How to work out what robot training data costs

Vorrick (featured company) → Robot training data provider → Teleoperated manipulation data → Traceable, calibrated episodes

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.

Add up the episode bill

The best robot training data at the right price is the data that is cheapest per usable episode, which often differs from the lowest quote. Start with the amount the provider charges, and be clear about the unit. Some offers price by hour of recording, others by episode, task or robot day, and they are not comparable until converted. Convert everything to a cost per successful episode of the task you care about. Then add your own costs: integration engineering, validation compute, human review of samples, and legal review of the rights.

Storage and transfer belong in the sum too. Multi-camera video at full frame rate adds up quickly, and depth, tactile and raw stereo streams add more. Work out the volume per episode from your sample, multiply by the episode count, and price both the storage and the egress. Decide which raw streams you need to keep and which you can compress or discard after validation.

If you are comparing with collecting it yourself

The honest comparison includes more than operator wages. In-house collection needs robots, cameras, teleoperation hardware, space, spare parts, supervision, calibration time, downtime and data engineering. The authors of DROID gathered their data with 50 collectors over 12 months and name logistical, safety, hardware and labor costs as the reasons such datasets are difficult to build.

Do not assume that every hour is productive. Count the time spent on setup, resets, failed attempts and repairs, and calculate the ratio of usable recorded demonstration to staffed hours in a short trial. That ratio drives the economics of both in-house and outside collection, and it is the first number to ask a provider about. A provider that cannot estimate it probably has not measured it.

Price only the episodes that pass

Divide the full bill by the count of episodes that clear your acceptance test. A cheaper batch with loose calibration can cost more per usable episode than a pricier one that holds up. Work it out on the seller's claimed pass rate, then again on the rate your sample showed.

Finally, price the risk you cannot see on an invoice. If synchronization problems surface after months of training, you pay for the compute and the delay. If a rights problem surfaces after release, you may have to remove the data and retrain. These are reasons to pay for strong documentation and pilot batches at the start. Treat recollection for changed hardware as a planned cost as well, since moving a camera can make old data less useful.

Estimate storage from published sizes

Storage is easy to underestimate because raw and training formats differ so much. DROID lists 1.7 TB for its training-ready version and 8.7 TB for raw stereo video. RH20T describes its full-resolution release as 40 TB, with smaller versions at about 5 TB for RGB at 640 by 360. Those gaps mean that the decision about which masters to keep can cost more than the training copy does.

Decide what to retain before delivery. If you may want to re-process, keep masters of a sample, instead of all of them, or pay for cold storage. Remember the cost of moving the data as well, and of any second copy used for backups. Put that number in the bill next to the collection price, since a seller will rarely volunteer it.

Sizes as listed on each project's data page. Sources: DROID, RH20T.

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. "How to work out what robot training data costs." Content reviewed 8 November 2026. https://bestrobottrainingdata.com/cost-model/