns that around: instead of buying what already exists, someone commissions what does not exist yet. They describe the task, name the robot arms they can use, set an hourly rate and deposit a budget. Contributors record on their own hardware.
How it works
- 1
Someone posts a collection
A company, a lab or a private person describes the task, picks the robot arms that qualify, and deposits the budget up front. The money is held before the first episode is recorded.
- 2
You set up and record
You read the specification, build the scene on your own arm and record episodes the way you record any LeRobot dataset. Nothing about your recording workflow changes.
- 3
Your dataset stays yours
Each contributor keeps their own dataset. It is only linked to the collection - no episodes are merged into someone else's tree, no renumbering, no shared upload target.
- 4
The requester reviews, then you get paid
They see one collection made of many datasets and can download the whole thing as one archive. Once your contribution is accepted, the hours you delivered are paid out of the deposited budget.
The recording itself works the way it always does here - with the desktop client, no Python.How recording works
Open collections
No open collection at the moment
Nothing is being collected right now. That changes as soon as someone posts a collection - which takes about a minute: pick the arm, describe the task, set the rate and the budget.
What you need to take part
One of the listed arms
The joint layout has to match - a different arm has different joints and different value ranges, and that cannot be converted after the fact. Which arms qualify is on every collection page.
The camera roles it asks for
A collection names roles, not numbers: wrist, front, side. You assign your cameras to those roles explicitly. A wrist camera filed as a front camera loads fine and ruins the training run later.
A normal LeRobot recording setup
Same recording, same upload as any dataset you already sell here. Frame rate and resolution are part of the specification, and you can check your setup against it before you start.
An account with payouts set up
10% platform share, on collection work and on dataset sales alike. Payouts from EUR 10.
How the money works
The budget is deposited first
A collection only goes live once the full budget is covered by the requester's credit. Nobody records against a promise.
Billed by the hour
Every collection states what an hour is worth and how much working time it allows per episode - setup, recording, putting the objects back, retries. That is the number the payment is built on, and it is public before you start.
Paid on acceptance
The requester reviews your contribution and accepts it; then the money moves from the deposit to your balance. If the review deadline passes, it counts as accepted.
Fees, plainly
10% platform share, on collection work and on dataset sales alike. Payouts from EUR 10.
Questions
- Who is behind a collection?
- Nobody knows, and that is deliberate. The requester is always anonymous: no company name, no account, no logo, no country. Attachments are stripped of their metadata for the same reason.
- Do I have to hand over my dataset?
- No. Every contributor keeps their own dataset exactly as today. It is linked to the collection, not absorbed by it. The requester sees one collection and can download everything as one archive, but underneath it stays your dataset in your account.
- What does the exclusivity tick mean?
- Some collections are exclusive. If a collection is marked exclusive, a dataset that was accepted and paid for cannot be listed for sale in the marketplace afterwards - not as a rule in the terms, but blocked. It is shown on the collection page before you start, and it cannot be switched on or off after the first contribution.
- How much can I contribute?
- As much as is left in the pot, minus whatever cap the collection sets per contributor. Both numbers are on the collection page, so you can see before you build the scene whether it is still worth starting.
- What if my setup does not match?
- You can check it against the specification before you record: frame rate, camera roles and resolutions, joint count, episode length. What does not match is named field by field, so you can fix it instead of guessing.
- What happens when a collection is finished?
- The page stays online and is marked as completed. Dead links help nobody, and a finished collection is the most honest thing we can show about what this actually pays.
- Can I commission a collection myself?
- Yes. The form is deliberately short: the robot arms, a description of what you need, the number of episodes and the budget. Images, videos and documents are optional extras, not a requirement.
10% platform share, on collection work and on dataset sales alike. Payouts from EUR 10.Plans and fees in full
What else you can do here
A collection is one way in. If none of the open ones fits your arm - or if you need the data today rather than in three weeks - these are the other three.
Buy a dataset that already exists
Not everything has to be recorded first. The marketplace lists finished LeRobot datasets with the arm, the episode count and the recording format stated up front, and the download starts once you have paid.
Browse the marketplaceSell a dataset you already recorded
What you recorded on your own arm can be listed for sale - unless it came out of a collection marked exclusive, where that is blocked. Listing happens in your account under Marketplace, so this one needs you signed in.
See what is listedCreate an account and take part
Every contribution is tied to an account, and that is where payouts are set up. Registration and sign-in are the same form; you pick there whether you record data or commission it.
Create an account
A look at the marketplace
This is the other half of the platform: data that is already recorded. The three tiles below are examples, not live listings - they show what a dataset listing looks like. The real catalogue is one click away.
- Example39.00 EUR
Pick a part out of the bin and place it in the fixture
SO-101
180 episodes
- Example24.00 EUR
Fold a towel and put it on the stack
SO-100
120 episodes
- Example15.00 EUR
Put the mug on the saucer, handle to the right
Koch v1.1
60 episodes
Need data that does not exist yet?
Describe it once, and contributors on their own hardware record it for you. You can fill the whole form in without an account; you sign up at the end, and the request goes live once its budget is covered by credit - reserved, not charged.
Would rather hand the whole thing over - specification, fleet, review?Managed data packages
Read on
Six pages around this directory, each written for a different person. Every one of them explains a part of what a collection is made of, or what to do with what it pays.
How a data request works
The long version of this page from the requester's side: what you fill in, what a working hour covers, how the money moves, and what stays invisible.
Read moreYou own an arm hardly anyone has
The rarer the arm, the fewer recordings exist for it - and the harder it is for a requester to find anyone who can record them. Check how rare yours is against the catalogue.
Read moreUniversity labs and student projects
An arm that sits idle between two theses can record paid episodes in those hours. Includes the questions a supervisor asks before signing anything off.
Read moreYou already have recordings to sell
For data providers and brokers: list what you already sourced instead of trading it around the marketplace, and see what the fee leaves you.
Read moreTurn hours at the arm into GPU hours
What an hour of collection work earns, what an hour of cloud training costs, and the one step between the two that nobody automates for you.
Read moreNo graphics card of your own
Training and inference run on rented cards, billed by the minute. Where the credit for the first run comes from, including the door that costs no money at all.
Read more