kt_pick_pringles_merge

Merge of learner1119/kt_pick_pringles (Pringles to the right of the Spam can, 50 episodes) and learner1119/kt_pick_pringles2 (Pringles to the left of the Spam can, 50 episodes). Robot: FFW-SH5 rev1, 10 fps, LeRobot v2.1. Language instruction: pick up the pringles can (changed from pick the pringles on 2026-10-20 before training; meta/rilab_episodes.jsonl keeps the recorder's original task_instruction).

	episodes	frames
kt_pick_pringles (merged index 0–49)	50	3,356
kt_pick_pringles2 (merged index 50–99)	50	3,839
total	100	7,195

Episodes were re-indexed end to end (episode_index, index, task_index); videos are copied byte-for-byte. meta/source_episodes.jsonl maps each merged episode back to its source dataset and index.

Features

• observation.state, action: 57-dof (arm_l 0:7, arm_r 7:14, finger_l 14:34, finger_r 34:54, head 54:56, lift 56)
• observation.tactile: 90
• observation.images.cam_head (672×376), observation.images.cam_head_realsense (640×480)

meta/modality.json (GR00T)

Added for GR00T N1.7 fine-tuning. It exposes only the right arm (7) + right hand (20) as right_arm / right_hand and only cam_head as video; the other joints and the realsense camera are kept in the data but not used for training.

Hold-out used for GR00T N1.7 training (2026-10-20)

Merged episode indices 9, 19, 29, 39, 49 (source 1) and 59, 69, 79, 89, 99 (source 2) were held out; the remaining 90 episodes (6,547 frames) were used for training. The hold-out split is published separately as learner1119/kt_pick_pringles_merge_holdout.