How FPGAs become the gatekeepers of Physical AI Security

0:00Hi, I'm Mike Waitzman, co-host of the Robot Report podcast. If your company wants to reach decision makers in robotics, automation, and AI, you're in the right place. Each week, thousands of industry professionals listen to this podcast for the latest news, expert insights, and conversations with leaders building the future of robotics. This spot could be yours. To learn more about podcast sponsorship opportunities, visit therobotreport.com slash podcast.

0:34What's up, everybody? Welcome in episode 260 of the Robot Report podcast. I'm Steve Crow, executive editor of the Robot Report. Thank you, folks, as always, for tuning into the show here. Joined by Mike Oitzman. He's the senior editor of the Robot Report. Back from a trip to Iowa in the cornfields. What the heck were you doing out in Iowa this week? Yeah, back to my roots. I'm an Iowa boy. Born in Iowa, raised in California, but went back to Iowa to see John Deere. Yeah, it was hot and humid and glad to be back in California. But we were on the campus of Iowa State near Ames, Iowa, where John Deere has donated millions of dollars to the university, Steve.

1:18It's just a really incredible facility that the university has back there for hosting all things ag-related. and so i got a chance to tour all that it's very cool they must have they i know you've written about this on the robot report so what was the big product announcement that you and everybody else was there to see yeah well this week's the the farm progress show it's one of the biggest farm shows in the midwest for row crops and john deere announced a new product called jd jd two letters yep it's a it's a new ai llm um generative ai that operates uh you know on a corpus of a farmer's historical operational data that they house today in their product called operation center so every farmer runs operation center to track you know everything that happens on the farm you know all of the inputs what you know what did they spend on seed what did you spend on all the you know fertilizer herbicides all that kind of stuff what did you do when did you deploy all that stuff's in there so basically they they worked on building this llm interface on on top of all that data.

2:25So now a farmer can just query, you know, what's going on. I'm sure farmers keep track of their business like any good business person would, but like spreadsheets, are we talking spreadsheets here with farmers or what are we talking about here? Yeah, well, I think spreadsheets were very popular for the last two decades. And before that, you know, generations of centuries of just notebooks, right? Keeping track of stuff in notebooks. But now, you know, in this age, everything's instrumented and logged. And, you know, there's basically it's information overload for all the farmers in terms of being able to see, you know, this signal through some of the noise.

3:04And so that's the concept behind GED, trying to, you know, turn that into a conversation with the data. I think that's the innovation that they announced here. And I think, you know, it makes sense, Steve, it that where I think where Gen AI is today, there's plenty of opportunity across every market vertical for, you know, some type of LLM type interface on the, you know, the data that exists, some private corpus of data rather than, you know, basically what they're promising is to keep every farmer's data private to them. So, you know, they're not sharing their data with John, well, John Deere can see it, of course, but John Deere doesn't put it into one big bucket and look across all of the different farm inputs.

3:48It's really each farmer's data for their own use. And, you know, imagine, and I've talked to several farmers, you know, over the last couple of years of covering this market, a lot of the modern day farmers, the farmers that are actually going to college and then going back to, you know, farming after they get some ag degree, a big chunk of that is, you know, They're actually SQL, right in SQL, you know, because they've got all this data to try and crunch across all their farms to try and make choices and decisions in today's world. So it's a very interesting, you know, future from this perspective.

4:23Yeah, well, hopefully the AI can help them, right, with all the historical data for farmers, but for many other businesses as well. I think I saw a picture of you. Were you riding in a tractor? What were some of the fun things that you got to do while you were there? Well, yeah, we got an update on all of the latest product releases for farm and farming. And that's all embedded in the story. You know, the new seed robots that they have that actually plant seeds and build furrows for that. There's some advancements there. But I didn't actually ride any tractors this time. I feel like an old dog here in terms of having done that before.

5:05Or I ran into one of our colleagues, Lisa Itel. She's executive editor over at Design World. She was out for the event as well. So I didn't know she was going to be there. So she'd never been on an automated tractor before. So I let her have my seat in terms of when they were asking for volunteers to go ride. So it was a good time. It was fun to see all the latest at John Deere. Awesome stuff there. Glad you had a good trip. So a great show for you today. We'll play an interview where you chatted with Eric Sivertson. He's the VP of the security business over at Lattice. We'll play that later on, but why don't you give us a little preview about what we'll hear from the conversation with Eric.

5:45So in this interview with Eric, we're going to go deep discussing how FPGAs provide deterministic performance, secure boot, runtime monitoring, and cyber resilience for robots and other physical systems. It's a good look under the hood at a ubiquitous piece of technology that exists in every autonomous system today. I mean, FPGAs and firmware, every electronics engineer, electrical engineer knows what those do. And so we're going to take a look at what Lattice is doing now to combat security and other issues that are of core importance to every robotics company today. Awesome. So we'll play that later on.

6:27We'll recap the news of the week, the bigger stories, and there's some good ones. We'll do that here in a minute. But before we do that, of course, I want to tell you about our upcoming Robo Business event, Mike. Less than two months away. It takes place October 20th and the 21st out in Santa Clara, California. This is our 20th anniversary show. We'll have 100 exhibitors on the show floor, over 2 ,000 attendees. We'll have a startup alley. Most of the agenda is done and posted and up there. So go peruse the agenda and see all the things that we'll be talking about. But we're talking physical AI deployed in the real world, not physical AI in some research lab that might or might not work.

7:06We're a keynote panel with Teradyne Robotics, Collaborative Robotics, Amazon Robotics, and ABB is new to this keynote panel on how they're deploying physical AI out in the real world today. So that should be a great one. We're doing the future of robotics with Rodney Brooks, Ken Goldberg, and Jonathan Hurst. We're going to be talking humanoids. We're going to be talking with Morgan Pope, formerly of Disney, now at Familiar Machines and Magic, right, trying to bring robotic characters to life. So there's a lot of great stuff happening, but really focusing on how to build and deploy robots in the real world.

