GPU cluster-friendly version

[aleSuglia]

Hello @nicklashansen,

Many thanks for this release. This is a fantastic effort and something that will definitely pave the way towards a lot of research in reproducible world models for Embodied AI.

I understand that Docker is useful to facilitate reproduction of the experiments; however, this might be cumbersome for running experiments on research clusters that do not support Docker. Moreover, this could be a barrier to entry for those who are not familiar with Docker and want to debug easily. Do you think that switching to something like uv would make the installation easier? I have a few students who would be potentially interested in this effort, so let me know if this is something we can help with.

Thanks a lot,
Alessandro

[nicklashansen]

Hi @aleSuglia, thanks for your interest and for raising this issue! I am still working on expanding documentation and code features so please don't hesitate to let me know if you run into any issues :-)

Regarding installation specifically: I have updated the readme with a more detailed guide for installation. If your infrastructure does not support docker, it is also possible to install dependencies directly using conda (this is what the dockerfile does, for the most part); I have added instructions on how to do that. Would that address your / your students' concerns? If not, I will also be happy to work with you to figure out another installation option!

[aleSuglia]

Thanks a lot @nicklashansen. We will give it a spin and let you know asap!

[keaganchs]

Hey @nicklashansen, I'm also looking to use a conda/pip install for our cluster and figure I'd reuse this issue. I ran into an problem with installing gymnasium[box2d]. License classifiers were deprecated with pip 25.0 and setuptools 77.0.0, which means installing this package now requires the pip --use-pep517 flag. As this flag can't be passed to the environment.yaml directly, I solved this issue by:

1. Removing - gymnasium[box2d] and - swig=4.3.1 from environment.yaml
2. Adding - conda-forge::swig=4.3.1 under the dependencies: in environment.yaml (before the - pip: section). This is necessary as building box2d requires the swig executable, not the python package.
3. Using the following modified install script:

conda env create -f docker/environment.yaml
conda activate newt
pip install --no-cache-dir --use-pep517 'ale_py==0.10' 'gymnasium[box2d]'

I would open a pull request but as you can't accept these until the publication is accepted, I figured an issue would be easier. Tested on Ubuntu 24.04. Cheers!

[nicklashansen]

@keaganchs interesting, thanks for letting me know! The contribution guidelines were a leftover from our anonymized code repository, PRs are very welcome :-) But I'm also happy to commit these changes myself, thanks for providing detailed instructions on how to solve it!