The atlas: what our data can show

Chapter 1

What we can see

How many robotics companies are there in France?

Counting robotics companies begins with deciding what belongs in robotics. A manufacturer, a software company and an integrator may all contribute, while appearing under different business classifications: no code means “robotics company”. The French classification files certain industrial robots with other special-purpose machinery, robot software with software publishers, and installation with engineering offices. So a count needs a scope rule, and no register supplies ours. Our collection follows a defined set of organisations and the connections we find around them. It is not a census of the industry.

Four sets of organisations appear on this page. Each has one name here, and the name is used everywhere.

The catalogue

53,445 organisations. Every organisation any source names, documented or not: a project’s participant, a paper’s institution, a funder, an exhibitor on a show’s list. The widest set on this page, and the loosest.

The organisations we searched

5,366 organisations. The ones we set out to document in every source. They came from six lists: the exhibitor lists of trade shows and conferences (2,647), the base list from our deep-research sessions (1,674), the most-tied organisations we found on the way and queued (1,124), the largest Chinese applicants in worldwide patents (420), the companies in the robotics press of 2026 (26), and two pilots that tested the process (70). An organisation can be on several lists; 538 are, so the counts add up to more than the total.

The French robotics companies our stories follow

89 companies. A chosen set, drawn from the organisations we searched; each story says which of them it uses. Not a census, and not a ranking.

France in the catalogue

4,791 organisations. Every organisation a source places in France: universities, hospitals, public agencies and companies of any sector that a robotics-relevant record names. 384 of them are there only because an EU-funded project names them, and no other source does. 1,115 are named in a paper, 1,034 in the French public funding registers and 1,125 in ROR, the register of research organisations; an organisation can be in several of these. Only 366 of them are organisations we searched. Of the French organisations in our directory, it types 256 as companies, 86 as research bodies and 36 as funders.

What we can count is what we hold, and we count it exactly. The catalogue names 53,445 organisations. Of those, 4,791 are placed in France by a source. We searched 5,366 organisations in every source; 366 of them are French. Our stories follow 89 French robotics companies, a chosen set. About all of these organisations we hold 810,497 event records and 250,437 ties, from 22 source collections. When this site says “French robotics”, it means the organisations we searched and what these sources say about them. It is not a census.

Where the facts come from

An event record — a dated fact, in the rest of this page — is one thing one source says about one or two organisations on a date: a grant, a shared project, a paper, a patent, a legal notice. The largest sources are Worldwide patent register (EPO Open Patent Services), CORDIS (EU-funded projects), US patent register (granted robotics patents) and OpenAlex (academic papers), with 18 smaller ones behind them. The length of a bar measures how many records we collected and kept from a source. That depends on the source’s scope, on how far we read it and on how it records things: a register that publishes every participant of every project gives thousands of records for one question, while a company’s website has to be read for the one sentence that says who works with whom. Size proves neither importance nor unimportance. The papers and the patents are counted apart, in chapters 4 and 6, never inside these records.

How many sources document each organisation

Of the 53,122 organisations in the catalogue that at least one source documents — an event from the source names them, or a search of it found them — 43,824 are documented by one source only and 9,298 by two or more. The chart counts every combination of sources. Read it by column: a column is one combination, its dots say which sources are in it, and its bar how many organisations have exactly that combination. One real column: 3,905 organisations appear in OpenAlex, OpenAlex ids and ROR and in no other source here.

Several sources can document the same organisation while missing the same parts of the industry: the EU’s project register and the French public funding registers both see organisations that take public research money, and both miss a company that never applied. This chart measures overlap between our sources; it does not score whether an organisation is real.

What a gap means

A missing result can mean that we have not checked, that a check failed, or that a completed search found no match. Those are different kinds of uncertainty, and we keep them apart. One check is one organisation we searched in one source: 118,052 checks, over the 5,366 organisations we searched and 22 sources. Each check is in one of six states. Found: the source names the organisation, or a search of it found the organisation, and the basis of the match is kept. Searched, nothing found: a completed search of the source found no match — that is information. Not checked: no check of the source was attempted for the organisation. The check failed or did not finish: a check was attempted and did not establish a result — a request failed, or the run stopped before it answered. Candidates found, identity unresolved: the source returned records that might be the organisation, and which of them is has not been settled; until it is, nothing is counted as found. The source cannot hold it: the source’s own rule excludes the organisation — the US small-business awards go by law to US-owned companies, so that check is never made for a French one. A single unsuccessful search is not a comprehensive investigation: an organisation searched once in one source and not found may well be in the next.