7:42So go check out the agenda. Plenty of time to register and buy your tickets. If you're looking to bring a group, reach out. If you're from an academic institution and want to bring a group of professors or students, we have special academic rates as well. So any questions, please reach out to any one of us here at the Robot Report and we'll get you straight. Again, if you've ever been to the Robotics Summit in Boston, same folks put on RoboBusiness, a little bit of a different feel, some different topics, but going to be another great show. Again, the 20th anniversary show of RoboBusiness takes place October 20th to the 21st out in Santa Clara.

8:16Hope to see everybody there. Again, go to robobusiness.com if you want to learn more. All right, on to the news of the week, and we have you covered with all the latest news and information about the robotics industry. Make sure you check us out at The Robot Report and Automated Warehouse. Follow us on all the social channels. Make sure you sign up for our weekly newsletters so you don't miss a beat. Big story of the week, just broke today. This has been rumored now for, what, a couple of weeks, I think, Mike, but just was confirmed today. It hasn't closed yet, but NVIDIA announced today plans to acquire Hugging Face for, what was it,$13 billion?

8:51And they're going to keep the AI development platform open. At least that's what they're saying at this point. We know that stuff can and often does change. But this just announced today. What was your reaction? No, I think this is interesting development, interesting acquisition for NVIDIA and the rest of the AI community. I mean, look, everybody knows what hugging face is in the AI world. So it's an interesting strategic play for NVIDIA to pick it up. I guess the CEO and the founding team will stay on following the acquisition, you know, continue to run things. And, you know, to keep the ecosystem open, I think it is important.

9:33a multi-cloud, multi-accelerators platform and to support all the model builders, I think it's right in line with what NVIDIA wants to do and should be doing here to continue to foster innovation across all the various opportunities. They want to dominate the AI ecosystem, right? Yeah. Is that fair to say? They're doing a pretty damn good job of dominating the computing side of it, right? The hardware side of it, the simulation. side of it. I think maybe hugging face is much more known when you're diving into the models and the data sets and discussions amongst the developers, right? So they're maybe a little bit closer and tied into how that stuff is going.

10:16But yeah, I think it just makes a lot of sense for NVIDIA to do this, right? So on top of selling all the chips and the hardware, they have maybe, I don't know if control is the right word, but more influence maybe about where AI is ultimately going to be built on at the end of the day. But we'll see how this plays out. Again, huge story. Everybody's saying that the developers will not be required to use NVIDIA hardware or compute to build or deploy on Hugging Face. So they're trying to maintain hardware independence. Again, we've seen not this exact fit, but with others, I believe when Amazon first acquired Kiva, it's another big acquisition in the robotics space.

10:58But I think initially the idea was to keep supporting Kiva robots that were used by third-party companies. Of course, Amazon did not. That didn't last long. Amazon pulled everything from the market, right? So we'll see how this goes. But Hugging Face is a huge company. I think the micro-duck is really what kind of threw this acquisition over the finish line, right? Jensen wasn't really sure if they should do this, right? And then they see the micro duck comes out. They, I forget how many thousands of orders of the micro duck they have and a few million bucks in pre-sales, Mike. But I think micro duck is what got this deal over the finish line.

11:32We'll have to see if everybody that goes to GTC walks home with a micro duck. Free micro duck. Yeah. Very cool. Very cool robot. I love that robot. Right. I'm going to try to get my, try to get my hands on one. Cause it is, it is very cool. I will say there, there, and again, I understand why the folks at Hugging Face are doing this, but there's this growing sentiment that it's the most important consumer robotics launch in the history of the industry. And it's an interesting discussion, but I don't really consider Microduck a consumer robot. No. Right? It's a developer robot. It doesn't really do anything out of the box.

12:07It's not a Roomba. The Roomba, I think, is by far the most significant consumer robotics launch ever. I don't know how you position it as a consumer, but that's a little bit of a misnomer. Somebody's just looking for headlines. I mean, it's going to be entertaining. and yeah, it's affordable. So, you know, it's a perfect price for any kid that's smart, high schooler, you know, they could do something with that. Absolutely. Yeah. If you don't have any of these coding skills or you're not interested in learning those types of things, I think would be a pretty difficult toy or device to play with.

12:40But very cool nonetheless. I hope I get my hands on one. Yeah. So we'll keep an eye on NVIDIA and Hugging Face moving forward. Skilled AI this week, Mike, unveiled its S1 flagship robot foundation model. Yeah, it's a story Brianna reported this week. And, you know, recently Skilled raised almost, I mean,$1.7 billion since its founding to build a general purpose brain. And I think this is the first serious step, you know, towards that with S1. And, you know, S1 allows robots to learn and execute complex long horizon tasks simply by watching a single video prompt. That's the promise of this one.

13:20Now, how good that response is, it's a different question. But, you know, they had to get to some point here of pushing some product out. And so this is that. And I think it's, you know, it's, again, throwing another gauntlet down here on the AI frontier model side for physical AI. I know lots of people are working on this. So it's a good measuring stick against what's likely to happen here in the next 12 months. So I think we're sort of at the cusp here of lots of releases and lots of innovation and lots just keep pushing the capability stack up with each subsequent release. So it's interesting to see.

14:01I know we've been waiting for skill to get to this point. It's taken them a couple of years, which would be expected to get to a point you can release a product that's this capable. Yeah, I'd really like to continue to learn more about this. A lot of what they're – and you can read Brianna's great story on the robot report. We'll link to it. There's a couple of videos in there showcasing some of what the model can do. Obviously, I'm sure it's not all of the model can do, but it focuses a lot on tasks or they're showing tasks that are done in a home, for example, making eggs, repotting a plant.

14:38I think flipping pancakes, making coffee, things of that nature, which I think we've seen all of these things be done before with various AI techniques. What I'm trying to figure out is Skilled AI is the company that acquired the assets from Fetch Robotics when Zebra Technologies was offloading the Fetch Robotics business earlier this year. What was it, four or five months ago, something like that? And I know for a fact one of the issues with the Fetch Robot over the years is the software and how, I think, difficult or not up to par it was. So if Skilled is able to do all this stuff in the home and certainly other industries with this new S1 model, I do wonder what are they doing on the mobile robot side of things, right?

