We fine-tuned the same policy twice on the ARX Lift-2 portion of InternData-A1, once on the release as published and once with 2,328 episodes removed, and the edited run scored 40.5% against 35.2%, a gain of 5.2 points over 1,200 rollouts.
A same-size random cut scored 35.4%, so the gain is from which episodes were removed, not from removing episodes.
Fold short shirts, fold towels and the cloth task carry the gain, which is where the cut lands. Pouring water, whose subset the cut never touches, moves the other way. The pooled lift clears zero, most single tasks at 120 rollouts do not, and the chart says so.
Measured on the release, every episode was reduced to one fingerprint, the mean, standard deviation, minimum and maximum of each of its commanded action channels, and compared byte for byte against every other episode in its own subset. Over 382 ARX Lift-2 subsets and 236,191 episodes, 2,328 episodes carry an action record that is bit identical to at least one other episode beside it. The 1,708 episodes of fold short pants, fold short shirts and fold towels hold only 42 distinct action records between them, one of them stamped on 50 episodes, and yet all 1,708 head camera streams have different image statistics, so the pictures move while the recorded motion repeats. Those three are also the only subsets in all 382 where every episode is exactly the same length, 1,200 or 1,900 frames, against a median of 152 distinct lengths in the 55 other subsets of comparable size, the tidiest of which still carries 79.
| Removed | episodes | distinct action records |
|---|---|---|
| fold short pants, every episode exactly 1,200 frames | 699 | 15 |
| fold short shirts, every episode exactly 1,200 frames | 559 | 14 |
| fold towels, every episode exactly 1,900 frames | 450 | 13 |
| track the targets, records written twice and three times over | 486 | 240 |
| nine more subsets, every repeated record written twice | 134 | 67 |
| Removed | 2,328 | 349 |
| Kept | 233,863 | 233,863 |
Measured on the release. Each clip is the window that episode occupies inside its packed head camera mp4, read from the from_timestamp column of meta/episodes. The number in every caption is how many episodes in that subset carry the same action record, every mean, standard deviation, minimum and maximum bit identical on every commanded channel.