Your arm stands still most of the day.Someone pays for its data.
- You can buy a dataset, or have one recorded. Both run over the same marketplace.
- A lab, a company or a private person posts the job. They name the arm, the task, the number of episodes and the budget.
- You record it yourself. You build the scene at your place, record on your own arm, and are paid for the time you record.
- The rarer your arm, the smaller the field. Few people own the hardware, so few can take the job at all.
Jobs are in preparation. This page explains how they will work.
- 10%
- platform share on paid recording time, the rest is yours
- €10
- has to add up before you can request a payout
- 353
- arm models in our catalogue, and datasets exist for few of them
- LeRobot
- the format you deliver in
How rare is your arm?
Pick your arm. We count live how many published datasets on this marketplace were recorded on that hardware.
Counting ...
We count published datasets on this marketplace, not every dataset in the world. The marketplace groups hardware into families, so your number covers the family your arm belongs to, not just the single model. These are the same numbers buyers see in their filter.
Data is exactly as rare as the hardware it was recorded on. There are many SO-100 recordings because there are many SO-100 arms: the parts cost roughly 110 to 150 euro, the build is open, and this marketplace itself was seeded with public SO-100 and SO-101 datasets. There are few recordings from an arm that lives in a cell, because few people own one and even fewer set it aside for recording.
That is why jobs for rare hardware are hard to fill. The group of people who can take them is small. What a job pays is decided by the client and not by this platform, and we are not promising you a premium here. We are telling you where one could come from.
Why data sticks to the hardware
The reason a job is posted for a piece of hardware and not for a task.
A robot that is meant to do a task is usually not programmed today, it is shown. A person carries out the motion while camera images and joint positions are written down in the same beat. One such recording is called an episode, and a model learns the task from many episodes. So the goods here are not the robot and not the model, they are the recording.
The joint positions sit in the recording as the action vector: the list of numbers that set the robot in that moment. An SO-100 has five joints and a gripper, six actuators, so six numbers per time step, and our catalogue also lists arms with five and with seven axes. That row of numbers is why a recording does not carry from one machine to the next.
Whether a model can learn from foreign hardware anyway is an open research question. Someone who runs a machine and has to get a task working does not wait for the answer. They have it recorded on their own hardware. That is the whole reason jobs exist.
| Difference | Does it affect the data? | How you can tell |
|---|---|---|
| Different servos, different voltage, different driver | No | Koch v1.1 against SO-100: six joints stay six joints, and everything behind the serial port stays the same |
| Extra axes or wheels | Yes | LeKiwi: the wheel velocities join the action vector, so datasets and checkpoints no longer fit together |
| Different reach, different camera position | Yes | The same task looks different from another viewpoint, and the model learns the viewpoint along with it |
What that means for you
Three sliders. Below them is the same calculation the platform settles with, platform share already deducted.
Our country rates for remote recording run from 8 to 22 euros; Germany sits at 17. Above 22 the slider shows what a better paying job would mean. We have no documented rate for rare hardware.
Every job sets a minimum and a maximum. The shipped default window is 20 to 180 seconds; 5 seconds to 30 minutes is possible.
- Price per episode
- €0.28
- Episodes per hour
- 60
- Platform share per episode (10 %)
- -€0.03
- Your share per episode
- €0.25
- Your rate per hour
- €15.00
What is left for you (88.2 % of the rate)
€90
Recording time until your first payout of €10: 40 min
Your earnings pass through a holding pot. They become payable once the client accepts or the review period runs out. That is why a late rejection can no longer take away what has already been released.
Because every episode is rounded to whole cents, your hourly value does not land exactly on the rate. It misses it in both directions, depending on episode length. Drag the middle slider and you will see it.
A calculation, not an earnings promise. Whether a job for your arm is open is up to the clients. How much you record is up to you.
What the calculator leaves out
You are paid for the time you record. Seven things to know up front.
- Setup
- Building the scene does not count as recording time.
- Before the first episode runs, the scene has to stand: the same objects, a sensible camera position, a setup that matches the description. A job has to show at least three good and one bad example, so you are not guessing what is meant. Setting it up is still not paid.
- Resetting
- After every episode the setup starts over.
- Object back in place, arm back to its starting pose. That often takes longer than the recording itself, and it comes up again for every single episode.
- Discards
- Not every recording is usable.
- The gripper misses, someone walks through the frame, the arm hits an end stop. You delete those episodes yourself before you hand anything in. That is why many episodes on one job pay off more than a few on many.
- Time window
- Too short or too long is not paid.
- Every job fixes a minimum and a maximum duration per episode. Anything outside is not billable, and you are shown the reason instead of a silent zero.
- Rejection
- Only held money can be taken back.
- After recording, your earnings are credited as held while the client reviews. If they reject, exactly that held amount goes away. Anything already released stays yours.
- Price lock
- Your rate freezes the moment you accept.
- Hourly rate, episode price and time window are captured when you take the job. If the client lowers the rate afterwards, that applies to the next run, not to yours.
- Exclusivity
- Some jobs close the resale route.
- A client can tick exclusivity on their collection. Then the dataset you hand in may not additionally be sold on the marketplace, and the sale route is closed for it rather than merely forbidden. The tick is visible before you start, and it cannot be flipped once the first contribution is in.
On a job you are doing for the first time, expect considerably more time than the calculator says, and on a job you already know, considerably less. We deliberately give no factor here: it depends on the task, on your setup and on your practice, and an invented number would be worse than none.
