Nobody Flies to Toulouse for a Reducer

Two robot shows opened on September 3. In Brussels, the European Parliament hosted a robotics showcase and Euronews ran the day as a race against the United States and China (https://www.euronews.com). In Guangzhou, XAIR Expo opened its doors to procurement teams from logistics groups, automakers, and hospital chains, and stayed open through the 6th. Add Enchanted Tools' Mirokaï rolling through hospital corridors, Pollen Robotics sitting inside Hugging Face, Wandercraft's Calvin-40 with a Renault logo behind it, and the arithmetic writes itself. France has the densest humanoid cluster in Europe. Europe is the third pole.

The arithmetic depends on what gets counted. Stage minutes and press mentions are one ledger. Serial numbers in service, hours per day per unit, industrial buyers outside the home market, and the country of origin stamped on the reducer inside the hip joint are another. The Brussels case runs on the first ledger. The Guangzhou case runs on the second.

Same week, two venues, one direction of traffic. Brussels was a display. Guangzhou was a market. LimX Dynamics spent the spring in Hannover and Paris as a vendor, selling into Europe, and went home to Guangzhou to sell some more. No procurement manager flew to Toulouse this week to source a harmonic reducer. Density is not gravity.

A pole is measured by what it pulls

A pole is a mass that bends flows. Capital flows toward it, because the returns are there. Demand flows toward it, because the buyers assume the best product is there. Components flow toward it, because the volume is there, and then the components start being made there, and then rival OEMs move there to be near the components. That is the whole mechanism. Everything else is marketing.

The United States pulls capital and training infrastructure. Figure, Agility, Apptronik, and Tesla raise on American terms and train on American GPUs, and a European founder with a serious humanoid plan still takes the Sand Hill Road meeting. China pulls demand and components. The WSJ tally summarized by aiweekly.co puts China at the majority of global humanoid shipments in the first half of 2026, counted in thousands of units, with Unitree, UBTech, AgiBot, and Leju doing most of the shipping. China also makes its own motors, its own batteries, and its own reducers, and the reducer part matters more than it sounds. A humanoid has thirty to forty joints. Each joint needs a precision reducer. Whoever makes reducers at volume decides whose bill of materials closes.

Against that, the French firms have publicly placed units in the dozens, not the low thousands. That is one order of magnitude gap on the most generous reading and two on the honest one.

The tests for a pole are four, and they stay fixed:

1. Serial numbers in service, not demos shipped to events
2. Duty cycle per unit (hours per day, days per week, in a real facility)
3. Industrial buyers outside the home market
4. A supply chain others source from (actuators, reducers, hands, tooling)

Density is a headcount. Gravity is a direction.

Where Europe sits in the layer map

Before any French logo gets graded, look at the stack a humanoid is built on. Five layers, bottom to top.

Actuators and reducers. Motors, planetary and harmonic gearboxes, encoders, the joint module that combines them. The volume supply here runs through Shenzhen and Suzhou. Europe has premium names: maxon in Switzerland, Harmonic Drive AG in Limburg (a subsidiary of the Japanese parent), Wittenstein in Igersheim. They sell into surgical robots, satellites, and semiconductor tools at prices a humanoid built to a five-figure target cannot carry. None of them sits in France. The French humanoid firms buy their joints, or design them and have them cut abroad.

Hands and sensing. Tactile fingertips, event cameras, force-torque sensors, lidar. Europe is real here. Prophesee in Paris makes event-based vision sensors that ship globally. SICK in Waldkirch owns a large share of industrial safety scanning. Schunk grippers are on cobot arms everywhere. This is a layer Europe partly owns, although the dexterous humanoid hands that matter most in 2026 come from Shenzhen and from a handful of American labs.

Compute and control software. Nvidia Jetson or a Chinese alternative on the robot, PyTorch and increasingly LeRobot in the training loop. Europe owns a slice of the software through Hugging Face's robotics stack and through the ROS community that grew up partly in German and Swiss labs. Chips, no.

Integration and certification. ISO 13482, ISO 10218, the EU Machinery Regulation that takes full effect in January 2027, TÜV and Bureau Veritas audits, CE marking, and now the AI Act's classification of robots that work near people. Europe owns this layer outright. Anyone who wants to sell a humanoid into a European hospital or a German auto plant walks through this gate.

End-customer demand. Logistics, automotive, healthcare, hospitality. Europe has the buyers but they are slow, unionized, and procurement-heavy, and they have not yet ordered in volume from anyone.

