The Technology for Fully Autonomous Attack Is Already Here | Alex Liannyi, NORDA Dynamics

The AI systems guiding Ukrainian combat drones aren't running on expensive Nvidia chips. They're running on a Raspberry Pi Zero - a $15 hobby computer - and that single detail tells you more about how Ukraine's drone war is actually being fought than any headline about AI weapons. Craig Smith sits down with Alex Liannoy, CTO of NORDA Dynamics, for a technically specific and operationally candid conversation about what autonomous drone guidance looks like from inside the team building it. NORDA builds the avionics and targeting modules that plug into drones made by approximately 40 of Ukraine's hundreds of manufacturers, solving the two problems that make drone warfare hard at range: GPS denial from jamming, and the radio horizon that causes pilots to lose contact with their drone in the final 300 to 400 meters before ground level. NORDA's terminal guidance module takes over at up to one kilometer from the target, completing the approach autonomously while a human pilot's earlier lock defines the destination. The most counterintuitive finding in the episode is about what keeps humans in the loop, and it isn't ethics. Alex's argument is pragmatic: military doctrine requires proof that a drone hit its intended target. If you're monitoring the drone to verify the hit, you have enough connection to control it, which makes full autonomy less necessary than it might appear. That said, he's clear that the technology is "very close", one pilot already controls up to 32 drones simultaneously in some deployed systems, and target detection AI for notifying pilots of candidate targets is already in the field. The episode closes with a hardware finding that should recalibrate anyone's assumptions about AI compute requirements: after engineers mocked NORDA for using Raspberry Pi 4, they moved to the smaller, less powerful Raspberry Pi Zero. Both run real-time computer vision in active combat. The lesson Alex draws is precise: good engineering overcomes chip limitations, bad engineering can overload even a high-end processor. Subscribe to Eye on A. I. for weekly conversations with the people building and deploying the future of AI.

Speaker names marked ? are inferred from speaking time and the show’s known hosts — they are not voice-verified.

• 0:00Craig S. Smith?

  What I'm really interested in is how artificial intelligence is being introduced to the drone operations. Do your modules have any artificial intelligence in them?

• 0:12Speaker 01

  Artificial intelligence is very broad term, it's very speculative. In our case, our last mile targeting is based on classical algorithms. It is mostly math and control theory and so on.

• 0:21Craig S. Smith?

  How many drones can you launch at a time right now? I mean, how many are you?

• 0:27Speaker 01

  One of the systems we use, it allows connection from one pilot to thirty two drones simultaneously. And as I'm being deployed in all military organizations, you have to prove that you've had hit the target. You need to monitor what had happened with drone, whether it engaged in a proper target or you engage in a false positive scenario. And if you can monitor them, you can control them. Technology is close, but I don't think it is completely necessary to use it.

• 0:52Craig S. Smith?

  How close do you think the technology is to moving the human earlier in the process to the launch of the drone? Once the drones are off, it searches, acquires and attacks.

• 1:03Speaker 01

  Technology is very close to that.

• 1:07Speaker 01

  My name is Alex and I'm CTO of Norda Dynamics.

• 1:12Speaker 01

  Before the war, I was a regular founder of a tiny IT business. I was doing like software development stuff with a tiny team of fifteen developers. When the war started, we started to search for ways to help. Tried all kinds of different things, from volunteering to developing drones. Not successful.

• 1:42Speaker 01

  And at some point of time, we

• 1:47Speaker 01

  I focused mostly on specific parts of drones, mostly avionics. Started developing different algorithms and at some point of time met other guys, Nazar, Dima and Roman.

• 2:04Speaker 01

  Actually, and we founded Norda.

• 2:06Craig S. Smith?

  And when you said that your initial efforts at making drones were not successful, what was the problem?

• 2:14Speaker 01

  Actually, not really problem because there were other guys who already done huge amount of work developing fixed wing drones and developing quadcopter drones. And it was just more

• 2:29Speaker 01

  it was more efficient to help them in things they are not done yet.

• 2:36Craig S. Smith?

  Right.

• 2:38Speaker 01

  That's it. So there were no reason to compete with them. There was a reason, there was it was much more efficient just to collaborate with them.

• 2:48Craig S. Smith?

  Yeah. And when did you start working on drones? When did all of this drone work start?

• 2:53Speaker 01

  Actually, first efforts, I think, started like mid 2022.

• 2:59Craig S. Smith?

  Yeah.

• 3:01Speaker 01

  Like some research and so on. But the work, work on avionics and bring order stuff started 2023.

• 3:12Craig S. Smith?

  And the avionics, you're talking about navigation as well as last mile target?

• 3:18Speaker 01

  Yeah, exactly, exactly. It's started with last mile targeting, researching on flight controllers and so on.

• 3:24Craig S. Smith?

  What is the problem, first of all, with navigation and second of all, with last mile?

• 3:29Speaker 01

  First of all, the first problem, the absence of GPS navigation. Like it's, it's the first issue which appears on all drones in Ukraine war.

• 3:41Craig S. Smith?

  Yeah.

