Building the physical-intelligence layer for robots. 🤖

Beginnings

My love for robots came from stories, films and books all mixed together. I, Robot, Big Hero 6, Andy Weir’s Project Hail Mary, and Isaac Asimov’s robot collection, which I worked my way through. Asimov shaped how I think about machines more than any film has.

I stopped just watching robots on screen and started building them. I was 13. The competitions came soon after, and once I started winning, something shifted. It went from a hobby I happened to be good at to the thing I wanted to do for the rest of my life. I ended up winning 30+ of them at national level 🏆, across RoboWar, Robo Soccer, Maze Solver, Call of Duty, coding championships and mechatronics design.

A robot is really three things stacked together: something mechanical that moves, electronics to drive it, and code to decide what it does. I did not have access to a robotics degree, so I had to work out how to get each of those three another way.

I picked electronics for a simple reason. Code was free. I could learn it online, at night, on my own. Mechanics I could learn with my hands, by turning up at a workshop. Electronics was the one part that needed equipment I could never afford. An oscilloscope. A signal generator. A bench where I was allowed to break things. That kit lives inside a university, and you have to be a student to use it. So I studied the part I could not get anywhere else, and took a second degree in signals and communications alongside it.

For the mechanical side I started working in garages at weekends, fixing cars and bikes. I worked for free, because what better way to learn than the real thing. You pick things up there that are not written down anywhere: how a bearing sounds before it fails, why a bolt shears where it does, how much force is actually a lot. Between the lab, the garages and whatever I could teach myself at night, I put together the degree that did not exist.

The competitions were where the three came back together, and I kept entering them the whole way through.

Eventually I ran out of room. I had taken what the infrastructure around me could give, and going further meant better hardware and more robots than I could get near. That is what brought me to the UK.

Founding

I studied Advanced Robotics & Artificial Intelligence at the University of Bristol and spent years at the Bristol Robotics Laboratory, one of the biggest robotics labs in the UK. I pointed myself at AI, and robot learning in particular. Having what felt like every kind of robot under one roof meant I could finally try things that had been out of reach.

It started as a dorm room project. I wanted to build the best intelligent bionic leg I could, for people who had lost a leg to war, to accidents, or to any other physical trauma. It grew enough that I made it my MSc thesis.

The thesis drew far more attention than I expected. People started reaching out, and it became clear the work had outgrown a degree. So I turned it into a company.

I founded Chisel Robotics and ran it for five years, building that leg properly: a socket that understands fit, gait and movement in real time, with feedback under 50 ms. We validated it across four NHS trials and built performance legs for Paralympic athletes, including a Rio Paralympic gold medallist 🥇. It was backed by Innovate UK and grew out of NHS research partnerships.

The recognition followed. I was one of 23 people across the UK given a Young Innovators Award, Chisel was one of three businesses taken onto SETsquared’s Breakthrough Bursary, and there were a couple of Santander entrepreneurship awards and a disruptive-startup finalist place along the way.

That changed how I thought about robotics. The best robots don’t always look like sci-fi. Sometimes they just help someone get a piece of their life back.

Production

Long before manufacturing, I was training and deploying computer vision at scale in places where a wrong prediction costs real money. Each time it was a different family of models.

At Q-Free I built the perception for smart motorways. Not one model but several, fused into a single roadside pipeline doing detection, classification, multi-object tracking, speed estimation and traffic-flow analysis across four, six and eight-lane roads in the UK and the US, all of it optimised to run on NVIDIA Jetson and Xavier at the roadside rather than in a data centre. On Xavier that held at 99.7% detection accuracy, 93.5% mAP and 65+ FPS, deployed across ten roadside units.

At BCA I led vision for automated vehicle inspection, and it was the first time I trained a domain foundation model rather than a set of task-specific ones, in that case for vehicles: damage detection, interior classification, segmentation and depth-based post-processing, and a text-to-image diffusion pipeline that synthesised the rare damage cases real data never covered. I designed the AWS infrastructure that ran it in production, which cut inspection from roughly four hours to under 90 minutes per car and contributed to a 1.3% complaint rate.

Demos are easy. Systems that run every day, on real hardware, in bad weather, with nobody watching, are not. By the time I reached manufacturing I had shipped detection, tracking, segmentation and generative models into production. What came next was a genuinely different modelling problem, but the discipline of making a model survive contact with a real site was the same, and that part takes the longest to learn.

Perception

I went after perception on purpose. A robot cannot act on a world it cannot read, so scene understanding is the step everything else waits on, and getting it wrong leaves even a good policy guessing.

