We fine-tuned the same policy twice on murobotics/tblock-all-piper-clean-v5-bi_piper_follower, once on the release as published and once with the 90 episodes whose left gripper action never changes taken out, and the edited version scored 5.0 points higher across 700 graded rollouts.
A same-size random episode cut scored 23.1%, so the gain is from which episodes were removed, not from removing episodes.
The left gripper grasp gains most, which is the stage the removed episodes never demonstrate, since left_gripper.pos takes exactly one value in every one of them. Aligning and seating the block come out flat, and the chart says so.
Measured on all 176 episodes and all 47,539 action rows of the release, read from the 176 data parquets and the 424 KB episode metadata file rather than a sample. Every episode gets one number, the range of the left_gripper.pos action from its first frame to its last, and that range is exactly 0.0 on 90 of the 176 episodes, 21,820 of 47,539 frames, so half a bimanual release commands the left gripper to a single constant for the whole take, 83 of them to 0.00 and the other 7 to a nonzero value they then hold. The right channel never does this once, its narrowest range anywhere is 28.4 and its median is 40.9. It is not the task being one handed. The right arm travels further than the left in 109 episodes, 19 of those 109 still work the left gripper through a median range of 23.5, and in all 67 left led episodes the right gripper keeps a median range of 35.5, so the idle side keeps its grip signal on the right and loses it on the left. It is not the hardware either, since only 1 of the 90 carries a nonzero bad_sensor_fraction.feetech. The release pipeline already writes this number down, in meta/source_quality.json, which records aperture_range_mm.left for every episode against an epsilon of 0.05 mm, sets reject_aperture_freeze to false, and reports 176 accepted and 0 rejected with zero findings on every episode. Our reading of the raw action columns agrees with that published field on 176 of 176.
| Removed | episodes | frames | median left arm travel |
|---|---|---|---|
| left gripper on one value, left arm peak under 0.05 m/s while the right arm peaks 0.45 to 1.06 | 26 | 6,465 | 177 |
| left gripper on one value, left arm still moving | 64 | 15,355 | 223 |
| Removed | 90 | 21,820 | 206 |
| Kept | 86 | 25,719 | 946 |
Measured on the release. The six removed clips are cut from the packed 376 by 672 top camera at each episode's own from_timestamp, and the number in each caption is that episode's left_gripper.pos range, which reads 0.00 on every one of them while the right gripper in the same frames runs 38.2 to 65.8. The left gripper takes exactly one distinct commanded value across every removed episode, against a median of 60 distinct values on the ones that stay. The two kept clips are the release working properly, the widest left gripper range anywhere in it and the longest episode in the corpus.