How can I test GR00T or Pi0 without owning a robot?
Three things are genuinely possible before you buy an arm: drive a real one over the internet, compare the models on published benchmarks, and fine tune on a GPU you rent by the hour. This page says which is which, and where the line really is.
- Real SO-100, no signup
- 85 models compared
- Cloud GPUs by the hour
Three ways in, in order of effort
Each of these costs you either nothing or a few dollars, and none of them needs a robot in the room. They answer different questions, so the order below is the order that wastes the least time.
Drive a real arm, right now
A physical SO-100 is wired up and streaming. Open the page, wait for the queue and move the joints yourself with the keyboard or the on-screen pad. No account, no installation, no simulator. It takes about a minute to learn what a 5 degree of freedom arm can and cannot reach, which is knowledge no datasheet gives you.
Open the live armCompare the models on published numbers
The arena holds 85 vision language action models with 332 benchmark results between them, each value linked to the paper or model card it came from. Parameters, GPU memory, inference latency and licence sit in one sortable table, so you can rule out most of the field before installing a single package.
Open the VLA arenaFine tune on a GPU you do not own
Training runs on rented cloud GPUs, selected by the VRAM the model needs. The form takes a dataset from your account, from your own machine, or from the public Hugging Face hub, so you can watch a complete run finish before you have recorded anything yourself. You pay for the hours the run takes and nothing else.
Read the training guidesWhat a cloud run actually costs
The pool buys GPUs on a spot market and picks them by the VRAM a model needs, so the hourly rate moves day to day. These are the two ends of the range across the five trainable policies.
| Model | Card the run needs | Typical run | Episodes before it is useful |
|---|---|---|---|
| GR00T N1.7 | A100 80 GB or H100 80 GB | 3 to 6 hours, about 4 to 12 USD | from about 50 |
| GR00T N1.5 | A100 80 GB or H100 80 GB | 3 to 6 hours, about 4 to 12 USD | from about 50 |
| Pi0.5 | A100 80 GB or H100 80 GB | 3 to 6 hours, about 4 to 12 USD | from about 50 |
| SmolVLA | RTX 4090 or any card with 24 GB | 2 to 5 hours, about 1 to 3 USD | from about 30 |
| ACT | RTX 4090 or any card with 24 GB | 2 to 5 hours, about 1 to 3 USD | from about 50 |
In practice that means the cheap experiment and the serious one differ by about one order of magnitude: SmolVLA on a RTX 4090 or any card with 24 GB comes to about 1 to 3 USD for 2 to 5 hours, while GR00T N1.7 on an A100 80 GB or H100 80 GB comes to about 4 to 12 USD. Running SmolVLA first is the normal way to find out whether a dataset is worth the bigger run.
When you really do need your own hardware
Two things stop working the moment the arm is not yours. The first is recording. A policy learns the task you demonstrate, in the workspace you demonstrate it in, seen through the cameras you mounted. No public dataset contains your table, your lighting or your objects, so as soon as you want the arm to do your task, you have to record it yourself, which means an arm on your desk and a leader arm or a controller to drive it.
The second is running the trained policy on hardware. Inference has to sit right next to the servos: the control loop for these models runs between about 20 and 485 milliseconds per action step depending on the model, and putting the public internet in the middle of that loop turns a working policy into a jerky one.
The entry ticket is smaller than people expect. An SO-100 is roughly 110 to 150 euro per arm in parts, the SO-101 roughly 130 to 170 euro per arm in parts, both driven by Feetech STS3215 bus servos at 7.4 V. That voltage is the single most expensive detail to get wrong: these servos are not 12 V parts, and feeding them 12 V ends the project before the first recording. The platform also supports Koch v1.1 and LeKiwi, which cost more and change the servo bus or add a mobile base.
The alternatives, fairly
None of these is wrong. They answer different questions, and the honest difference is what you must already own before they tell you anything.
| Route | You need | You get | The catch |
|---|---|---|---|
| NVIDIA Isaac-GR00T example notebook | Your own GPU, plus the CUDA and Python setup around it | A fine-tuning run on the demo dataset shipped with the repository | The closest thing to a first-party tutorial, and the reference for how the model is meant to be trained. It assumes you already have the card, and the demo dataset is not your task. |
| LIBERO and other simulation benchmarks | A simulator install, no robot at all | Comparable success rates across models on standard tasks | The right tool for reading papers and reproducing numbers. It says nothing about how a policy behaves on a real arm with your lighting, your camera angles and your servo backlash. |
| Rent a cloud GPU and drive LeRobot yourself | A GPU rental account and time on the command line | Full control over every training flag | The most flexible option and the one with the most manual work: provisioning, dataset transfer, dependency pinning and cleanup are all yours, and an instance left running bills all night. |
| This platform | A browser, an account only once you train | A real arm to drive, a sourced model comparison and managed cloud runs | Narrower on purpose: five trainable policies and four supported arms, not every model in the field. What is here is wired end to end. |
Questions
Can I try GR00T online?
Partly, and it is worth knowing which part. You can compare GR00T N1.7 and GR00T N1.5 against every other vision language action model in the arena without installing anything, and you can start a fine-tuning run on a rented A100 from the browser, so you never touch a GPU yourself. What you cannot do here is run GR00T inference on the live arm: that arm is teleoperated, so the motions you see are somebody driving it, not a policy.
Do I need a GPU to fine-tune Pi0.5?
Not your own one. Pi0.5 needs an 80 GB card, an A100 or H100, which rules out consumer hardware, so the training pool rents one for the duration of the run. A typical run lands at 3 to 6 hours and roughly 4 to 12 USD at current spot prices. If you would rather start cheap, SmolVLA trains on a 24 GB card for about 1 to 3 USD and tells you within a couple of hours whether your dataset is worth scaling.
Is the /live robot real or simulated?
Real. It is a physical SO-100 with Feetech STS3215 bus servos at 7.4 V, filmed by the cameras you see in the stream. Nothing on that page is rendered or replayed. That also means it can be offline, jam, or hand control to the next visitor in the queue while you are still typing.
Which robot arm should I buy to start?
The SO-100 or its successor the SO-101, at roughly 110 to 150 euro and 130 to 170 euro in parts. Both use the same servos, the same calibration and the same recording workflow, and everything on this platform is tested on an SO-100 first. Koch v1.1 and LeKiwi work too, but they cost more and add a different servo bus or a mobile base, which is complexity you do not need for a first dataset.
Can I train on a dataset I did not record myself?
Yes. The training form takes a dataset from your account, from the machine running the desktop client, or from a search over public datasets on the Hugging Face hub. That is the fastest way to see a full run end to end before you own an arm. Just remember that a policy trained on a different robot and a different camera setup will not transfer cleanly to yours.
Do I need an account to drive the live arm?
No. The live page needs no signup, no download and no payment details. An account only becomes necessary when you want to upload your own recordings, start a training run or rate models in the arena.
The live SO-100 needs nothing from you: no account, no install, no card on file. Everything else on this site makes more sense once you have moved a real joint.