#226 Nathan Michael: The Future of AI-Powered Drones and Autonomous Warfare (with Shield AI)

0:00What we've really focused on is creating the tools and the means to specialize in autonomy and resilient intelligence capabilities to different platforms, different mission sets, different needs. So whether it's one platform or multiple platforms working together, taking all of the different parts of robotics, autonomy that you need to create intelligent systems that can really think for themselves, using the best methodologies possible in terms of the state of the art for all of those different things, how you enable those systems to perceive the world, think about that world, take action within that world, do so as individuals and as a team.

0:35And we've created basically a framework for the autonomy development life cycle and the specialization of that kind of intelligence to different platforms. When I started working on my own and I didn't have the structure of an organization behind me, I knew that I needed to incorporate and file taxes and all of that stuff. I looked into getting a lawyer, but it was prohibitively expensive. At that point, I really wasn't making much money. So I turned to LegalZoom to get the corporation officially registered and legally compliant. This episode's sponsor is LegalZoom, a trusted leader when it comes to setting up a business right.

1:20I'm a true believer because I still use LegalZoom to keep myself. Business owners, have you set up your company properly? The pain-free way is LegalZoom. It's the most trusted way to set up your business and stay out of legal trouble. So launch, run, and protect your business to make it official today at LegalZoom.com and use the promotional code IONAI to get 10 % off any LegalZoom business formation product, excluding subscriptions and renewals until the end of the year. Get everything you need from setup to success at LegalZoom.com. That's L-E-G-A-L-Z-O-O-M.com and use the promo code IONAI. That's E-Y-E-O-N-A-I.

2:17LegalZoom.com with the promo code IONAI. LegalZoom provides access to independent attorneys and self-service tools, but it's not a law firm and does not provide legal advice, except where authorized through its subsidiary law firm, LZ Legal Services, LLC. Can we start by having you introduce yourself, how you got to Shield AI, what Shield AI's mission is, and then I'll start asking questions. Great. Yes. And it's great to talk to you today. So my name is Nathan Michael. I am the CTO of Shield AI. I've worked in the area of AI and robotics, primarily robotics, for about 20 years. I was a faculty member in the Robotics Institute at Carnegie Mellon University up until about a year ago.

3:10And in the Robotics Institute, I built out a reasonably sized lab that focused on this topic of resilient intelligence, where we were really trying to create systems that could think for themselves and operate in extremely challenging environments, both single platforms going into unknown environments, as well as teams of intelligent agents operating and exploring unknown challenging conditions which within in today's language are always degraded and denied limited to no communications underground in tunnel networks caves those kinds of things or other types of really challenging conditions where access to external information did not exist no prior information or limited prior information existed and so now you needed to create these systems that could really think for themselves operate with very limited prior information and go in, explore, understand, model, interact, and engage.

4:09In that work, I largely looked at air platforms, but also worked with some surface vehicles and underwater vehicles, as well as land vehicles. And homogeneous teams, teams of similar characteristics, as well as heterogeneous teams, and teams with different characteristics working with each other. So around 2016, 2015 to 2016, I started to see what we were creating within a laboratory environment, really kind of stacking up on the shelf in terms of getting stuck in this technological valley of death where you have these great innovations, but these innovations are not being fielded. They're not being operationalized.

4:47They're not being used and leveraged by end users. And for a lot of what we were creating, this would have added a lot of value and really delivered tremendous outcomes. From a laboratory context, then, within an academic context, I really focused then on working with different parts of, well, different funding agencies and laboratories, like the Army Research Lab, Air Force Research Lab, of course, the National Science Foundation, and a number of other parts of the U.S. funding sources for academia. Ultimately, I came to the conclusion that in order to really have impact and reach scale, I would need to be able to bridge this gap and started looking at either starting my own company or working with an existing company.

5:36I started to engage and advise technically Shield AI, which was very small, just a few people at the time back in 2016, and then ended up agreeing to join the team, the leadership team back in 2017. At this point, you know, fast forwarding over time, what you can see is a lot of the vision that was laid out back in, say, 2017 toward the future of building out this high degree of intelligence running on board these systems, building out this capacity to not just develop, integrate, test, evaluate, deploy, operationalize this high degree of intelligence, but doing it across a number of different classes of platforms and then fully seeing those capabilities fielded.

6:23has been something that we've been working on for quite a few years, and we've seen through successfully several times over or many times over. And so that really brings us to the present. Today, Shield AI's mission is to protect service members and civilians with artificial intelligence systems. We deploy our capabilities on everything, primarily in the air, but everything from really small systems, the quadcopters, all the way to more medium-sized systems. We have a platform called the VBAT Group 3 system, and then all the way up to uncrewed jets. So these would be Group 4, Group 5 platforms.

