lerobot/community_dataset_v3, measured from the public release on
16 September 2026.
Every row is read out of the release. Nothing here is estimated and nothing is rounded up.
| LeRobot datasets inside the release | 1,755 |
| contributors behind them | 488 |
| episodes in total | 72,418 |
| frames in total | 38,160,530 |
| robot time in total | 365.0 hours |
| episodes in the median dataset | 20 |
| datasets holding fewer than 5 episodes | 436, 24.8% |
| datasets holding exactly 1 episode | 101, 5.8% |
| datasets holding fewer than 50 episodes | 1,226, 69.9% |
| share of all episodes held by the largest tenth | 50.7% |
| share of all episodes held by the smallest quarter | 1.2% |
| datasets with a 6 dimensional action space | 1,576, 89.8% |
| datasets on an so100 or so101 arm | 1,546, 88.1% |
| datasets recorded at 30 frames a second | 1,545, 88.0% |
| datasets that name no action dimensions | 40, 2.3% |
| datasets with no camera at all | 0 |
Nothing here is a defect. A one episode upload is someone trying the format, which is what a community corpus is for. The point is what the distribution does to anyone who treats the 1,755 as 1,755.
Counted by dataset this looks like 1,755 datasets and 488 contributors. Counted by episode it is close to one arm family at one frame rate, and half of it sits in a tenth of the releases. Anyone sampling uniformly across datasets, rather than across episodes, will spend a quarter of their sampling budget on 1.2% of the data.
Read meta/info.json under any root for fps, robot_type and the action shape, and meta/episodes.jsonl for the episode count. No credentials, and no parquet has to be downloaded.
This page is what the release tells you on its own. A full grading removes the part that scores worst, retrains on what is left, and reports whether the smaller corpus scores higher, with rollout counts and intervals. The one result measured on physical hardware is Stera-10M, two arms on a four arm SO-ARM101 bench, 21.9% against 36.9% task success across 1,270 rollouts. The Stera-10M report.