Read the map and the third-pole claim fails before a single company is named. Europe owns a gate and part of a sensing layer. It imports the layer that determines cost and the layer that determines scale. A gate is valuable. Every robot in the world that wants European revenue passes through it. But a gate does not bend flows toward itself. It stands where it stands and charges admission.

Enchanted Tools: the most commercial, and the least humanoid

Mirokaï is the one French robot with a real duty cycle. Enchanted Tools (https://enchanted.tools) has units running at AP-HM in Marseille, at the Broca geriatric hospital in Paris, in Paris airports, and in hotels, moving linens, meals, and medication carts through corridors built for humans. Units run shifts. Staff know the robots by name. Some placements are outside France, in Belgium, Switzerland, and the Gulf. The company has said publicly that it wants to build toward six figures of units by 2030, which is a fundraising number, but the current fleet is a real one in the tens.

Grade it against the four tests:

1. Serial numbers in service, not demos shipped to events: passes, at small scale, in the tens.
2. Duty cycle per unit: passes, and by a wide margin over most humanoid startups anywhere. Hours per day in a working hospital is the hardest test on the list and Mirokaï clears it.
3. Industrial buyers outside the home market: passes, narrowly, through hospitality and healthcare rather than industry.
4. A supply chain others source from: fails. Mirokaï's ballbot base, its actuators, and its compute are bought in. Nobody designs in an Enchanted Tools part.

And then the category problem. Mirokaï rides on a single-wheel ballbot base. It has two arms, a projected animated face, and a service-robot job description. Enchanted Tools calls it a humanoid and the press agrees. On the shop floor at Renault Flins or at a BMW plant, a humanoid means legs, bimanual manipulation, and lifting, and Mirokaï does none of that. It is the best-run robot product in France and it competes in a different race from the one Brussels was staging.

That is what a real product looks like against the tests. Small numbers. Real hours. A customer in another country. And a bill of materials someone else owns.

Pollen Robotics: the open-hardware bet

Pollen (https://www.pollen-robotics.com) is in Bordeaux and, since April 2025, inside Hugging Face (https://huggingface.co). Reachy 2 is a wheeled bimanual research platform with a distinctive head and open-source arms. Below it sits a line of cheap open hardware: Reachy Mini on the desk, the HopeJR humanoid, the SO-101 arm that thousands of people have printed and assembled to run LeRobot policies. The CAD files, the firmware, and the training code are all public.

Grade it:

1. Serial numbers in service, not demos shipped to events: weak. Reachy 2 ships to labs. Reachy Mini shipped in the thousands as a consumer and developer device, which is a real number but not an industrial one.
2. Duty cycle per unit: fails as industrial deployment. Reachy runs in labs for hours, then gets reset for the next experiment.
3. Industrial buyers outside the home market: fails. The buyers are universities and hobbyists on four continents.
4. A supply chain others source from: passes, by a different route than the test imagines. Pollen does not sell actuators. It sells designs, tooling, and a training stack that other people build robots on top of. The SO-101 arm has become a de facto standard bench for imitation learning. That is a supply chain of a kind.

This is the interesting case. A pole built from downloads rather than shipments. Hugging Face already proved that a repository can bend flows in software; every model lab uploads there because every developer downloads from there. Whether the same works for hardware is an open experiment. Downloads are cheap. A robot built from Pollen files still needs motors, and those motors come from Feetech in Shenzhen and Dynamixel in Seoul. The gravity runs to the file server and stops at the shipping dock. But if a non-European OEM ships a commercial robot next year with a Pollen-derived hand or arm and says so, that changes the grade. Track that, and not the star count on GitHub.

Wandercraft and UMA: hardware ambition, thin serial history

Wandercraft (https://www.wandercraft.eu) spent a decade building exoskeletons. Atalante X has CE marking and FDA clearance for rehabilitation, and Eve, the personal exoskeleton, is in clinical use. That history gives the company something no other French humanoid firm has: a certification team that has walked a bipedal machine through a medical regulator twice. Calvin-40, the industrial humanoid built in forty days in 2025, is the turn toward factories. Renault has invested and has said Calvin units will work at Renault plants.

UMA is younger. Northstar is a full-size humanoid that exists as a small number of engineering prototypes and a demo reel. The team is credible. The product is early.

Grade both:

1. Serial numbers in service, not demos shipped to events: Wandercraft passes on exoskeletons, fails so far on Calvin, which is in pilot at Renault. UMA fails.
2. Duty cycle per unit: unproven for both humanoids. Wandercraft's clinical exos log real hours but that is a different product.
3. Industrial buyers outside the home market: Wandercraft sells exos into American clinics, which counts for the company and not for the humanoid. Renault is a French buyer. UMA has none disclosed.
4. A supply chain others source from: neither has said anything about it, and silence on a supply chain is an answer.