7:03All of these platforms are leveraging what we call our HiveMind capabilities, but really HiveMind is, let's say, jargon or marketing term for really our AI and robotics capabilities, which consist of sort of an edge-level intelligence that runs on each platform. So you could think of this as an autonomy stack that uses everything from more traditional autonomy capabilities to stay-of-the-art and cutting-edge AI capabilities. So everything that makes the most sense to really deliver that high degree of intelligence. It includes capabilities related to building. So HiveMind includes capabilities that you could call like an autonomy factory in order to really create high performance capabilities depending on the scenarios or the mission, the environment, the context, the needs, and then to apply and specialize those capabilities to these variety of different classes of platforms.

7:56So that's sort of what we do and what we're doing. Probably just to wrap up that intro, we've had a degree of success in terms of creating systems that are able to operate in these degraded and denied conditions where all the intelligence is running on board the platform. We've seen some recent recognitions for the fact that we're deploying these systems in a variety of different conditions, particularly where it's not possible to get external information, not to be able to consistently communicate with the platform or to have the platform consistently access external information like GNSS data or some way.

8:34Yeah. Well, and you're a software company. You're not building hardware, are you? Yes. Yes and no. So in fact, Shield AI has two businesses. Really, one business is focused on hardware, and that's building out our aircraft, and that's the VBAT. That's that medium-sized Group 3 platform. And then the other business is software, and that's HiveMind, and that's our AI and robotics ecosystem. them. And so I primarily focus at this point within the company on the hive mind side of the business. But one thing that we really focus on and specialize toward one area of strength is we've really organized our hive mind business to not just create a product, which I described a little bit about earlier, but also to very efficiently and effectively apply that product, specialize it to meet needs.

9:29And so we do that with our VBAT platform. And so that allows us to deploy our VBAT platform into a variety of conditions. So you can almost look at our VBAT platform as sort of the tip of the sphere, leading edge capability, where we can very efficiently and effectively through a vertical integration, deploy capabilities onto that platform. But then we also are organized to specialize and deploy those same types of AI pilot capabilities to other classes of platforms. And we do so with a variety of other companies that want to be able to have this type of resilient intelligence and autonomy capabilities on their systems.

10:07Okay. And you're saying VBAT, like virtual BAT. Is that right? No, it's VBAT is vertical BAT. Actually, there's no explanation for the acronym. if you look. So we've tried to retroactively come up with something that makes sense for that acronym. But actually, there's not really an explanation for the acronym. But what it is, is it's a tail sitter. And you can look it up online. It's very readily accessible. But it's a tail sitter that will take off sort of in a hover state. And then once it gets to a certain height, it'll transition to on wing. And then that allows it to fly longer distances, operate different kinds of surveillance, intelligence surveillance, reconnaissance type missions, and then come back and then land in a tail sitter mode, which allows it to take off and land from fairly confined environments under a variety of different conditions.

11:06Yeah. When you talk about a couple of things, you guys have been fairly vocal about opposing lethal autonomous weapons laws. Am I correct on that? I'm not sure that we've been either one way or another vocal about it, so I'm not sure what you're referencing. So perhaps you're thinking of another entity? Well, in any case, that's been a debate. When I was at the National Security Commission on AI, that was a big debate. and the U.S. military has been very adamant that it doesn't deploy laws or lethal autonomous weapons. But then you have like Palmer Luckey who's saying, yeah, we're ready. You know, if the Chinese have them, we'll go for it.

12:08So HiveMind is, and correct me if I'm wrong, is AI that allows multiple VBATs, for example, to operate in a swarm environment. environment or a yeah let me talk let me talk about it a little bit because it's interesting when you're building out autonomy capabilities you're building out and bringing together many different types of technologies and there's a lot of flexibility and a lot of ways that you can use that technology and there is no one-size-fits-all solution and there's no one-size-fits-all uh methodology if you will for the type of work that we do so what we've really focused on is creating the tools and the means to specialize autonomy capabilities and resilient intelligence capabilities to different platforms different mission sets different means so whether it's one platform or multiple platforms working together so teaming or what you're you're referencing as swarm whether it's v bats or other classes of systems and so what we've really focused on is taking all of the different parts of robotics autonomy that you need to create intelligent systems that can really think for themselves using the best methodologies possible in terms of the state of the art for all of those different things how you enable those systems to perceive the world think about that world take action within that world do so as individuals and as a team.

