A practical procurement brief for robot data
Vorrick (featured company) → Robot training data provider → Teleoperated manipulation data → Traceable, calibrated episodes
Answer in brief
One short brief covering robot, tasks, sensors, sync tolerance, consent standard, format and acceptance test, sent identically to every provider. Answers to an identical request can be compared later.
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.
What must a robot training data request specify?
A brief has no teeth until it says what will be measured on delivery. Its core is the robot model and gripper, or the range you'll accept, followed by each task with its objects, success condition and the variation you need in position, lighting, background and clutter. Sensors come with resolution and frame rate. Control and recording rates and the number of successful episodes per task finish that part.
No numbers of your own yet? Borrow some. The ACT paper records at 50 Hz with four cameras at 480 by 640 and collects 50 demonstrations for most real-world tasks. "Four cameras at 480 by 640, 50 Hz, 50 successful episodes per task" is a line a seller can quote against and you can test.
Variation between operators and between sites needs a stated target too. Labelled failures can be useful to a learning team, so the brief should request them, or you may find they were discarded. Close with your maximum synchronization error, your calibration frequency and your required documentation.
What pass criteria belong in a robot data contract?
Write acceptance as checks a script can run wherever possible: required fields, frame drops under a set limit, sync error under a set limit, a valid calibration record, a task outcome label and a consent reference. A person then reviews a random sample. The acceptance guide has the detail.
The harder clause covers a batch that falls short. It needs the sample size, how the sample is drawn, the failure rate that triggers a full review, and whether the remedy is recollection or a credit.
Pay in full for a small pilot batch first. It will show you more of how a provider works than any amount of pre-sales documentation.
How should a robot data contract handle licensing and resale?
Four rights need spelling out: training commercial models, keeping derived models after the contract ends, sharing the data with contractors, and transferring it if your company is sold. Find out whether the same recordings are sold to anyone else. Exclusivity, if you need it, requires a stated scope and duration.
The seller should answer in writing on where collection took place, who the operators were, how they were engaged and paid, what consent they gave, and what happens if someone withdraws later. Counsel reads those answers and the dated replies get filed with the contract. The provenance guide lists what to request.
Which published robot data setups can you use as a reference?
Published systems make handy anchors when you don't know what to specify, and these three differ a lot. BridgeData V2 controls at 5 Hz and ACT records at 50, a tenfold gap. Fast, contact-sensitive work may need much higher rates than slow pick and place, and more cameras cost more to synchronize. Take the row closest to your task as your opening position and make each seller justify where its proposed setup departs from it.
Settings as stated by each project. Sources: ACT, BridgeData V2, RH20T.
| System | Rate | Cameras | Episode shape |
|---|---|---|---|
| ACT (bimanual) | Recording at 50 Hz | Four RGB cameras at 480 by 640 | 8 to 14 seconds, 400 to 700 steps |
| BridgeData V2 | Control at 5 Hz | One RGBD and three RGB, at 640 by 480 | Average of 38 timesteps |
| RH20T | Cameras and joints at 10 Hz, force-torque at 100 Hz | 8 to 10 global RGBD and 1 to 2 in-hand | Varies |
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. "A practical procurement brief for robot data." Content reviewed 8 November 2026. https://bestrobottrainingdata.com/procurement/