We fine-tuned the same policy twice on FreeTacMan, once on all 6,428 episodes as released and once with the 799 that write position, gripper and orientation on a different scale from the rest removed together with one episode under a second long, and the edited run scored 29.6% against 24.7%, a gain of 4.9 points over 1,100 rollouts.
A same-size random cut scored 24.2%, so the gain is from which episodes were removed, not from removing episodes.
UsbPlug and WipeBoard move most at 9.1 and 8.2 points, and neither lost a single episode, so what they gain is a shared action head that is no longer being pulled toward a scale three orders of magnitude smaller. TextureClassify and Write gain 6.4 each from a low base even though the cut took every episode they had, because the full run learned a command size for them that the gripper reads in millimetres. Stamp is flat and ArrangeFruit moves 0.9. The pooled lift clears zero, most single tasks at 110 rollouts do not, and the chart says so.
Measured on the release. Every one of the 6,428 episodes reduces to a single number, the largest absolute TCP position value its own trajectory file writes. Across 5,629 episodes that number runs from 197.92 to 542.91. Across the other 799 it runs from 0.271 to 0.594, and nothing at all lands in between, so 799 episodes, 12.4 percent of the release, record position and gripper in metres and orientation in radians while every other episode records millimetres and degrees. The other channels split the same way, gripper distance peaks between 0.026 and 0.333 on those 799 against 22.43 to 125.84 everywhere else, and the euler columns peak between 3.025 and 3.142 against 144.15 to 1,383.68, while the quaternion columns carry unit norm on both sides, so the rotation itself is stored correctly and only the euler copy disagrees. Column names, folder layout and the dataset card are identical for the two, and the clearest visible marker is a capital letter, since those 799 episodes name their video files _Camera and the other 5,629 name them _camera, 1,598 files against 16,886. They also ship two camera views where all but one of the other 5,629 ship three, and the view they drop is one of the two tactile fingertips on a release whose subject is contact-rich visuo-tactile manipulation. Every one of them was recorded on 6 or 15 April 2025 UTC, and every episode that stays was recorded between 22 August and 8 September 2025. One further episode goes, ScrewInTheBulb_70, whose trajectory covers 0.05 seconds in 1,033 bytes against a 10.14 second and 54,064 byte median and whose three videos are four frames each. What stays includes the tasks that look flat and are not, since WipeBoard_46 opens its gripper by 3.45 mm across the whole take while the arm travels 1,082 mm, and holding an eraser at a fixed opening is the task rather than a fault. It also includes every episode whose clock repeats, 1,171 of the 5,628 kept episodes carry at least one non-increasing timestamp and CutTofu_6 runs 89 of its 390 rows on a frozen clock, because the pose keeps moving through them and the trajectory is recoverable.
| Removed | episodes | share of the release |
|---|---|---|
| TextureClassify, metres and radians, peak position 0.538 to 0.594 | 200 | 3.1% |
| FragileCup, metres and radians, peak position 0.348 to 0.430 | 200 | 3.1% |
| Stamp, metres and radians, peak position 0.271 to 0.446 | 200 | 3.1% |
| Write, metres and radians, peak position 0.422 to 0.555 | 199 | 3.1% |
| ScrewInTheBulb_70, 0.05 s of trajectory and four frames of video | 1 | 0.02% |
| Removed | 800 | 12.4% |
| Kept | 5,628 | 87.6% |
Measured on the release. Each clip is the wrist camera of one episode, the view that shows both tactile fingertips. The number in every caption is that episode's peak absolute TCP position value as written in its own trajectory file, against 214.04 to 542.91 across the 5,628 episodes that stay.