Curatrix
Blog

Five metrics for judging a robotics dataset before you buy it

Hours of video is the least informative number on a datasheet. Contributor diversity, environment spread, episode completeness, annotation density and rights coverage tell you more.

1 min readBy Curatrix Research

Dataset marketing leads with hours. Hours are easy to count and easy to inflate: one contributor repeating one task in one room for a week produces a large number and a narrow dataset. Before committing budget to a release, look at five other numbers.

1. Contributor diversity

How many distinct people performed the task? Skilled workers differ in handedness, pace, grip and error-recovery strategy. A policy that has seen twenty electricians generalises to the twenty-first far better than one that has memorised a single person’s habits.

2. Environment spread

Workshop, construction site, residential, industrial: each has its own lighting, clutter and background motion. Check how many sites and environment types are represented and whether that matches where your robot will actually operate.

3. Episode completeness

What fraction of recorded hours became accepted episodes with a clear start and end state? A low ratio is not necessarily bad, since strict acceptance criteria reject a lot, but you should know the ratio and the criteria behind it.

4. Annotation density

Per episode, how many phases, tools, objects and outcome states are labelled, and by whom? Annotations by people who know the trade are worth more than crowd-sourced bounding boxes, and the difference is visible in label consistency across contributors.

5. Rights coverage

Finally: for what share of episodes can the provider show a contributor release, a site authorisation and a privacy review record? Anything below one hundred percent is a liability you inherit. Every Curatrix release publishes these figures in its datasheet; the catalogue shows the headline numbers, and the FAQ explains how they are calculated. If you want to discuss a specific release, get in touch.

Related articles

All articles
2 min read

Why Physical AI needs field data, not lab benchmarks

Robots trained on tidy lab demonstrations fail on real job sites. A practical look at what field-captured manipulation data adds and how to judge it.

  • physical-ai
  • data-quality
  • field-capture
1 min read

Egocentric video for manipulation learning: head, helmet or chest camera?

Where the camera sits changes what a model can learn. Trade-offs between head-mounted, helmet-mounted and chest-mounted capture for skilled manual work.

  • field-capture
  • sensors
  • physical-ai
2 min read

What “rights-cleared” training data actually means

Consent, location releases, GDPR and licence scope: the four layers that separate usable robotics data from footage you cannot ship a product on.

  • licensing
  • compliance
  • data-quality
Next step

Start with a scoped pilot

Bring a task. On the call we scope the capture protocol, the episode volume, the annotation schema and the delivery format.

Pilots typically start from approximately €25,000. This is an indicative figure only—final scope and pricing are set after a technical qualification call.