We fine-tuned the same inverse kinematics network twice, once on IKNet as released and once on an edited IKNet with the 8.9% most duplicated seconds removed, and the edited one scored 5.2 points higher over 960 held out reach targets.
The eight regions are the measured workspace split three ways by end effector height, low under 50 mm, mid 50 to 200 mm, top over 200 mm, and three ways by reach radius, near under 150 mm, mid 150 to 250 mm, far over 250 mm, keeping the eight regions that hold more than 1,000 pairs each. A same-size random cut scored 22.9%, so the gain is from which seconds were removed, not from removing seconds.
Both low band regions gain, and that is where the cut landed, since 1,926 of the 3,744 removed pairs sit below 50 mm. The pooled lift clears zero, most single regions at 120 targets do not, and the chart says so.
Measured on the release. The two recordings, 419.4 seconds at 100 Hz, were cut into 420 one second windows, each scored by how many distinct 5 mm end effector cells its samples land in, and the 38 windows landing in 20 or fewer were dropped against a median window of 37. The last 10.4 seconds of the training recording, 1,040 of its 33,186 pairs, sit inside two cells, drift at most 4.45 mm and carry 28 distinct joint vectors between them, the arm resting on the table with the recorder still running. Slow is kept and only still is cut, so 9,100 kept pairs sit in a second carrying under 120 mm of end effector travel, every one of those seconds still landing in more than 20 cells. Removing all 3,744 flagged pairs costs 296 of the 14,655 distinct cells the corpus covers.
| Removed | pairs | cells per 100 | distinct joint vectors |
|---|---|---|---|
| mid recording dwell, 67 mm of travel inside a 42 mm box | 2,200 | 17 | 2,085 |
| end of a recording, arm set down | 1,344 | 1 | 89 |
| first second, before the hand reaches the arm | 200 | 2 | 27 |
| Removed | 3,744 | 11 | 2,201 |
| Kept | 38,200 | 38 | 38,053 |