Release RAPTOR Foundation Policy on Hugging Face

[NielsRogge]

Hi @jonas-eschmann 🤗

I'm Niels and work as part of the open-source team at Hugging Face. I discovered your work through Hugging Face's daily papers as yours got featured: https://huggingface.co/papers/2509.11481.
The paper page lets people discuss about your paper and lets them find artifacts about it (your models for instance),
you can also claim the paper as yours which will show up on your public profile at HF, add Github and project page URLs.

Your work on RAPTOR, a foundation policy for quadrotor control, is incredibly impressive! It's fantastic to see the policy released via PyPI as a Python package and as C++ headers for deployment.

Would you like to host the RAPTOR foundation policy on https://huggingface.co/models?
Hosting on Hugging Face will give you more visibility and enable better discoverability. We can add tags in the model cards so that people find the models easier, link it to the paper page, and it would be a great resource for the robotics community.

If you're interested, here's a guide for uploading models: https://huggingface.co/docs/hub/models-uploading. For the Python package, you could potentially use the PyTorchModelHubMixin class which adds from_pretrained and push_to_hub to any custom nn.Module. This would allow users to easily pip install foundation-policy and then from_pretrained from the Hugging Face Hub. For the C++ header, people can also use hf_hub_download to easily fetch the policy files.

We encourage researchers to push each model checkpoint to a separate model repository, so that things like download stats also work. After uploading, we can also link the models to the paper page (read here) so people can discover your work.

You can also build a demo for your model on Spaces to showcase its capabilities. We can even provide you a ZeroGPU grant, which gives you A100 GPUs for free.

Let me know if you're interested or need any guidance!

Kind regards,

Niels

[jonas-eschmann]

Hey @NielsRogge!

Thank you for the interest and for offering guidance! I'm not sure how well of a fit it is in this case, though. The policy only has 2084 parameters and hence the checkpoint is ~8kb plus some test data so ~30kb in total. This is quite easy to ship via PyPI or git directly. As for HF spaces, I love them, but in our case the inference is so lightweight, it can just run inside the browser, client-side (e.g. https://raptor.rl.tools).

Using spaces (with a CPU backend) for training might be interesting, though. With a workflow like: user submits the dynamics parameters of his platform and gets a trained checkpoint out of it. But then again, the RAPTOR policy kind of eliminates the requirement for users to conduct system identification because it can just adapt zero-shot.

Thanks for reaching out again, I will definitely keep it in mind (has been there already) and come back if I stumble upon a better use case

Best regards!
Jonas