How many open robot datasets can actually train a robot? A census of 400 lerobot datasets on hugging face

Abstract

Robots learn manipulation from recorded demonstrations, and the LeRobot project has made it easy for anyone to record such demonstrations and share them on Hugging Face. On 23 October 2026, 81,881 datasets outside the LeRobot team's own account carried the LeRobot tag. This study asks how many of them can realistically be used to train a real robot. A uniform random sample of 400 was drawn and checked against a threshold proposed by the author, at least 10 recorded episodes, combined with three further filters: the name does not mark a test or an evaluation, there is at least one camera, and the name does not mark a simulation. Of the 389 datasets whose description could be read, 57% had ten episodes or fewer and the median number of episodes was 10. After these filters and a reading of the dataset cards, 148 datasets, 38% of those that could be read (95% confidence interval 33% to 43%), remained; for 9 of them a possible sign of simulation could not be resolved. The share depends on the threshold: it was 42% at 5 episodes and 30% at 20. Three of the 148 were re-uploads of large academic datasets that together held 95% of all episodes in the usable group; among the other 145, the median dataset had 48 episodes, 61% were recorded on the low-cost SO-100 or SO-101 arm family, and the ten largest held 68% of the episodes. In this sample, most uploads are too small, are tests or policy evaluations, or are simulations, and the usable part is dominated by one type of arm.

Keywords: robot learning, datasets, LeRobot, Hugging Face, data quality

Список литературы

1. Open X-Embodiment Collaboration. Open X-Embodiment: Robotic learning datasets and RT-X models // Proceedings of the IEEE International Conference on Robotics and Automation (ICRA). 2024. P. 6892-6903. https://doi.org/10.1109/ICRA57147.2024.10611477
2. Hugging Face. LeRobot, GitHub repository. https://github.com/huggingface/lerobot (accessed October 2026).
3. Gebru T., Morgenstern J., Vecchione B., Vaughan J.W., Wallach H., Daumé III H., Crawford K. Datasheets for datasets // Communications of the ACM. 2021. Vol. 64, No. 12. P. 86-92. https://doi.org/10.1145/3458723
4. Khazatsky A., Pertsch K., Nair S., et al. DROID: A large-scale in-the-wild robot manipulation dataset // arXiv preprint. 2024. arXiv:2403.12945. https://arxiv.org/abs/2403.12945
5. Walke H., Black K., Lee A., et al. BridgeData V2: A dataset for robot learning at scale // arXiv preprint. 2023. arXiv:2308.12952. https://arxiv.org/abs/2308.12952
6. Mandlekar A., Xu D., Wong J., et al. What matters in learning from offline human demonstrations for robot manipulation // Conference on Robot Learning (CoRL). 2021. arXiv:2108.03298. https://arxiv.org/abs/2108.03298

Как цитировать

Azamat Sagadat How many open robot datasets can actually train a robot? A census of 400 lerobot datasets on hugging face // Central Asian Scientific Journal. – 2026. – No. 4(32), Vol. 2. – Pp. 34–38. – DOI: 10.5281/zenodo.23763950.