15:32Is this foundation model some other foundation model that they're using or developing? Are they able to apply that to these Fetch AMRs? Is that something that they even want to do at this point? Is there another reason it acquired everything from Fetch Robotics and Zebra? So I'm going to poke around and see what I can learn about that. But if these models can really learn anything in a short period of time just by watching videos of people do it, I would think there is some way that then that could benefit the fetch AMRs that are out in the world. Yeah, I mean, that's an interesting thesis, Steve.

16:07I follow along here and see if they have any sort of other physical product to actually release. We just talked about the duck. so you know there's an opportunity uh you know here for for for this uh to also grow in that way so i guess we we open a kashi um bet on this is that illegal now didn't they didn't that just get shot down or something kashi get get your get it in while you can't get the right in while you can't um last story and this is an interesting one uh again not the last that we'll hear on this but it was uh i posted about this we broke the story last week that teradyne robotics was suing jaca robotics jaca is a chinese collaborative robot arm maker oh there's my boss an accent coming out looking at uh robot arm maker and teradyne robotics sued them for alleged patent infringement both against the hardware and software for collaborative robot arms from universal robots which is one of the teradyne robotics subsidiaries.

17:11So we wrote this story last week, broke the story, got a bunch of interaction on my LinkedIn page. And we reached out to Jaka, as you're supposed to do when we learned about this and gave them plenty of opportunity to respond to the story. They just responded to us this morning. I live on the East Coast. We got the email at 3.08. I woke up later in the morning, but saw the email came through at 3.08 in the morning with Jaka's response to the lawsuit. And Terodyne's really looking to have Jaka stop selling these robots, right? They're not looking for financial damages to be repaid or anything like that.

17:50They're really just looking for Jaka to stop selling these products, I believe, over in Germany. That's a harsh response, right? That's the harshest of the remedies. Right. So Jaka hit back today saying it has gone to court in Germany seeking injunctive relief over what it calls Teradyne's, quote, false, misleading and damaging, end quote, public statements about the lawsuit. So Jaka told us it hasn't received any direct communication from Teradyne Robotics or Teradyne Robotics legal reps about the alleged infringement. Jaka also said it hasn't seen any legal documents or court notices related to this.

18:32They said they learned of the allegations only through Teradyne Robotics social media announcement and subsequent press inquiries. So they're also saying that, Jocka is saying that one of the concerns for Teradyne Robotics is safety over this cobot. And Jocka came out and said characterizing its products as being less safe is tied to, quote, the company's Chinese origin, end quote, and calling those sentiments unsubstantiated. So Jocka Robotics is a Chinese company. The German subsidiary is being sued by Teradyne. But I think what was great here was the Teradyne robotics response. And this was from David Brandt.

19:17He's the VP of R &D and he's also the CTO of Universal Robots. He's the person we spoke to last week to break this story. But again, we're trying to be fair here and reach out to all parties when something happens in this story. So Jaka sent us their response. So we have to give Terodyne a chance to respond. And they did. And here's what they said. Again, it's all on the robot report dot com. But I thought this was a pretty interesting response. Quote, I'll just read the whole thing. I can assure you, this is again David Branch from Universal Robots, I can assure you that we have filed a case with the courts in Copenhagen.

19:52If it has not yet been served to Jaca, it may be due to the time required by the courts, but they should receive it shortly. Regarding their freedom to operate analysis, they have apparently not done a very good job if they failed to identify our patents. We also find the number of patents Jocko claims to hold irrelevant to this case. This matter concerns their infringement of our patents. Furthermore, the number of patents is generally irrelevant as it is not about the quantity of patents but the strength of the patents. And then finally, in relation to their statements about legal action in Germany, we await further information, but we are not concerned.

20:28We have every right to defend our intellectual property and will continue to do so, end quote. So pretty to the point. Doesn't seem like Teradyne Robotics is backing down. Doesn't seem like Jocka Robotics is backing down either. But again, as we reported last week, this is the second lawsuit that Teradyne Robotics is taking against a Chinese collaborative robot arm maker. It's also in a battle with elite robots over copyright infringement. So interesting, it seems, Terradine is trying to defend its turf, Mike, right? Yeah, that's as every holder of IP should do and would be expected to do, right?

21:10I don't know how you operate in today's environment without having a good patent lawyer and, you know, trying to find some white space between all of that. But I've got my own opinions about what should be patentable and what shouldn't, but that's no issue here. It's just a matter of half of the battle of taking a new product to market is making sure that the lawyers have said everything is lined up. It's interesting they're doing it in Germany. I think Germany is one of the regions where very strong IP opportunities there. So they certainly are picking their court where they want to defend this.

21:50Yep. So we'll obviously continue to follow us. Again, of note, we've tried to find court documents and actual paperwork and the actual case filing for this as well. We have not seen it yet either. Again, Teradyne has told us it is coming. So once we find that, again, we will pass along. We'll learn more from actually seeing what's in the lawsuit. We've asked Teradyne Robotics for specifics about what they think or what they're alleging Jocka has specifically infringed on when it comes to the hardware and the software of the universal robots. Cobot arms, they haven't told us yet. So, again, if we're able to get our hands on those actual documents to learn more, we'll pass it along to everybody as soon as we find out.

22:36But just wanted to pass along that update. Again, we have you covered with all the latest news and information about the Global Robotics Network with our websites, The Robot Report and Automated Warehouse. If you don't already, make sure you follow us on social media. Make sure you sign up for our weekly newsletters. No better way to stay updated on the industry than doing that. On to the featured interview now. Again, joining us, you'll hear from Eric Sivertson. He's the VP of Security Business at Lattice. Mike, set this up for us. Yeah. So we've reported recently about, you know, the foreign robot bans in the U.S.

