pickplace_real100 — 100 of the 101 pickplace_real101 demonstrations

Derived locally on 2026-09-11 from LeeHakHo/pickplace_real101 (kept untouched at /home/hyeonhoo/real_data/pickplace_real101/). Everything about the recording, the layout of data/demo_<i>, the calibration and the two caveats (no simulator behind it; actions are reconstructed at the measured 14.4 Hz) is unchanged — see that dataset's README.

Two things differ.

demo_8 is dropped, leaving 100 demonstrations

Ranked by the fraction of valid entries in obs/aux_valid, demo_8 is the worst of the 101: the cylinder tracks perfectly (100%) but the cube is valid on 2.4% of its 536 frames. The next worst is demo_26 at 7.5%. Demo names are NOT renumbered, so demo_8 is simply absent and the remaining names still match source_traj in the original file.

mask/valid is the 10 worst-tracked demonstrations, not a random split

	demos	frames	cylinder valid	cube valid
mask/train	90	see file	high	high
mask/valid	10	see file	high	low (7.5–40.6% per demo)

mask/valid = demo_11, 21, 25, 26, 27, 33, 45, 64, 76, 78 — the 10 next-worst demonstrations after the dropped one. The ranking is the same whether demos are scored by the mean over both aux_valid columns, by the cube column alone, or by the worse of the two columns, so the split does not depend on that choice.

The point is to keep the cleanest object labels for training: the auxiliary head is supervised on obs/object, and these ten carry the least trustworthy version of it. Note the consequence for reading validation numbers — the validation policy loss is computed on demonstrations whose aux labels are poor, but the action labels (actions, reconstructed from the arm's own motion) are unaffected by tracking quality, so validation action loss stays comparable across arms.