Most of what we do not know, we have not checked. CORDIS, the EU’s register of the projects it funds, documents 826 of the organisations we searched; we searched it for 1,306 more and found nothing; for 3,233 the check was not made. 320 of the organisations we searched are documented by no source at all; 320 of them were searched in at least one source and came back empty. Those absences were looked for. It does not follow that the organisations are absent from every source.

How much rests on one register

40.9% of our event records come from one place: CORDIS, the European Union’s register of the projects it funds and of the organisations that take part in them. Records pile up there for a simple reason. One project with thirty participants gives thirty records, one per participant, and a large project gives more. So one share cannot say how much of the collection is really the register. We need three shares, each of its own total.

A direct tie is a tie where the evidence says the two organisations did something together: ownership, affiliation, trade, money, people moving or co-owned patents — a customer, an investor, a co-owned patent. Direct ties come from company websites, the US federal spending register, co-owned patent families worldwide (the European Patent Office's register), co-owned US patents, CORDIS (the EU's register of the projects it funds), US public research awards, Wikidata's records of which company owns which, French public funding registers, the GLEIF register of which company owns which and the ROR registry of research organisations and their parent institutions. Of the 5,392 direct ties the record keeps, only 0.7% rest on CORDIS alone. The other kind of tie only puts two organisations in the same place — the same project consortium, most often — and says nothing about whether they worked together directly.

So the register widens the catalogue: it names 14.5% of the 222,706 entities — organisations and people — that appear in any record. But it is almost absent from the ties that say who works with whom. It also sees only whoever takes European money: research bodies, large firms with a grants office, Europe; it says nothing about a company that never applied. Hence the test: remove every tie that only CORDIS supports, and see what remains.

Removing the ties only CORDIS supports changes the network. The organisations still holding a tie fall from 6,476 to 5,127, the largest connected group from 5,768 to 4,302, and five of the ten biggest brokers remain among the top ten. An organisation left with no tie is dropped from the count, which is why the number of separate groups can rise as the network shrinks. Many ties supported by other sources remain. This test shows which parts of our network depend on that source. It does not tell us whether the removed relationships exist outside our collection.

The base list

Before we collected anything, we needed a list of who to look for. We built it by deep research: long sessions with Perplexity, Gemini and ChatGPT, asking each for the robotics ecosystem in France, then in Europe, then in the world. That gave the base list, 1,658 organisations: a starting point, not a finding, and never checked entry by entry. Many of our searches started from it. Finding a record for one of its organisations is still useful evidence; what it cannot give is an estimate of how many organisations the list missed. What the figure tells you is how much of the list the evidence backs.

We found at least one dated record for 1,354 of the 1,658 organisations on the base list. That supports their presence in the collection. It does not by itself verify every name, organisation type or connection to robotics.

Lists other people keep

How much of what exists we hold can only be measured against lists we did not build. Each list establishes one thing about its members, and only that: a MassRobotics sponsor pays to sponsor a hub in Boston; a Y Combinator company is in that portfolio; a top patent holder holds many granted US patents; an exhibitor is on a show’s list. So we measure each list on its own, never pooled, in three groups by what being on the list establishes. Known from another source means the organisation is documented in a source collection other than the list itself, so a show’s own exhibitor list does not count as knowing its exhibitors. The worldwide list is the largest applicants in the patent extract we have collected so far, not a complete world ranking. The lists of CES 2027 and Global Industrie 2027 are advance exhibitor lists, captured between 20 September and 25 September 2026; those shows have not taken place.

What each list establishes is narrower than a blind spot, so the conclusion is narrow too. We know most of the organisations on the lists kept by the robotics community itself: 90.3% of the MassRobotics sponsors, 82.9% of Wikipedia’s US robotics companies. We know 34% of the largest holders of US robotics patents (102 of 300). Of the exhibitors at the six large shows, we know between 5.7% (Global Industrie 2027) and 39.9% (WRC Beijing 2026). An entry we do not match may be an organisation we have not reached, a brand or subsidiary under another name, or a name our matching missed; we have not yet checked a sample. Whatever this site says about who matters in robotics, it says from inside the specialists’ world.

