Release UMI-10k dataset and additional RDT2 artifacts on Hugging Face #25
Description
Hi @LBG21 🤗
Niels here from the open-source team at Hugging Face. I discovered your work on Arxiv regarding RDT2 and the massive UMI dataset. It's an incredible contribution to the robotics community—the zero-shot generalization results and the dynamic tasks like table tennis and archery are really impressive!
I noticed that you've already started hosting several model checkpoints (RDT2-VQ, RDT2-FM) and an example dataset on the Hub under the robotics-diffusion-transformer organization. That's fantastic for accessibility and discoverability.
I'm reaching out to see if you plan to also host the full 10,000-hour UMI dataset on Hugging Face. Given its scale and importance as a potential foundation for the community, having it on the Hub would significantly improve its visibility. It would allow people to easily stream/explore it using the dataset viewer and the datasets library:
from datasets import load_dataset
dataset = load_dataset("robotics-diffusion-transformer/UMI-10k")If the dataset is structured in a way that suits it, we also support Webdataset natively.
Additionally, I noticed the action normalizer is currently on a custom server and the RDT2-UltraFast model is coming soon. Moving these to the Hugging Face Hub would centralize all artifacts in one place, making deployment even more seamless for users.
Once uploaded, we can also link all these artifacts to your paper page (https://huggingface.co/papers/2602.03310) so researchers can find them right away.
Let me know if you're interested or if you need any help with the upload process for these large-scale artifacts!
Kind regards,
Niels
ML Engineer @ Hugging Face 🤗