Updated 2026-08-23 · Rechecked with every build

Robot Training Data Statistics

Numbers from Datum's checks over the 1000 most downloaded LeRobot datasets on the Hugging Face Hub. Every stat stands alone. Cite freely with a link.

Stat 1Of the 1000 most downloaded LeRobot datasets, 42% fail at least one basic quality check.

Stat 2230 of 1000 datasets (23%) are junk: missing metadata, fewer than 5 episodes, or placeholder task labels.

Stat 3172 of 1000 datasets (17%) have no usable license, so nobody can safely build on them.

Stat 4147 of 1000 datasets have no readable meta/info.json file.

Stat 5582 of 1000 datasets (58%) pass every check Datum can run from metadata.

Stat 6The most common robot in the checked set is so_follower, with 237 datasets that pass the junk gate.

Stat 7662 of 1000 checked datasets use the new LeRobot v3.0 format.

Stat 8The checked datasets hold 2,477,509 episodes in total.

Stat 980 datasets declare no camera streams at all in their schema.

Stat 10809 of 1000 datasets (81%) carry a license that allows commercial use.