A humanoid robot 1.60 metres tall and weighing 50 kilos stands in a Paris laboratory. Somebody pushes it. It wobbles, corrects its posture, finds its balance again. That is all it does. And it is already considerable.
The startup that built it is called UMA, for Universal Mechanical Assistant. It came out of stealth on 1 December 2025, and claims nine months between its creation and a working prototype, designed and assembled in Paris. Its teams have just named it Astro, after Osamu Tezuka’s manga — the little nuclear-powered robot which, back in 1952, was already asking what we owe a machine that resembles us.
There is a detail the images show without ever commenting on it: during demonstrations, a safety cable stays attached to the robot. A fall could damage it, slow its development, and cost the team time it does not have.
That cable is not an admission of weakness. It is the most honest summary of the real state of French humanoid robotics in 2026: a genuine technical feat, held on the end of a rope, in a race where the best-funded European competitor has raised thirty-five times more, and where China’s Unitree took twenty-two times that sum in a single stock market day, on 19 August 2026.
What exactly happened, and the two names of the same robot
On 7 July 2026, at the Machina Summit held at Station F in Paris, UMA lifted the veil on its first prototype and on the learning method that goes with it. The specialist press took away one name: Northstar. The official visuals show a human-scale silhouette, a neutral visor with no eyes or mouth — a deliberate choice so that the machine is never mistaken for a person — and an interchangeable technical-textile covering, closer to clothing than to bodywork.
The caption accompanying the presentation was sober: “Prototype Version 0 — AI, software, hardware. A small team, 9 months. Designed and assembled in Paris.”
A television report broadcast in August 2026 adds a detail that appears, to date, in no specification sheet published by the company: that first prototype has just been named Astro, and it measures 1.60 m for 50 kg. We flag this caveat because it matters: those dimensions come from a journalist’s commentary, not a manufacturer’s document. No official UMA specification mentions degrees of freedom, battery life or payload to date.
That is already, in itself, information about the sector. A startup communicating on a vision rather than on figures is a startup that does not yet have figures to defend.
Why “standing up” is both a feat and a commonplace
Bipedal balance is one of the oldest problems in robotics. Honda was already working on it in the 1980s. What changes here is the method.
“It is thanks to artificial intelligence that we can keep this robot balanced,” the company explains. “It is an artificial neural network connected to the robot’s sensors, in particular to the motors. And we trained it to hold the robot upright.”
The distinction is technical but decisive. In the classic approach, an engineer writes the equations of walking and postural control: the robot follows a hand-programmed physical model. In the learned approach, nobody writes those equations. The controller is a neural network that discovered for itself, through trial and error, what commands to send the motors given what the sensors say. It is more robust in the face of the unexpected — uneven ground, a sideways push — and far quicker to develop.
It has also become, in 2026, a rite of passage rather than a competitive advantage. Chinese, American and European teams all know how to make a humanoid stand up and walk. Some make them do backflips.
UMA has the clear-sightedness not to play on that ground. Nobody there is trying to do kung fu or run a marathon. The stated ambition is more down to earth: to build a robot that is genuinely useful. Which, in a sector where a viral video stands in for proof, is already a position.
We have documented elsewhere how far that gap between the demonstration and real use shapes the whole market: in our wide-angle on what humanoid robots really do when nobody is driving them, a home robot sold at $20,000 had completed no task autonomously during a trial supervised by a journalist.
UMA’s real bet is not mechanical, it is pedagogical
“70% of what we do is AI,” the company says. “That is where progress is greatest right now. It is what will unlock new capabilities in these robots.”
The rest — 30% — is hardware. That split, which UMA has claimed publicly since July, says everything about its strategy: the company thinks of itself first as an artificial intelligence laboratory that incidentally builds robots, and not as a manufacturer adding AI.
Direct consequence: “We are developing an AI that is generalist, so it can be applied to a whole range of robots, not just humanoid ones. But we also co-design the robots to make them even better suited.”
Translation: the final product may not be Astro. The final product is the brain — and it is designed to run on other bodies. That is consistent with the first commercial milestone announced, which is not a humanoid at all: a wheeled robot, with dexterity and learning capabilities, targeted for late 2026.
A wheeled robot does not need to solve bipedal balance. It needs to handle objects without breaking them. That is a useful admission about what is ready and what is not.
”Real-Time Learning”: what it means, and what it does not yet prove
Here is the heart of the bet, in the company’s own words:
“Before even scaling up in terms of data and compute, we developed a method that is radically efficient, a learning method we call real-time learning, which is a human specificity, and which lets robots learn very quickly by watching a whole range of gestures a human can make.”
You have to understand what that sentence is aimed against. Since 2023, robotics has imported the recipe of large language models: pile up data, increase compute, let the capability emerge. That is the logic of the so-called vision-language-action models — π0 from Physical Intelligence, Gemini Robotics at Google, GR00T at Nvidia.
