Guides · Updated 2026-10-15

Robot data, explained with real numbers

Plain-English guides built on Datum's own checks. Every number refreshes with every build.

How to tell if a robot dataset is goodSix checks that catch most bad robot training data, from a scan of the 1000 most downloaded LeRobot datasets.
Which robot arm has the best public training data?SO-101, SO-100, ALOHA, and Koch compared by real numbers: clean datasets, total episodes, and commercial-use share.
LeRobot v2.1 vs v3.0: what changed and what breaksThe v3.0 format rewrite explained in plain English, with real adoption numbers from 1000 checked datasets.
Placeholder task labels are silently poisoning robot trainingWhy labels like "task desc" hurt language conditioned models, and the ten minute fix.
Robot dataset licenses, explained in plain EnglishWhat you can actually build on: the real license breakdown of the 1000 most downloaded LeRobot datasets.
How much data do you need to train ACT or SmolVLA?Honest reference points from real projects instead of fake thresholds.
How to clean robot teleop data before trainingA practical cleanup pass mapped to the six checks that catch most problems.
Who checks the quality of robot datasets?What Datum is, how the checks work, and why the referee cannot also be a player.