Bigenlight/bowl_stack_triple_raw

Triple bowl stacking — raw recordings

Task: stack the orange bowls and place them into the purple bowl

Stack the two orange bowls together, then place them into the purple bowl. Collected on 2026-09-19 using UR7e, GELLO and a Robotiq gripper with two RGB cameras.

The collector supplied this photograph as the task overview. Original filename: 20260919_200321(1).jpg.

Contents

60 episodes; 36,030 cam1 master frames at 30 FPS; approximately 20 minutes. Raw cam2 has 36,027 frames and is aligned by timestamp in LeRobot. RGB: 1280×720; no depth. State/action: six UR joints plus gripper (7-D). Original recordings are retained without editing.

Collection tags

Folder suffixes are preserved verbatim in source_takes.json as folder_tag and folder_labels. recovery_label="recover" groups all recovery variants for compatibility with previous lighting releases. Counts: {"untagged": 46, "recover": 10, "recover_many": 3, "recover_twice": 1}.

These are collection annotations, not verified outcome labels. Normal completion requires multiple grasps, so multiple gripper closures alone do not imply recovery. No take is excluded because of its tag. No paper train/validation/test partition is assigned; LeRobot's default train range merely exposes the entire corpus.

Alignment and provenance

Uses the existing fixed-recorder convert_takes_to_lerobot.py recipe: cam1 capture stamps are the master clock; camera and joint streams use ROS stamp_s; command and gripper use callback t_rel_s. No timestamp-lag correction or frame dropping is applied. The older two-bowl lighting corpus is a distinct task and is not merged here.

Companion dataset: Bigenlight/bowl_stack_triple_lerobot_v3.