13:48And we've created basically a framework for the autonomy development lifecycle and the specialization of that kind of intelligence to different platforms. And what we've really indexed on to your point of sort of this more social question or the question of lethality, if you will, is we focused on enabling an immense amount of flexibility to design exactly the systems you want to be able to deploy to meet your needs and to do so as efficiently and effectively as possible. And then on the topic of lethality, our view and the general approach is we really are living within the guardrails of society.

14:30And so our society says what they do or do not accept. And that's true for every country. And so what we've designed is really a means to develop very quickly and efficiently intelligent autonomous systems with the flexibility and the controls, if you will, from an engineering perspective in place to make sure that whatever you're building meets those expectations and guardrails set forth by society, policy, standards, and so forth. And that will differ and vary between different countries. So in terms of taking one stance or another, I don't think that we're really taking a stance per se, but more deferring to what are the policies, the guardrails, the guidance that our politicians are, and whatever that is set out in whatever country.

15:21And then what we're focused on is creating the tools, the technologies, the frameworks that make it possible to design exactly the technology that meets those expectations and constraints, as well as the needs of the different operators and users that would be working with the system. Right. Uh, the, uh, uh, and, and the software you, um, deployed in, uh, VBAT for example, but it could be, I mean, you, you talked about, uh, fighter jets. It could be, uh, deployed in a, in a fighter jet, uh, manufactured by one of the big, uh, Oh, it has already. Okay. Yeah. And so this is public knowledge. We deployed what we call our high mind pilot, or sometimes I'll just say AI pilot, but our high mind pilot onto what's called a, well, onto what one in F-16 Vista, which is a modified F-16 that's designed to support autonomy.

16:26autonomy. And so that allows then that AI pilot to live on top of the platform autonomy. And so then, even though it wasn't uncrewed, there was a pilot inside the platform flying it. They flipped a switch and then the AI pilot took over and engaged in various maneuvers and so forth. And then of course, at any time they could flip the switch back and take back control of the platform. So that's one example of where we've done that. Again, this is public knowledge we've deployed on more uncrewed pardon me uncrewed jets and so these are um and and there's uh for example something called the fire jet from a company um called kratos we integrated and deployed on that integrating the capabilities not just onto or one platform but also multiple platforms working together and so there are a lot of examples of where we've deployed our hive mind pilot onto other classes of systems on the other side of the the spectrum we deployed the hive mind pilot onto small quadcopters.

17:25A few years ago, we used to make our own quadcopters. They were called the Nova, and we made a few generations of that. But we had a Hivemind pilot instance that was deployed to those. Those platforms actually were recently used in some operational context as well. So these are systems that have been operationalized and actually used in combat for many high-profile missions. In the quadcopter context, it really focuses on how do you build a system that can really operate in a completely unknown environment? So imagine an underground tunnel network that is far too dangerous to enter for a person, but sending that platform in and being able to navigate within that environment with or without communication, with or without external information, to be able to build up a model and understand both the layout, but also what's in that environment.

18:15And so we built those capabilities out as well. Yeah. In the quadcopter case, how is that information then transmitted back if you're in a jammed environment? Sure. Yeah. So the way to think about it is it's as if the robot starts with no prior information. There's no presumption of access of information. So it's sort of, if you will, opening its eyes, turning on its sensors and starting to build that model of the world. All that while, as it's building out that model, it is sharing information back to an operator. But as it starts to enter that environment, it's able to tell that it's no longer able to talk to the operator.

18:59It's continuing on its mission. The way to think about the way the operator interacts with the system is it's more or less the operator is giving recommendations to the platform. And so if the operator is not heard by the platform, the platform will continue to execute its mission without receiving those recommendations. Now, of course, if the operator wants to just take control, they can if they're communicating with the platform. But once it gets into that environment where there's no communication, there's no external information, it's really just executing against its mission at that time.

19:32which may be to explore the environment and just build up a map of whatever it can during that time, to go through the environment and find specific or certain things, and so forth. And so it's just executing against that. There is the ability to kind of tell the platform that if you stop communicating and you see something important or some amount of time has passed, go back to where you could communicate before and try to find a means to share the information. But that's sort of how it handles it. And then from a technical perspective, we have a whole set of technologies that allow for really high performance compression of information, kind of storage and compression and analysis of saliency of information so that when communication does become available, the most important information is shared to that operator while that opportunity for communication exists.

20:26and um you were saying that uh this has been deployed in a combat uh situation can you talk about one of those situations um with with the quadcopter it was um used in in a few high profile mission sets over the last few years i can't really say more than that probably the um most recent example that I can comment on was a recent Wall Street Journal article just came out. And so this is very much public knowledge of where the VBAT was being used in a demonstration, if you will, in Ukraine. And it happened to be the case that while it was operating, it saw some items or things of relevance. And then that information was used as well as the location of those items by the Ukrainians to be able to proceed forward with a certain mission set that was sort of unplanned.

