We fine-tuned the same policy twice on the InstructMove release, once on all 500 episodes and once with the 300 dual arm episodes removed, and the edited run scored 24.3% against 19.7%, a gain of 4.6 points over 960 rollouts.
A same-size random cut scored 19.9%, so the gain is from which episodes were removed, not from removing episodes.
The two biggest gains are both single object pick instructions on the dual arm robot, and one place instruction and one spatial pick instruction come out flat or below. The pooled lift clears zero, most single tasks at 120 rollouts do not, and the chart says so.
Measured on the release, every episode was scored on how many of its own action dimensions hold at most two distinct values from its first frame to its last. That count is 7 of 14 in all 300 PiperX dual arm episodes, one entire arm, and 0 of 8 in all 200 Franka episodes. The still arm's total variation runs between 0.03 and 1.15% of the moving arm's, a swing of at most 0.014 rad over a whole episode, and which arm is still flips from episode to episode, 160 left and 140 right, so the constant half of the label can only be told apart from the video. The cut removes every episode carrying a dimension that never moves, which is the three PiperX subsets, 300 episodes and 44,562 of 78,980 frames.
| Removed | episodes | still arm motion |
|---|---|---|
| pick by category, PiperX, one arm still from first frame to last | 100 | 0.13% |
| pick by attribute, PiperX, one arm still from first frame to last | 100 | 0.13% |
| place A to B, PiperX, one arm still from first frame to last | 100 | 0.04% |
| Removed | 300 | 44,562 frames |
| Kept | 200 | 34,418 frames |
Measured on the release. Each clip is one episode read straight out of its subset zip by byte range, the dual arm ones from the static camera that sees both arms and the kept ones from the scene camera. The last two are kept on purpose, a scripted fixed duration trajectory is still a trajectory a policy can follow, while an arm that never leaves its start pose teaches the policy to hold seven of its fourteen outputs at a constant.