metadata
license: apache-2.0
task_categories:
- robotics
tags:
- LeRobot
- lance
- lancedb
berkeley_autolab_ur5, stored in Lance
lerobot/berkeley_autolab_ur5 (1000 episodes, 97,939 frames, 3 cameras) converted to the Lance layout that LeRobot's LanceDBDataset reads natively. Same meta/ sidecar as the original, tabular features in frames.lance, encoded videos in a videos.lance blob v2 column.
from lerobot.datasets import LanceDBDataset
ds = LanceDBDataset("lance-format/berkeley_autolab_ur5-lance") # streams tables from the Hub, only meta/ is downloaded
item = ds[0]
Or just point training at it. The dataset factory detects the format automatically:
lerobot-train --dataset.repo_id=lance-format/berkeley_autolab_ur5-lance ...
Requires lancedb>=0.35.0b2 (pip install --pre --extra-index-url https://pypi.fury.io/lancedb/ lancedb) and the lance-dataloader branch of LeRobot until the upstream PR lands.