The Robot Is Waiting on Your Data.

For the last few years, we've experienced AI through a screen. Things like chatbots, code tools, or content generators. But the next frontier isn't something you're going to prompt. It's something that sees, senses, and acts on its own.

That's physical AI, and it's already moving from concept to operating reality. In fact, Gartner has named physical AI as one of the top strategic technology trends for 2026. And leaders like NVIDIA are building the simulation, robotics, and world model stack to help bring physical AI into factories, warehouses, and autonomous systems. But as always, the challenge is whether your environment is ready for AI to operate inside of it.

Because unlike chatbot fails, where you get a bad answer, when physical AI fails, operations stop, safety breaks, real-world consequences tend to follow. So in today's episode, we're looking to deliver clarity on what physical AI is and isn't. And our guests, WWT's page reader and Wes Long, will stress the important groundwork like data infrastructure, security, and systems that make physical AI possible. So let's jump in.

Pretty good, pretty good. I'm actually sitting in a warehouse right now working on one of my robots, so the timing is perfect. Nice, yeah. It's a nice Wednesday, and I am not stuck in a garage somewhere working on a robot.

So yeah. We've spent the last few years talking about AI as chatbots, things that summarize our meetings or output an article or a poem or things like that, things that react to prompts. But lately, specifically over the last several months, I feel like I've been hearing the word physical AI a lot more. Most of the time it has to do with going into robots.

But Paige, we can just start with you. That feels like a very vague term in its highest sense. So let's narrow the discussion. Tell me what physical AI means before we get into the rest of the topic here.

Yeah, so physical AI is an interesting term. And it's actually, there's sometimes other terms that kind of mean the same thing. You've often heard maybe embodied AI. Essentially, what it means is anytime that artificial intelligence, AI comes into the physical world.

So whether that's because we're looking at things in the physical world, because there's perception happening, or if you're controlling something in the physical world, that also falls within this category. So here within WWT, we tend to put several things within the physical AI bucket. So there's robotics, digital twins, computer vision, and extended reality, which is like AR and VR. Those are the things that we tend to talk about when we say physical AI.

And there's some interesting nuance there. There's not necessarily a requirement that all of those sub-genres even need AI to function, but a lot of the times AI can be used to kind of take them to the next level, make them be a little bit smarter, a little more connected to the world. You know, what I think is really interesting is a lot of folks, when they start considering physical AI as opposed to generative AI or traditional machine learning, they think that there's this new incredible, crazy danger, right?

Obviously, if you're controlling a robot that can thwack you in the head, then yes, that's kind of dangerous. But if we really think about AI kind of holistically, the same kind of concerns we should have around generative AI and other kinds of AI still exists with physical AI in terms of the way machines are making decisions and the way the output of those decisions could affect you as a person looking at that output. In addition to the robot arm that might thwack you in the head.

So you get this kind of a broader field of things to worry about, but you don't lose the other one either. So it's kind of like a compounding effect as opposed to a some completely new thing. It's interesting because a lot of the same guardrails that we have to have in place for Gen AI, gentic AI, we want to have in place for physical AI as well. The thing that, you know, kind of building upon what Wes was saying, where maybe some of that extra fear comes from, sometimes rightfully so, is if you are truly giving control of things to your AI.

So if you're saying, all right, I want to start using more artificial intelligence in my manufacturing plant, okay. Does that mean that you actually want it to have control of your PLCs that control the production lines? And then if it makes a mistake, which as we know, AI can make mistakes and it often does. If we don't have the right sort of guardrails in place, you might be losing time because your production line is shut down, or it might make a mistake of where it puts something together.

So there's definitely some interesting considerations that come into play from, especially in the early days. We talk a lot about like proof of concept where people want to just start exploring this technology versus when they actually want to deploy it at scale. And so, kind of those baby steps of what it takes to get everybody feeling confident with what the AI is doing in their physical spaces, is a big part of what we've been working on recently. When a physical AI system fails, like you just mentioned, it can bop you on the head.

