A new hire watches you do a job once, then copies you. A robot, until now, has needed a coder to spell out every move. A Paris startup founded by a Tesla veteran says its humanoid learns the first way, by watching. It is a bold promise, and the hardware does not exist yet.
Think about the last time someone new started where you work. You did not hand them a thick manual and walk off. You showed them. Watch me once. Now you try. I will fix what you get wrong. By day's end they were useful.
Now try teaching a robot that same task. You cannot show it anything. You have to tell it, in code, exactly where to move its hand, how fast, how hard to grip, and what to do when the part sits crooked. Change it even slightly and you write it all over again.
That gap, between showing and telling, is the central problem in robotics right now. A Paris startup says it has built a robot that learns the first way.
The robot that watches you work
This July, at a technology summit in Paris, a company called UMA, short for Universal Mechanical Assistant, pulled the cover off a robot named Northstar. It is a lightweight humanoid with a blank visor for a face. The body is not the pitch. The brain is.
UMA calls it Real-Time Learning. A person does a task in front of the robot. The robot watches, copies, and improves each time it repeats the motion. No engineer scripts the steps.
The founder is Rémi Cadène. He spent three years at Tesla on the AI behind Autopilot and the first neural networks for its Optimus humanoid, then led LeRobot, an open-source robot-learning project, at Hugging Face. UMA has raised money from well known venture funds and lists Yann LeCun, a giant of modern AI, as an adviser. It says it is in talks with roughly 50 potential customers and wants to run trials in factories, warehouses and hospitals.
That is the news. Why it matters runs deeper, because teaching a robot to watch and copy is both very hard and very valuable.
How robots are taught today
Almost every working robot on the planet is taught in one of three ways. None of them is watching.
Scripting. An engineer specifies every movement by hand. Fast and rock solid on an assembly line, where the same part arrives in the same spot a million times. It falls apart the instant anything is different.
Teleoperation. A human wears controllers and puppets the robot through a task, often hundreds of times, to record examples for it to copy. It works, but it is slow and ties up a skilled operator for every robot.
Simulation. The robot practises millions of times inside a video-game copy of the world, then takes what it learned into the real one. Cheap to run, but the simulation is never quite reality, and that mismatch, the sim-to-real gap, is where many trained robots stumble.
The common thread: every method needs an expert and hours of setup before the robot is useful. The expert is the bottleneck, and it is expensive.
What "learning from demonstration" really means
The dream is to delete that bottleneck. In the field it has a name: imitation learning, or learning from demonstration. Instead of coding a task, you do it, and the robot works out the goal you were aiming at and builds its own movements to reach it.
The word carrying the weight is generalise. A scripted robot taught to lift a red cup from the left knows nothing about a blue cup on the right. A robot that has truly learned the idea of lifting the cup can handle one it has never seen, in a spot it was never shown. One is memorising; the other is closer to understanding.
Get this right and the economics flip. A robot stops being a machine that needs a programmer for every new job and starts to look like a new hire: show it, and it learns. That is why watch once, then do it is the holy grail, the line between a single-purpose tool and a general worker you can point at almost anything.
Why Europe, and why light
Most of the loud names in humanoid robots are American or Chinese: Tesla, Figure, and well funded Chinese firms. A credible European entry is rare, and this one is run by someone who helped build Tesla's robot before leaving to do it his own way. Cadène says Europe comes first, ahead of the US or Asia. For a region that worries about relying on foreign technology, a home-grown humanoid is a statement.
Then there is lightweight, which reads like a dull spec and is really a choice about safety and cost. UMA says Northstar is built to be light, a target it puts near 40 kilograms, well under many rivals. A heavy robot that topples or swings an arm into a bystander is dangerous; a lighter one carries less force and is easier to make safe around people. Less metal, smaller motors and a smaller battery also cost less to build and run. If these machines are ever to work beside humans in a hospital corridor or a warehouse aisle, light is not a luxury. It is the point.
The honest catch
Now the cold water.
Northstar is a reveal, not a product. What UMA showed is a "version zero" prototype, built by a small team in about nine months. Nothing is shipping. The road from a slick stage demo to a robot that works a full shift with nobody hovering nearby is long, and littered with companies that did not make it.
And learns by watching is far easier to say than to prove. Imitation learning is real and improving fast, but doing it reliably, in a cluttered real room, on tasks nobody rehearsed, is one of the hardest open problems in robotics. A smooth demonstration can hide a lot of retries and careful staging.
The rest of the numbers are the company's own. In talks with about 50 customers means conversations, not orders. The 40 kilogram figure is a design target, not a measured spec. Take it as the ambition of a serious team, not a machine you can buy today.
EDITOR'S TAKE
The interesting thing about UMA is not the robot. It is where its founder came from and what he chose to build. Cadène helped teach Tesla's humanoid, then bet that the winning move is not a stronger body but a faster-learning brain, and that Europe should own one. He may be right that learning by watching is the switch that turns humanoids from costly demos into real workers. He is also selling a promise, on a stage, with no product in a customer's hands, which is exactly how this field has burned believers before. Watch the pilots, not the reveal. The day a Northstar picks up a task nobody scripted, on a real factory floor, on an ordinary weekday, is the day this stops being a pitch.
Quick questions
How is UMA's robot different from Tesla's Optimus or other humanoids?
The body is fairly ordinary: a lightweight humanoid with two arms and a blank visor instead of a face. The claimed difference is how it learns. Rather than being hand-programmed or trained mostly in simulation, UMA says Northstar picks up new tasks by watching a person perform them and then practising, an approach called imitation learning or learning from demonstration. If that works reliably, one robot could take on many different jobs without an engineer rewriting its code each time. For now, though, it is a prototype, not a product you can order.
What does "imitation learning" mean?
It is teaching a robot by example instead of by code. A person performs a task, the robot observes the demonstration, infers the underlying goal, and generates its own movements to achieve it, improving with each repetition. The real prize is generalisation: a robot that grasps the idea behind a task can handle new objects and new positions it was never explicitly shown, instead of blindly repeating a single memorised motion. It is one of the most active and most difficult areas in robotics today.
Can I buy a Northstar?
No. What UMA revealed on 7 July 2026 is an early "version zero" prototype, not a product on sale. The company says it is in talks with roughly 50 potential customers and plans pilot programmes in manufacturing, logistics and healthcare, but those are trials, not orders. The customer count and the robot's specifications, including its weight, are figures stated by the company and not yet independently verified. Expect years between a reveal like this and a robot working dependably alongside people.
Sources
Electrek: ex-Tesla Optimus scientist unveils UMA and its Northstar humanoid, 7 July 2026.
Bloomberg: ex-Tesla scientist unveils plans for a European humanoid robot, in talks with about 50 customers, 7 July 2026.
The Next Web: Rémi Cadène launches UMA to build Europe's humanoid robot, July 2026.
Interesting Engineering: how UMA's real-time learning lets Northstar learn from human demonstrations, July 2026.
Related from Frontier Signal: last week's deep dive on why GPS is really about time. Frontier Signal explains frontier technology in plain English. This is general information, not investment advice.
