object_retrieval-taco-gemini

Real-robot teleoperation episodes of the Object Identification task (internal id instance_retrieval) 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 and visual memory 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 the full task context — the subtask catalogue with its DONE WHEN criteria, the sequence rule, the tick rules and the memory program. The labels were then reviewed and corrected by the benchmark author frame by frame; 2,084 of 3,657 annotated ticks carry a corrected label.

The Baseline companion object_retrieval-preset-gemini holds the same episodes labelled from the subtask preset alone, with no memory operations.

Task instruction: Collect every object that matches those shown inside the reference tray, place them into the collection box, 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 memory slot of the low-level policy (see below)
  meta/dense_annotation_map.json                        per-episode subtask spans
retrieval_map.json                per episode: the stored keyframe and the ticks that retrieved one
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.

The keyframe stream

Unlike the Baseline variant, whose keyframe slot is a copy of the exterior view, this dataset carries the memory slot the policy sees at training time. It is produced by replaying the annotation's memory operations through the same bank the online planner uses:

• a tick with is_keyframe stores its exterior frame under the tick's caption (consecutive identical captions are deduplicated),
• a tick with a rag_query retrieves the single stored frame whose caption best matches the query (cosine similarity of Qwen3-Embedding-0.6B embeddings, asymmetric query/document prompts),
• every frame governed by a retrieving tick shows the retrieved frame in the keyframe slot; all other frames mirror the exterior view.

In this task the planner stores what it saw inside each left drawer compartment and retrieves it while picking the matching object from the center tray. retrieval_map.json records, per episode, which frame was stored and which ticks retrieved it.

Annotation rows

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

{"episode_index": 0, "decision_frame": 120, "subtask": "Pick up the white holder from the center tray and place it into the right container.",
 "updated_memory": "near: white holder far: metal ring", "is_keyframe": false, "caption": "", "rag_query": "contents of the near compartment"}

updated_memory is the textual memory the planner carries. The prompts, the task context (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 and the memory slot only.