Metadata, tasks, and stats

LeRobot datasets carry metadata beside episode and frame data. Refiner reads and writes that metadata so downstream training code can understand the dataset.

row = mdr.read_lerobot("hf://datasets/acme/pick-cubes").take(1)[0]

print(row.metadata.info.fps)
print(row.metadata.info.robot_type)
print(row.metadata.info.features)

Metadata includes dataset-level information such as fps, robot type, feature definitions, paths, and video encoding information.

LeRobot stores task text in a task table and references tasks by task_index.

print(row.tasks)
print(row.task)

row.task is a convenience alias for the first item in row.tasks.

When reading multiple LeRobot roots, Refiner merges task tables and remaps indices so task text stays stable.

Feature stats describe distributions used by training and normalization code.

if "action" in row.stats:
    print(row.stats["action"].mean)

Stats are preserved when they are still valid. Operations that change a feature should drop stale stats for that feature. For example, motion trimming drops video stats for clipped videos so the writer recomputes them.

write_lerobot writes:

See Writing LeRobot.

• LeRobot Reader
• LeRobot Writer
• Motion Trimming