Perception has been the constant since the TIAGo robot at the Bristol Robotics Laboratory, where I wrote the object recognition and ran SLAM-based navigation for the European Robotics League. Manufacturing asked a harder question of it. Detection tells you what is in a frame; understanding a production line means knowing what a person is doing, and for how long.

I joined Almetra (previously Deltia) as a 📷 Staff Computer Vision Engineer, and was later promoted to 🤖 lead robotics and build that division from the ground up. The vision work came first, and it went to production scale. I trained action recognition models and a 7B vision-language model for spatio-temporal action segmentation: show it a video and it breaks the work down action by action, over short and long time horizons.

The harder part was making that hold up at scale. I ran large-scale pre-training on real production video to build a cross-customer foundation model, then distilled the teacher into compact student models small enough to run on-station. Those VLMs went into production across hundreds of stations and 60+ factory sites for ABB, Bosch, Continental, Thermo Fisher, Viessmann, Grundfos and Siemens Energy, among others, helping teams see patterns on the factory floor that were previously invisible.

Deployed with

That work helped carry Almetra from seed stage through to its $19M Series A, and it became the foundation the robotics programme grew out of.

Physical AI

That move happened in January 2026. As Robotics Lead at Almetra I build the division alongside our Co-founder & CTO, and in practice it is a startup inside the startup: setting the research direction, choosing the stack, building the team and owning delivery. It is close to the job a founder does, on the robotics side of the company. I had done a version of it before at Chisel Robotics, which helps more than I expected.

What I architect is the perception-to-action stack, turning that understanding of real manufacturing work into policies that can plan and execute physical tasks.

Two programmes took that work further. Google DeepMind selected 15 early-stage companies from 10 European countries for the inaugural cohort of its Gemini Robotics Accelerator, running June to September 2026. MassRobotics took 9 startups into the 2026 Physical AI Fellowship with AWS and NVIDIA. I lead the technical work on both.

Through the accelerator I have had the chance to run the Gemini Robotics family on our own hardware: Gemini Robotics-ER for embodied reasoning and planning, the vision-language-action model for control, and the on-device model for the cases where latency or a dropped network makes a round trip to the cloud unsafe.

Working directly with the robotics teams at Google DeepMind, AWS, NVIDIA and MassRobotics was never about the badge. The point was getting pilots onto real production lines, at customers I already knew from the vision side: same sites, same processes, same people on the floor, now with a robot in the loop. One of those pilots has now run through to completion, with a customer I am not able to name.

The MassRobotics side took me to Boston, touring the incubator and representing the company at the Robotics Summit & Expo.

Three companies of perception work, across motorways, vehicles and factories, is what made the action side approachable.

The work itself is robot learning, and it is still the part that feels magical: watching a robot acquire a behaviour instead of being told every step. Once a system can read a line reliably, the question becomes what to do with that. Can you go from pixels straight to actions in a single network, or does the answer sit somewhere in between?

None of it is settled, though. General-purpose robotics is a long way off, and no single model solves every line, so I work customer by customer and test what survives contact with their process. That means pulling from frontier labs and open source alike: manipulation policies trained on human demonstrations and interaction data, diffusion policies for contact-rich tasks, vision-language-action models, world action models, and classical control like high-frequency impedance control where physics demands a guarantee rather than a learned guess. Every site has a different process and a different tolerance for failure, and most of them need handholding to get there. So far the honest answer is that the best systems sit in the middle. Robots need learning, but they also need structure, safety, feedback, and respect for physics.

I chose manufacturing because the economics are legible. A line has a known cost per hour and a known cycle time, so a robot either pays for itself in the customer’s own numbers or it does not. No market forecast required, which is a rare luxury in robotics and the reason I would rather prove this on a production line than in a demo video.

A demo proves a robot can do a thing once, on a good day, with someone watching. Production means it runs for months on a line that does not stop for a dropped frame or a slow policy. That recipe has not been written yet, and closing that gap is the work I care about.

What actually works at scale?

If it does not hold at scale, it does not matter. That is why I do this on real production lines rather than in a lab.

Fun facts

I’ve been a student athlete my whole life. If I’m not building robots, I’m in the water.

• Started competing at eight, and somehow won a gold that first year 🥇.
• Kept at it until I finished my Masters, picking up a few more along the way, then retired.
• Swam for my university at national level a handful of times.
• Long sea swims of 10, 15, 20 and 30 km 🏊, in the Indian Ocean and off England.
• These days it's open water swimming 🌊 and scuba 🤿, and I'm working toward the professional side of diving.
• PADI certified to 30 m: Open Water, Advanced Open Water, and deep and technical diving.