Lance now natively supported in Hugging Face LeRobot, enabling direct training from object storage

lhoestq · x · 2026-09-15

Hugging Face's LeRobot has merged PR #4363 adding native support for robotics datasets stored as Lance tables. With LanceDB's map-style reader, training can stream directly from S3, GCS, or HF Storage Buckets at near local-disk speed while keeping a true global shuffle — no need to download the full robotics dataset first, unlike the previous IterableDataset approach limited to windowed reservoir shuffle. LeRobotDataset remains the only public dataset class.

Original post →

More from Embodied

Embodied channel →