SciEdu_pauses_3exp_210epi_leafdir

leapshared/SciEdu_pauses_3exp_210epi with the 60 leaf episodes labelled card by card. Same frames, videos, states and actions; only the leaf episodes' task_index and the two task sentences they point at differ.

• Episodes 0-59 (leaf sorting). A person stands one leaf card at a time on the orange line; the left arm puts what goes left, the right arm what goes right. Each frame now says which side the card being moved goes to: task 0 Do the leaf sorting experiment. put the leaf on the left side, task 1 Do the leaf sorting experiment. put the leaf on the right side. In the source they carried one rule sentence per episode ("compound leaves go to the left side and simple leaves go to the right side" or the other way round), and a model trained on that put the cards on either side whatever the rule. The kind of leaf is now for a separate scene verifier; the policy hears only the side.
• Where a card's sentence starts. 50 frames before the arm that takes it starts to move (the end of its last rest 9 cm or more from where it closes on the card), never before the previous card was let go; the first card's at frame 0. The side is read from the grippers (a closure in the middle, an opening over that arm's side), which matched the cards counted on video in 60 of 60 episodes: 325 cards, 169 left and 156 right. Spec: config/labels/leaf_direction_episodes.yaml, script: scripts/dataset/leaf_direction_labels.py in the OpenArm SciEdu robot repository.
• Episodes 60-209 (Redox, electroscope). Unchanged, frame for frame; task indices 2-17 as in the source.
• Training list. train_episodes.json holds 198 episodes. Held out, as for leapshared/SciEdu_pauses_3exp_210epi_pi05: 6, 7, 46, 47 (leaf), 67, 87, 107, 127 (Redox), 147, 167, 187, 207 (electroscope).