Machenta exits stealth with $40M and one-demo learning
The robot foundation model was trained without a single pixel of real-world data, the founder says.
Machenta came out of stealth with a $40m seed round co-led by Creandum and USV. Founder Edward Johns announced it on X, pitching what the company calls Instant Robot Learning, a foundation model that he says learns a new task from one demonstration with no further training.
The part that will get argued about is where the training data came from. Johns says the model was trained entirely in simulation, "without a single pixel of real-world data." That is a sim-to-real claim, meaning everything was learned in a simulator and then run on real hardware with no real robot data in the loop.
What the launch thread shows
Johns posted four demo categories. One shows the robot reacting to new situations and recovering from disturbances using a closed-loop, end-to-end policy. Another, which he calls a one-shot, one-person factory, shows three types of robots working together on a task, with Johns saying the same model works on any robot. A third covers steerability, where the same set of objects can be handled three different ways depending on what you show the robot. The fourth he calls physical vibe coding, where the team found a "double handover" strategy by iterating through different physical prompts rather than programming it.
Lukas Ziegler, an angel investor in the round, described that shoe demo as the robot passing a shoe between its hands twice to flip it the right way around, and said the team did not pre-program it, they tried things until it worked. Ziegler also called the round the largest robotics seed ever raised in the UK, and named Johns and @vitalisvos19 as founders out of a team that spent years on the concept at Imperial College London.
What is not verified
Everything above comes from the company's own launch thread and from an investor in it. Nobody outside Machenta has tested the one demonstration claim, and the thread does not say how many tasks, how long the demos run uncut, or what the success rates are.
Ted Xiao, who is not listed as an investor, offered outside praise for the direction rather than the results, saying he has been a fan of demos as a test-time modality, meaning the demonstration itself is the input you give the model at the moment you want a task done, instead of a text instruction.
Shown versus claimed matters here. The videos are what the company chose to post. The interesting bet is the sim-only training, because if the gap really is closed that tightly, a new task becomes a prompt instead of a data collection project.
A model trained entirely in simulation, without a single pixel of real-world data.
Teach a robot a new manipulation task in a minute. Not hours. Not days.
I’ve been a big fan of demos as a test-time modality
Sound like you were at the demo
One email a day. Under three minutes. The demos, the deployments and the part the press release left out.