Except there is a structural problem, and it is well identified in the literature: there is no internet of gesture. The text of the web made it possible to train language models; no equivalent exists for the movements of an arm grasping a soft cardboard box. Every hour of manipulation data has to be produced by hand, robot by robot, through teleoperation. Robotics’ bottleneck is not compute, it is data.
Two strategies then compete. Nvidia generates synthetic trajectories with a world model to manufacture the missing data artificially. UMA takes the other path: not needing as much data.
The company has published no scientific paper describing its method, so we cannot verify it. But it does not come from nowhere. Remi Cadene and Simon Alibert, two of the four co-founders, launched and led LeRobot, Hugging Face’s open source robotics library — which went from zero to more than 12,000 stars on GitHub in its first year, and became the reference platform for open robotics.
And LeRobot incorporates one precise algorithm: HIL-SERL, published on 20 August 2025 by Jianlan Luo, Charles Xu, Jeffrey Wu and Sergey Levine (University of California, Berkeley) in Science Robotics. Its principle: start from a handful of human demonstrations, derive a reward classifier from them, then let the robot train for real while an operator steps in to correct dangerous or unproductive trajectories.
Its measured results are public: near-optimal policies in 1 to 2.5 hours of real-world training, with on average a doubled success rate and execution 1.8 times faster than classic imitation learning — on dynamic manipulation, precision assembly and two-arm coordination.
We are not asserting that UMA uses HIL-SERL. We note that the method the company describes — demonstration, practice, correction under an operator’s supervision, reward — belongs exactly to that family, and that two of its founders carried it into the open source ecosystem. It is the best publicly available indication of what “Real-Time Learning” covers.
To that is added a second claimed building block: a predictive world model, tasked with anticipating the future state of the environment and the behaviour of people around the robot. It is presented as a safety function — and it is also, as things stand, the least demonstrated part.
$40 million against $1.4 billion: the real balance of power
This is where the heroic story runs into arithmetic.
UMA raised around $40 million in seed funding, led by the fund Greycroft, with Relentless, Unity Growth Fund, Red River West, Factorial, Drysdale, ALM Ventures and Kima Ventures — Xavier Niel’s fund. Around them, a line-up that reads like a directory of European AI, as individual investors or advisers: Yann LeCun, Thomas Wolf (Hugging Face), Matthieu Cord, Olivier Pomel (Datadog). Around thirty people, split between the Paris AI laboratory, a “robotic hands” team in Geneva and a commercial team in London.
Now, the landscape around it:
| Player | Country | Money raised / valuation | Industrial status |
|---|---|---|---|
| Figure AI | United States | $39 billion post-money (September 2025), series C of more than $1 billion | Figure 03 deployed at the BMW plant in Spartanburg; the Figure 02 had accompanied production of more than 30,000 BMW X3s in ten months |
| 1X Technologies | Norway / United States | $820 million confirmed (series B); round targeting at least $10 billion not closed | NEO announced at $20,000, no verified customer delivery by mid-July 2026 |
| Unitree | China | listed on 19 August 2026: $904 million raised, around $50 billion market capitalisation at the first day's close | 5,632 humanoids and 33,294 quadrupeds shipped in 2025; €217.6 M revenue, €35.6 M net profit |
| Apptronik | United States | $935 million in series A ($520 M extension in February 2026), valued at around $5.3 billion | Apollo, logistics and industry |
| Agility Robotics | United States | listing by SPAC (Churchill Capital XI, agreement of 24 June 2026), $2.5 billion | Digit: more than 65,000 hours of operation across nine customer sites (GXO, Schaeffler, Toyota Canada, Mercado Libre) |
| Neura Robotics | Germany | up to $1.4 billion raised on 10 June 2026 | 4NE-1, target of "several million robots by 2030" |
| UMA | France | around $40 million | prototype on a harness, pilot programmes targeted |
A note is needed on the nature of these amounts: private valuations are figures announced at fundraising rounds, not market prices. Only one line of this table has been validated by real, public investors — and it is Chinese.
On 19 August, the stock market gave a humanoid its first price
On 19 August 2026, Unitree became the world’s first pure-play listed humanoid robotics company. The stock was introduced on Shanghai’s STAR Market at 150.80 yuan. It opened up 629%, then closed at 845 yuan, that is +460% on the very first session. Market capitalisation at the end of the day: around €43.8 billion, close to $50 billion. The operation raised $904 million for around 10% of the capital.
Two figures deserve to be set side by side. Unitree last year posted €217.6 million of revenue — quadrupling it — for €35.6 million of net profit, shipping 5,632 humanoids and 33,294 quadruped robots. The market therefore valued it, in one session, at around two hundred times its annual revenue.
