In the wide universe of artificial intelligence, Physical AI is a confluence of perception, reasoning, and execution - as Michael Kenworthy, CTO of VulcanForms helpfully explains. Physical AI builds on foundations like large language models (LLMs) and vision-language-action frameworks - and bridges internet text intelligence with physical robotics.

Now Physical AI is spreading out from the realm of automation and into more pragmatic applications. In this episode of Great Question: A Manufacturing Podcast, Kenworthy uses his insights in metal additive manufacturing to describe the potential for Physical AI, how it is maturing, how to prepare to apply it, and where to watch for its impact.

Below is an excerpt from the podcast:

Robert Brooks: Hello and welcome to a new installment of the Great Question podcast presented by Endeavor Business Media's Manufacturing Group. I'm Robert Brooks with American Machinist and Foundry Management & Technology. The subject of our discussion here is physical artificial intelligence, or physical AI. And this choice stems first from my curiosity after seeing some demos by robotics developers, but more specifically from an introduction to today's guest. He's Michael Kenworthy, the chief technology officer with VulcanForms. VulcanForms produces critical components for aerospace systems and medical implants and defense hardware using metal additive manufacturing and precision machining. And I'm confident that's the sort of manufacturing business that has a sense of how to proceed with physical AI. Michael, thank you for joining me today.

Michael Kenworthy: Glad to be here, it’s exciting.

RB: I want to discuss physical AI in a practical rather than in a conceptual way, but I'll start by putting it in a context that I think most of our listeners will understand, starting with our understanding of artificial intelligence, such as generative AI or agentic AI. From that, can you explain to us what we should understand about physical AI?

MK: So there's a lot of, I think, some core framing that's needed. Like, this term is relatively new. It really only appeared in the last year or 18 months. But effectively, what we mean when we talk about physical AI is we are perceiving the world through a set of sensors. So there's a perception layer. There's a reasoning layer, right? We need to sort of assess the sensor feedback and what is the right kind of course of action. And then actually execution layer, right, whatever the actuating mechanism is, different kinds of mechanical systems or other types of systems, how do we act? And so it's that combination of the three that makes this AI term an appropriate label for what we're talking about.

RB: Do you have a timeline? Can you tell us when and how it emerged?

MK: Yeah, I think it's really Jensen (Huang) at NVIDIA, of course, kind of almost pushed this into the mainstream, but it goes a little bit of the ways back into the foundation models, and people thinking about, okay, what is this combination of kind of the foundation AI models, what we think of as the large language models, the ChatGPT and the tools you probably are using to make your life a little easier already. And what is this sort of trend as it combines and collides with robotics and the ability to bring these essentially language or internet text models into the physical world. There was a lot of work around the 2023 time frame, the vision language action models. There was a bunch of work happening at Google, DeepMind and others. And effectively, that's the really exciting part is, how do we leverage these advanced language models and what I described as that sensor perception-reasoning-actuation loop to do some really interesting things here in the real world.

So that's kind of how it's proceeded. Like I said, it is relatively new. Of course, we think about the AI term broadly and how folks have kind of leveraged it. The label itself kind of confuses a lot of people like, oh, I used to do that. I used to do statistical, like I used to do programming and do these kind of feedback. What makes this different is that kind of ability for the system to act autonomously, which is kind of the scary part of this, right? And so, how is the engineer specifically programming a set of actions versus effectively programming and managing a set of behaviors and reactions? That's how it's emerged, and this is truly like frontier stuff, right? There are very few examples of folks that have deployed physical AI in the real world, but tying it back to what my company, Vulcan Forms, is doing and kind of my history in industry, it's been really through this additive manufacturing lens, which I think is really the most exciting opportunity in a lot of ways for what we talk about as physical AI.

RB: I'm glad you said that because I do want to get away from the conceptual discussion here and into the practical understanding of it. As we were talking earlier, I pointed out that I'm aware of how this is being played out by robotics developers. But could you untangle that for me, how the development is going on the robotic side, separate from on the manufacturing side, and especially for additive manufacturing, which is where your company is active?

MK: Absolutely. I think the core framing that we need to talk about and think about is most people, when we think about this, they immediately snap to sort of a robot, a humanoid robot most likely, or maybe a robot arm, versus sort of these industrial systems that might look a lot more like a box that might have a spindle in it, or in the case of 3D printing, the lasers and sort of motion systems for directing the progress of those lasers, whether it's a gantry-based or a galvo-based system, which VulcanForms does both.

But effectively, thinking about the term more broadly, really any system that can generate process data has this kind of intelligence layer of sensors monitoring the process, can actually become a physical AI opportunity. And so in the 3D printing world, what we call additive manufacturing in the metal side of things, it was born as a digital process, right? So these systems are highly exquisite, lasers interacting with fine powders, and it looks a little bit like black magic, right? We're making these metal objects, complex metal objects using literally lasers and powder in these boxes.

It was born from that as a purely digital technology. It's always been that. And then we've continuously evolved the technology and applied sensors and additional types of in situ monitoring tools to enhance the process. Being a relatively recent technology, it's had to survive a higher level of scrutiny, frankly, than legacy technologies have. If you think about castings and the quality that we accept in that industry, if we use that as the baseline for additive manufacturing, it's almost hard to make a bad part. But that's not the reality. We're being viewed through this lens of higher criticality, more regulated industries, things like medical components or aerospace, turbine engines. Of course, we want to apply the right regulatory and qualification frameworks to accepting the process for these types, whether it's a human safety factor or safety of flight type concern.

And so because of that, we've done a lot of work to make the process smarter. I mentioned this is like really frontier territory, like how do we close the loop on the process today? The process effectively runs open loop. We pre-program, we do a lot of simulation to figure out how the part's going to distort in the welding process. Of course, it's an intense thermal process, so there's distortion that we need to manage and other types of potential for defects. And so we continually have been kind of stacking more intelligence on top of these systems to assert control and unlock these critical applications. So when I talk about additive, I always think about it as one of the purest use cases and opportunities for physical AI, because we've already heavily instrumented these systems and we can act on that data in real time. So that's how we're applying it, yeah.

RB: It's fascinating. If I may go off road here for just a minute, for the listeners’ benefit, in additive manufacturing, all of the intelligence, all of the decisions are made in the CAD stage, right? And so are you suggesting that more decision making can be done beyond the CAD stage, in the production of the additive manufacturing part?

MK: That's exactly the change that we are driving here at VulcanForms, and I think where we see the broader unlock for the technology happening in industry. There continues to be a high level of scrutiny and questions about acceptance of the technology. We look at the Navy and the submarine industrial base and kind of how they're trying to embrace the technology. With these tools, I think we can actually demonstrate to them that not only is this an acceptable process, it's actually a superior process, right? Because we have unprecedented level of visibility and quality control that we can apply at almost an atom by atom level, if you will, but really voxel by voxel level, right? We are able to generate these thermal data, these other types of quality data, optical data, whatever it is, and combine those into this multi-sensor fusion look on the health of the process and give real-time feedback on, were we successful? Did we meet the gold standard product that we qualified, or where are we at in the process window? That's what's changing.