docs/data_format.md

# Data format and usage notes

[Back to the dataset card](../README.md)

## Directory structure

Each task/variation instance contains 30 stored demonstrations and five review videos:

```text
data/<suite>/task_<id>/<case>/
  episodes.hdf5
  metadata.json
  source_index.json
  review_videos/
    success_000.mp4
    ...
    success_004.mp4
viewer/
  episodes/<suite>/*.parquet
  review_videos/*.parquet
manifests/
scripts/inspect_episode.py
```

The case directory identifies a base task (`BASE`) or a perturbation variant. Coverage varies by task.

| Suite directory | Included base task IDs |
|---|---|
| `libero_spatial` | 1, 2, 3, 4, 5, 6, 7, 9 |
| `libero_object` | 0, 2, 3, 4, 5, 6, 7, 9 |
| `libero_goal` | 0, 1, 2, 4, 5, 7, 8, 9 |
| `libero_10` | 0, 1, 2, 3, 4, 6, 7, 9 |

Paths in original `source_index.json` files are relative to `data/`; a suffix after `::` identifies an HDF5 group. Video references are relative to the instance directory. Referenced simulator and BDDL assets are not bundled with this dataset.

## HDF5 schema

Each HDF5 file contains groups `data/demo_0` through `data/demo_29`. In the table below, `T` is the number of steps in a demonstration.

| Field under a demo group | Shape | Description |
|---|---|---|
| `actions` | `(T, 7)` | float32 relative control commands |
| `obs/agentview_image` | `(T, 128, 128, 3)` | uint8 RGB scene images |
| `obs/robot0_eye_in_hand_image` | `(T, 128, 128, 3)` | uint8 RGB wrist images |
| `obs/robot0_eef_pos` | `(T, 3)` | End-effector position |
| `obs/robot0_eef_quat_xyzw` | `(T, 4)` | End-effector quaternion, xyzw order |
| `obs/robot0_gripper_qpos` | `(T, 2)` | Gripper joint positions |
| `obs/robot0_joint_pos` | `(T, 7)` | Robot joint positions |
| `obs/robot0_joint_vel` | `(T, 7)` | Robot joint velocities |
| `rewards` | `(T,)` | Rewards |
| `dones` | `(T,)` | Episode termination flags |
| `server_step_seconds` | `(T,)` | Server timing, where present |

HDF5 attributes include environment and generation metadata. Demo attributes contain success and source metadata, including the language instruction. Consult [schema_example.json](../manifests/schema_example.json) and the original attributes for controller, control frequency, and gripper semantics before integrating a training loader.

The HDF5 observations and actions retain their original values. Training code should select observation/action fields explicitly and apply any normalization required by the policy.

To inspect a downloaded instance:

```bash
pip install h5py
python LIBERO-Pro2.0/scripts/inspect_episode.py \
  LIBERO-Pro2.0/data/libero_spatial/task_1/BASE/episodes.hdf5
```

## Episode index

The `episodes` Viewer configuration contains one row per stored demonstration. Each row includes the initial scene and wrist images, instruction, suite, task, variation, step count, success label, and HDF5 file/group. The four suite configurations are filtered views of these same records.

You can stream the index without downloading the full HDF5 data:

```python
from datasets import load_dataset

# Install the datasets package first: pip install datasets
episodes = load_dataset(
    "zhouxueyang/LIBERO-Pro2.0-Train", "episodes", split="train", streaming=True
)
row = next(iter(episodes))
print(row["instruction"], row["num_steps"])
print(row["hdf5_path"], row["hdf5_group"])
row["agentview_image"].save("example.png")
```

Camera previews are lossless PNG encodings of the first HDF5 frame. Full temporal observations and actions remain in HDF5.

The `review_videos` configuration contains five successful example clips per instance. It identifies each clip's instance, but does not provide a verified clip-to-demo mapping. These clips are examples, not evaluation results or videos for every stored demonstration.

## Trajectory provenance

The demonstrations were generated by a privileged-state oracle, rather than human teleoperation. The supplied metadata states that oracle privileged state is not recorded in the observations.

Of the 1,349 instances, 1,243 have 30 distinct source trajectories each. The remaining 106 replicate one successful source trajectory into 30 stored demonstrations. The source audit therefore reports **37,396 unique source trajectories** among **40,470 stored demonstrations**.

The Viewer index and [episodes.jsonl](../manifests/episodes.jsonl) expose `keep_without_declared_fanout`. Selecting `true` retains all ordinary records and one representative from each declared replicated instance:

```python
episodes = episodes.filter(lambda row: row["keep_without_declared_fanout"])
```

This retains 37,396 records and removes declared copies only; it is not a new global content-based deduplication. Original source fingerprints are preserved when available. See [fanout_instances.json](../manifests/fanout_instances.json) for the affected instances and [source_final_audit.json](../manifests/source_final_audit.json) for the source audit.

This is a training-only dataset. When constructing your own evaluation split, keep related and replicated trajectories together to avoid leakage.

## Statistics and integrity

[statistics.json](../manifests/statistics.json) contains the dataset counts. [release_validation.json](../manifests/release_validation.json) describes structural checks; these do not independently replay the simulator or establish task success.

For a complete local download, verify payload checksums with:

```bash
cd LIBERO-Pro2.0
sha256sum -c manifests/SHA256SUMS
```

Subset downloads will omit files listed in the full checksum manifest. Viewer tables have separate hashes in [viewer_files.jsonl](../manifests/viewer_files.jsonl).