Methodology
Our methodology
How the Physical AI Atlas collects, verifies and updates its data, what it refuses to publish, and what you are allowed to do with it.
Principles
The Atlas follows two simple rules. First: every figure is sourced and dated. A price, payload, runtime or deployment count only appears in an entry if it is tied to an identifiable source and a publication date. Second: null rather than invented. When a data point is not public or not verifiable, the field stays empty and says so plainly. We prefer a table with holes to a table that is wrong: an empty cell tells you something true about the state of available information, an undeclared estimate lies to you.
We systematically distinguish three statuses: verified fact (corroborated by at least one source independent of the vendor), vendor claim (uncorroborated announcement, flagged as such) and estimate (third-party calculation, always attributed).
Source hierarchy
Not all sources are equal. Our confidence scale, from weakest to strongest when a source is taken in isolation:
- Vendor communication (press releases, product pages, executive statements): essential for specifications, but promotional by nature. Never sufficient on its own for a deployment or performance figure.
- Specialised press (The Robot Report, IEEE Spectrum, Robotics 24/7 and peers): provides corroboration, context and sometimes contradiction. This is our reference level for validating an announcement.
- Trackers and cross-cutting databases (deployment trackers, official regulatory data such as EUR-Lex or ISO catalogues, financial filings): the strongest level, because it is structured, comparable over time and independent of announcement cycles.
A figure gains confidence when confirmed at several levels of this scale; it is downgraded when a higher-level source contradicts it.
Dated verification per entry
Every Atlas entry carries a "last verified" badge with a date. The badge means that on that date, each claim in the entry was checked against its original source and at least one cross-check was attempted for sensitive figures (prices, deployments, regulatory timelines). An older entry is not necessarily wrong, but the badge lets you judge the freshness of the information before quoting it in a board meeting. If you spot an error, report it: the correction is dated and visible in the entry's history.
Weekly updates with guardrails
The Atlas is updated on an automated weekly cycle: monitoring agents collect announcements, regulatory publications and standards updates, then propose changes to entries. This automation runs inside strict guardrails: schema validation on every field, a ban on overwriting verified data with lower-confidence data, and human review of every diff before publication. No change goes live without being read. Automation gives us the cadence; review gives us the reliability.
Known limitations
Three structural limitations of the field apply to us and deserve to be stated:
- Humanoid prices are largely speculative. Very few manufacturers publish firm pricing; most figures in circulation are statements of intent or third-party estimates. We flag them as such or leave them null. Find this confidence level displayed entry by entry in our robots comparator.
- Specifications are not standardised. Payload, runtime or speed are measured under conditions each manufacturer chooses; comparing two entries requires reading the measurement notes.
- Benchmarks are not comparable. There is no shared evaluation protocol yet for general-purpose AI robots; a success rate published on an internal demo does not compare with a rate measured in production. We therefore build no rankings on that basis.
Licence and reuse
The Atlas's structured data is published under the Creative Commons Attribution 4.0 (CC BY 4.0) licence. You may copy, redistribute, adapt and build upon it, including commercially, under three conditions: credit "Physical AI Atlas, D-Fairy Consulting, d-fairy.fr", provide a link to the licence, and indicate whether changes were made. Attribution must not suggest that D-Fairy Consulting endorses your use. For large-scale reuse or a regular data feed, get in touch: we would rather provide a clean export than watch stale copies circulate.
Keep reading
- See the definition that frames the Atlas's vocabulary: our what is physical AI pillar.
- Find short, sourced answers in our FAQ.
- Browse the companies we track in our companies comparator.
Sources: CC BY 4.0 licence (Creative Commons), EUR-Lex (official regulatory texts), ISO catalogue (standards status). Methodology maintained by D-Fairy Consulting, last revised 9 July 2026.