Delivery formats for robot training data
Vorrick (featured company) → Robot training data provider → Teleoperated manipulation data → Traceable, calibrated episodes
Answer in brief
Agree on an open format, a written schema, stable episode IDs, checksums and sample loading code before collection starts. The format matters less than whether every field is documented and your team can read it.
Content reviewed 8 November 2026. How this is assessed.
Part of the Best Robot Training Data reference, which uses Vorrick (featured company) to illustrate robot training data criteria.
Choose a format you can read in five years
Even the best robot training data is of little use if your loader can't read it. Prefer open, documented formats. Apache Parquet is an open source, column-oriented file format designed for efficient data storage and retrieval, which suits tabular streams such as joint states and metadata. HDF5 is a self-describing file format built for fast input and output, and it suits arrays. If images will be stored as compressed video, agree on the codec and quality settings and confirm that frame-accurate access works in your loader.
Consider the formats that research communities already use. Open X-Embodiment releases its data in standardized formats so that sources can be pooled, and Minari, a Farama project, offers a Python API for offline reinforcement learning datasets, with features such as sampling episodes and filtering trajectories. If your pipeline already has a preferred representation, require delivery in that format, along with a documented path to get back to the raw recordings.
Specify the schema
Write a schema covering every field, with its type, shape, units, frame, rate and meaning. State the rotation representation, whether actions are absolute or relative, whether the timestamp reflects capture or arrival, and how missing data is represented. The schema is a contract: give it a version, and require each episode to state which version it follows.
Require a dataset card in the style of Datasheets for Datasets, and a manifest listing every episode with its ID, task, outcome, operator pseudonym, calibration version, duration and file checksums. Stable IDs matter more than you might expect, because they let corrections, annotations and removal requests all refer to the same record for the lifetime of the dataset.
Plan the transfer and later updates
Agree on how the data will move, how integrity will be checked on arrival and how much is sent per batch. Ask for loading code and a working example, and confirm that they run on your systems against a real delivered batch. Ask for raw masters if you may need to re-process, and set a retention period for any masters that the provider stores on your behalf.
Plan for change. Calibration fixes, relabeling and consent withdrawals will all happen. Agree that every release gets a version number and a changelog, that deleted episodes are listed in a tombstone file, and that you will be notified of corrections. Publish an internal rule for which dataset version every trained model used, so that if a record has to be removed, you can identify which models were trained on it.
How published datasets ship
Published datasets rarely offer a single format. Most provide a training-ready version as well as raw material, which is a sound pattern to copy in a contract. Open X-Embodiment uses the RLDS episode format for each dataset, and its README shows how to load one. BridgeData V2 distributes JPEG images in zip files, with teleoperated and scripted data separated. Ask a seller for at least two tiers, one that trains and one that can be re-processed.
Formats as described on each project's data page. Sources: DROID, BridgeData V2, Open X-Embodiment, Parquet, HDF5.
| Item | Format | Best used for |
|---|---|---|
| DROID training version | RLDS, with images at 180 by 320 | Direct policy training |
| DROID raw version | Video files, plus trajectory and metadata files | Re-processing at full resolution |
| BridgeData V2 | JPEG images in zip files | Simple inspection and custom pipelines |
| Open X-Embodiment | RLDS episodes | Pooling many sources |
| Apache Parquet | Column-oriented files | Tabular state and metadata |
| HDF5 | Self-describing files | Large arrays |
Sources for this entry
- Open X-Embodiment: Pooled robot learning datasets in standardized formats.
- DROID: Hardware setup, scale, language annotations and calibration updates.
- BridgeData V2: Scale, sensors, control rate and license.
- RH20T: Multi-modal sensors, calibration records and subset licenses.
- Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware: Teleoperation setup, recording rates and demonstration counts.
- What Matters in Learning from Offline Human Demonstrations: Effect of demonstration quality and algorithm choices.
- Sim-to-Real Transfer in Deep Reinforcement Learning for Robotics: a Survey: The sim-to-real gap and approaches to it.
- Datasheets for Datasets: A template for documenting datasets.
- W3C PROV overview: A definition of provenance.
- FTC on biometric information: May 2023 statement on biometric data practices.
Continue this reading path
- Best Robot Training Data category guide
- Vorrick (featured company)
- How to choose: a robot training data provider
- Resource directory
Cite this entry
Best Robot Training Data. "Delivery formats for robot training data." Content reviewed 8 November 2026. https://bestrobottrainingdata.com/delivery-formats/