Physical Vibe Coding: Our Investment in Machenta
When I first met Ed Johns and the Machenta team, Ed said something that now sounds obvious but has shifted my thinking: intelligence isn't about existing scale of knowledge and capability. It's how quickly you can learn to do something new. By that measure, the most impressive thing about humans isn’t our set or depth of skill but how well we can pick up new ones.
Bringing intelligence into the physical world will be the defining innovation of the coming decades, making existing industries dramatically more efficient and creating new ones we haven't thought of yet. Watching the world come to life through automation will be the hallmark of the era we're entering.
Much of that depends on robots that can learn with both speed and scale. So far, robotics has primarily run on a belief in abundant data and the correlation between scale of data and strength and breadth of skill. More data collected, more days of demonstration, more hours of training, better robots. Every new skill starts another round of collecting.
Machenta is pioneering a different approach, modeled on how people actually learn. For many tasks, we only need to see something done once to know how to do it (anyone with a toddler has marveled at this.) This ability to learn, and learn quickly, is a deeper form of intelligence and unlocks whole new sets of capabilities and pace.
Machenta brings that to robots. Show the robot a task once, and it starts doing it. The model is trained entirely on synthetic data, so there's no real-world training data to collect at all, let alone days of it. With this approach, we can bring robots to life in deployment settings far faster and cheaper, across a huge array of tasks and formats.
The Machenta team's videos are worth watching: robots completing tasks with objects they have never seen before, after a single demonstration. And if you happen to be in London, go try it for yourself and see how quickly you can teach a new skill.
We've watched easy-to-use LLMs unlock use cases no one could have dreamed up in advance. The same is now on the horizon for the physical world. We have been investing heavily behind the idea of data at the edge and the net new opportunities unlocked when we can access, process, and utilize new formats of data in novel ways. With Machenta, when robots are fast and easy to train, imagination instead of capability becomes the constraint. The team calls this “physical vibe coding.”
Ed and his team are world experts in this approach. As Director of the Robot Learning Lab at Imperial College London and Head of Robot Learning at Dyson, Ed invented the methods and concepts they are now building into accessible, widely usable models.
We're thrilled to co-lead Machenta's $40M seed round alongside our friends at Creandum, and to work with Ed and the team as they build and release this model to bring fast learning to the next era of robotics.