That is the first public price ever put on humanoid robotics, and it is also a warning. Analysts quoted the day after the listing judged that the surge lacked rational grounding and came from retail investor enthusiasm — the offering had been oversubscribed around 8,000 times. Market forecasts, for their part, remain widely dispersed: Barclays puts the market between $10 and $25 billion in 2030, JPMorgan expects 1.75 million units by the same date, and Morgan Stanley projects $5 trillion of annual revenue in 2050. When the gap between forecasts runs to orders of magnitude, it means nobody knows.
That is the setting in which UMA raises $40 million.
Neura, the real European benchmark
The Neura case is the most instructive, because it is European. On 10 June 2026, the German company closed the largest robotics round ever raised in Europe: up to $1.4 billion, around €1.2 billion, with Nvidia, Amazon, Qualcomm, Bosch, Schaeffler, Tether and the European Investment Bank at the table.
In other words: the EIB put money into a German humanoid while the French champion was raising thirty-five times less.
That does not condemn UMA. Remi Cadene says he is not constrained by his funding and is in no hurry. And a small team very well placed on the software layer can outrun a heavy industrial player — that is even the story of Mistral AI against the American giants, which we examined in our investigation into who actually processes your data when an AI arrives at your bank or your doctor’s.
But it sets the nature of the bet. UMA cannot win by making more robots. It can only win by having a better brain — and one that can be sold for other bodies.
The French humanoid robot that is already working is not this one
This is the point coverage of UMA systematically leaves in the shade.
While Astro stands on the end of a cable in a laboratory, another French humanoid is already on a production line. It is called Calvin-40, it is built by Wandercraft — the French company known for its rehabilitation exoskeletons — and it has been on trial since February 2026 at the Renault plant in Douai, in northern France.
Its specifications are public: around 1.70 metres, a payload of 40 kilos, and no head — all the engineering is concentrated in the arms, the legs and perception. At the Douai site, it bends, plunges both arms in, picks up tyres and puts them on a conveyor. Around ten tasks are being trialled. Wandercraft says the robot doubled its execution speed in six months thanks to the training of its AI — a performance announced by the manufacturer, not independently verified.
Renault invested $75 million in Wandercraft in June 2025. The two partners are targeting around ten robots by the end of 2026, then 350 units by the end of 2027 in the group’s French and Spanish plants, across around ten use cases. Those too are announced intentions: no independent count of units actually installed is published.
This contrast deserves to be set out calmly, because it says something precise about the French sector:
- What France can do today: specialised humanoids, on tightly framed tasks, backed by a large industrial client that funds and absorbs the risk.
- What it cannot do yet: the versatile robot that changes task without reprogramming — precisely what UMA is aiming at.
Those are two different bets, on different horizons. Confusing the two means telling yourself France is further ahead than it is — or less.
The argument about 85 million missing workers does not say what it is made to say
The whole humanoid sector, UMA included, rests on the same bedrock of justification: Europe is said to be massively short of hands, and the robots would be filling a void, not taking a place.
The figure cited is almost always the same: 85 million unfilled posts in 2030, for $8.5 trillion of unrealised revenue.
That figure exists. It comes from a Korn Ferry study titled The Global Talent Crunch. But when you open its methodology, three elements change the reading:
- It dates from 2018. It therefore predates the pandemic, the inflation wave, and the arrival of generative AI — the three events that have most shaken the labour market since.
- Its scope is narrow. It covers 20 economies and three sectors: financial and business services, technology-media-telecoms, manufacturing. It does not measure warehouse handling, which is nonetheless the first market humanoids are aimed at.
- It is about skilled workers. The shortage it models is a skills shortage — engineers, technicians, managers — not warehouse operatives.
Finally, the modelling was designed by Korn Ferry, a recruitment firm, with Man Bites Dog, a communications agency, and carried out by Oxford Analytica. That does not make the study false. It means it was produced in a promotional setting, and that it does not have the status of a public statistic.
Using that figure to justify deploying robots in warehouses amounts to mobilising an engineer shortage to explain that order-picking posts are going to be automated. That is neither the same population nor the same job.
It is the same mechanism we took apart on employment: between what the studies announce and what the figures show, the gap is systematically in the story’s favour. We set it out in what the studies really say about jobs at risk from AI and in our analysis of the reversal behind layoffs “because of AI”.
They will not sell you the robot, they will rent it to you
The business model UMA announces is explicit: “You pay a monthly subscription to access a fleet of robots that automates a task. You do not buy the robot.”
That choice is not trivial, and it has three concrete consequences.
For the customer, it removes the purchase price barrier and turns an investment into an operating expense. A company can rent robots to absorb an order peak, then give them back.
For UMA, it guarantees recurring revenue — but it also transfers all the reliability risk onto the company. If the robot breaks down, it is no longer the customer’s problem: it is a revenue line that stops. It is a model that does not forgive technical approximation.