23:08that was for security issues. One of the things brought up, you know, in that foreign ban. So, you know, this conversation with Eric is timely because he's going to remind us of the role and the function that FPGAs have in a robot system architecture. Eric says that a compromised humanoid robot could certainly exfiltrate sensitive information. That's one of the likely scenarios. It could observe classified conversations or intentionally damage equipment. And so he treats that as both information warfare as well as potentially physical warfare. That's the potential outcome. But in this context, we'll come back to how FPGAs can deliver deterministic outcomes in a new world where AI could deliver uncertain outcomes or be compromised.

23:59It's one of those things everybody should be worried about. So it's a good sort of recentering about how any system architecture plays together and the role that firmware has. So let's listen to this conversation with Eric now.

24:17All right. On the show today, Eric Sibbertson, he's VP of Security Business at Lattice Semiconductor. Eric, welcome to the show. Thanks, Mike. Thanks, Gene. Happy to be here. Yeah. So we're interested to have you on the show today, talk about what's going on at the physical level with everything that Lattice represents. And so maybe in the context of starting out here, tell us a little bit about the types of products that Lattice is producing today, and specifically for our audience of robotics developers, sort of the relevant things that you guys are doing. Yeah, sure. Happy to do that. So Lattice makes field programmable gate arrays, FPGAs.

25:01And that's really, you know, it's a little different than your traditional software microcontroller product. We really are programmable hardware. And we've been in this business for, I think, 42 years now. We're one of the early FPGA vendors along with what was at the time Xilinx, Altera, and I forgot Actel, I think, was the original name. They've all changed hands. We're kind of last man standing of those guys that were all independent. Altera went over into Intel. Now they're spun out. Their private equity, Xilinx got picked up by AMD. They're part of AMD now. And Actel morphed multiple times through acquisitions and now part of Microchip.

25:49So we really are the last pure play FPGA guy out there. And FPGAs, like I said, are programmable hardware. So it's very different in software. In software, you're locked into an instruction set. You've probably heard the term ISA, instruction set architecture. And some common ones are x86. That's what pretty much all your laptops and PCs ran on for decades, started by Intel. And then along came the RISC products, RISC, for reduced instruction set computers. ARM was the leader in that. And then Sci-5 is a good example of kind of the red hat of the RISC-V community. But RISC-V was also a big project that came out of academia to have more of an open source RISC processor.

26:43But the foundation for all of those is a fixed set of instructions that software commands. And in FPGA, we don't have that. It's literally programmable logic gates. So you're going in and configuring how those logic gates operate. We can emulate with those logic gates these instruction set architectures. In fact, at Lattice, we have a RISC-V based soft implementation that can run in our gates. So you have really a lot less limitations with an FPGA versus software. But it acts a little bit like software because you do, quote unquote, program it, which is really you're configuring how those gates are going to operate.

27:29And the benefit of that technology, particularly in the robotics side of things, is we can give you very strong determinism and we can typically run things faster. Common term is to unroll an algorithm. So if you have an algorithm that's running in a instruction set architecture, it could take multiple clocks to run through all the instructions to perform it. Because we can program that hardware to do whatever we want, we literally can unroll the algorithm, if you will, and put it in more gates than using software instructions. And by doing that, we can sometimes have an algorithm that could take hundreds of clock cycles, execute in one clock cycle.

28:14So you can get more determinism and you can typically get better performance when you put something in an FPGA. Yeah, certainly for system designers, they're very familiar with FPGAs, but I think, you know, for the broader audience and for folks that don't go under the hood to speak so often, it's important to know how important they are in systems design today and the role that they play. And when we first considered having you on the podcast, given your role and as well as the role that the FPGAs play in the systems design and control architecture, we wanted to explore the concept of safe operation in the world of physical AI.

28:54I just got off a podcast, I mean, a webinar about physical AI and we talked about safety there. So, you know, you've said recently that if you have a safe humanoid, you can have a safe humanoid, but if it's connected and vulnerable, it can be exploited. I think that's the other aspect of, you know, what FPGAs represent in the entire architecture. So can you explain to us a little bit more about how physical AI changes the threat model, you know, for these systems? You know, we've had the recent FCC banned because of potential, we'll just say, call it potential malfeasance in the future. And, you know, we want to make sure we have safe architecture.

29:39So compare the, you know, traditionally IT or cloud AI, all those sort of things that have happened in the last decade to where we are with, you know, physical AI and robotics today. Yeah, you know, this is a good question. And it really, I like to use humanoids as the kind of standard bearer of this intersection of AI and the physical world. You can look at robots too, but humanoids really highlight this problem and the situation. And the way I like to say it is that you can't trust a humanoid if it's not both safe and secure. And this is really the big paradigm shift that's about to hit this industry.

30:26If you look at robots today, a lot of robots are, you know, in a cage or they're locked to the floor, if you will. And so they have limited mobility. When you get to a humanoid, and some robots fall in this category, too. You can think of, you know, think of actually a Tesla is really a rolling robot in a way. So, but when you think about these intersections of what I would call cyber with the physical world, you know, another term that's used is cyber physical systems. The humanoid is an easy one for people to wrap their head around because it really is this intersection of AI and the physical world.

31:12And that's why I say it needs to be both safe and secure. If that humanoid is just safe, right? In other words, if it follows the typical safety paradigm that it does certain functions within its limits, that's good. You have to have that. We want the humanoid to be able to detect the force properly and have a way to stop when it might be going to bump into a human or have some guardrails to keep it from doing bad behavior. But because it's typically run by multiple different components, some are software-based, some are FPGAs or hardware-based or ASICs, particularly those that are programmable.

31:59Now the security part kicks in because it's very easy for bad guys now to come in and attack that software structure. So you've built all the safety parameters. But if someone can go in now and adjust the data set that those safety parameters are using to make decisions, it's not really safe anymore because this now could be used for bad purposes. So I think, you know, when you have a cyber attack on infrastructure today, everyone thinks about my system goes down, I lose my ransomware, I lose my data, I get locked out. I mean, that's not a good thing. But imagine now there's a thousand humanoids working in a particular area and they all get hacked and now they start doing really bad behavior and they're moving 50, 100 kilogram things around and now they turn into weapons, right?

