We fine-tuned the same policy twice on robocasa-PnPCounterToSink-MG_100_F_472, once on all 572 episodes as released and once with the 316 in which the gripper never holds the object for as long as 1.60 s removed, and the edited run scored 40.1% against 34.7%, a gain of 5.4 points over 700 rollouts.
A same-size random cut scored 33.6%, so the gain is from which episodes were removed, not from removing episodes.
The boxed drink and milk instructions move most, and the cut emptied both hard, taking 10 of their 12 episodes and 9 of their 11. Garlic and mug sit flat, and they are the two the cut barely touched, losing 4 of 14 and 2 of 13. The pooled lift clears zero, most single instructions at 70 rollouts do not, and the chart says so.
Measured on the release. Every episode reduces to one number, the longest unbroken stretch in which the gripper is commanded shut while the two finger joints stay more than 5 mm apart, which is the arm holding an object rather than closing on air, and all 100 episodes that end before the horizon hold for at least 1.60 s. 316 of the 572 never reach that, and all 316 are episodes that run the full 700 frame horizon, 35.0 s, and stop only because the clock does. The tail is a spike rather than a slope, 202 of the 316 peak under a single second and 3 hold nothing at all, against a median hold of 5.43 s in what stays, and it is legible in one channel only, because action dims 7 to 10 are numerically zero across all 375,092 frames and dim 11 is pinned at -1.0, leaving 5 of the 12 action dimensions carrying nothing. What stays is every episode where the arm actually held the object, including the 156 that held it and still ran the clock out, because a robot that grips the jug and cannot land it in the sink is a mistake the next model can see.
| Removed | episodes | median longest grasp |
|---|---|---|
| Holds nothing at any point, and in all 3 the gripper is never commanded shut in 35.0 s | 3 | 0.00 s |
| Longest hold under a second, the fingers shut and whatever was there is gone before the next second | 199 | 0.70 s |
| Longest hold between 1.0 and 1.6 s, short of the weakest episode that finishes on its own | 114 | 1.20 s |
| Removed | 316 | 0.80 s |
| Kept | 256 | 5.43 s |
Measured on the release. Each clip is the left agent view of one episode, the camera that has a file for all 572 of them, cut from the stretch around the grasp attempt, and the number in every caption is that episode's longest unbroken grasp, the seconds the gripper spent shut with its fingers held more than 5 mm apart.