For the question of work, it is the most important and least discussed point. A robot rented by the month compares directly with a monthly wage cost. It is no longer an industrial investment amortised over ten years, but a quarterly budget decision. That makes substitution much easier to decide — and much easier to reverse.
Here too the legal subject is already open, and it does not arise the same way everywhere: we have shown why Chinese courts refuse to let a company dismiss on the ground that it has adopted AI, while French law puts technological change among the valid grounds.
What a humanoid robot costs in 2026
Since the question keeps coming up, here are the verifiable orders of magnitude, category by category:
| Category | Model | Indicative price |
|---|---|---|
| Entry-level Chinese platform | Unitree | around $16,000 |
| Announced home robot | 1X NEO | around $20,000, or $499/month |
| Research robot | Reachy 2 (Pollen Robotics / Hugging Face) | around $70,000 |
| Low-cost open source humanoid | Hope JR | under €3,000 |
| Open source desktop kit | Reachy Mini | from $299 to $449 |
| Rented industrial humanoid | announced UMA model | not disclosed, monthly subscription |
A warning is in order: these are announced prices or research platform prices. They include neither on-site integration, nor maintenance, nor team training, nor compliance work — which, for industrial use, often weigh more than the machine itself.
What France has, and what it lacks
The numerical setting is unflattering, and it is necessary in order to judge UMA’s bet.
According to International Federation of Robotics data, France has a density of around 195 industrial robots per 10,000 employees, which puts it around 18th in the world. Germany is above 400. In 2024, France fell back to fourth in Europe for annual installations, behind Germany (close to 27,000 robots installed), Italy (around 8,800) and Spain (around 5,100).
On public funding, the robotics strand of France 2030 mobilised around €800 million: half for industrial site projects, the other half for laboratories, technology platforms and innovative companies.
The imbalance is therefore double. France has recognised research and teams able to produce a working prototype in nine months. It has neither the installed base that would let it absorb these robots at scale, nor the private capital that lets a Neura aim at millions of units.
That is also why UMA’s “software first” positioning is rational rather than modest: it is the only layer where a French player can compete without a giant factory. A reasoning close to the one we examined about the Chinese strategy of giving the best models away free to occupy the ground rather than the market.
The five questions Astro does not yet answer
At this stage, here is what remains open — and what will have to be watched to know whether the promise holds.
1. Real autonomy. No public demonstration shows the robot carrying out a complete task, with no operator, over a measured period. The company mentions videos of AI performing industrial tasks “for hours on end”, but with no published measurement protocol. That is exactly the point on which American players were caught out by reality, teleoperation often hiding behind advertised autonomy — a mechanism whose most literal version we showed with those platforms where AI, unable to act physically, simply hires humans.
2. Regulatory safety. The machinery regulation (EU) 2023/1230 becomes mandatory on 20 January 2027. Any humanoid sold or rented to work near European employees will have to comply. It is a heavy constraint, and paradoxically a possible advantage for a European player that designs for it from the start rather than afterwards.
3. Industrial scale-up. Designing a prototype in Paris and producing hundreds of reliable units are two different trades. No industrial production partner has been announced.
4. Proof of transfer. UMA’s central argument is that a generalist AI will apply to several types of robot. That remains to be demonstrated publicly on at least two different platforms.
5. Customers. Around fifty potential customers are “in discussions”. No firm order has been announced. And a commercial discussion is not a purchase order, nor is a prototype that regains its balance under the lights a fleet running across successive shifts.
What to take away
UMA is genuinely good industrial news for France, and that should be said without reservation: four very high-level founders, a learning method consistent with the state of the art, a working prototype in nine months, and a software positioning that is the only rational one for a European player.
It is also a thirty-person company, funded thirty-five times less than its German competitor, whose first commercial product will not be a humanoid but a wheeled robot, and whose flagship prototype stays attached to a cable.
Both sentences are true at the same time. The problem in this sector is never that the prototypes are disappointing: it is that the story systematically runs faster than the machines. Remi Cadene says so himself: deploying humanoids at scale will take years, as the internet and the smartphone did.
To place that timetable within the wider landscape, our analysis of what really awaits AI in 1, 2 and 5 years applies the same method: separating what is already measured from what is still only projected. And for service robotics in public space, the French regulatory framework is already settling: we set it out in autonomous delivery robots and what France actually allowed.
Astro stands up. That is the easy part. The hard part starts when you take the cable off, put the robot in a warehouse, and count how many totes it moves in eight hours without anybody stepping in.
The figure to keep
$40 million against $1.4 billion. That is the gap between UMA’s seed round and the round closed in June 2026 by Germany’s Neura Robotics, with Nvidia, Amazon, Qualcomm, Bosch and the European Investment Bank at the table. The question put to France is not whether it has the engineers. It has them, and Astro proves it. It is whether it is willing to fund them at the scale of the race they are in.