33:01So I think the intersection of the AI with the physical is going to drive this safe and secure concept. Whereas traditionally, we've had secure systems and safe systems, but not everyone has really put those two together. And I think this new paradigm we're going to go into is going to force that.

33:23Eric Sivertson: Eric, I know we recently spoke with you and Lattice for our download on humanoids and mobile manipulators, but let's talk about endpoint security and specifically about IoT devices and robotics on the factory or warehouse floor. A lot of the systems today depend on cloud connections to function, right? They're reporting data, they're processing workloads. For a lot of automation users, this creates exposure in their networks that maybe wasn't around a decade ago. How is hardware, as you've discussed, sort of foundational to having that trust or that level of security that you've been discussing?

34:08Yeah, well, there's a fundamental weakness in today's AI models that not everyone talks about. The biggest weakness, especially when you get into safe and secure type operations, AI models today are non-deterministic. So you really, at a fundamental level, AI is not deterministic. When you build a safe system, one of the most critical elements of it is it's deterministic. And what do I mean by that? That means if a series of events occur and a decision is made on those events, you always get the same outcome. With AI, and you've probably seen hallucinations, right? Everybody's experienced that.

34:49You ask your favorite AI engine a question, and it comes back with a nonsense answer. And you're like, what the heck? And sometimes you can go back and ask it that same question, and it's not as nonsensical. Or you change one word and it becomes less nonsensical. AI must run to completion. And it's really a probabilistic system, not a deterministic system. And so that's a huge problem for safety. So where FPGAs really are handy is, and this is what we see evolving back to that safe and secure. the cloud does a lot of these AI inferencing and information. So you can feed it images and you can almost feed it real time images now.

35:39And it comes back and can modify those or make decisions on those. And so you can use the AI as a first order filter, for lack of a better way to put it. So the AI is now going out and giving the humanoid or the robot or the distributed factory set of robots, if you will, kind of instructions on what to do based on the environment. Like if it's a production line, there's multiple sensors and cameras reading lots of information about stuff coming through. AI can very quickly process that and make some decisions to do this or that with, let's say, you want to throw out bad potatoes on a potato conveyor belt.

36:26But the problem is what happens in the second order when that AI hallucinates or does something really stupid and it's non-deterministic and it's a safety hazard. So this is the beauty of the FPGA. We can now use the FPGA as a second order filter. Let AI make some big decisions, but we can impose deterministic guardrails on those decisions and we can do it in real time. And that's probably going to be one of the biggest benefits of using an FPGA and the cyber physical autonomous AI type systems. While we're still on this topic of secure, I'm curious, you know, from the hardware level, because again, you talk about, you know, communication at the hardware level with all the sensors, but also again, with the network and communicating with other devices, So what are some of the issues that hardware developers might need to consider when it comes to securing communications in today's world across all of these devices and the cloud?

37:26Yeah, I think a couple of key security features need to happen. Attestation is a big one. But am I authorized? Am I what I think? The system looks out and says, hey, is this PLC controller or is this sensor the one I think it is? So you need to be able to test that these are trusted entities, right? And it goes the other direction as well. That sensor needs to know, am I talking to this manufacturer's true cloud? or the enterprise data setup? Am I doing that? Is that who I'm able to talk to? Or wait a minute, I don't think this is who I think it is. I can't talk to them. And then the next thing is, are both ends running the right stuff?

38:19So this is where you get into digital signatures. And this is where, is this the right authenticated code to be able to run? So do I know who I'm talking to? And think of it, it's a human process, right? You go in, you're going to get on an airplane and the guy walks in and he doesn't look like a pilot. He says, I'm your pilot. And you're like, wait a minute, there's no way that guy's the pilot for this plane. It's the same concept, right? But let's say, okay, now I know it's a pilot. Then the next issue you've got to know is, is all the systems properly configured to do what I expect it to do?

38:56And so I need authentication. And if you notice a pilot, he'll go through a checklist. You sit there at the gate for a long time while they're going through a whole bunch of checklists. They're really authenticating the aircraft to make sure everything on the aircraft is set to perform properly. So you need to do that process. And then once you have all that, now you can start to run. And the next layer in protecting on the run is that we actually could encrypt that data. So now it's hard to intercept that data or spoof that data and give bad input to the system. So we need to know the entities that are running the system or who should be there.

39:34The stuff it's running is what should be there. And then once it's running, we have confidence that it's running properly and safe and secure. So those are kind of the concepts that you want to look at. And the beauty of the FPGA is it can fit in all of these pieces to the system.

39:52Eric Sivertson: And you talked about the networks. how challenging is it to deal with like internal networks whether it's complex devices like a humanoid a mobile robot or the totem sport or even you know with these in-house networks where people say oh it's in a secure facility but is that still a risk? Is there still some issues there? Yeah that's huge. I mean you have open and closed networks right? When you have a closed network, and I like to say inside what traditionally in security we've called guns, guards, and gates, right? So an OT network inside a factory where there's a lot of physical controls to get into that factory and to get into that OT network.

40:42Security on that network can be a little lighter because you know that the only people that can get into that building, because you have guards, And sometimes if it's the most important one, they're armed guards and only authorized personnel can come in. And the only part of the network that's running is inside that protected perimeter. So a lot of traditional OT systems have not worried about security as much because they have that guns, guards, gate model. Now, when you bring that OT network and connect it to the IT network, because you want all the benefits of the IT system, right? better interfaces, a rich set of data tools so you can analyze all the data that's coming out, all these SaaS models that give you all these services that were never built for the OT world.

41:35Well, once you cross that barrier and connect up to the IT networks, I mean, just do a Google search or your favorite AI engine and say, tell me about the latest network attacks to the IT world. And it won't stop streaming on your screen, right? So it's when we cross that barrier that we're no longer guns, guards, and gates. We are now exposed to the big World Wide Web and all the nastiness that comes from it. So it's that network crossover where the problem happens. And the key to being able to survive that is you need all the good network security tools applied. But this is where a concept called cyber resilience really kicks in, where all these and zero trust.