The records. Every organisation the map draws, and every organisation we searched with at least one dated fact, is in the directory, with the sources we searched for it and the sources that found it; the story What does this map leave out? walks through these limits and what the map's default view leaves out.

Chapter 2

Who is in it

What kinds of organisation are in our directory, and what do they build?

A register says what an organisation is only for the organisations it holds, and no register holds them all. So the directory types its organisations by one rule, and the rule is provisional. An organisation we searched keeps the type its list gave it; any other is a funder when it only ever pays, a research body when it co-signs papers, and a company otherwise. That reads what we observed an organisation doing — paying, publishing — as what it is, and the two are not the same field: a company can fund projects and publish papers, and a public body can do neither in what we collected. So the figure counts the rule’s result, not verified legal forms; the error runs both ways, and we do not yet trust the base list’s type over better evidence, since that list was never checked entry by entry. The repair — one field for the organisation type from an identity source, one for the roles we observed, one for the robotics activity its own description states — is under way in the data, not on this page.

Beside the matrix, one gap of the directory itself: the country is missing for 2,769 of its 5,659 organisations ( 48.9%), the “Not stated” column. Chapter 3 counts 12,481 organisations without a country in the whole catalogue; that is a different population, and neither share stands in for the other.

What an organisation builds is written in no register we read whole; some company profiles and product registers say it, none completely. So we read it only where we hold the organisation’s own description — its website, or a register’s profile of it — and a language model filed each description by the form of robot it identifies. That is 572 organisations, 1.1% of the catalogue and 9.9% of the directory’s companies ( 496 of 5,029): a small sample, heavy with the companies whose websites we read. For everyone else we leave the question open rather than guess from a name. The figure is a reading, not a measurement, and it carries a method note in place of a claim chip.

Our documentation is concentrated around a small number of organisations. The next chart starts with the most frequently mentioned organisation and adds the others in order, reading off how much of everything we have written down a few of them hold. The unit is a mention: one organisation named in one event record, so a record naming two organisations contributes one mention to each. The median organisation has 3 mentions; nine in ten have 13 or fewer; one in a hundred has more than 509, and the most mentioned 18,122. A median of two does not mean that most organisations have only one record, and a mention count is not knowledge: repeated records and the detailed membership of one project are not independent information.

Two counts of organisations with nothing behind them appear on this page, and they count different sets. 325 organisations of the catalogue carry no mention at all: they are organisations we searched that no source has documented yet. Chapter 1 counted 320 organisations we searched that no source documents, over the organisations we searched alone, with a documented check counting as documentation even without a dated record; the two are cohorts of the same update, not one figure written twice.

The records. The directory gives every organisation its type, with a filter to keep one type; a row opens to what we hold on it and a door into the map.

Chapter 3

Where

Which countries appear in our organisation, paper and patent records?

France is the most common country in our organisation catalogue. That tells us where our collection is strongest. It does not tell us where the most robotics companies exist. A country here is the one a source states for an organisation, and the sources’ fields differ: a company register gives an incorporation address, a project register the address of the participant, a paper index the location of an institution; none of them is necessarily where the work is done. 12,481 organisations have no country in any source. CORDIS, one of our two largest sources, is centred on EU-funded projects, which is one reason the European countries stand where they do.

We can also compare two activities: publishing research and obtaining patents in the United States. They produce different country patterns. The paper collection follows citations from our starting papers; the patent collection follows our selected patent classes, in one patent office. Starting from papers or patent classes instead of a company list removes one of our choices, not all of them — the starting papers, how far we followed citations, the classes and the single office remain ours — and neither collection is a census of where robotics is invented. Each country gets two bars on one scale, over the same five years: its share of the robotics papers byfractional credit (each paper split equally among the countries of its institutions, so a paper counts once) and its share of the US patent grants, by the country of the first-named patent owner: the first organisation the patent office lists as holding the rights. The last window is2020–24; 2025, a single year, is in the table apart from it.

