object_retrieval-preset-gemini

Real-robot teleoperation episodes of the instance_retrieval task on a single-arm Franka Research 3 cell (80 episodes, 36,239 frames at 10 fps, released as Myungkyu/object_retrieval) with dense high-level labels produced by the TACOR offline annotator: Gemini 3.7 Flash reads each whole episode as one video clip (one sample every 10 frames = 1.0 s) and labels every sampled frame given only the subtask preset of the task - the label list below, without boundary criteria, sequence rule or tick rules. The companion dataset layout_reconstruction-preset-gemini covers the second task of the cell with the same recipe.

Task instruction: Collect every object that matches those shown inside the left drawer, place them into the right container, and close the lid. In words, the robot opens and closes the near compartment of the left drawer, then the far one, and for every compartment that held an object it picks the matching object from the center tray and places it into the right container, finally closing the lid. Objects: white holder, metal ring, black bracket, black round base.

Layout

instance_retrieval/                 LeRobot v2.1 dataset (data/, meta/, videos/), episodes 0..79 as in the source release
  videos/chunk-000/observation.image.exterior_camera/   exterior camera (448x252, 10 fps) - the planner's view
  videos/chunk-000/observation.image.wrist_camera/      wrist camera (448x252, 10 fps)
  videos/chunk-000/observation.image.keyframe/          the keyframe slot of the low-level policy; the preset variant uses no visual memory, so it is a copy of the exterior camera
  meta/dense_annotation_map.json                        per-episode subtask spans
retrieval_map.json            per episode (no retrievals in this variant)
annotations/instance_retrieval/ep%06d.jsonl   the raw per-tick labels (see below)

Per-frame features: observation.state (8: 7 joints + gripper), action (8: absolute joint targets + gripper, the lerobot_v21_abs_joint layout of the source release), subtask (the label of the frame), timestamp, frame_index, episode_index, index, task_index, the two camera views and the keyframe slot. The subtask label changes at the tick boundaries of the annotation (tick = 10 frames = 1.0 s). State and action follow the source release unchanged.

Subtask preset (9 labels)

• Open the near compartment of the left drawer.
• Close the near compartment of the left drawer.
• Open the far compartment of the left drawer.
• Close the far compartment of the left drawer.
• Pick up the white holder from the center tray and place it into the right container.
• Pick up the metal ring from the center tray and place it into the right container.
• Pick up the black bracket from the center tray and place it into the right container.
• Pick up the black round base from the center tray and place it into the right container.
• Close the right container lid.

Annotation rows

One JSON object per tick (decision_frame = 0, 10, 20, ...):

{"episode_index": 0, "decision_frame": 120, "subtask": "Open the near compartment of the left drawer.",
 "updated_memory": "...", "is_keyframe": false, "caption": "", "rag_query": ""}

subtask is one line of the preset. is_keyframe / caption / rag_query are always false / empty in this variant. updated_memory is the textual memory the planner carries. The prompts, the task contexts (round r1, built around the benchmark author's subtask presets) and the annotator are part of the TACOR code release.

Source

Episodes: Myungkyu/object_retrieval (DROID-style teleoperation captures, Meta Quest controller, cartesian-velocity control at 10 Hz, two ZED cameras), Apache-2.0. This dataset adds labels only.