So is this a structurally different engineering equation, or is it stay true to some of the principles on how we've typically built systems to ingest and digest AI? I think you're kind of foundationally, it's going to be very similar to what we had before. The the big advantage, of course, is that now with your own eyes, you can see when things start going haywire, right? If you're looking at the output of a chat bot, you have to discern the fact that some of the data it's giving you back in that text output is maybe not accurate, maybe it's a little off.

When you're looking out across your entire plant floor, you can see that bottles of ketchup are falling on the floor, and clearly something has gone horrible here. So I don't think you lose the diligence. I think maybe you have a little easier time with it with the real physical components of it, but that's not to say that you don't have that secondary concern still too, right? When we start talking about physical AI with the data we're bringing in from sensors and the decisions we're making, that could still be ethereal things.

This could be digital decisions that are on an Excel spreadsheet in the back of the house somewhere, or orders going out to your supplier. So you still have to have that same level of concern about the digital effects, and now you have maybe a little easier time of seeing the physical effects as well. Paige, maybe before we go any further, we should get a little bit more practical in terms of where the use cases are here. I think a lot of people hear physical AI, then they automatically think robots, and then you automatically think of the robot that's uh vacuuming your floor on a consistent basis or whatever it might be.

But in the real world, where what type of industries are picking this stuff up and what are the use cases that we're seeing best fit for physical AI? You've mentioned, you know, the factory floor, the product line. Is that typically where we see these happening first? That's a big place where a lot of these conversations are happening, but certainly not the only place.

And I'd love to hear Wes's opinion on this as well. But what I think is interesting and worth explaining about physical AI and kind of almost its maturity curve is that oftentimes the cool sexy thing out there, for example, right now, is humanoid robots. Um, everybody wants to know when can I have a humanoid robot that can go out there and do a task from A to B. And we're a bit of ways away from that.

And we can talk about that sometime if we'd like. But what's actually useful is that all these other components we talked about from physical AI build up to that eventuality. So, for example, in order to have a successful robot, your robot needs to understand where it is in the world. So, understanding where it is in the world might benefit from a digital twin of your environment, getting inputs from different sensors, getting all that information sent over to your robot, getting input from your different systems that are running your plant floor, running your warehousing locations, things of that nature.

And then a step back even further from that, computer vision is all about using different sensors in order to understand our environments, things that are happening there, and essentially teaching a computer how to understand like what we do with our eyes. And that is one of those inputs into a digital twin, which, like I said, informs the robot. So they all kind of build upon each other. So a lot of the times our customers come and they want to talk about robots, and I usually have to step them back a couple steps and say, okay, well, do you understand your facility such that we can teach the robot to understand the facility?

Do you have sensors that we can take advantage of so that we can do this understanding? And they really build upon each other. So that's one thing that I like to share. So the second part of that is I'd like to talk about all the different industry verticals where we've been having these conversations.

So one of the most obvious ones is manufacturing. You know, we talked about how essentially you could digitize a lot of what you're doing on the manufacturing floor. You can do automation in both a traditional sense with robot arms and things like that, but also start doing some more intelligent things like digitizing and understanding how to optimize your supply chains. There's some interesting opportunities happening in that space as well.

Warehousing is somewhat related to manufacturing, where it's very useful if you can understand where all the things are that you want to actually interact with and how many you have and how often they're being moved through the facilities. Let's see, another great one is healthcare. So a lot of our healthcare professionals are coming to us trying to improve the patient experience within healthcare environments. So, in order to do that, first we need to understand what is their experience right now.

So, for example, how long are people sitting in an ER before they're seen? So, in order to do that, we might want to use some computer vision to see when does a person come in, when does somebody approach them? When are they taken into the back room? And we can do these things anonymized.

So we don't necessarily have to, you know, collect any sort of personal information, but we can do a general, okay, I see a person there, and this is how long they've been there. Retail is another really big space where this is all coming into play. Tackling things like shrinkage or inventory management, understanding how long people are dwelling in a certain area might indicate, oh, somebody's standing there a long time, maybe they want to purchase it. Why didn't they put it in their basket?

And that might give you some insight into customer buying habits and things like that. I think another big one we've been hearing a lot of is more in the city scale. So, you know, we talked in most of these things about fairly small, you know, within the walls of an of an organization, you know, warehouse or something like that. But we've actually been talking to a fair number of cities about trying to understand, especially, you know, maybe after a big game, something like that, where are their guests going?