A map shows where, and only where. The bars above rank the countries; the maps below show how far our catalogue reaches and, inside France, where each organisation sits. A place here is a head office: we take the address from the highest source that records one — the company register first, then ROR, CORDIS, OpenAlex, Wikidata and the organisation’s own website — and read its town and region from that address. 37,066of the 53,405 organisations we hold have a place that way (69.4%); 3,897 have only a country, and12,442 not even that. The world map shades each country by the organisations our sources place there: a map of our collection, not of robotics.

Most of the catalogue is European: 81.7% of the organisations with a stated country are in one of the 40 European countries of the next map. We draw them as circles rather than shades, because a count is not a density: on a shaded map a large country looks full simply for being large.

Inside France we can go down to the region. 4,110 of the4,791 organisations our sources place in France have a place (85.8%); the rest have only a country. The first map shades every region by the organisations whose head office is there; the second draws the 89 French robotics companies our stories follow, 75 of which we could place.Île-de-France leads both. That is partly what a head office means: it is where an organisation is registered, and many organisations that work elsewhere are registered in Paris.

Organisations in different regions are tied too: they share a public project, sign a paper together, or one runs a laboratory in the other’s region. 43.3% of the ties between organisations placed in France stay inside one region. The table lists the ten pairs of regions tied most often, each kind of tie in its own column. It is a table rather than a map of lines: the 80pairs of regions drawn as lines would cross until none could be read.

Papers leave a trace of place of their own: the institutions of their authors. The last map shades each region by the robotics papers we gathered with at least one author at an institution there. A paper with authors in two regions counts in both, and 102 papers with a French institution could not be placed in any region.

The records. The directory filters by country; the map draws the French core with its ties abroad; the story Why do China and the US change places between these charts? follows the contrast through time.

Chapter 4

Research and invention

Which organisations appear in both the papers and the patents, and how concentrated is each?

Papers and patents leave different traces of research. A paper names its authors and their institutions. A patent names inventors and may name an organisation that owns the rights. Here we ask which organisations appear in both collections, and we count only the organisations we searched on both sides — in the papers and in the US patent register — so that every blank cell is a search that came back empty, never a search we did not make.

Within each collection, a few organisations hold most of the output, and the two curves rest on two different populations. The papers are led by universities — Chinese Academy of Sciences, Tsinghua University and aff_harbin_institute_of_technology — and the patents by companies — Intuitive Surgical Operations, Inc, FANUC and Honda Motor. Each curve ranks its own organisations from the one holding the most and adds up their share; the unit is a counted paper affiliation (a paper counted once for each of its institutions) on one line and a counted patent ownership (a co-owned patent counted once for each owner) on the other, and the marked point is the top tenth of each, with how many organisations that is.

The records. Each organisation’s row in the directory says whether the papers and the patent register were searched for it and whether they found it; the story When a university patent becomes a company patent: what do the documents establish? follows the inventors.

Chapter 5

How they connect

Which relationships tie these organisations together, and who holds each one together?

The measures of this chapter describe the relationships we have recorded, one kind at a time. A measure of who is central, taken over every kind at once, would rank a funder’s list of grants and a consortium’s list of members on one scale, which measures nothing in particular. So the chapter opens on what each kind of tie means, compares the kinds as separate networks, and only then turns to the map, which draws them together: what its display rules leave out, and, under “Graph methods and diagnostics”, the calculations on that mixed view.

One network per kind of tie

A grant, a joint paper and a customer contract connect organisations in different ways. We keep those relationships separate so you can see what each connection actually means, and every measure below names the kind of tie it was taken on. A tie is one recorded relationship between two organisations; the record files each under one of nine kinds, and a relationship can legitimately belong to two — an investment that buys a stake is funding and ownership at once. The rows say what the evidence behind each kind establishes, what it does not, and give one tie of that kind you can open on the map with its records.

Before you read any ranking, look at what a tie of each kind actually records. Take ownership and its neighbour, affiliation. An ownership tie says one organisation holds or bought another, as a register or the company itself states it: 195 ties, 134 of them a parent and its subsidiary and 61 an acquisition announced on a website; German Research Center for Artificial Intelligence comes first there. An affiliation tie says one organisation is part of another — a laboratory listed under the institution that runs it, as the registry of research organisations records it: 2,479 ties, and CNRS comes first there, with the joint laboratories it runs with universities. We keep the two apart because the first is property and the second is not: the CNRS owns none of its laboratories the way a group owns a subsidiary, and a joint laboratory belongs to each of its parents.