42:24So cyber resilience and zero trust are the two security concepts you want to think about when these OT networks now cross out into a much larger, basically exposed attack surface. And that means each component in the system doesn't trust anything else. So you're always doing that attestation authentication process. But then while you're running, this is where the cyber resilience part kicks in, you are constantly looking to see if you're under attack. So you're not trusting anything. And as you run, you're under the assumption somebody in that IT world got into my OT world and is going to wreak havoc.

43:01I need to be looking for that constantly. And the minute I think, or the first command, I think I see where that's happening. I want to stop that attack and bring the system back into a safer, secure state that I trust. So that's kind of, you know, said a lot there, but that's kind of what's happening as we transition from the guns, guards, and gates of a closed OT into the big wide world of bringing IT over to give us all the benefits of that world. Yeah. And thanks for sort of the stepping back, pulling our heads up to the bigger view, picture view of where many of those risks are. But now I want to ask you to dive down a little bit deeper because you've said, you know, multiple times you've said, this is where FPGAs come in.

43:49So let's talk about where FPGAs come in to this model of risk that you landscape that you just described. And what are some of the importance of guidelines, designs for firmware safety today that are in place or should be considered as hardware designers making decisions about their architectures? And what should robot manufacturers specifically be concerned with for their system designs? Yeah. So the first point, I'm going to pick on the cyber resilience piece because it's well established today with secure microcontrollers and with FPGAs to do that attestation and that initial authentication.

44:32We'll call it a secure boot, if you will. Microcontrollers, FPGAs, system components, all pretty well have established how to do a secure boot. So that's how you start everything up. And I know I'm in a secure, trusted state. The part that's new and the part where the FPGA really kicks in is the cyber resilience. And so this is the runtime side of things, where we're now, we trusted we started properly, but because we have this vector coming in from the outside world that could come in at any time on any component in any way to disrupt its operation. And so where the FPGA kicks in, and a lot better typically than some of these, what I'd call one or two core, like 32-bit secure microcontrollers that are at a low cost point, they get interrupt star very quickly when they're trying to monitor many things because they basically have a limited instruction set that can dig deeply on multiple channels at the same time.

45:35Whereas with an FPGA, we can unroll everything, we can run everything in parallel. And so we can like, I'll give you some detailed specifics. A lot of buses for controlling these things are on the I2C bus or I3C bus or on a SPI bus. we can take with the FPGA and we can look at multiple channels, I2C, I3C, SPY, all at the same time. And because the efficiency of being at a hardware level, we can look on a command by command basis on any of those channels. So let's say we have a sensor that should only be reading information from its non-volatile memory store to tell it what to do. We're watching that spy channel where it's doing that.

46:25And we know that at this particular time, it should only be reading. It should not be writing. All of a sudden, this thing now issues a write command to its non-volatile memory. Well, we can catch that in real time. We can see that command is not a read, it's a write. And because we're outside of that system, we can now stop that sensor, go and interrogate its non-validable memory and see if it's still in an authenticated state. Oh, wait, it's not. Obviously, something hacked it. Something got put in there. But now we've held the sensor in reset. We've told the system, you can't trust this sensor right now.

47:02And we're going to recover and put a known good image back into its non-validable memory store and let it start back up. So that's kind of at a lower level what you can do with the FPGA that is not as easy to do with some of these other, what I would call low-cost secure controllers, because they were really built for that OT environment where once you started, you trusted the system. You didn't conceive that, oh, no, I'm under attack. Every single command I need to be watching.

47:32Eric Sivertson: Well, and we mentioned a moment ago, the FCC trade restrictions that are based on national security and concerns about cybersecurity, like Chinese-made robots being hacked or having back doors. How serious do you consider that threat to be? Oh, I think with humanoids, the nature of espionage and warfare are going to change dramatically. And I'll take the espionage angle, for lack of a better way to put it, information war. It really, I hate to say it, people don't like to hear this, but it is warfare. Usually you have two forms of warfare. You have information warfare and you have kinetic warfare.

48:23Kinetic warfare is understood by everybody. Get in a tank, you drop a bomb, you shoot, that's very easy to understand. Information warfare, which is really the cyber world we're in, is not as easy for people to wrap their heads around. There's all kinds of levels to that. And I'm going to use the espionage angle. Let's say you have these humanoids that you bought from a country that you don't trust and you're afraid there's a backdoor. Those humanoids are now in your factory building your most secure fighter aircraft. So we're going to cross from the info to the kinetic side, right? And those things now have vision, they have sound, and they have control.

49:05So they can do multiple traditional espionage things. They can exfiltrate how the design and how the thing is built. They can exfiltrate conversations by the employees that are having classified conversations, how they're doing things. and they can now also be trained to go in and use the wrong torque on a screw. And so now the aircraft, after so many hours, starts falling apart in the air. So the fact that you have to worry about these things from a foreign nation that's going to inject a tool to build all this stuff, it makes sense, right? So yeah, I think it's a very real threat and a very real problem.

49:48Eric Sivertson: And given that it is a significant threat, what should robot manufacturers do to harden their systems against these threats? Or robot designers even? How should it be thinking about these things? Because security is often an afterthought. Safety is often an afterthought. We've talked about that before. But given the magnitude of both the global espionage and even corporate espionage dangers here, what are some measures that they can take? Well, I'm going to put a shameful plug in for Lattice. Use Lattice FPGAs and their secure capabilities. That'll help. But I think the number one thing is we are moving out of that guns, guards, and gates model.

50:39We're in a very, as you guys well said, distributed AI cyber physical world. That's where we're moving into. And so when you do that, you really have to do this safety and this security at the very beginning of the design cycle and take it all the way to the end. So for companies today that are going to build robots or humanoids or any AI cyber physical type system, they really need to think about how do I attest my system, not just to my customer, but to myself first, probably. How do I know none of the things I designed for purpose have been compromised? And then how do I prove that to my customer so he can trust my humanoid to do what it needs to do and nothing else?