Where are the people that were at that game going? Where do they need to move the police officers in order to help with the congestion and maybe change the lights as well? But yeah, Wes, I'd be curious what other ideas you have of the spaces we've been seeing, these conversations. Given your background, we also can't leave out entertainment, right?

So think theme parks. A lot of the stuff that you're mentioning about, you know, retail directly transfers over into entertainment venues. So people uh dwelling in lines or outside gift shops or stadiums, that kind of thing. We've been talking with uh a soccer team very recently about how we can start applying CB to understand the challenges they're having inside their facility to do crowd control as well.

So it's not not just the city saying, How do I deal with this group? But it's the individual businesses. How do I deal with these people that are coming in here? And the interesting thing to me about all this is these problems are not new.

And if you've been working in these industries for a while, you're aware that there are already ways that these things are being tackled today. And that I think points out that when we say physical AI, we don't necessarily mean cutting-edge new types of AI. We don't necessarily mean generative AI. It can be, but it doesn't have to be.

So you look at a lot of these solutions and the things that we're talking about and dealing with, some of them are tried and true things we've been doing for a long time. Internally in WWT, you go back and look at some of the first AI work we did 16 years ago, 15 years ago, with these more classical AI approaches of like machine learning. We're still gonna be using those things and these CV opportunities that exist in all these verticals pages talking about. So it's like it's kind of a moment where you're getting this publicity around physical AI that's kind of pointing out the cool things we've been doing forever.

So, like some of those older nerds are finally getting their recognition they've deserved this entire time. Yeah, Wes, I was gonna ask you that exact question. What is conversion right now that's making physical AI pop more onto my collective radar here? Are there advancements in the technology that are accelerating what we're doing with physical AI and robotics?

Or, and maybe it's gonna be both, but or is does it just have AI in its name and therefore we're, you know, we're putting a laser beam on it? I mean, the answer is yes, right. You you can get away from the fact that we have some very big players in the AI space that have amazing marketing teams, right? So they're gonna get to start talking about these things.

But that kind of does downplay the fact that you look over the last you know five, 10 years in terms of the hardware available to us to run these kinds of workloads, it has skyrocketed. So a lot of these classical solutions for AI are riding on the coattails of the generative AI explosion. So I might need crazy, crazy GPUs to develop generative AI. Well, now I can do large numbers of the non-generative AI and the same hardware, especially when you start talking about uh year over year upgrades to the hardware.

Suddenly I have old racks and servers laying around. Well, let's do more cool CD things with that. How can we explore things that were out of touch before because we didn't have the budget or the room? We've had to make the room for these other things.

Now we have that ability and that and that attention. And because that's becoming such a major player in the IT space, you're starting to see uh buy-in from maybe parts of the executive suites that weren't really to talk about these things before. Maybe there was a little bit of concern about budget or security or liability that you're now breaking down these doors with other technology that's important, and suddenly these things can slip in as well. So you get this kind of conglomeration of the market as a whole is paying attention, so that kind of feeds itself.

You have this new hardware and new technology showing up, and now you have more experts for digging into it. So it's all just kind of coming together in one good explosion of neat things. Yeah, and to expand upon that just a moment, I think one other really interesting thing is that kind of comes out of that is as more and more powerful computers have become available, people have been realizing how important getting their data in order is so that you can do powerful things with those new computers.

Because if they're all scattered around and in places that we can't work with it, then it's worthless to us. We can't start doing neat artificial intelligence applications. And so that's been another thing that's really helped explode the kind of physical AI conversations. One more point going on to that is I think what's interesting is not just that the places where you're training these models, where you're hosting most of this data, have gotten better.

It's also the edge compute that's gotten better as well. So, for example, we'll use a really powerful server farm in order to quickly train a computer vision model. But then we don't want to have to run our computer vision model from this server farm. It might be far away from the application where we actually want it to be, and it might be important for us to have low latency in our system.