Funding is the kind to read most carefully, because it is where the record is most uneven. The public side — 1,760 ties from 34 funders — comes from registers that list every grant they make: the EU’s project register, the US federal spending register, the French funding registers. The private side — 902 ties from 570 investors — comes from what companies say on their own websites, because nobody publishes a register of venture deals. So the European Commission is first, with 1,074 recipients (194 of them French, 739 among the organisations we searched), because CORDIS names every participant of every project it funds; and Bpifrance shows 9 ties because no register we read lists its investments one by one. The amounts tell the same story the other way round: the EU’s grant records carry the money for 15,107 of 18,116 records, while 791 of the investor ties from websites name the investor and no sum, and only 151 rounds carry a disclosed amount. A funder that is central here is a funder whose register we could read. Whether public money outweighs private money in French robotics is exactly what this record cannot say.

So read every ranking below with its register in mind. Each kind of tie is a different instrument: co-authorship sees whoever publishes, shared projects see whoever applies for EU money, commercial ties see whoever names customers on a website, patents see whoever files in the United States. An organisation first in one kind is first in what that instrument records. The rankings are not comparable across kinds, and none of them measures who matters in robotics: they measure who is documented as tied, and by whom.

Ownership · the evidence names the relationship

Establishes: One organisation holds or controls another, as a register or the company itself states it: a parent and its subsidiary, or an acquirer and what it bought.

Does not establish: It does not separate an equity stake, majority control, a subsidiary relationship and an asset purchase; the date and the source’s exact claim stay with the record. A laboratory and the institution that runs it are not here: that is affiliation, the next kind.

Example: Google ownership Boston Dynamics — show on the map, with the records behind it.

Affiliation · the evidence names the relationship

Establishes: One organisation is part of another, as the registry of research organisations lists it: a laboratory, institute or unit under the institution that runs it.

Does not establish: It does not establish ownership, funding, or that the two work together beyond that; a joint laboratory belongs to each of its parents, and the tie says so for each.

Example: the default map draws no tie of this kind; the directory lists the organisations that hold one.

Commercial · some records name the relationship, others only a shared place

Establishes: A business relationship a company states on its own website or a register records — customer and supplier, partners, a distributor, a product that works with another — with the stated subtype kept.

Does not establish: “Works with” or a logo on a partners page does not establish a sale, its direction or its size.

Example: GLS France commercial Actemium (Vinci Energies) — show on the map, with the records behind it.

Funding · the evidence names the relationship

Establishes: Money announced or recorded as going from one organisation to another: a public grant awarded, or a private investment announced; the kind of money and its direction are kept.

Does not establish: An award or an announcement does not establish that cash moved, when, or how much; amounts are not added up here. An investment that buys a stake belongs here and in ownership at once.

Example: European Commission funding EasyMile — show on the map, with the records behind it.

People · some records name the relationship, others only a shared place

Establishes: The same named person seen first at one organisation and later at another, in the papers or the patents; or a company recorded as spun out of another.

Does not establish: A matched name is not a verified change of employer, and a person moving does not establish that know-how moved with them.

Example: the default map draws no tie of this kind; the directory lists the organisations that hold one.

Patent co-ownership · the evidence names the relationship

Establishes: Both organisations are named as owners of the same granted US patent, or as applicants in the same worldwide patent family.

Does not establish: Applicants on different members of one family are not necessarily joint owners of the same patent; the two cases are not yet kept apart here.

Example: AcuSurgical co-owns a patent with CNRS — show on the map, with the records behind it.

Public project partners · the evidence puts them in the same place

Establishes: Both organisations are listed as partners in the same publicly funded project: the EU framework programmes, the French national programmes.

Does not establish: A project with many partners makes a tie between every pair of them; being listed together does not establish that the two worked directly together.

Example: Navya is a partner in the same public project as University of Geneva — show on the map, with the records behind it.

Research co-authorship · the evidence puts them in the same place

Establishes: People affiliated to both organisations signed the same publication; the publication is kept with the tie.

Does not establish: One author with two affiliations links two organisations by themself; a shared paper does not establish that two research teams collaborated.