51:33So that's really a mindset change. And many companies today are not there. You can see them talk about safety or they talk about security. and not many of them talk about these two together over the life cycle of their product. And I would say if you as a consumer went to one of these companies and said, prove to me that everything in this system came from you and only does what it's supposed to do in a deterministic way, many of them will struggle with it, especially the startups on the humanoids because they're gonna be very focused in giving you a humanoid that does cool stuff. But these other components are very expensive to do.

52:18Not everybody wants to build that in in the beginning. So what usually happens, and you see this all the time, a bad event occurs, the humanoid kills a bunch of people in a factory, and then all of a sudden the government agencies say, we must have safety protocols so this can't happen, right? So unfortunately, that is the nature of the cycle. We haven't had that, I would say, moment yet in the industry. And hopefully we'll see more leaders in that space push this concept of, I have a safe and secure robot you can trust or safe and secure humanoid you can trust. And I do think that will become a selling point, but it's a fractured industry right now.

53:03You're going to see pieces of this come out, you know, evolve, if you will. Yeah. Well, I think what we've seen evolve is the development of these early systems, testing, a lot of great demos. Everybody's in a race to try and create a functional system and has done amazing things with gait and walking and dancing and fighting and all that sort of stuff. But the thing that they haven't done is optimize the bill of materials costs, right? So up to this point, let's get it working. Then we'll figure out, again, how to reduce bomb costs. And you just said something that scared me a bit, which is, you know, in the race to reduce bomb costs, now there'll be decisions made about how do we optimize, you know, everything that goes into, you know, building the system.

53:52So I want to bring you back to your guns, gates, and guards analogy back down on the chip board. Again, let's come back to what FPGAs can do to be a security guard on board around all of the system on chips, the NVIDIA chips and systems that are the main processing units. Again, what's the role from an architecture perspective of deploying FPGAs across the various subcomponents of a system like a humanoid robot? Yeah. So particular to Lattice, we call them our root of trust ROT products, if you will. And so in each of our major FPGA families, we have ROT components. And one of the key differentiations from most of our competitors in those ROT products, we have the non-volatile memory contained within the chip.

54:52And we have it in such a way that we can boot from two different flash images and we can lock those flash images. So you effectively, once you have that chip leave your factory, you can lock down its configuration. It's reprogrammable, but yet you can lock it so it can't be corrupted. And then because we have two of these flashes, once you have a known, good, working, secure, safe image, it can supervise the update of the other image. And so it's very difficult to attack that. And now back to my cyber resilience piece, most of the main processors or the AI engines do not have internal flash. It's very hard to do that in these very deep, you know, small nanometer.

55:47You don't have a three nanometer flash today, right? But all the GPUs and all the CPUs that are the highest capacity are moving into that two and three nanometer geometries. So it's impossible for them to get flash into that same system. And the systems have been designed not to have that. I mean, look at the memory wars that are happening, right? Memory is considered a commodity, whereas the GPUs, the CPUs, the MCUs are not. They're distinguished by what their capabilities are. It's why NVIDIA is what they are today. It's not a commodity product. They all have to go and get their instructions because it's an instruction set architecture.

56:29They have to go get that software to run from an external non-volatile memory, typically flash, but it could come from the cloud as a system download. But that process is extremely vulnerable to attack. That is the place you can hit in a big, big way. So the GPU, MCU, CPU can securely boot itself, right? It can make sure, hey, I'm not going to run code that has not basically been signed for me to run. But the bad guy is smart. He's like, I don't care about that. Let's just not let it run. Let's make sure that thing doesn't start. That's your classic ransomware. That's your denial of service. And man, you don't want denial of service when your car's driving down the road and going to automatically turn and denial of service stops and it keeps going and off the cliff you go, right?

57:18So the whole idea with cyber resilience and why particularly lattice FPGAs can be very helpful to these systems, we can stop that from happening. So we can be monitoring, like I was saying earlier, that GPU reading its instructions. And if we see anything behaviorally where a problem is occurring, we can even have a channel to that GPU because it can securely boot. It can tell us, hey, I'm in a booted state, but I can't run because I'm not getting authenticated code anymore. We can now take action and go recover the image for that GPU and get it back online. So even though that GPU is much more sophisticated than our part and even a lot more expensive, you can use some of our parts are sub$10.

58:08You can use our sub$10 parts to guarantee that$100 or$1 ,000 part is always going to function and do what you want it to. So that's a big part of the cyber resilience story and why using Lattice, and we call it as companionship to a lot of these systems, will help you stay secure and safe. It's the gatekeeper.

58:27Eric Sivertson: Well, as we're looking at, you know, different types of robots, and I think you're absolutely right that this applies to any robot that's going to be in sensitive settings, but also as robots become either more autonomous, more driven by AI, that need for that baseline of security increases. Do you think that's going to affect the rate either of development or spread the adoption of humanoids and other systems beyond their existing industrial base? Maybe. I think, you know, having done this, you know, I've been doing security since actually high school, to be honest, probably since a toddler. Um, the, the, the pattern is always the new technology and its promise captivates everyone.

59:21And you, you end up in a race to get capacity and capability, security, um, and even safety. I mean, look at seatbelts. I mean, God, how many accidents and dead bodies had to happen before finally a law got passed. You have to have a seatbelt, right? So I think the technology promise for efficiency or life improvement and all that is going to outweigh the safety side and the security side of things. So I don't think it'll stop the progression at all. I think it's going to move very quickly. In fact, this is probably the fastest technology cycle we've seen in recorded human history. So I don't think we're going to see getting safe and secure slowing that down.

1:00:10I do think that the guys that get that right, though, will get market advantage in certain areas. I have a Tesla myself. I love self-driving. They call it full self-driving, FSD, right? But then in parentheses, supervised. Why is it supervised? Because you can't trust it yet. Right. So it's the easy one in the car. Exactly. So I think now if it would be so nice because I get a strike every once in a while because I grab my cell phone and I look at it and the camera in the car is like, you need to pay attention. Right. It would be nice if it was truly full self-driving without the supervised part.

