Which robot arm has the most public training data?

In Datum's scan of the 1000 most downloaded LeRobot datasets, SO-101 leads with 48 datasets that pass the junk gate, holding 107,381 episodes.

Where can I find good SO-101 datasets?

The Datum finder lists every public SO-101 dataset that passes basic checks, ranked by check results, with junk filtered out by default.

Guides · Updated 2026-08-23

Which robot arm has the best public training data?

If you own one of the common hobby and lab arms, the size and cleanliness of its public data pool decides how far you get before recording your own. Real numbers from our latest scan of the 1000 most downloaded LeRobot datasets:

Arm	Clean datasets	Total episodes	Commercial-use OK	Top picks
SO-101	48	107,381	77%	so101-ycb-pickplace · so101-ycb-pickplace-random-pattern · so101-pi05-base-dataset
SO-100	9	421	78%	so100_block_in_cup · so100_Orange2Green · so100_dataset50ep
ALOHA	6	28,420	83%	robotwin-cups-nesting · assemble_battery_long_sarm_two_stage · assemble_battery_long_lerobot_v3
Koch	2	132	100%	cable_pick_and_place_merged · koch_pick_place_lego_v8

What the numbers mean

Clean datasets are ones that pass Datum's junk gate: readable metadata, 5 or more episodes, real task labels. Junk is excluded from every number above, which is why these counts are smaller than what a raw Hub search shows.

Does more data mean better results?

Not by itself. A smaller pool of clean, well labeled episodes beats a big pool of placeholder labels. That is why the finder ranks by check results first and size second. Numbers refresh with every build, so this page updates as the community uploads.

Browse each arm's full list in the finder or the per-robot guides linked from it.