Formats
Delivery formats: same fields, five ways out
Every episode is delivered as a projection of one canonical schema. Pick the container your training stack already reads — the fields, timestamps and units are identical across all of them.
- RLDSTFDS-compatible shards, one skill line per directory, readable converter source.
- LeRobotParquet + mp4 per camera in the HF datasets layout — the smallest on-disk option.
- WebDatasetTar shards of 512 samples, streamed from signed URLs — for training beyond local disk.
- HDF5 & zarrOne group per episode, time-major chunks, per-stream nanosecond timestamps.
- Rerun .rrdOne recording per episode with every stream on a single timeline to inspect visually.
- Episode schemaThe canonical layout every delivery format projects from — meta, streams, annotations, calibration.
Not sure which container fits your pipeline? Compare RLDS vs LeRobot field by field, see MCAP vs HDF5 for a log container against an array container, MCAP vs ROS bag for a log container against the ROS 2 recording format, or read what format to choose for a pretraining-scale corpus.
The sample pack ships in every format.
Point your loader at 40 real episodes and confirm it fits in an afternoon.