Please support "stand alone" or "air gap" installation!

[mpgussert]

Proposal

Isaac Lab needs a method for installation that does not require an internet connection in order to support usage on secure intranets and air gapped networks. It should be possible to download a single package, load it to a USB drive, and install a fully working version of Isaac Lab from that storage device. Presently, even when using a binary install of Isaac Sim, an internet connection is required to install the actual RL dependencies of Isaac Lab.

I think the binary installation process should look something like ...

1. unpack the isaac-lab zip file
2. setup environment variables for Isaac Sim
3. setup environment variables for your RL library of choice
4. run some setup script if necessary,
5. use Isaac Lab

Motivation

Many organizations have isolated intranets in which proprietary research and development occurs. A "stand-alone" or "air gap" installation method would help support work in these kinds of secured environments, and alleviate the need to manage a cache of dependencies.

Alternatives

An alternative would be to detail how to "install" Isaac lab without running the install script.

Another one would be to build one large binary that packages all the RL deps together.

Build Info

N/A

Additional context

Add any other context or screenshots about the feature request here.

Checklist

• I have checked that there is no similar issue in the repo (required)

[AntoineRichard]

Hey Gus! We'll look into this, could you maybe help us narrow down this request scope?

• The air gap installation will likely also need some asset infrastructure changes. I don't think our assets are air gapped right now, and likely also the kit extensions that end up getting pulled.
  • Kitless workflows might be better suited for that. Would it be acceptable?
  • Would it be fine for it to not be able to pull any assets?
• Even if we install from pip wheel, it'd still need to download all the dependency packages.
  • unless we pre-download everything and package all those together?

[mpgussert]

Heya man! Oh how the tables have turned lol Now I'm on the outside looking in, and you're fielding tickets! XD

A Kitless workflow would be ideal I think! The runtime extension download and installation is definitely a problem in the airgap scenario.

I think assets can be kept separate, with the exception of the mujoco classics like ant and humanoid. This would allow the user to immediately test the install without needing to setup assets separately. The Isaac Sim binary points to asset files already, so it would be reasonable to follow that system (just zipped directories with an expected structure provided in the docs).

Downloading all the dependencies and packaging them together might be the only way, The Isaac Sim binary install somehow manages all the python deps and environment via the python.sh script, so a binary install of Isaac Lab might need to do similar? Maybe a python.sh like script that expects to be pointed to the python.sh from Isaac Sim and contains all the logic for the additional dependencies for Lab?

[thetenthbox]

How would this impact licensing rules if we were to distribute this by downloading the library and its dependencies (omniverse) with no app built on top? Would it still be permissible?

[StafaH]

Hi @mpgussert,

With the added support of uv and kit-less installation path, the option we recommend is a pre-seeded uv cache. On a connected machine that matches the target OS and architecture, you can populate a dedicated cache for the exact Isaac Lab revision and extras they need:

UV_CACHE_DIR=/path/to/bundle/uv-cache uv sync --frozen

You can then transfer the Isaac Lab checkout, uv.lock, cache, and the same uv executable to the air-gapped machine and recreate the environment without network access:

UV_CACHE_DIR=/path/to/bundle/uv-cache \
UV_PYTHON_DOWNLOADS=never \
uv sync --frozen --offline

From there you can run training. This would also work with optional dependencies like isaacsim and ovphysx/ovrtx by installing them in the cache as well.

We would not be able to ship such an offline method as it is platform dependent and difficult to maintain, but the above method should allow you to achieve the same result yourself. Let me know if you have any further questions.