How to choose a robotics data collection provider

Editorial guidance by Robotics Data Collection · Research method

Choose between custom capture, existing datasets and marketplaces, then define a pilot that tests your actual training pipeline.

A useful shortlist starts with what the robot needs to learn. “Robotics data” can mean unlabelled human video, synchronized robot observations and actions, or annotations added to recordings you already own. Those are different purchases.

1. Choose the supply model first

Custom collection lets you specify tasks, equipment and collection rules. It also puts more responsibility on your team to define resets, scene variation and accepted episodes. Use it when your embodiment or operating environment is central to the learning problem.

Existing datasets can be evaluated before you commission a full campaign. Ask for an inventory that separates raw video, annotations, robot state and actions. A dataset about the same task may still have the wrong camera placement or control representation.

Marketplaces introduce another party between you and the original collector. Establish who is responsible for the license, factual listing accuracy and a delivery that fails your agreed checks.

The directory separates these supply models. Its labels describe published offerings rather than an assurance of suitability.

2. Write an accepted-episode specification

Describe the task as an observable start state, action and end state. For a drawer-opening task, specify drawer types, handles, starting positions, objects inside and what counts as completion. State whether failed attempts and recoveries should be retained.

Define the minimum streams your loader needs. Record camera placement, resolution, frame rate, timestamps, coordinate frames, calibration and the meaning of each action dimension. “Multimodal” is too broad to be an acceptance criterion.

3. Test the handoff before buying volume

Ask for a small, representative pilot rather than only a curated highlight clip. Load it with the same code you expect to use for the final delivery. Check manifest counts against files, decode the recordings and inspect time alignment at task transitions.

Give each issue a clear outcome: accepted, reworked or rejected. Decide who pays for unusable episodes and what happens when a task proves harder than the initial estimate. Compare cost per accepted unit, with a consistent definition of “hour” or “episode”.

4. Check rights and collection conditions

Data that is technically useful may have restrictions on training, model distribution or reuse across products. Use the licensing checklist to prepare questions for the provider and your legal reviewer. This guide does not establish anyone’s legal rights.

5. Make the inquiry specific

Include task, environment, embodiment, required streams, volume and unit, delivery format and deadline. Describe an acceptance test. Leave credentials, subject identities and confidential recordings out of the first inquiry.

Build a collection brief to put these requirements in one document. It is generated locally for you to review and share yourself.

Profiles to use in your research

Compare Deepen AI’s operator-led collection, Robotics Center’s custom campaigns and Kinetic Blocks’ marketplace model. These are different sourcing approaches, not a ranking. Read the profiles and confirm the proposed scope directly.