# Converted Robotics Dataset

This dataset stores authoritative timestamps and metadata in Parquet tables and
stores frame payloads in video containers. Video presentation timestamps are not
used for synchronization; use `tables/frames.parquet` and `video_frame_idx`.

Core loading rule:

```text
decoded video frame k == frames.parquet row where video_frame_idx == k
```

Useful files:

- `dataset_manifest.parquet`: converted sessions in this dataset root
- `sessions/<session_id>/metadata/streams.parquet`: stream IDs and modalities
- `sessions/<session_id>/metadata/videos.parquet`: stream-to-video mapping
- `sessions/<session_id>/tables/frames.parquet`: authoritative timestamps,
  frame indices, clock domains, and hashes
- `sessions/<session_id>/tables/robot_states.parquet`: robot states, if present
- `sessions/<session_id>/metadata/calibration.json`: calibration and transform
  conventions
- `sessions/<session_id>/metadata/validation_report.json`: validation status

Downstream reading requires Python, NumPy, PyArrow, and PyAV/FFmpeg-compatible
decoders. Camera SDKs are conversion-only dependencies.