| # ShakeBottle selected 16,000 trajectories |
|
|
| This release contains exactly 16,000 selected simulated ShakeBottle trajectories, |
| 8,000 using the left arm and 8,000 using the right arm, collected with the Piper |
| user real-robot URDF (piper_ros@ac41fcbc). These are simulated trajectories; |
| the parent directory name `real_data` does not mean real-world recordings. |
|
|
| Each shard contains byte-identical original HDF5 episodes under |
| `shard_*/data/episode*.hdf5`, their original configurations under |
| `shard_*/configs/episode*.json`, and episode metadata under `shard_*/manifests/`. |
| Three original RGB views, measured joint states, absolute joint control targets, |
| the raw 250 Hz stream, timestamps, calibration, and scene metadata are preserved. |
| Physics runs at 250 Hz with save_freq=15 (nominal observations at 16.6667 Hz). |
| No image conversion, resampling, normalization, or action realignment is applied. |
| The 5,256 unselected successful candidates and the processed NumPy training cache |
| are excluded. This matches the selected-trajectory publication used for ClickBell. |
| |
| During publication, UPLOAD_STATUS.json says `uploading`. Wait until it says |
| `complete`, then check SHA256SUMS before extracting all shards into one directory: |
|
|
| ```python |
| from pathlib import Path |
| import tarfile |
| from huggingface_hub import snapshot_download |
| |
| root = Path(snapshot_download("INTACT-JEPA/INTACT", |
| allow_patterns="real_data/shakebottle/*")) / "real_data/shakebottle" |
| for archive in sorted(root.glob("shakebottle-*-of-*.tar.gz")): |
| with tarfile.open(archive, "r:gz") as f: |
| f.extractall("shakebottle_episodes", filter="data") |
| ``` |
|
|
| The final selection_manifest.jsonl maps all 16,000 selections to their published |
| archive, HDF5 member, byte count, and SHA-256, as well as the paired config hash. |
| source_selection_manifest.jsonl and source_summary.json preserve the original |
| selection. DATASET_INFO.json describes the release. SHA256SUMS covers all other |
| published files. Source completion markers are prefixed `source_` and describe |
| the source dataset, not the upload progress. Downloaded HDF5 SHA-256 hashes can |
| be checked against selection_manifest.jsonl to verify exact byte preservation. |
| |
