Analog's Mechatronics Engineer: Hardware Powering AI Inference | Researcher Conversations at GTC

Everyone, welcome back. I’m here for SemiAnalysis and Cell Media at GTC 2026 with Misha F Musa. We’re going to have a little chat about what’s new.

Thank you.

You had an interesting demo on the show floor this week, and I know you’re part of a new division within Analog Devices. Can you give me a high-level summary of what you’re working on?

I’m with the Emerging Tech Hub at Analog Devices. We’re a relatively new group focused on robotics and edge computing within the company. We’re a 60-year-old company providing semiconductor and chip-level solutions. However, our group works a little bit higher up, at the system level, trying to see how we can integrate our existing technologies into robotic solutions and push the edge of technology there.

Cool. Take me through some of the concepts or ways you see robotics being useful in these applications.

Specifically, talking about one of the demos we’re showcasing here at GTC, we’re using our in-house-built tactile sensing solution. Essentially, it’s a sensor that fits into the fingertips of robot grippers or hands, or whatever types of dexterous manipulation we’re trying to accomplish. It gives us a sense of the forces we’re applying as we grasp objects. It allows us to be delicate when we need to handle, for example, fiber connectors in a data center application.

What’s nice about the sensor as well is that it’s multimodal. Not only are you getting force feedback and pressure sensitivity, but it also has accelerometers and some vibration sensors that can give you a better picture of what’s happening during that interaction.

So, literally plugging in cables?

That’s one of the applications. We work very closely with a lot of different partners on our team. It’s a co-innovation space within Analog Devices, and one of the key applications we’re noticing is trying to automate some of the maintenance that’s done inside data centers. That includes going in and disconnecting cables, polishing fibers, and all of that requires a high level of precision as well as sensitivity to these types of connectors.

You’re going to need a robot using one of the clickers right now to clean?

At the moment, no, but that’s the step we’re trying to push toward. We’re also trying to engage more with the robotics community. We’re developing a set of benchmarks that will be physical boards analogous to the types of problems we see in data centers, automotive cable-harness assembly, gearbox assembly, and so on.

It’ll all be released to the public as open source, so researchers and people in industry can work on these problems and share knowledge about how we go about solving them.

But wait, the benchmark leaderboard is going to be a robot-in-a-room setup?

Not necessarily the robot. You’re welcome to choose whichever robot you want to use to solve the problem.

It’s just a board that sits on the table, and the robot can interact with it. You program it however you choose, whether that’s using classical control or newer AI-driven policies, in order to solve these problems and hopefully allow others to learn from what you’re doing as well.

A little bit different from grade-school math benchmarks, though, I think.

A little bit. What’s nice about what we’re doing, too, is that not only are we providing the physical device itself, but we’re also working with key partners to develop high-fidelity simulation assets that allow people to train their policies in simulation.

We’ll be able to provide simulations for our sensors as well, such as our tactile sensing or time-of-flight solution, which is able to create a depth scene, or a depth estimation, of the scene we’re looking at. All of that can be used together to create really nice robotic solutions.

What does this look like as a product that you guys would sell? Would it be a software product people can run on generic robots, or would you sell the robot itself?

We’re not a robotics company by any means. We provide the solutions that enable roboticists to do what they do best. We control a very precise signal chain all the way down through our analog-to-digital converter, CMOS sensors, and now tactile sensing and time-of-flight sensing.

We’re really pushing the edge of what these devices can do to enable roboticists to solve these problems using our sensor and our other types of technologies.

Got it. Makes sense. Where do you think you go from here? More benchmarks, more examples?

I think robotics is a long-standing challenge in terms of being able to collect data with robots. Being able to provide this type of solution, where we have very accurate simulation models, minimizes what we call the sim-to-real gap.

From there, people can choose how they want to solve those problems. We just want to be the company that enables people to do that. If you choose to use our sensors, we think that’s probably what’s going to be best. But again, it’s all about enabling people to do what they do best with their robots.

Makes sense. What are some of the applications you’re most excited about? The cleaning of fibers is one that comes up all the time in data centers.

Definitely. We’re pretty excited about the data center in particular. Within our group, we have some key partners that we’re working with to see how our group can integrate some of these solutions and test them out.

A big thing for us is being able to understand the problems people are having in the industry, what they need to solve, and how we can provide the technology that will enable them to solve those problems.

How do they do that? Do they provide data, or just a specification that you have to design toward?

Yes, and that’s the unique feature about our team. It’s designed around this idea of being a co-innovation hub. They’re more than welcome to bring their engineers into our lab, and we work together to try to solve the problem.

They might bring in new technology from their company and see how we can support that, or we might show off some of the new technology we have and how we can enable them to solve their problems as well.

Do you have any examples of collaborations that you can talk about?

At the moment, no, but it’s ongoing, and we’re pretty excited about what we’re doing.

How about GTC in general? There’s lots of work with NVIDIA and GPUs, AI, and training models. Are you doing anything unique on the model-training side, in terms of how you run the system infrastructure or data center for yourself?

I wouldn’t necessarily say we’re doing anything unique on the training side per se. But when it comes to actually deploying the model and doing inference, we have a team dedicated to edge computing that’s trying to push the technology to the point where we can take these large models and distill them down into a form factor that allows people to run them on their robots or in their autonomous vehicles.

Yeah, so what does that work like? The chips are getting a little more powerful every year and a little more energy-efficient. Meanwhile, model distillation is getting a little bit better, so they’re kind of converging together.

That’s the idea. We’re taking these large models and putting them into a compact form factor that can sit on the edge in a more power-efficient way. That enables you not to have a super-big GPU box sticking off the side of your robot or whatever. All of it can be put into a really small form factor.

Are these NVIDIA GPUs, or are they custom chips?

It kind of varies. It depends. I’m not too involved on that side of the story when it comes to edge computing.

It can be pretty complicated just to get a model to run on some of these custom chips that people are building. I’m curious if there’s a way in which the chip itself might be co-designed with the model it’s actually going to be running.

I think that’s a big part of it, too. Again, when it comes to the co-innovation I talked about, that’s not only in the robotics space but also in edge computing. We talk to the big players in the space and see how we can integrate their technologies—or, sorry, their models—into the fabric of whatever we’re trying to design.

And of course, it costs them SoCs or ASICs and all that, right?

Right.

I guess there are a lot of cases where people might be training really big models and have no idea how they’re going to run them. The model solves the task, but then it’s, “What? Wait 3 years for the chips to catch up?”

Exactly. We’re thinking a little further ahead: Where do we see these models getting deployed—in the robotic space, autonomous cars, home robots?

You’ve got me thinking about something to do my laundry now instead of just the fiber ends. I think that’s a good way to wrap. I want to thank everybody for watching, and I appreciate you coming in and taking the time.

Thank you.