Both are strong on engineering. Wandercraft's legs are probably the best balance controllers in Europe. Neither has a humanoid that a buyer outside France runs on a schedule. That is the whole grade.

What LimX came to Europe to do

LimX Dynamics (https://www.limxdynamics.com) is in Shenzhen. In April it put Luna on the floor at Hannover Messe, the first time the platform had left China. In June it was at the MACHINA Summit in Paris, on a stage a few metro stops from Enchanted Tools' office. And the week Brussels held its show, LimX was at home at XAIR Expo in Guangzhou, on the same floor as UBTech and AgiBot, in front of buyers.

The European trips had a job. Chinese humanoid OEMs treat Europe as a customer geography with two features: a certification gate worth passing early, because the CE mark opens the whole continent at once, and buyers who move slowly and pay well once they move. Hannover gets you in front of German plant managers. Paris gets you in front of French media and the handful of French industrial groups willing to pilot. Both trips were sales calls.

What LimX did not do in Europe is source parts. No motor deal in Switzerland, no reducer deal in Germany, no hand from a French lab. The traffic ran one way. Product came west. Purchase orders, if they come, will go east. That is what it looks like to sit downstream of a pole.

The docket

The reader who has followed the cluster closely will push back, and the pushback is fair in tone if not in conclusion.

You are grading a three-year-old cluster against a national industrial policy with a decade of subsidy behind it. Unfair test. Europe's edge is safety, certification, and open tooling, not volume, and that edge compounds.

Agree the test is hard. That is the point. "Third pole" is a volume and gravity claim, and the people using the phrase in Brussels and on Euronews chose that frame. Grade the claim on the terms it set.

Certification is a real asset. It is also a gate that other people's robots walk through. Unitree, UBTech, and LimX will all carry CE marks within the year, and the TÜV invoice will be a rounding error against their unit economics. A gate is a layer in the map. It is not where the mass sits.

Open tooling can build gravity. Pollen is the live experiment, and Hugging Face has the distribution to make it work. But downloads only become a pole when they convert into components and deployments that other companies depend on. Track the conversion.

Subsidy history explains the size of China's lead. It does not change what Europe should be counted on today, and the buyers at XAIR do not discount a Chinese robot because it was cheaper to build.

And one clean concession. Enchanted Tools has more paid duty cycles than most humanoid startups anywhere, including several in the Bay Area with ten times the funding. Small, real, and in the wrong category for the race being described. Although if the race turns out to be about service robots in hospitals rather than bipeds in factories, that concession gets bigger.

What would make it a pole

Convert the four tests into three events a reader can check in 2027 without trusting anyone's press release.

One. A foreign industrial customer running a French humanoid two shifts a day. A German or Spanish or American plant rather than a French hospital or Renault, a named site, a stated shift pattern, and a robot with legs or at least bimanual lifting. Wandercraft is the likeliest candidate. UMA is the long shot.

Two. A European actuator, reducer, or hand that a non-European OEM designs in and names. If LimX or Agility or Figure ships a product with a maxon joint module or a Prophesee sensor or a Pollen-derived hand and says so in the spec sheet, Europe has moved a layer. If a French firm starts selling a joint module, that would be the bigger signal.

Three. A public serial-number count above a stated floor. Pick one hundred units per company in customer facilities, disclosed, with the customer's name attached. Enchanted Tools may cross it first and will cross it on wheels. Nobody else in France is close.

Two shows opened on September 3. One had a stage. One had buyers. France has one company that passes three of the four tests at small scale, one that passes the fourth by an unusual route, and two with excellent engineering and no serial history. The traffic this week ran from Shenzhen to Hannover to Paris and back to Guangzhou. When a buyer flies to Toulouse for a reducer, write the third-pole piece. Not before.

Sources

• Euronews, European Parliament robotics showcase, Sep 3, 2026
• XAIR Expo Guangzhou, Sep 3–6, 2026
• Enchanted Tools, Mirokaï deployments
• Pollen Robotics, Reachy 2 and open hardware
• Hugging Face, LeRobot and robotics
• Wandercraft, Calvin-40 and Atalante X
• LimX Dynamics, Luna at Hannover Messe and MACHINA Summit Paris
• WSJ humanoid shipment tally via aiweekly.co, H1 2026
• Vader Research, French humanoid cluster list