Bertie Highmore

Science is done by hand. We’re giving it eyes! I’ve always been a science fanboy. After Covid, Chris and I started looking at deploying robots into labs, and we were blown away by what we found: there was no foundation for powerful, flexible automation that could actually support human scientists. Machines couldn’t see or understand the work. So we set out to build that foundation. Here’s what we’ve been up to at @lumi_systems We’ve built a family of ultra-lightweight, high-performance vision models that understand lab work as it happens: every action, every step, every sequence, down to the fine-grained hand movements at the bench. They run in real time, on compact hardware, inside working labs, and turn what happens at the bench into structured data. Why that matters: we instrument everything in the lab, the machines, the samples, the readouts, but the work itself has stayed almost entirely unseen. You can’t automate what you can’t see, and you can’t train physical AI on work that was never captured. That data: → shows what really happens, not what the protocol says should happen → gives automation a ground truth to learn from and build on → creates an independent layer of understanding across humans, instruments and robots. This is already live. Lumi is running today in CROs and in labs handling the most valuable samples there are: IVF embryos and human tissue, where there’s no second chance. Over the coming weeks I’ll share real footage of our models at work, what we’ve learned building them, and where I think physical AI in the lab is heading. If you build robots, run a lab, or think about the future of scientific work, let’s chat! #PhysicalAI