No GPU, no robot

You do not need a graphics card to train a robot policy.You need an account with credit on it.

Training and inference here run on rented cloud GPUs, billed by the minute they actually run. Nothing is installed on your machine, nothing is bought. The only question left is where the credit comes from, and this page answers exactly that.

See what credit costs Try it with no account first

10.00 EUR: Smallest top-up. That is the entry ticket.
0.30 EUR: Estimated cloud inference, per hour
10 %: Platform share when you earn here
none: Hardware you have to buy first

Start here

Three things work right now, with no account and no payment.

Before any of the money questions: find out whether this is even for you. Nothing below asks for a card, and nothing below installs anything.

Drive a real arm from the browser

A physical SO-100 stands in a room and takes commands from a web page. You get a queue slot, a camera view and the joint controls. No account, no installation, no simulator.

It is teleoperation: you drive the arm, a policy does not. There is a queue, and the arm is sometimes offline.

Open the live arm

Compare the models before you pick one

A directory of the vision-language-action families: what each one is built for, who released it, whether the weights are open, and how it scores on the public benchmarks. The 5 that can be fine-tuned here are marked as such. It is the cheapest way to decide which one is worth your first training run.

Published benchmark numbers and spec sheets, not a model you steer.

Open the arena

Move a simulated arm on the front page

A 3D arm runs in the browser with keyboard control, joint sliders and a leader-follower mode. It is a real physics scene, and it is the fastest way to understand what an episode looks like before you pay for anything.

The record button in that studio is a mock-up. No dataset comes out of it.

Open the browser studio

And two things that people expect and do not get

You cannot record a usable dataset in the browser simulator, and you cannot run a policy on the live arm. Both are stated here rather than discovered later: the simulator studio labels its own recording as sample data, and the live arm is a teleoperation station. Everything on this page that involves a real dataset involves either a real arm or somebody else's recording.

Not sure it is worth an account yet? The live arm and the arena need neither.

Try it for free

The circle

Collect, earn, compute. On one account, in two separate books.

This is the reason the page exists, so here is how it really runs, including the step that is not automatic.

You deliver work

A dataset you list in the marketplace, or episodes you record for an open Collection. Both need an arm, yours or a borrowed one. This is the one step on the whole page that hardware is required for.

The buyer accepts, and you are credited

A dataset sale and a released Collection contribution both land in your seller balance, minus the platform share of 10 percent. Nothing is released before the buyer has accepted; nothing is deducted twice.

You move it into training credit

The seller balance is a payout book, not a compute wallet. You request a payout from 10.00 EUR upwards, the transfer is executed, and you top your training credit back up from 10.00 EUR upwards. Same account, two ledgers, one manual step in between.

What does not happen, and we would rather say it here

Your earnings do not turn into GPU time by themselves. There is no button that converts a seller balance into training credit, and there is no automatic transfer behind the scenes. The money goes out and comes back in, and both directions have a floor of 10.00 EUR. Anyone telling you that selling data buys you free compute is describing a product that does not exist here.

Where the circle really is closed

Once credit is on the account, one pot pays for everything: a training run, an inference pod, and the budget of a Collection you commission yourself. And the Starter subscription comes back in full as training credit, month after month - the amount you pay is the amount you get to spend on compute.

Three doors

There are exactly three ways credit gets onto the account.

Two of them cost money up front. The third is the one this page is about, and it is the slowest.

Fastest

Top up

Smallest amount 10.00 EUR, largest 10000.00 EUR. No subscription, no term, no monthly fee. The credit sits on the account until something consumes it.

Best if you train regularly

Subscribe to Starter

49.99 EUR per month, and 49.99 EUR of training credit lands on the account every month. The fee comes back in full and the credit adds up. If you train more than once a month, this is the cheaper door.

Unused credit carries into the next month. It has no expiry date.

No money of your own

Earn it here

Sell a dataset in the marketplace, or record for an open Collection. The platform keeps 10 percent, the rest is yours. Then take the payout from 10.00 EUR upwards and put it back in as credit. Slow, manual, and the only door that does not start with your own money.

What an hour at the arm buys in GPU time

And this is how far it goes

Two numbers decide whether a beginner budget is enough: what an hour of inference costs, and what one training run costs. Both come straight out of the billing code, not out of a sales deck.

