We fine-tuned the same policy twice on v30_apple_storage, once on the release as published and once with every frame recorded after the apple is already in the plate taken out, and the edited version scored 5.3 points higher across 600 graded rollouts.
A same-size random frame cut scored 20.6%, so the gain is from which frames were removed, not from removing frames.
The release stage gains most, and it is the stage the removed frames sit immediately after, so the policy stops spending capacity on what the arm does once the apple is already down. Carry is flat. Only release clears zero on its own at 150 rollouts, the pooled lift clears it, and the chart says so.
Measured on the whole release, all three episodes and all 899 frames, from 239 KB of metadata and the one 23.8 MB video rather than a sample. The gripper command sits at one of two values on 884 of the 899 frames, 130.0 degrees open and -5.0 closed, so the frame where the apple is let go over the plate is exact rather than a judgement, and it lands at 4.60 s, 4.13 s and 3.73 s inside three takes that each run a flat 10.0 seconds. 525 of the 899 frames, 58.4% of the release, are recorded after the task is already finished, 17.50 seconds out of 29.97, and every take loses more than half of itself at 53.8%, 58.7% and 62.7%. The arm is still retracting through part of that, so the cut drops 58.4% of the frames but only 40.6% of the commanded arm travel, 396 degrees of 974.
| Removed | frames | seconds | share of the release |
|---|---|---|---|
| the arm retracting once the apple is on the plate | 348 | 11.60 | 38.7% |
| the same retract with the head turned off the workspace | 41 | 1.37 | 4.6% |
| the arm parked, no commanded joint moving for a second or more | 136 | 4.53 | 15.1% |
| Removed | 525 | 17.50 | 58.4% |
| Kept | 374 | 12.47 | 41.6% |
Measured on the release. The first three clips each open on the exact frame the gripper reopens over the plate and run to the end of that take, so everything in them is what the cut removes, and the next three are the moments inside that tail worth naming on their own. The 189 frames where the commanded gripper sits 67 to 126 degrees away from the measured gripper are kept, not cut, because the apple is holding the jaws open and a policy can see it in the hand, and on the other eighteen channels the same measure never exceeds 15.1 degrees.