So we want to use some sort of edge compute in order to essentially move that closer to the place where we're actually receiving information maybe from our cameras or other sensors and making the decision and doing something with that decision. So I think that's been another really interesting trend that's been kind of continuing to grow over the last couple of years. This episode is supported by Weka. Weka's high performance data platform uses parallel file system technology to deliver extreme throughput and low latency for AI, machine learning, and HPC workloads.

Unlock the full performance potential of your AI infrastructure with Weka's enterprise data platform. Yeah, and it occurs to me now that if you know if if physical AI or robotics or whatever we called it a few years ago has is now starting to become more in vogue, and you both are saying, well, this is stuff we've been doing for for a while now. There are probably a lot of good lessons to be learned from what you've been doing over the last several years for what more of the advanced or newer Gen AI type stuff is doing.

So, how what can we learn from how the two of you and your teams have worked with data, data pipelines, portability, moving from edge to on-prem, you know, things like that. What are the key lessons learned that are really going to help put other enterprise leaders on the right track? Because that's what everybody's dealing with right now is how do I get my data into a spot where it's usable, digestible, and ultimately a way for AI to use it and make a decision? One thing that's really important is this concept of, and it's not super exciting to talk about, but very important functional safety.

So when we start talking about robots getting out there and being controlled by AI or using computer vision to start making safety decisions, things of that nature, knowing where to draw the line between trusting the technology to absolutely perform, and knowing that it is something that it's better to have it than to not have it is a really interesting line that a term and a discipline called functional safety can really help us dive into deeper. It's it's a discipline of figuring out how to essentially deploy new technology, deploy automation, deploy essentially how people work with machinery into these different types of use cases, whether that's in a factory or some sort of logistics automation warehouse facility, those types of things.

And what's been really interesting as the conversations around computer vision and robotics have been ramping up, the standards, the standards that are used to control functional safety requirements have been lagging behind. And this is actually something that people have been watching happen within the autonomous vehicle segment for a while now, where all these companies came out and they said, Hey, we have this cool autonomous vehicle technology we'd like to test out. And the various governing bodies that allow you to do testing on public roads were actually lagging behind the technology in certain cases.

So they'd actually have this technology they wanted to try, but nowhere to actually try it. As we can see from the fact that there are now some autonomous vehicle tests happening out on the road, things are starting to catch up, but it's still typically in fairly small little places. So that same sort of revolution is happening with physical AI within this functional safety context. Computer vision's a great example of this.

Computer vision is a type of there are multiple different types of it, but the one I'm talking about is more deep learning based and artificial intelligence subcategory. And it's essentially a black box where there's something you put in, which is your raw image, it goes through and does some processing via its neural network. And on the other side, it comes out and it tells you what it thinks it sees. And, you know, if you train your computer vision algorithm well, it's normally gonna do a nice job of translating that.

But it's not gonna do a great job of understanding a novel visual that it's never seen before, that it's never been taught what to understand about this image. And so things like that start to break down when you try and figure out, okay, how do I use this then in a safety context? If I can't be 100% sure I know what I'm gonna get at the end of the day, that it's gonna be able to recognize everything I'm talking about, how do I then tie it into all these other things that I've been using for years and years in order to safely control these processes?

So it's an exciting time, but it's also a time where some folks are having to kind of take their steps slowly in order to figure out how do you actually take it out of the proof of concept phase and scale it and actually start trusting it to take over some of the responsibilities out on the show floor. So and I the the kind of a scary thing and almost kind of uh exciting thing to go along with that is that we talk for years about data readiness, data usability, data availability.

We also, of course, talk about data security, right? So we need to have access controls and decide who can access these things and use them and see them for trade secrets. But now we have to start talking about PII, I think, more than ever, especially when we start talking about the physical AI part with sensors. So if we have we have cameras bringing in images to people's faces, right?

Or identifiable information, suddenly there's this layer of concern that maybe we were a little laxadaisical about before. We've thought about it, of course, but now there's more of that data and it's more front and center. And now when you think about the security around that and the access point. To it in a way that we maybe didn't think before.

Similarly, obviously, we've said before, like you're controlling things that are physical. So now we need to start thinking about the security around those systems. Who should be able to get to the things that are controlling these robots or theme park rides or whatever giant things that are dangerous out in the real world are? Who can influence the data that's coming into that?