What	How long it runs	What it costs
Cloud inference, one session	Up to 4 hours per session, then it stops	0.30 EUR / h
One training run, ACT or SmolVLA	2 to 5 hours	about 1 to 3 USD(0.30 to 0.60 USD / h)
One training run, GR00T or Pi0.5	3 to 6 hours	about 4 to 12 USD(1.20 to 2.00 USD / h)

The hourly figure for inference is the estimate shown before you start, rounded up on purpose. What actually gets booked is the real rent of the pod that was found, and that has been lower. The training figures are the observed range for a full fine-tune; the run is billed by its real runtime, with no markup on top.

Which means the smallest top-up of 10.00 EUR is worth roughly 33 hours of inference at the estimated rate, or a handful of small training runs. That is the honest size of the entry ticket.

Two things nobody mentions until the bill arrives: a pod needs about 8 minutes to come up before it answers anything, and it is kept warm for 30 minutes after you stop clicking so the next attempt does not pay that wait again. Both are on the meter.

Credit on the account is the only thing standing between you and a training run.

See plans and credit

The real obstacle

Without an arm, the hard part is not the GPU. It is the data.

A rented GPU is a card number away. A dataset of your task, with your gripper, in your light, is not. Here are the three routes that need no hardware at all.

Point at a public repository

Training takes a Hugging Face dataset id and checks it before the job starts. Any public LeRobot dataset works, including the large community sets. Costs nothing, and it is how most first runs happen.

Somebody else's table, somebody else's light. Good for learning the pipeline, rarely good enough for a real task.

Take a free listing

Some marketplace listings are priced at zero. Those you can acquire outright with an account, no payment step, and train on them like any other dataset.

The selection is whatever people have decided to give away.

Buy one that already fits

The marketplace lists finished LeRobot datasets with their episode count, cameras and arm. You see what is inside before you pay, and you download once. The seller keeps the price minus 10 percent.

Costs money, and the exact task you want may not be there yet.

Or have it recorded for you

If nothing fits, post a data request: describe the arm, the task, the scope and the hourly rate, and people with the hardware record it on their side. Every delivery is linked into one Collection that you download once, and nobody is paid out until you have accepted. It is the route that costs the most and fits the best.

How data requests work Or fill one in right away

How much data is enough to start

ACT becomes usable from around 30 episodes of one task; the larger families want about 50. That is a weekend of recording for somebody with an arm, and it is the number to check a dataset against before you spend credit training on it.

The honest list

What you need, what you do not, and where it stops.

Written out so nobody has to find the wall by walking into it.

You do need

An account and credit on it. Training and inference both refuse to start at zero, with a clear message rather than a silent failure.
A dataset you are allowed to train on: a public repository, a free listing, one you bought, or one you had recorded.
A browser and patience for the cold start. A pod is not instant.
An arm - but only if you want to earn here rather than pay here. That is the one hardware requirement on the whole page.

You do not need

A graphics card. The training and inference GPUs are rented per job and released afterwards.
A robot, to train, to evaluate, or to compare models.
A Python environment, CUDA, or a working local install. The job is configured in a form.
A subscription. Credit works on its own, and a top-up carries no term.

And here is where without hardware stops

Recording your own dataset needs a real arm, cameras and the desktop client. There is no browser route around it.
Running a trained policy on a physical robot needs a physical robot. The live arm on this site is teleoperated and does not run your checkpoint.
Local inference is free, but free is the whole point of owning the card. Without one, inference is a rented pod and rented pods are billed.

Pick one

Six things you can do on this platform today.

Three of them cost nothing, two of them earn, one of them spends. All six are live.

Buy a dataset

Finished LeRobot datasets with episode counts, cameras and arm listed. Download once, train on it as often as you like.

Browse the marketplace

Sell your dataset

Already recorded something on an arm? List it. You set the price, the platform keeps 10 percent, and the rest goes to your seller balance.

See what sells

List a data request

Describe the recording you need and let people with the hardware make it. One request, many collectors, one Collection.

See how requests work

Collect for an open request

If you do have an arm, this is the door into the earning side: pick an open Collection, record the task at your place, get paid by recorded time minus 10 percent.

See open Collections

Train a model

Pick a policy family, point at a dataset, start the run. The GPU is rented for the job and released when it finishes.

How training works

Serve a checkpoint

Put a trained policy on a cloud GPU and drive a robot against it. Billed per hour of pod, stopped automatically when you walk away.

How cloud inference works

Or take the desktop client and run it on your own machine

If you do end up with a card, local inference costs nothing at all - and the same client is what records datasets once you have an arm.

Download the client

The card is not the barrier. It never was.

Rent the compute for the hours you use it, get the data from somebody who already has the arm, and keep the 10.00 EUR entry ticket instead of the two thousand euro one.

Create an account Try it without one first