Machenta raises $40M seed to teach robots from one demonstration
A London lab of eleven people raised $40 million before it has anything to sell. Machenta's model is pretrained only in simulation and learns a new physical task from a single demonstration.
Also see more funding coverage: /funding/
Founded 2025 · London, United Kingdom
Machenta, a London robot-learning company spun out of Imperial College London, has raised a $40 million seed round co-led by Creandum and Union Square Ventures to build a model that teaches a robot a new physical task from a single demonstration.
Forty million dollars and eleven people
The round was announced on October 27, 2026, and it is how Machenta came out of stealth. Creandum and Union Square Ventures co-led it. No other investors were named in the announcement.
Get the Weekly brief
The week’s rounds, last calls and events, every Monday.
The money goes to the team, to compute, to generating more simulated training data, and to getting the model through tasks that need precision. The company is moving into a new office and lab in King’s Cross, London. Tech Funding News counted eleven people on the day of the announcement (Tech Funding News, October 27, 2026).
Machenta was founded in 2025 by Edward Johns, professor of robot learning at Imperial College London, and Vitalis Vosylius, his former PhD student, now chief technology officer (Tech Funding News, October 27, 2026). The research under it goes back six years to Johns’s lab at Imperial.
One demonstration, then the robot gets on with it
The model is pretrained entirely in simulation, with no real-world data in the pretraining set. Johns describes it as vision-demo-action: it takes one demonstration plus a live camera feed and predicts the next action in a closed loop, with no task-specific retraining in between. Rival approaches generally need hundreds or thousands of human demonstrations and days or weeks of training (Tech Funding News, October 27, 2026).
Johns’s early work at Imperial found that a robot shown a task still needed roughly thirty minutes of extra data collection before it could run on its own. Machenta’s claim is that the thirty minutes are gone.
The company reports hour-long trials on three simple self-resetting tasks: 74 of 74 block placements, 78 of 78 screwdriver handovers, and 97 of 99 attempts at standing a bottle upright, with the two misses attributed to collision risk and kinematic limits (Machenta figures, reported by Unite.AI, October 27, 2026). Those are the company’s own numbers on easy tasks, and Machenta says as much itself: precise motions and complex movement remain hard, and it calls itself a research company with no near-term plan to commercialize. One model has been shown on packaging, assembly and tidying across several robot types, humanoids included.
“Intelligence is not so much about what you can already do, it’s about how quickly you can learn to do something new,” Johns said in the announcement.
The European robot market is growing slower than the checks
Europe’s industrial robotics market was worth $6.9 billion in 2025 and is forecast to reach $13.2 billion by 2033, an 8.3% CAGR over 2026-2033 (Grand View Research, June 2026). That is a real market growing at a solid industrial pace. The capital is moving considerably faster: global venture funding into physical AI reached $47.4 billion across 521 deals in the first half of 2026 (Crunchbase, cited by Tech Funding News, October 2026).
Our own European funding records show the same squeeze. Five days before Machenta, Enact Intelligence raised a $30 million pre-seed for robots that sense force. In April, Sereact closed a €93 million Series B for robot learning in warehouses. A $40 million seed now sits between a pre-seed and a Series B, and the company it funds has no product to sell.
What a simulation-first bet is telling us
So what does an investor see here that the robot market does not yet show? The honest answer is that the expensive part of robotics has always been data collection, and almost every well-funded competitor is paying for it. Physical Intelligence and FieldAI are Machenta’s named rivals; Skild AI was valued at over $14 billion after a $1.4 billion round in January led by SoftBank (Tech Funding News, October 2026). One American startup, Maxinsights, raised $30 million to pay more than 5,000 contractors to film themselves working, purely as robot training data.
The good news for Europe is that simulation is cheap and compute is buyable, and a lab of eleven people in King’s Cross can compete on that axis in a way it never could against a 5,000-person filming operation. Imagine if the cost curve of robot learning bends inside a simulator. The pattern we keep seeing in European deep tech, where a university spinout automates what everyone else is doing by hand, would find its clearest case yet in robotics.
After ten years watching European research labs hand their best work to American balance sheets, I read the Creandum and USV co-lead as the interesting detail. Two investors, one European and one from New York, wrote a seed check this size into a company that will not sell anything soon. What to watch is whether the benchmarks hold up when someone else runs them, on tasks the robot has not been designed around.
Eleven people in King’s Cross, no product, a model that learns in one shot. The goal was never to out-film the Americans on data collection. Let’s see what Europe can build when it refuses to.