Like Paige is saying, How can we make sure that we have good data that drives to the right decisions? Similarly, how do we make sure we have access control so that data is not poisoned to make the wrong decisions, right? So, again, like we've said before, this is not new. It's a new horizon to think about this, like another dimension to your poor security hat person's horrible, horrible life.

Now they think about more things that can go wrong. But really, I think the takeaway when you look at lessons learned and what you're going to do with physical AI, you look at what you've done for decades already in the IT space, and you just start thinking about how is this more dangerous? How is this more open to liability issues? And how do I control that?

And how do I train my folks and how do I invest in this in a way that maybe I wouldn't think about immediately. Don't rush into it. We've been talking about AI forever and saying figure out what your foundation is, build that foundation, and get to it. Don't think that tomorrow you're going to have an answer.

Nothing is different here. In fact, maybe you should go a little slower because of the real world implications of what you're trying to do. And I don't want this to all sound too doom and gloom. We we of course want people to be careful and responsible, just like with any type of AI application.

So hopefully that's not how it's coming across. But I did just want to add one more fun thing that I think transitions from kind of the past of where we were with some of these things, some of these tools, robotics and vision systems, things like that, to the future. I like to say that we're making, in the case of robotics, we're making the robots context aware. So when I used to work in the automotive industry, there were plenty of robots out there helping to assemble vehicles, but they were typically a robot arm and they were doing the same type of motion over and over again.

It was repeated, it went to this waypoint, then it went to the next waypoint, and then you strung those together into a program. Now what we're trying to do is make those robots smarter. We want to give them real-time information so that they can on the fly change how they're moving, change how they're moving objects around and how they're interacting with things. So I think that's another really big important part of this kind of physical AI revolution that's different than it used to be and kind of where we're trying to go into the future.

Well, and Paige, at what point do we give these these robots who have typically over the you know the past several years or decades, even have been trained to go from point A to point B? But at what point do we start to give them agency? At what point does agentic AI catch up to where physical AI is? And now we have robots that are able to roam the streets and so people definitely want to get there.

They want to get to can I have this robot that has the ability to do all the things that a human can do? Can I give it the agency in order to figure out how to do a task? Instead of just telling it, hey, I want you to move from point A to point B, I want you to grab this thing and drop it off over here. Can you instead tell a robot, hey, I need you to get this general thing done, you figure out how to do it.

And that's where we're definitely seeing a lot of POCs that people are doing are heading in that direction. A lot of the data prep that we're doing right now is in preparation for some of those different things. What becomes really interesting and is worth balancing is how what does the trade-off look like between a very specialized machine or robot or device versus a more generalized solution. So theoretically, a generalized solution, call it like a humanoid robot.

That's the fun thing that lots of folks are talking about these days. You can make it do many different things. But you could make the argument that because it can do many different things, it might not do any of those particularly particular things very quickly or very efficiently. It might not be optimized for that solution.

So there's some really interesting trade-offs that are being discussed within that space, but it's definitely the place a lot of our customers want to get to. I think we're definitely, especially in the terms of robotics, we're still a little ways away from that. It's it's one of those revolutions that we think is coming. And we're definitely seeing, frankly, hundreds of new robotic manufacturers coming out with very interesting and advanced robots that we are starting to get in, we're starting to test.

But what's interesting, and it goes back to some of the things that Wes was talking about earlier, making things robust and actually able to repeat these types of things and be able to handle real-world challenges. The race to getting the coolest humanoid sometimes means that those manufacturers miss some of those requirements of making it robust, knowing how to handle the data governance concerns that Wes brought up earlier, making sure that those robots are gonna actually be able to handle the challenges of a manufacturing environment in terms of their battery life, in terms of making sure that they have a high IP rating.

So making sure that they are resistant to dust and water and things like that. So it's this really interesting convergence that's gonna have to happen in the future of this cool new form factor, agentic AI all coming together with all the things we've learned in the past of how to make these robust solutions. And I'm excited to see what that looks like because I feel like that's that's kind of the holy grail a lot of people are trying to get to right now. I I think you're going to see kind of across the board some people sprinting into it, like Paige was saying.

