ctr-scan-object-200ep-uniform

200 episodes, 56984 frames, 100 unique scene seeds, LeRobot v3,25FPS.

Episode order / 数据集备注

Original cohort first, then E742 cohort; the two seed sets are disjoint.

Seeds and timing

The complete seed list and per-seed episode/timing mapping are in seed_timing_map.json, episode_seed_timing.csv and merge_episode_map.json. retime.source_seed and retime.source_cohort are also stored on every numeric row. Positive offset_s delays the right arm; negative values delay the left arm. Nominal offsets are rounded to the nearest250Hz physical step. Workspace waits are separately recorded.

Each of100seeds has exactly two original timings. Each50-seed source cohort used u=slot/100 and u=(slot+50)/100. Thus the merged200episodes retain two copies of the0%–99% grid, not a newly sampled200-point grid. The per-seed map explicitly lists both u values, requested/quantized offsets and episode indices. For Scan Object, timing varies only grasp/lift; the fixed cooperative scan tail starts after both arms arrive. For shoes, shared workspace waits do not advance frozen controls and remain supervised.

Cameras, labels and provenance

All sources use New FOV centered_fovy90:320×240, vertical90degrees, calibrated centered optical-axis mount; head plus both wrist cameras,250Hz physics,25FPS recordings. Original raster rendering is retained. This is a physical-video merge, with no resimulation, cross-time image composition or retiming of controls. Source videos are copied without re-encoding, and video files remain separate across source boundaries.

The first cohort is the20260911 paired750 collection; the second is E742 paired750. Requested gripper targets are checked against frozen controls. Any stale first-cohort pre-command zero labels are corrected only to the verified initial-open command; all other action values and all observation values are preserved. First-cohort observations retain their original capture/accessor provenance; unavailable historical raw physical qpos is not reconstructed. E742 records measured qpos and normalized measured grippers. Per-source corrections and hashes are in meta/source_provenance.json. Uniform alone includes observation.arm_active_mask:0 only for the initial artificial start delay,1 for all later responsibilities including finished holds, scan barriers and workspace waits. Other methods have no IdleMask. Numeric statistics are exact global linear quantiles over all valid published rows once, without padding or mask exclusion; no mean of per-episode quantiles. Image quantiles are omitted.

The E742 Scan Uniform left-wrist video was restored from original HDF5 after a native exporter dropped one frame; the verified repaired parent is used. Physical success and collision checks follow source receipts (2mm penetration gate). Publication does not imply user audit approval or authorization for model training. No train/test split is added.

Full seed list: 5, 7, 9, 12, 21, 26, 30, 34, 46, 50, 58, 59, 71, 72, 75, 76, 78, 82, 83, 84, 86, 87, 90, 95, 97, 98, 107, 118, 119, 127, 128, 130, 134, 135, 138, 141, 142, 146, 150, 151, 152, 154, 155, 159, 160, 164, 170, 172, 174, 185, 187, 189, 190, 191, 199, 200, 201, 202, 205, 211, 214, 218, 224, 231, 235, 237, 244, 247, 249, 253, 255, 257, 258, 261, 262, 265, 267, 275, 283, 290, 294, 298, 301, 306, 311, 312, 313, 319, 323, 326, 333, 344, 347, 353, 355, 356, 358, 362, 363, 365

Timestamp repair in this merge

The repaired E742 left-wrist video contained all frames, but its historical metadata retained a one-frame-short episode. Three timestamp fields in merged episodes198–199 were rebuilt from exact cumulative frame counts; images and numeric actions are unchanged. See timestamp_repair.json.