Backing Machenta: the era of Instant Robot Learning

We never expected Creandum caps to become a diligence tool.

When we visited Machenta’s lab earlier this year, we got impressive demos. But the real test was bringing our own tasks to see how the robot would react to items it hadn’t seen before. So we turned up with our own props: a handful of Creandum caps, a roll of paper and a Chinese lucky cat. We showed the robot the task once, with one object. It executed flawlessly - even when we moved the objects around. 

That moment captured what Ed had been telling us since our first meeting, with a question that sounds disarmingly simple but, in robotics, is deeply contrarian: humans don’t need post-training, so why should robots?

Most of the field has spent the past few years asking a different question: how do we collect more data? Companies pay people to teleoperate robots from across the world, or to film themselves doing chores at home. With most models, teaching a robot arm to stack a few blocks still takes hundreds of demonstrations, a training run, and days or weeks of waiting. But if you think about how a child learns, you show a young child once, and often they can immediately conduct the task.

Six years ago, when this idea sat well outside the mainstream, he made it the focus of his lab at Imperial College London. Their research made it to the Science Robotics cover paper, in which a robot learned 1,000 everyday tasks in under 24 hours from one demonstration each, and won Best Paper Award at ICLR 2025 Robot Learning Workshop.

We believe the hardest problems are often solved by people who bring a different lens. Ed is a true polymath, and his lens is a deeply human one. The son of two artists, he spent a year as a singer-songwriter in London after his PhD, playing gigs alongside a then-unknown Ed Sheeran, before going on to found Imperial’s Robot Learning Lab and lead robot learning at Dyson. In a band, when someone shows you a new chord, you rarely need to see it twice. Perhaps that’s why Ed has always started from a different question than most of his field: not how to make robots learn more, but how humans learn so fast.

Ed is perfectly complemented by his co-founder Vitalis. Originally from Lithuania, he was Ed’s most senior PhD student and was first author of the Instant Policy paper that brought in-context learning to robots. When the two started Machenta, their research team voted with their feet and almost the entire lab came with them. 

What they’ve built inverts the industry’s logic. Machenta’s foundation model is trained entirely in simulation, not to learn every task a robot might one day meet, but to learn how to learn. If you show it one demonstration, the robot does the task straight away, adapting to new variations with no further training. While others are bounded by how much real-world data they can collect; Machenta is bounded by compute. And because the learning is instant, every Machenta video shows the teaching and the robot’s first attempt in a single uncut take.

Itl entirely changes who gets to teach a robot. If showing it once is enough, you no longer need to be a robotics expert. Instead anyone can walk up, show it what they want, and the robot gets to work: the physical equivalent of vibe-coding.

We’re proud to co-lead Machenta’s $40M seed round alongside Union Square Ventures, and to back Ed, Vitalis and the team as they build from their new home in King’s Cross to bring robots to the people.