Hugging Face released LanceDB integration for robot training on September 24, 2026, saying it enables direct access to datasets stored in object storage with global shuffling capabilities. It is the company's first major update to its LeRobot framework since its initial release in 2024.
Hugging Face reported a 1.7× faster data loading speed compared to streaming on DROID, measured across six public datasets. That compares with a 11.6× faster speed on other tested datasets.
LanceDB integration is built on Hugging Face Storage Buckets and targets robot learning by allowing teams to train directly from object storage without local copies. Availability begins with the release of the LeRobot framework, initially for robotics researchers and developers.
"LanceDB reads random frames straight from the Hugging Face Hub or any popular object store like Hugging Face Storage Buckets, fetching only the bytes each batch needs," said Caroline Pascal, Hugging Face blog author. The integration reduces data transfer overhead and enables efficient global shuffling across datasets.
The announcement follows Hugging Face's launch of Storage Buckets, which provide native object storage for large robotics datasets. Hugging Face said the integration streamlines data access and reduces the need for local data duplication.
Hugging Face did not say how the integration will scale with larger datasets, and raised the question of how metadata complexity affects performance. The company said it will continue to refine the integration for broader use cases.
Source: huggingface