21:22In the context of the VVAT operating in Ukraine, certainly those are degraded and denied conditions because it's being jammed, spoofed, and similar. And so there you have something very similar to what I described with the quadcopter, where you have onboard the intelligence required to be able to understand the world without the presumption of external information and to be able to then make decisions based on what's seen. In that context, information was shared back to humans, and so communications were regained, and then humans made decisions based on what was seen, where it was located, on how to proceed from there.

21:59Yeah. The VBAT flies at what altitude? um it can vary uh anything from a lower altitude a few hundred feet all the way up to um a much higher you know 10 15 000 feet that that kind of altitude yeah or even higher depending on on the needs of the of the mission set and uh is it uh primarily an intelligence gathering vehicle or or can it carry armaments? It's primarily an intelligence gathering vehicle at this time. It is the case that in a kind of prototypical sense, we've integrated some munitions on the platform, but that was in an R &D style approach. So what is produced and operationalized and currently out there being used at this time is more ISR oriented.

22:54Yeah. Because So I was at a conference in D.C. Actually, you might have been there, the Special Competitive Studies Project, kind of the follow-on organization that Eric Schmidt's funding from the National Security Commission on AI. And Mark Milley made a comment that struck me, and I don't remember the numbers, but he said something like, you know, the aircraft carriers that are being built today will be obsolete as soon as they're put in the water, and that in some short number of years, the U.S. military will be, again, some large percentage entirely robotic. and then Schmidt was saying you know that the war in Ukraine we're watching this technology and the doctrine develop in real time because at the beginning of the war when you had these long convoys of armor headed toward Kiev and there weren't sufficient drones available that could happen.

24:22But today that wouldn't happen because there are enough drones that would take that out. And so how do you see this kind of technology changing the nature of warfare? And how long do you think before tanks and destroyers really are obsolete?

24:50It's an interesting question. It certainly is transforming the way that we approach warfare. And for the reasons that you called out, it is certainly the case that these types of intelligence systems can navigate and enter into areas where previously they could not. And not only that, but we're seeing these technologies, at least subsets of them, particularly relevant for Ukraine, move from being something that is less common to commoditized. And you're really starting to see that happen as well. And so with that commoditization of the capability, it's going to be proliferated and everywhere. With that said, a lot of the technologies that are coming together in Ukraine are at various levels of sophistication.

25:37Some work, some do not. And so you're always going to start to see this sort of cat and mouse game emerge, which is exactly what one would expect, where now you have to think about what protections can you place on the tanks in order to protect against these types of systems. And then you have to build the systems that can navigate and mitigate those types of protections. And so I think what you're going to start to see is that merge or not start, we're already seeing it. And you're going to see the same thing with your aircraft carriers as well. And so I think that is going to emerge pretty rapidly.

26:11I definitely agree with the statement that autonomy is becoming a commodity. The level of sophistication of the autonomy is going to be varied. A lot of what we're really seeing scaled out there that is being commoditized is fairly simplistic autonomy, but you will see that ramp incrementally as more and more capabilities are drawn in. And so then, as I noted, you're going to start to see more and more techniques used to mitigate and remove those autonomous systems. And so that's what I really expect. I think the key challenge will not be the role of autonomy. The key challenge will be how quickly can we update, modify, adapt, evolve, and improve upon our autonomy within this cat and mouse game.

27:01where we used to be able to move at 18 months and a typical iteration cycle would be one of we have this problem, we need to solve it. Here's the statement of the problem, engage with industry or academia to come up with a solution, go through a process, potentially implement that problem and end to end from ideation to real world realization, 18 months. Now it's going to be far, far faster, you know, at a pace of on a daily basis. Here's what happened yesterday. We iterate. Now we move forward today. We iterate. We move forward today. And once we hit that daily basis, it's going to become even more compressed such that you are updating it on an almost near real-time basis as you're adapting and mitigating these techniques.

