We fine-tuned the same policy twice on so101-pick-and-place, once on the release as published and once with every frame inside a still run of a second or more taken out, and the edited version scored 5.4 points higher across 720 graded rollouts.
The four stages are the graded stages of the one task string in this release, pick up a tape and put it on top of an orange pingpong ball. A same-size random frame cut scored 27.6%, so the gain is from which frames were removed, not from removing frames.
The three stages that sit next to removed time all gain. Carry is the one stage the cut never borders, because the still runs open and close a take rather than interrupt the middle of it, and carry is the one stage that comes out flat. Only reach clears zero on its own at 180 rollouts, and the pooled lift clears it.
Measured on the release, all 72 takes and all 52,796 frames, from the 1.3 MB of tabular metadata rather than a sample. A frame counts as still when the six commanded joint angles move under 0.5 degrees in total from the frame before, and 25,618 frames, 48.5% of this release, are still by that test, still 41.8% at a 0.1 degree threshold worth about one servo step. The cut keeps every take and removes only the frames inside a still run of a second or more, 20,712 of them across 148 runs, and 68.3% of those sit at the end of the take after the tape is already on the ball. The cause is in the metadata, 70 of the 72 takes are locked inside a 702 to 736 frame window with a standard deviation of 4.8 frames, so the recorder rolls a fixed 24.3 seconds whether or not the operator has finished, and the clock is clean, the largest deviation from a 1/30 second step anywhere in the release is 1.4 microseconds, so this is dead time and not dropped logging.
| Removed | frames | median run | share of the release |
|---|---|---|---|
| trailing run, the recorder still going after the last commanded motion | 14,145 | 7.0 s | 26.8% |
| leading run, before the arm first moves | 6,188 | 2.9 s | 11.7% |
| freeze inside the take | 379 | 1.3 s | 0.7% |
| Removed | 20,712 | 148 runs | 39.2% |
| Kept | 32,084 | 4,976 still frames | 60.8% |
Measured on the release. The five trailing clips each open a second and a half before the still run starts, so the arm is moving when the clip begins and stopped when it ends, and the sixth opens on the first frame of a take that does not move for 5.4 seconds. All 72 takes lose time to this cut, a median of 9.4 seconds out of 24.3, so no take is dropped whole. What stays is every frame of commanded motion plus the 4,976 still frames sitting in pauses shorter than a second, because a hesitation before a grasp is something a policy can learn from and an arm parked at its stop is not.