They're gonna have some hard-learned lessons from a variety of different sources. I think you're gonna see some folks that are okay with being not on the bleeding edge, but the leading edge, and they're gonna let those other people learn their lessons and they're gonna learn from them secondhand. I don't necessarily think in that way that this is something special. We see this with every new huge technology as it comes out.

There are folks that are willing to set everything on fire, trying something new, and there are folks that want to lean back and be safe about it. And that's okay, and that makes sense. I think probably the publicly known and shown examples of this physical AI are those people being on the leading edge. And from the conversations that we have with potential clients and partners, you can see people looking at that, saying, that looks cool.

When can I really use that? How far away are we? And you see these big companies having these sobering conversations with us when we're like, yeah, that's years away. We're not gonna actually put that dancing robot into your factory tomorrow.

Turns out that's just not where we're at. But the folks that we we have these conversations with, they're grounded in reality enough to know that yes, this is gonna be a slow burn for some of these things. And then the conversation becomes, well, what part of that can we do today? How can we start building towards the future?

How can we start putting our data together? How can we start building new facilities that'll have room for these things in the future? So I think like everything we see in IT and technology in general, you're gonna see this wave of preparation of folks wanting to be able to jump on as soon as it makes sense to not burn down their house and start making use of it. What are some of those stepping stones?

What do we need to do to get ready for a future where we can accomplish some of that? So we're actually having some very similar conversations right now with one of our customers that is looking towards the humanoid future, is trying to understand what it takes, but they also understand that it's not quite ready yet. And so, as preparation for that, one of the main things that we're working on is creating a digital representation of their facility. So the reason we want to create, and I don't want to call it quite a digital twin, because it's not necessarily getting real-time information from those sensors, but a kind of a 3D representation.

And one of the reasons that's useful is because we can start doing testing of robots in a virtual environment before ever moving them out to a real world environment. So that allows us to do some very early investigation of can this robot physically move its joints in such a way that allows it to actually do the tasks that we need it to do. And there's some groundwork that these people can be doing to start preparing for those things that Paige is talking about as well. So digital twins, a good example.

If we're going to build this digital representation of your facility or your theme park or whatever to do this training in, we have to be able to create that 3D visualization. How do we do that? Do you have good access to the architecture diagrams? Do we need to come in and do LiDAR scans?

Do we need to do video scans? Do you have governance in place where I can request access to your facility to do those things? Fun fact, not always. We've learned that.

So thinking beyond the technology and thinking about where the data comes from that drives the sensors that we use in this technology, that's stuff you can do today. You can start talking to your corporate facilities folks. You can talk about managing people coming on site, really starting to think about the kinds of people maybe you haven't interacted with before. We were at a manufacturer's plant not too long ago, and we were out there with a LiDAR scanner walking around.

And I will bet you dollars to donuts. Nobody has ever said, Can I get permission to bring a LiDAR scanner into your factory before? So it doesn't have to necessarily be even high-tech stuff you're prepping for. It's just about really understanding your business and how you're going to transform it to use these things in the future.

Another great one we've run into is we want to use cameras. We want to use cameras to see what's going on. And unsurprisingly, lots of facilities are like, I would prefer you don't put me on camera. And you got to work through the legal part of that.

You got to work through the personnel part of that. There's a lot of steps to go through there. So think beyond high-tech, think beyond AI, and really just think about the things I'm trying to bring this technology into. Are they prepared from a physical location standpoint and a data about that physical location standpoint?

Well, lots of work to get to, but an exciting future nonetheless on the horizon. Paige, Wes, thank you so much for taking the time to join us on the show today. Yeah, absolutely. Thanks for having us.

Yeah, thank you very much. All right, thanks to Paige and Wes for joining. The big takeaway from this conversation is that physical AI is not a leap straight into humanoid robots, it's a progression. First, you need to understand the environments, then connect the data, then build the guardrails, security, and infrastructure that allow AI to act safely in the real world.

Because the robot may be what captures our imagination, but readiness is what determines whether physical AI actually works. This episode of the AI Proven Ground Podcast was co produced by Nas Baker and Kara Kuhn. Our audio and video engineer is John Noblock. My name is Brian Phelps.

Thanks for watching. See you next time.