Example: Schneider Electric co-authors with Toshiba — show on the map, with the records behind it.

Conference co-presence · the evidence puts them in the same place

Establishes: Both organisations were listed — as exhibitor, sponsor or speaker — at two or more of the same trade-show or conference editions; the editions and roles are kept.

Does not establish: Exhibiting at the same event is not sharing a stand or meeting; with fewer than two shared editions no tie is recorded.

Example: Boston Dynamics was listed at two or more of the same events as PSYONIC — show on the map, with the records behind it.

Three measures compare the nine networks. Organisations: how many hold at least one tie of the kind. Connected pairs: how many pairs of organisations are tied that way, several records between the same two counted once. Share in the largest connected group: of those organisations, the share reachable from one another through chains of ties of that kind — connected, not collaborating; a shared project with many partners ties every pair of them, which is why the project network is so joined. Every measure of every kind, thirteen in all, is in the methods and diagnostics below.

Who is central, kind by kind

One ranking at a time: choose a kind of tie and a measure, and the ten organisations with the highest value appear with a link to their place on the map. The measure’s meaning is written under the selectors, because a high number of connections, a high betweenness and a high PageRank say three different things.

Which funding relationships have we recorded?

About two-thirds of the funding relationships in this collection concern public grants. That is a count of recorded links, not a share of the money invested in robotics. The record keeps two kinds of funding tie apart: a public grant — a European programme, a national research agency, a government department — and a private investment, a fund or a company taking a stake. They come from different sources with very different reach: the public registers publish every award; a private round is known only where a company announces it on its own website and names its investors.

How the map’s display rules affect the graph

The map opens with a selected view. It combines the kinds of relationship and draws only the pairs allowed by its default filters: a pair is drawn when at least one organisation belongs to the small curated core of French robotics we assembled, and drawn faintly when one end is a large body. The calculations in the diagnostics below describe the graph those rules keep. Changing the filters can change the rankings and the apparent groups. Nothing is deleted: a hidden relationship is still in the record and in the directory, and a rule you cannot see reads as data that never existed, which is why it is drawn here.

The records. Every tie the map draws opens to its years, its records and their sources; the ranked organisations above link to their place on it; the story Who holds this map together? asks who holds each kind of tie together.

Chapter 6

What is changing

What is changing in robotics — as opposed to what is changing in our collection?

A growing database can make an industry look as if it is growing too. To study change, we need to know that the counting rules stayed comparable. Here we use US patent grants selected by the same patent-class rules across time: one register we hold whole for the classes we chose, so a change in it is not a change in what we collected — though a register held whole can still change with how classes are revised, how firms file, and how fast the office works.

The year is the year a patent was granted. It may be several years after the work began. These counts describe patents obtained in the United States, rather than every invention or every robot built, and several grants can belong to one patent family. Three series we hold and do not draw here, for the reason above: our own facts by year, which measure our collection; the papers by year, gathered by following citations, so a year’s count depends on where we started; and the founding years of the companies our stories follow, which describe a chosen list.

The growth is not shared evenly among the countries of the patent owners. The table gives each country’s grants in 2015 and 2025 with the change in patents; the rate of change and the share of 2025 are a second view, because a large rate on a small base is only a few patents, and a small rate on a large one can be thousands.

Two years say little about the years between them. The same seven countries, year by year and on one scale, show when each line turned.

Worldwide, by patent office

The US register sees only the inventions someone chose to patent in the United States. This is the chapter’s second series read whole: the European Patent Office’s worldwide register, counted by its own search service over every family in four robotics class roots, not over the records we fetched — so, like the US grants, a change in it is not a change in what we collected. A patent family is one invention, counted once however many offices publish it, and dated by its first filing. The last two years are still filling, so the chapter reads the series to 2023.

The step in the Chinese panel between 2015 and 2016 is partly the register changing, not only invention: from around then the Chinese office assigned these classes to its own filings, and before it fewer Chinese filings carried them at all. The Chinese years before 2016 are best read as a floor.

US grants by class

Yearly US grants in our classes were 3.3 times as many in 2025 as in 2015. Which classes account for the rise? A grant can carry several of the five classes, so the class panels overlap: a grant counts once in the total and once in each class it carries, and the classes’ rises add up to more than the total’s.