We fine-tuned the same policy twice on Gongsta/trlc_tshirt_folding, the DK1 recording the TRLC Data Explorer lists first, once on all 532 takes as released and once with the 21 that hold still for the largest share of their frames removed, and the edited run scored 5.0 points higher over 720 rollouts.
A same-size random cut scored 25.4%, so the gain is from which takes were removed, not from removing takes.
krishkrosh gains most, and the cut took 12 of his 89 takes, more than the other three teleoperators lost between them. kesu, who lost 2 of 97, comes out flat. The pooled lift clears zero, no single operator group of 180 rollouts does on its own, and the chart says so.
Measured on all 532 takes in the release. Each take carries one number, its held-still share, the fraction of its frames on which the twelve commanded arm joints move less than 0.01 radians in total from the frame before, and every take above twice the corpus median of 4.5% was dropped. 12 of the 21 removed takes come from one of the four teleoperators, krishkrosh, who recorded 89 of the 532, against 9 removed from the 443 the other three recorded. The 74 takes carrying the most gripper open and close commands all stay in, because a re-grasp is something a policy can learn from, and 53 of the 54 longest takes stay in, because slow is not wrong.
| Removed | takes | median held still |
|---|---|---|
| commanded gripper never opens or closes | 4 | 100% |
| arms frozen for 3 s or more at the start | 8 | 11.9% |
| dead air spread through the take | 9 | 9.6% |
| Removed | 21 | 10.2% |
| Kept | 511 | 4.5% |
Measured on the release. Six takes the cut removes and two it keeps, cut from the top camera at the per-take offsets carried in the release metadata. The number on each card is that take's own held-still share, the fraction of its frames on which the twelve commanded arm joints do not move, in a release whose median take runs 68.3 s.