We fine-tuned the same policy twice on the PlayWorld human demonstration release, once on all 925 episodes as shipped and once with the 140 whose instruction never names the manipulation or whose end effector sits still for a fifth of the recording removed, and the edited run scored 25.3% against 20.5%, a gain of 4.8 points over 750 rollouts.
A same-size random cut scored 20.7%, so the gain is from which episodes were removed, not from removing episodes.
Move object into or out of container holds 726 of the 925 episodes and 35 of the removals, and it moves furthest. Open or close curtain lost one episode of forty two and does not move, and move lid on or off of container lost none and gives two rollouts back. The pooled lift clears zero, most single tasks at 150 rollouts do not, and the chart says so.
Measured on the full release, all 925 episodes, not a sample. Every episode ships one annotation JSON holding an instruction string, a success flag and the end effector pose at each stored video frame, and the success flag reads 1 on all 925, in all six version and split directories, so it cannot separate a good demonstration from a bad one and the filter has to come out of the trajectory instead. Reading the pose, the share of an episode's steps on which the end effector translates under a millimetre, turns under a thousandth of a radian and leaves the gripper where it was runs from 0 to 50 percent with a median of 9.8, and covers 7,256 of the 63,529 frame to frame steps in the release, 24.2 of the 214.8 minutes. The cut takes the 102 episodes whose instruction is a DROID collection prompt such as Do anything you like that takes multiple steps to complete rather than a description of the manipulation, and the 38 more that hold still on a fifth or more of their frames, 140 episodes and 35.8 minutes together. What stays is every episode where the arm fumbles, overshoots or re-grasps, because a mistake the next model can see is training signal and a still arm is not.
| Removed | episodes | share of the release |
|---|---|---|
| Instruction is a collection prompt, not a description of the manipulation | 94 | 10.2% |
| End effector still on a fifth or more of the episode's stored frames | 38 | 4.1% |
| Both faults in the same episode | 8 | 0.9% |
| Removed | 140 | 15.1% |
| Kept | 785 | 84.9% |
Measured on the release. Each clip is the exterior_1 camera of one episode, stored at 320 by 192 and 5 frames a second, so one video frame is one row of the pose stream and the two line up exactly. The number in every caption is the share of that episode's steps on which the end effector does not move, and where it matters the longest unbroken run of them.