1:00:53And I think when you get to that and you can really offer that, you're going to sell more of your vehicles than the guys that have to have the supervised still with it. So I think the rate of adoption and the, let's call it the market leaders in this, they will get an economic advantage. They'll get more margin. They'll be considered a premium or better product in the marketplace. And for some situations, like these military things we were talking about, you won't be able to sell to that market without it, but it's a smaller market. But it could extract a higher premium, and it will bring the rest of the industry along.

1:01:32I think that's more likely what will happen than, hey, we better slow down because it's not safe and secure. Nah, it's going to go very fast. You're not going to see it slow the market down.

1:01:43Eric Sivertson: Well, and there's usually a regulatory lag, right? I mean, even let's say somebody, let's say the government said, oh, you know what? Everybody should be using FPGA. That would be great, but they would probably make that ruling five years from now. And so that's not necessarily helpful. The industry has to somewhat police itself. And also, I think you're right that the market dynamics will ensure that people start thinking about this more and more. And certainly people like yourself talking about it helps highlight that need in a way that otherwise can easily get overlooked. Yeah, and I'll use, we do a lot of business in the data centers today.

1:02:26In fact, we are, as a companionship, we have over 100 % attached on every server on the planet. And some of the, I don't want to get into our exact customers, but I can say this, that the leading data center and, you know, hyperscaler type companies, they understand the advantage we bring. And they highlight that as one of the reasons they have, you know, platforms you can trust. So I already see that in a different market, this concept is playing out very well. And I think it's going to happen in more of the cyber physical world as well. Well, Eric, it's been a great deep dive here into the world of FPGAs and security.

1:03:13You've really given me a lot to think about. I hope the audience appreciates as well, you know, the role that you play and the function that FPGAs play as well in system architecture from safety and security perspective. So once again, thanks for your time. Thanks for sharing your knowledge. And again, great conversation about a very important pedestrian part of the architecture that's very key to our future. Well, thanks, Mike. Thanks, Gene. Very happy to have been on this podcast with you guys. A lot of fun.

1:03:56Awesome stuff there. I think just to come back to how you set it up, Mike, right? AI models, they're probabilistic, not deterministic, and that creates the fundamental challenge for safety-critical systems. What stood out for me was where he described AI as being useful for first-order decisions, but that FPGAs can enforce deterministic second-order guardrails in real time. Interesting stuff there, but what were some takeaways for you? Yeah, well, just a reminder. I mean, Eric obviously works for one of the leading FPGA manufacturers, so he's certainly bringing it from that perspective. But I was just impressed by his vast amount of experience in this.

1:04:33And, you know, we became friends. I think he's going to become a resource in the future. Anytime I have questions about, you know, system architecture and how FPGAs play here. And, you know, firmware is critical for every modern computing system. We all know that. That's long before AI became a thing. And now that AI is becoming a part of every embedded systems, and as most in the audience know, It's good to take a look at all of the vulnerabilities in the architecture, both for safety use cases as well as security. And we harp on safety all the time, but it's also important to understand the role that it plays in security as well.

1:05:15That should also be important. And I think that a lot of the times as we talk to folks that are just launching products for the first time, Steve, that safety and security are often the second thing that they worry about. The product's functionality is of most importance, and then we'll make it secure and safe. And so, you know, embedding things like, you know, the FPGA and the firmware, you know, deep into the system architecture, it should be a primary use case to understand what vulnerabilities exist there. And so, Eric brought us some great experience to that topic. Yeah, great stuff there.

1:05:50Come back anytime, Eric. Really appreciate the time. Yeah, that's going to do it. But before we go again, just a quick reminder about Robo Business, upcoming event from us, October 20th and 21st out in Santa Clara, California. We're talking all things commercial robotics development and deployment, physical AI. We're doing humanoids. We're doing startups and business and investments. Eli Lilly is someone that we haven't announced yet. They're going to be there talking about human-robot interaction and challenges with that that they're trying to overcome on the show floor. Again, we'll have 70-plus speakers on our stages over the course of the two days talking all things commercial robotics development and deployment.

1:06:34If you're looking to scale your robotics business and deploy robots and get them out there in the real world, there's no better show on the planet than RoboBusiness to do this. This is our 20th anniversary show. We'll have over 100 exhibitors on the show floor, over 2 ,000 people at the show. So you can head to robobusiness.com. I think most of the agenda is done and up there. Again, we have to add Eli Lilly and a couple of other folks, which we'll do here in the coming hours. But a lot of amazing people are going to be there sharing their commercial robotics knowledge for two days. So take advantage of it.

1:07:06Again, head to robobusiness.com if you want to learn more. We now have a way for you folks to save our loyal podcast listeners. If you enter Robo 15, that's R-O-B-O 15 during checkout, you'll save 15 % off your registration. So thanks to Mike for getting that set up. But I appreciate all of our listeners for tuning in. And we hope to see you guys at Robo Business October 20th and 21st. I'll be out there. I think I'm going to try to get out there on the 18th, right? So a couple of days early. So if folks who are listening to this are located in the Silicon Valley area and want to show us around your place, let us know.

1:07:42Always looking for some. And I'm sticking around for the Thursday afterwards. So I'm setting up some visits on that Thursday. Yeah. So if you guys are doing some fun stuff and want to show us, we'd love to come by, swing by. But again, hope to see everybody at Robo Business October 20th and 21st out in Santa Clara. Head to robobusiness.com and enter robo15 at checkout to save 15%. All right. That's going to do it for episode 260. Thank you again to everybody for tuning into the show. Thanks to Eric Sivertson, the VP of the security business line at Lattice for joining us as well. New episodes of the Robot Report podcast drop each week.

1:08:20You can find us anywhere you listen to your podcast. Please subscribe, leave us a rating, give us a review. This episode was produced by Mike Oitzman. It was edited by Tyler Kubisti. Our executive producer is Amanda Buhner. I'm Steve Kroll. Thanks so much, folks, for tuning in. Have a great week. Have a great weekend. We'll talk to you again soon. Thank you.