2026-07-19 · methodology, classification, evidence · Robotics Census

How the Robotics Census classifies companies

The core, adjacent, participant-only and unassessed classes, and the evidence-first rules that decide which companies are published.

Every company page on this site states a classification: core or robotics-adjacent. This post explains the full taxonomy behind those labels, why two further classes never appear on the site, and the evidence rules that decide every assessment. It is the canonical reference for how the Robotics Census classifies companies.

The four classes

Core

The company designs, builds or sells robots or robotic systems. Robot makers, robot arm manufacturers, autonomous mobile robot vendors. Published on the site.

Adjacent

The company supplies components, sensors, actuators, drives, software, simulation, integration or services specifically for robotics or industrial automation. Published on the site.

Participant only

The company's actual business is something else, and robotics was only the context in which the census encountered it: a pilot site, an end user, a data provider, a consortium partner in a robotics project. Never published. A hospital that hosted a robotics trial is not a robotics company.

Unassessed

The census has no usable evidence either way, so it makes no claim and creates no assessment record. Never published.

Two tests separate the classes. Core versus adjacent asks what a company sells. Adjacent versus participant-only asks whether robotics is the business or merely the context. Unassessed is not a verdict at all: it is the census declining to guess.

The classes in numbers

The census currently tracks 5,937 companies. It publishes 509 of them: 58 core and 451 adjacent¹. A further 6 companies are assessed participant-only with high confidence and are deliberately not published¹. The remainder are unassessed: tracked, but without evidence the census will put its name to¹.

What counts as evidence

An assessment is only recorded when it can cite evidence, and every published profile shows that evidence verbatim with its URL. Two kinds of evidence decide assessments:

• The company's own website. Decisive signals are read only from a page's declared identity zones: its title, description and headings. Marketing copy in the page body never decides an assessment.
• A technology tag the company declared for itself on the A3 (Association for Advancing Automation) member roster⁴.

Three things are never evidence: appearing in an EU-funded robotics project²³, which proves participation but not identity; similarity of a company's name to another company; and a language model's opinion. Assessments are made by a deterministic procedure over recorded evidence spans.

Why publish less than the census holds

Publishing only 509 of 5,937 tracked companies¹ is the design, not a gap. A directory that lists everything it has heard of must either guess or go silent about its confidence. The census does neither: what it cannot prove, it does not print, and what it prints, it cites. Corrections are applied as an editorial layer with the correcting evidence recorded alongside them; the methodology documents the pipeline end to end.

Sources

Every numeric claim above is footnoted to one of these sources with its as-of date.

1. Robotics Census database, aggregated at build time: https://roboticscensus.com/stats , as of 2026-07-27.
2. EU CORDIS open data, Horizon Europe projects: https://cordis.europa.eu/data/cordis-HORIZONprojects-csv.zip , as of 2026-07-15.
3. EU CORDIS open data, H2020 projects: https://cordis.europa.eu/data/cordis-h2020projects-csv.zip , as of 2026-07-15.
4. A3 (Association for Advancing Automation) public member roster: https://www.automate.org/companies , as of 2026-07-15.

How the census produces these numbers is documented in the methodology; census-wide tables live on the statistics page.