Machenta raises $40 million seed to teach robots a task from one demonstration

Machenta, a London startup spun out of Imperial College London, has raised a $40 million seed round co-led by Creandum and Union Square Ventures. Imperial says it is the largest seed round ever raised by a UK robotics startup. Machenta declined to disclose its valuation, according to Business Insider.

The money backs a bet that runs against most of the field. Machenta pre-trains its robot foundation model entirely on simulated data, "without a single pixel of real-world training data", and says a robot can then learn a new task from one demonstration with no further training. Business Insider reports the company has 11 employees, remains focused on research and has no near-term plans to commercialize.

A demonstration replaces the language prompt

Machenta calls its model a vision-demo-action (VDA) model. Where a vision-language-action model is conditioned on images and a text instruction, the VDA model is conditioned on images and a single demonstration. Machenta says VLA models are usually post-trained on many real-world demonstrations, which can take days or weeks per task.

In the company's videos the same model runs on six different robots and accepts four teaching modes: leader-follower, VR controller, kinaesthetic teaching and a person's own hands. One clip shows three robots, each with a different gripper and camera setup, automating an assembly line within two minutes. Machenta's answer to the sim-to-real problem is that simulation is used to learn how to learn a task, while the single demonstration comes from the real world.

The published test results are on simple tasks

To stress-test the model, Machenta ran self-resetting tasks for an hour each over a weekend. The robots completed 74 of 74 block placements into a bowl, 78 of 78 screwdriver handovers and 97 of 99 attempts at standing a bottle upright. The company calls these tasks "relatively simple". Business Insider adds that more precise tasks and complex movements remain challenging.

Some rivals already sell. Business Insider says Skild AI, valued at over $14 billion, unveiled a model in August that it said could perform unseen tasks lasting up to 10 minutes after a single video demonstration, and last month reported $100 million in annual recurring revenue from more than 60 customers. Physical Intelligence is valued at roughly $11 billion, and FieldAI is raising at a $10 billion valuation, according to the same report.

Six years of lab work before the company

The research began six years ago in the Robot Learning Lab of co-founder and CEO Edward Johns, a professor of robot learning at Imperial. He founded Machenta late last year with his former PhD student Vitalis Vosylius and colleagues from the lab.

Machenta will spend the round on more compute and simulated data for its model, and on a larger team at its new office and lab in King's Cross, London. "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.