Lightwheel, LightwheelAI/lightwheel_tasks. Measured for this page from the public release.
21,500 episodes across 219 task names and five embodiment subsets. All 21,500 ship a video. 16,800 of them ship it as isaac_replay_state.mp4 and the remaining 4,700 as isaac_replay_state_product.mp4, and no episode carries both. A loader that globs the second name sees 21.9% of the corpus and reports no error.
| embodiment subset | episodes | _product.mp4 | share |
|---|---|---|---|
| lightwheel_libero_tasks_piper | 6,500 | 0 | 0.0% |
| lightwheel_libero_tasks_x7s | 5,850 | 0 | 0.0% |
| lightwheel_robocasa_tasks_x7s | 4,450 | 0 | 0.0% |
| lightwheel_libero_tasks_g1_controller | 3,100 | 3,100 | 100.0% |
| lightwheel_robocasa_tasks_g1_wbc | 1,600 | 1,600 | 100.0% |
| all | 21,500 | 4,700 | 21.9% |
The split is clean. 336 of the 430 task and embodiment groups have the plain name on every episode, 94 have the _product name on every episode, and not one group mixes the two.
Task sizes come in three fixed counts. 87 tasks ship exactly 100 episodes, 70 ship exactly 50, and 62 ship exactly 150. So the corpus is generated rather than collected, and the top 20% of tasks hold 30% of the episodes.
curl -s https://huggingface.co/api/datasets/LightwheelAI/lightwheel_tasks and count the rfilename entries ending in each of the two video names. The manifest is public and the counts are exact.
EgoDemo, EgoPro and EgoStandard are gated behind manual approval and were not read. Nothing on this page is about them.
Every report on the results page takes a public robot dataset, measures one property of it across every episode, removes the part that scores worst, and retrains on what is left. The corpus figures, the defect rates and the cuts are computed from the releases themselves. The lift column on that page is read from each report's fine tune table rather than from a training run, and the page says so.
The one result that comes off physical hardware is Stera-10M. Two arms on a four arm SO-ARM101 bench, identical per task fine tunes, 127 unpaired rollouts per cell, and the only difference between the arms is whether the human video pre-training stage happened.
This page is the part that can be computed from a published manifest and the labels a release ships. The full grading retrains on the corpus with the worst scoring part removed and reports whether the smaller version scores higher, with the rollout counts and the intervals. That needs the data itself and a conversation about which property is worth measuring on your corpus rather than ours.