Why Format Choice Matters
The dataset format determines which models you can train, how easily you can share data with collaborators, and how much engineering overhead you face in your pipeline. Choosing the wrong format means painful conversion later — or worse, losing metadata in translation.
Format Comparison
RLDS (Google): TFRecord-based, used by Open X-Embodiment and RT-X. Excellent for large-scale cross-embodiment training. LeRobot (Hugging Face): Parquet-based, easy to push to Hugging Face Hub, growing community adoption. HDF5: self-describing binary format, used by RoboMimic and many older projects. Zarr: chunked, cloud-friendly, gaining traction for large datasets.
- For cross-embodiment training: RLDS
- For community sharing and quick iteration: LeRobot
- For legacy compatibility: HDF5
- For cloud-native pipelines: Zarr
SVRC's Data Pipeline
SVRC's data collection infrastructure outputs both RLDS and LeRobot formats natively. Our Data Platform handles format conversion, versioning, and quality validation automatically.