Most LeRobot datasets on the Hub don’t have a licence anyone actually chose. I scanned the 3,000 most-downloaded LeRobot datasets: 56.5% carry apache-2.0 on an untouched template card, which is just push_to_hub’s default.
305 of them can be matched to one of 65 public robot datasets whose terms I checked at the source (DROID, BridgeData, LIBERO, Open X-Embodiment components, …). 201 of those, about a third of all LeRobot downloads, declare a different licence than the original, mostly CC BY 4.0 data relabelled apache-2.0 or mit. Another 45 claim a licence for data whose original release never stated one (for example the ALOHA sim demos and Push-T).
A tag on a copy can’t grant rights its up doesn’t hold, so if you train commercially on Hub robot data, check the source’s terms, not the tag.
Try it on any dataset, in your browser: rdq robot dataset licence check - a Hugging Face Space by rs1g6r
It gives a one-line answer (OK to use / check first / stop) with the evidence and the original’s licence statement.
Open source: GitHub - s1g6r/rdq · GitHub The command-line tool also catches broken recordings before training (desynced action/state streams, frozen logs, glitches, idle episodes, duplicates, empty task instructions) and writes a machine-readable data passport. In a controlled benchmark, 30% broken recordings cut behaviour-cloning success from 99.7% to 15.7%, and filtering with rdq’s checks restored it.
Most matches are made from the repository name, and the tool says so when that’s the only evidence. Corrections and feedback are very welcome.