27:51So you're going to see more and more, for one, software-defined vehicles, more and more emphasis on rapid iteration on the autonomy and the AI capabilities, more and more emphasis on guardrails and trustworthiness of what's being produced, because now you're pushing that to the forward deployed. You're pushing that to the non-experts that are in the field. And so those types of capabilities become a real relevant consideration in order to ensure that your iteration in the field based on current information that is being automated and sort of auto-generated if you will is going to perform as you expect it to on that next day so that's where we're going to start to see it really start to transform i would i would claim yeah uh and wherever the hive mind uh technology is deployed whether it's in vbat or quadcopters or you know fighter jets part of the the the technology allows these this teaming right across or what i like to refer to as swarms how large can those teams be and how do they communicate with each other if they're in an environment where there's jamming coming from the other side yeah so first off on size it can be somewhat arbitrary um it depends on the nature and the class of the platforms when you have more costly platforms you're going to have smaller numbers for obvious reasons and then when you have lower cost highly attributable platforms it'll be larger numbers um the um it it was historically the case but it's definitely changing with compute the level of sophistication and intelligence, it used to be the case that the more expensive the platform, the more you would end up seeing highly sophisticated, exquisite performance.

29:55But now we've gotten to the point where you can kind of see that Rolls-Royce performance, not just on the exquisite bespoke or like this highly capable performance on these systems, but you can see this also start to emerge when even the level of the charitable platforms that are much lower cost. And you can see that both in the defense context, but also in the commercial context. So that's pretty cool. In terms of numbers, therefore, it's kind of arbitrary. The types of behavior and performance changes as you change the numbers. typically what you would have is more tightly coordinated teams that are able to behave in a in a very kind of agile well synchronized coordinated manner but then as you start to go to larger teams then the techniques start to build on top of that so that's that's very similar to humans when we work together and we aggregate we work well in small groups but as we become larger we coordinate with each other but at hierarchical levels in terms of coordination across teams.

30:58So that's what you could think about. And then that kind of scales arbitrarily. Now, the other part of your question is, how do you think about communications? Well, we have communications between the systems, but you said, what happens if they're jammed? Once they're jammed, then you're constantly assessing the communication network. And there's an underlying structure to that network. And if it changes, then the system itself and the team can kind of change their behaviors. Ultimately, we have to design for teams of size equal one or size equal n, and we have to be able to, at any moment in time, change our performance based on what the communication network is, if we're losing communications, if the teams break apart, if platforms are attrited or something to that effect.

31:41Ultimately, once you do get to a point of an extremely high jammed condition, then you're going to go to a team of n equals one. no agents can truly talk to each other and then within that context the behaviors of the team itself changes the mission profile changes on what they seek to do and they'll start to engage in ways that then try to mitigate that loss of communication by repositioning themselves to share information to create communication relays these kinds of considerations yeah so but but but the base communication is is uh just radio frequencies it's not uh uh i don't know that much about uh but you're not using bluetooth between quadcopters that are in proximity uh to each other close proximity well there's um there are analogs to bluetooth but when they're in close proximity talking to each other so mesh networks radio communications of different forms uh satcom all of these are options and we'll switch between the different types of communication mechanisms based on what makes the most sense or based on how the conditions are changing um as as the mission progresses yeah uh can you could you for example position uh a net of or a network of a quadcopters over a battlefield in a stationary configuration or formation and just watch movement on the battlefield, you know, that we used to do with blimps or balloons in World War I and II?

33:28Or is that the sort of thing that you can do with teaming? So certainly we could do what you described, but probably we wouldn't use quadcopters because they're not typically long endurance platforms. And so you'd be having to change them out a lot. It probably wouldn't be the right choice of platform for that kind of need. So typically within the context of teaming, there are coordination roles that the teammates play. Some may be carrying some sensor payloads. Others may be carrying other payloads. They're working with each other in order to best leverage those capabilities. Some may have, like, all of their endurance levels will vary based on how much capacity they have on board from a fuel perspective.

34:20And they'll coordinate with each other to cover areas or regions. depending on the conditions from a degradation and denial perspective, they'll reposition and change their roles in order to share information or communicate effectively with external people. So those are some of the examples. But certainly one could use these platforms for situational awareness and monitoring particular regions. And that's certainly the way that they're sometimes used. again i don't think that you would use a quadcopter to do that because they just don't have the endurance often as contrasted with something that could give you just as much information but for many more hours yeah and another thing that's fascinated me is years ago i saw a chinese paper about a swarm flying through a forest, a swarm of quadcopters that were able to navigate through the forest, avoid the trees.

35:23Is that the sort of thing that HiveMine can do? Sure. I mean, we deployed our quadcopters in much harder conditions, I would claim, than that. So think of doing something like that, but in dust clouds under extremely confined environments with very small openings and holes. So, yeah, what you're describing is basically the ability for a system to, so in a single context, single intelligent agent context, build up an accurate 3D model of the environment with the onboard sensors, navigate through that environment, and then across multiple agents, build up a common shared picture of that environment and where they are relative to each other, and then control relative to that.