CoolFace
Datasetpublic

huzheyuan/yam3lego_whole_round1_baseline

3lego WHOLE-EPISODE reconversion — huzheyuan/yam3lego_whole_round1_baseline Dual-YAM 3lego block-sorting, reconverted from the raw session recordings (huzheyuan/3lego_round1_baseline_raw) with one LeRobot episode per raw session — the multi-cycle recording under ONE fixed block→tray permutation. The earlier 3lego_* LeRobot repos are chopped per subtask (one episode per place/remove primitive), which makes every episode start mid-session; this family restores the whole sessions… See the full description on the dataset page: https://huggingface.co/datasets/huzheyuan/yam3lego_whole_round1_baseline.

sourceHugging Faceapache-2.0updated 17d agoView on Hugging Face
0likes575downloads
Dataset Card

3lego WHOLE-EPISODE reconversion — huzheyuan/yam3lego_whole_round1_baseline

Dual-YAM 3lego block-sorting, reconverted from the raw session recordings (huzheyuan/3lego_round1_baseline_raw) with one LeRobot episode per raw session — the multi-cycle recording under ONE fixed block→tray permutation. The earlier 3lego_* LeRobot repos are chopped per subtask (one episode per place/remove primitive), which makes every episode start mid-session; this family restores the whole sessions for in-context-learning use, where the permutation is genuinely latent at episode start.

  • —103 episodes / 865,497 frames @ 60 Hz (4.01 arm-pair-hours), 19 arrangement(s)
  • —Cameras: observation.images.top / left_wrist / right_wrist, 256x256 bicubic squash, SVT-AV1 crf 30 / preset 12 / keyframe every 2 frames
  • —observation.state / action (20): 2 arms x [pos(3), 6D rot(6), grip(1)]; cartesian SYNTHESIZED by FK at the wrist flange. joint_state / joint_action (14) = [lj0..5, lgrip, rj0..5, rgrip], radians; grippers fraction-open (1 = open), clipped. state = MEASURED, action = COMMANDED (leader-follower teleop).
  • —Alignment: UNSHIFTED (state[i] with action[i]) — bit-identical to the vetted chopped corpus (verified: replaying the chopper over these raws reproduces all 1,563 chopped train episodes with exact state/action equality). A per-episode derived_shift field records the offset the current timestamp-alignment rule would apply (metadata only).
  • —Task string (LeRobot tasks) = the whole-session permutation prompt, e.g. "Take apart the lego structure and sort the blocks: put the red block in the red tray, ...". Per-segment subtask prompts live in meta/episodes.jsonl -> subtasks: [{text, start, end}] (foot-pedal boundaries from the raw language_annotations.json; [problematic] spans are kept as frames but flagged). Extra per-episode fields: task_id, bin (tray order left->right), block (structure order), source (raw session folder), chopper_skip, label_override.
  • —One known label correction is baked in: raw session 3lego_task11/3lego_task11_20260521_194258 (round1baseline) was recorded under task 11 but is actually task 35 (bin BGR) — task string + 3 place subtask texts corrected, originals preserved under `textoriginal / label_override`.
  • —Container aspect flag: sample_aspect_ratio 1:1 since 2026-09-09 (remuxed in place, pixels unchanged). The 2026-08-27 upload carried SAR 16:9 inherited from the 1920x1080 source, which made previews and browsers display 455x256 although the frames are 256x256.

Family: yam3lego_whole_round1/2/3/4, yam3lego_whole_round1_baseline (train), yam3lego_whole_val_adv, yam3lego_whole_val_base (val; no subtask annotations). Conversion provenance: meta/conversion.json.