Install¶
You install the small urdf2vla command on your machine. It runs the heavy work in two Docker containers: one with
Isaac Sim, Isaac Lab and Isaac Lab Arena (simulation), one with GR00T N1.7 (conversion, training, inference).
Where to run commands
Every command in these docs runs in a terminal on your machine, in the urdf2vla folder, after
source .venv/bin/activate. You never enter a container: urdf2vla starts the work in the right one for you.
Docker itself only appears once, in steps 2 and 3 below.
What you need¶
| OS | Linux x86_64 (Ubuntu 22.04 or 24.04) |
| GPU | NVIDIA RTX GPU with 16 GB or more and driver 580 or newer |
| Docker | Docker with the NVIDIA Container Toolkit |
| Disk | 150 GB free; 300 GB if you train on this machine (each checkpoint is 24 GB) |
| Python | Python 3.10 or newer and uv |
| Hugging Face | An account with the license of nvidia/Cosmos-Reason2-2B accepted, for training and evaluation (step 5) |
| Training | A GPU with 44 GB or more (L40S, A100, H100, RTX 6000 Ada). It can be another machine (Train and evaluate) or a cloud GPU |
1. Install the urdf2vla command¶
git clone https://github.com/momoiorg-repository/urdf2vla.git
cd urdf2vla
uv venv .venv
uv pip install -e '.[view]' # [view] adds the browser robot viewer
source .venv/bin/activate
2. Build the containers (once)¶
This fetches Isaac Lab Arena and Isaac-GR00T at fixed versions into external/ and builds two images,
urdf2vla/isaaclab_arena:b9e37c5 (38 GB) and urdf2vla/gr00t:23ace64 (26 GB). The first build downloads a lot; on an
L40S it took 13 and 4 minutes. bash docker/build.sh sim or gr00t builds one of them; --force rebuilds.
3. Start the containers¶
Running them again just makes sure the container is running. --rm removes it (downloads stay cached).
4. Check the machine¶
doctor checks the GPU driver, Docker and its access to the GPU, the images and containers, disk space, NVIDIA's asset
server and Hugging Face, and prints the command that fixes each problem. Run it whenever the environment looks wrong.
5. Hugging Face (for training and evaluation)¶
GR00T N1.7 uses the vision-language model nvidia/Cosmos-Reason2-2B, which needs its license accepted:
- Accept the license at huggingface.co/nvidia/Cosmos-Reason2-2B.
- Create a token with read access at huggingface.co/settings/tokens.
-
Log in (the token is saved in
~/.cache/huggingface, which the containers share):
Next¶
- Let an AI agent run the steps → Use it with an AI agent
- Run each step yourself with the bundled Franka → Tutorial
- On a cloud GPU (one script sets up the driver, Docker GPU access and disk) → Run on NVIDIA Brev