This Robot 'Spine' Could Slash R&D Time by Months

What we talked about.

In this episode, Sunny Ray talks with Raphael, founder of Tekton, about building the 'spine' for field robots. Raphael traces his path from a master's degree in electrical engineering and computer science to a chip design role at Synopsys, roughly five years at National Instruments, and then eleven years running his own systems integration business serving machinery makers in oil and gas, automotive and academia. Around 2022 to 2023, watching AI move into the physical world through synthetic data platforms and faster edge chips, he saw an opening to turn years of custom engineering into a single product. Tekton's hardware combines an Nvidia AI compute module for perception and reasoning with a fast AMD FPGA chip for deterministic motor control, mirroring a fast and slow thinking system. The company delivers free custom designs within a month or two and lets robot builders program via natural language, aiming to compress R&D timelines for field robots in agriculture, construction and heavy industry rather than humanoids.

A robotics founder explains how a dual brain hardware 'spine' could shrink years of robot R&D down to months.

The questions, and the answers.

Before you got into robotics, what does your backstory look like?

I'm an engineer by training, with a master's in electrical engineering and computer science. My first job was as a chip design engineer at Synopsys, then I worked about five years at National Instruments. In 2011 I started my own systems integration company, delivering custom automation solutions to machinery makers in oil and gas, automotive and academia.

What is enabling this leap from dumb machines to fully autonomous robots, what's the fundamental breakthrough?

The physical world is hard because it combines mechanical, electrical and software challenges. What changed in the last five years is AI entering the physical world, helped by synthetic data platforms like Omniverse that let us train models to understand physics. At the same time chips like Nvidia's Jetson became faster and more accurate, giving robots enough compute to sense, perceive and act in real environments.

Can you talk about the journey from a service and consulting business to seeing this product opportunity?

My co-founder and I noticed robot builders spend months or years designing the custom electronics inside every machine, and that R&D time could be drastically cut. Unlike factory automation, field robots need a permanent, non modular brain because once deployed they run for twenty years with the same sensors and actuators. We saw nobody was building an AI enabled version of that custom system, so we built it.

Are we talking about humanoids at all, or a specific type of robot?

We focus on industrial and field robotics, not humanoids. It's really a business decision rather than a technical one. Humanoid or high volume programs like Tesla can justify designing their own custom silicon internally once unit volumes pass a million. Field robotics like mining equipment, wind turbines or MRI machines never reach those volumes, so a universal spine actually makes economic sense there instead.

Can you give us the TLDR of that recent A16Z article on physical AI?

The piece argues there's a huge gap between a robot working in the lab and one reliably deployed in the field, and today's technology isn't ready to close it. It breaks down why, including the fact that vision language action models that perform well in the cloud run far too slowly, often only twenty hertz, when compressed onto edge compute, which isn't fast enough for real world control.

In this setup, is the robot talking to a central AI data center, or is it fully tethered and separate from the internet?

Take an autonomous crane picking up bricks for a building under construction. A human still sets the top level goal, like where to place the load, but the moment to moment joystick style control happens onboard without needing the cloud. Connectivity still matters for monitoring things like a remote oil pump or wind farm through a SCADA system, but not for the real time control loop itself.

So isn't it a spine with a brain also, not just a spine?

We call it a spine rather than a brain because we didn't build the Nvidia chip, that's removable and sits on our board as the reasoning brain running vision language models. Our system also includes a second AMD FPGA chip doing extremely fast, deterministic motor control, like an old style PI controller. Together they're the fast and slow thinking system that plugs into a robot's sensors and actuators.