Suggest to install conda via miniforge instead of miniconda?

[traversaro]

Currently, in lerobot instructions, the users are suggested to install conda via the miniconda installer. However, I am afraid that this may problematic for two main reasons:

• Legal challenges in using defaults conda channels: The miniconda installer install itself and by default create environments using the defaults conda channel, and according to the Anaconda Inc. their term of service is not free to use for all users (see https://www.anaconda.com/blog/is-conda-free and https://prefix.dev/blog/towards_a_vendor_lock_in_free_conda_experience). Indeed, several university and companies (see https://www.theregister.com/2024/08/08/anaconda_puts_the_squeeze_on/ and https://thehftguy.com/2025/04/07/anaconda-inc-has-entered-litigation-against-non-paying-user-of-anaconda-alibaba-intel-dell-airbus/) had legal problems as they used the defaults conda channel without being aware of the ToS. A simple alternative to avoid any kind of legal problem is to only use the conda-forge community channel, and so use an installer that only install packages from the conda-forge channel (including conda itself) like miniforge
• Technical challenges in mixing defaults and conda-forge conda channels: even if lerobot suggest to install conda via miniconda (and hence creating environments with packages from the defaults channels), then the instructions suggest to install packages from the conda-forge channel, see for example https://github.com/huggingface/lerobot/blame/882c80d446a63a44868c67ae535467af32ce0e80/docs/source/installation.mdx#L20 . As of 2025, mixing conda-forge and default packages is explicitly not supported, see conda-forge docs in https://conda-forge.org/docs/user/transitioning_from_defaults/ , so mixing conda-forge and defaults can lead to crash in the future (if this never happened, I guess is just due to lucky combination of packages used that did not result in any kind of ABI incompatibility, but this may change in the future).

For these two reasons, I think it could make sense to just suggest to use miniforge instead of miniconda in lerobot documentation.

[traversaro]

To be a bit more concrete, the suggested modification to fix this issue are proposed in #1821 .

[imstevenpmwork]

Thanks for the detailed explanation. I will take a look at this after the next release, thanks !

[traversaro]

Thanks @imstevenpmwork !

• As of 2025, mixing conda-forge and default packages is explicitly not supported, see conda-forge docs in https://conda-forge.org/docs/user/transitioning_from_defaults/ , so mixing conda-forge and defaults can lead to crash in the future (if this never happened, I guess is just due to lucky combination of packages used that did not result in any kind of ABI incompatibility, but this may change in the future).

To add more concrete details to this, there have been at least 3 crashes on the torchcodec side due to mixing defaults and conda-forge:

• C++ tests on FFmpeg 6 are broken meta-pytorch/torchcodec#518, fix cpp_tests: Switch to only use conda-forge and add back ffmpeg 6 meta-pytorch/torchcodec#658
• FFmpeg 4 is failing to install on CI meta-pytorch/torchcodec#974, fix Do not mix defaults and conda-forge in linux_wheel job meta-pytorch/torchcodec#975
• Support FFmpeg8 on Windows meta-pytorch/torchcodec#951, fix Enable tests with ffmpeg8 on Windows meta-pytorch/torchcodec#973

[imstevenpmwork]

Concerning, I agree this needs to be solved. Thanks for the heads up !