Many datasets, one collection
Move the sliders and look at the same job from both sides.
Your dataset stays yours. You record it, you upload it, it belongs to you, exactly as it does today. What a job adds is a link: your dataset is tied to the collection that the job is filling. A collection is not one big file, it is many separate datasets, one per collector.
The client sees that collection as one thing, with one fill level and one download. Because the datasets stay separate underneath, nobody ever writes into your files, and a different camera count on your setup cannot break anybody else recording.
A large job is not taken just because someone started before you. You keep filling the same pot as long as it is not full, and twelve episodes are twelve episodes even if the job runs to four hundred.
You never find out who posted the job. Clients stay anonymous.
You see your own dataset, your hours and your money. Everyone else is one grey area to you.
Split evenly, for illustration only. The numbers come from your two sliders.
See the collections that are open right nowFrom the tile to the money
Five steps, no application in between.
Pick a job
Jobs sit on the marketplace as tiles next to the finished datasets. Each tile shows which arm is needed, how many episodes are already collected, how many hours that is, what is still missing to fill the pot, what the hour pays and how big the budget is. Who posted it is not on the tile.
Read the specification
Open the tile. Inside are the task description in the client own words, the time window per episode, the camera roles the job asks for, whether the job is exclusive, and example recordings for good and bad. Images, video or documents can be attached, but none of that is required.
Build the scene
You rebuild what the job describes at your place. Your hardware, your table, your cameras, assigned to the roles the job asks for, so a wrist view never arrives as a front view.
Record
With the client for macOS, Windows and Linux. It detects arm and cameras, walks you through recording and review, and writes the LeRobot format directly. If you have your own recording pipeline, you export into that format.
Review and payout
Your dataset is linked to the collection of the job. The client reviews it. Once they accept or the review period runs out, your earnings become payable. From ten euros you can have them paid out.
Sell your own, or collect on a job
Both exist here. The difference is not the work, it is the risk.
| Criterion | Selling your own dataset | Collecting on a job |
|---|---|---|
| Who picks the topic | You do. You record what you find interesting. | The job does. Task, arm and time window are fixed. |
| When the money arrives | When someone buys. Possibly late, possibly never. | The client tops up the budget in advance, so what is posted is covered. |
| Platform share | 10 percent of the sale price. | 10 percent of the hourly rate. |
| Billing unit | One price for the whole dataset. | Per episode, derived from the hourly rate. |
| Your risk | You carry the risk that nobody buys. | You carry the risk that an episode is rejected. |
Ten percent either way: on a dataset sale and on paid recording time.
What you need
If you already record, you have most of it.
- Hardware
- An arm you can drive yourself.
- Which one is needed is stated in the job, precisely: not a similar one, the one named. Our catalogue knows 353 models from 37 manufacturers, and whatever is missing you enter as free text.
- Cameras
- At least one, better two.
- A recording without images is worthless for learning. The usual pair is one overview camera and one on the wrist; our own datasets run with up to five. A job states its cameras by role, wrist, front or side, and you assign yours to those roles when you set up.
- Software
- The client for macOS, Windows and Linux.
- Free, no Python. It detects arm and cameras and writes the LeRobot format directly, v2.1 or v3.0.
- Time in one go
- Yours, whenever you want.
- No shift plan and no minimum quantity. But a job pays off when you set it up once and then record a series. Ten minutes in between are not enough for the setup.
- Space
- A table you can leave standing for a while.
- A job describes a scene. Anyone who has to take it down in between rebuilds it every session.
- Payout
- PayPal, USDC, USDT or Bitcoin, or SEPA on request.
- For USDC and USDT you pick the network too: ERC20, BEP20 or TRC20. The platform checks that address and network match, because an address on the wrong network loses the money. Payable from ten euros.
So that no false expectation comes up
This is not employment.
You take a job or you leave it. Nobody assigns you shifts, and nobody guarantees there will be a job for your hardware next week.
Nobody drives your robot.
On a job you record yourself, on your machine, in your room. Strangers driving a robot remotely also exists on this platform, but that is a different route, and it always starts with the owner explicitly releasing it.
We are not telling you what your arm is worth.
What a job pays is stated in the job. This page explains how the amount is put together and what reaches you, nothing more.
Do not hand in anything that is not yours.
If the machine belongs to an employer or a university, check first whether you may hand out recordings of it. That goes for anything that happens to be in frame as well.
Questions
No. The arm stays with you, you record yourself, and nobody drives it remotely.
Read on
The same thing in a university lab
Idle hours between two theses, and what a supervisor wants answered first.
Read moreSelling instead of only recording
Recordings you already own can be listed for sale. What the fee leaves you, in one calculation.
Read moreRecording without Python
What the client does and how a recording session runs.
Read moreWorking remotely
The other route: no hardware of your own, on someone else arm.
Read moreWhat you can do here
Buy datasets
Finished recordings from other people, ready to train on, filtered by arm and task.
Open the marketplacePost a request
Need recordings on a particular arm? You can fill the whole request in without an account.
Write a requestBrowse open collections
Which collections are being filled right now, on which arm, and how full the pot already is.
See the collectionsCreate an account
You need one to take a job, to get paid, or to list a dataset of your own.
Sign upSee whether your arm is wanted
If none of the open jobs fits your arm right now, you still see which tasks are in demand and what they pay.