Datasets:
Oracle merged — gripper updated
Contains 702 episodes / 72,123 frames at 15 FPS after removing 15 episodes.
Current episode provenance
All episode IDs are zero-based. These are the current IDs after filtering and reindexing.
| Source dataset | Retained episodes | Current episode IDs | Source episode IDs |
|---|---|---|---|
| bkuster/oracle_pickups | 69 | 0–68 | 0–4, 12–20, 24–32, 38–83 |
| bkuster/oracle_pickups_filtered | 42 | 69–110 | 0–41 |
| bkuster/oracle_pickups_v0 | 101 | 111–211 | 0–100 |
| bkuster/oracle_basic_time_wrench_done_20261027_144325 | 490 | 212–701 | 0–489 |
Every current episode has its exact original source repository, pinned source revision,
source episode ID and previous merged episode ID in episode_provenance.jsonl.
Source episode IDs refer to the immediate datasets listed above. Earlier cleanup/merge
records remain under source_metadata/ for tracing their own ancestors.
Removed episodes
Removed from input revision 194a3bdcfa0d69e9df3e40406801f56a8abf1f8d: [5, 6, 7, 8, 9, 10, 11, 21, 22, 23, 33, 34, 35, 36, 37].
Removed 3,038 frames. These old IDs must not be applied again to the reindexed dataset.
episode_removal_manifest.json contains the complete old-to-new mapping and source checksums.
Gripper labels
Opening reference: 60 mm, closed target: 0 mm. Open remains 0.
Converted partial labels preserve the old 40 mm-reference grasp gaps:
0.4 → 0.6, 0.6 → 0.7333333, 0.8 → 0.8666667.
Exact execution requires matching finer/continuous width control; a 0.2 quantization
step cannot represent all converted gaps. Historical images remain unchanged in meaning.
The separate _v1 dataset is not included or modified.
Structure and validation
LeRobot v3.0; 31-dimensional state, 11-dimensional action, 640×480 RGB video.
Kept action/state/timestamp sequences are unchanged; global, episode and task indices
and numeric statistics are rebuilt. Mixed video files are trimmed/re-encoded; files
containing only kept episodes are copied. See episode_removal_validation.json.
Root completion/start records use current episode IDs, but recording parameters are historical.
Prior README, manifests and validation reports are archived in source_metadata/pre_episode_removal/.
Dataset statistics cover all retained data; compute training-only normalization after splitting.
- Downloads last month
- 254