• 3:42Speaker 01

  And the second problem is direct radio control. Distance, earth curvature, trees, any obstacles, they

• 3:56Speaker 01

  introduce issues with radio signal. This is why drones don't usually don't have

• 4:04Speaker 01

  pilot don't have connection to the drone on last three hundred meters, four hundred meters, depending on distance. Like when you fly further, you'll lose connection earlier before the ground.

• 4:18Craig S. Smith?

  And that connection is because of jamming, jamming of the signal?

• 4:22Speaker 01

  There is such a thing as a radio horizon. Because of earth curvature, the further you go, the higher need to be to get line of sight connection. When you lose altitude, you lose your connection with the drone.

• 4:37Craig S. Smith?

  Right.

• 4:38Speaker 01

  And this, the, the distance before the ground where you lose your connection, it increases with the distance you fly.

• 4:47Craig S. Smith?

  In the beginning, we're, we're, the drone range is much shorter and so pilots were

• 4:55Craig S. Smith?

  piloting all the way to the target?

• 4:57Speaker 01

  Yeah, it, it works on a, on a shorter distance, but it still won't work if you have a different ground altitude.

• 5:07Craig S. Smith?

  I would say, yeah.

• 5:08Speaker 01

  Basically, especially on Donbass, you can have like between pilot and drone, you can have this artificial mountain of like a berm. Yeah, yeah, I just don't know if it's a proper word.

• 5:25Craig S. Smith?

  So first of all, the navigation, a lot of people are now using Starlink for large drones, but for these smaller drones, either small fixed wing drones or quadcopter drones, there isn't enough, you can't fit a Starlink panel on it.

• 5:45Speaker 01

  Sure, yeah.

• 5:46Craig S. Smith?

  I mean, what I'm really interested in is how artificial intelligence is being introduced to the drone operations. So do your modules have any artificial intelligence in them?

• 6:03Speaker 01

  Artificial intelligence is very like broad term, you know? It's like, yeah, sure. It's very speculative, you know? In our case, our last mile targeting, it is based on classical algorithms. It is mostly math, control theory and so on. It doesn't have any like fancy AI. It has some neural networks which can be I don't know, someone can tell, name them AI. Their usage is quite isolated. Drone does not take any, it does not do anything without pilot approval.

• 6:45Speaker 01

  Basically, pilot selects target, drone does its work afterwards. You have some developments on target detection. When drone can search for the target itself, but still it won't engage without pilot approval.

• 7:03Craig S. Smith?

  Yeah, this is something that I'm really interested in because, I mean, currently, do you have object detection on the drones?

• 7:11Speaker 01

  On this, you've seen, no, but overall, yeah.

• 7:16Craig S. Smith?

  There are drones in the field with object detection that are trained to recognize a tank, for example, as opposed to a passenger car.

• 7:27Speaker 01

  Yeah.

• 7:27Craig S. Smith?

  And the drone then searches for that? It helps the operator by scanning for that and recognizing that?

• 7:37Speaker 01

  It highlights like estimated target.

• 7:40Craig S. Smith?

  And is that done in the cloud or on device?

• 7:43Speaker 01

  No, it's on a device.

• 7:44Craig S. Smith?

  How large is the model? Where do you get the models for that?

• 7:48Speaker 01

  Actually, these modules are not big and these are mostly, as I said before, it's a hobby level modules. Basically, we don't usually use high end edge devices because it's hard to make proper supply of them. It is always in short supply, you know?

• 8:10Craig S. Smith?

  Yeah, so what kind of chips are you, are you using right now?

• 8:14Speaker 01

  Actually, we are big fans of Raspberry Pi stuff.

• 8:17Craig S. Smith?

  Yeah.

• 8:18Speaker 01

  It's a hobby level, single board computers. We use them a lot.

• 8:23Craig S. Smith?

  Yeah, although I've been reading the, the different units are developing Jetson level chip modules that have higher accuracy in their targeting.

• 8:38Speaker 01

  Most of they have higher speed.

• 8:40Craig S. Smith?

  I see.

• 8:40Speaker 01

  Not accuracy, like there is no direct, direct connection between accuracy and chip type.

• 8:48Craig S. Smith?

  Right.

• 8:49Speaker 01

  It's just a number of operations you can perform on it. That's it.

• 8:54Craig S. Smith?

  How widespread it is? And when you say speed, it's a matter of speed, you're talking about the speed of the compute?

• 9:00Speaker 01

  Yeah, physical speed, power.

• 9:03Craig S. Smith?

  In what conditions is that speed of compute important? Is it how fast the drone is moving?

• 9:10Speaker 01

  I would say it's the complexity of environment. Yeah. And but from my, my experience, good engineering usually overcomes cheap limitations. Because if it's bad engineering, you can overload even high end Jetson speed, Jetson processor.

• 9:30Craig S. Smith?

  Right.

• 9:31Speaker 01

  So it still will perform very poorly.

• 9:34Craig S. Smith?

  Yeah.

• 9:34Speaker 01

  Actually, when we started developing our system, like a few years, two years ago, or something like that, people were making fun of us using Raspberry Pi four because they were saying that it is not possible to run proper computer vision algorithms on that. And after that, we started using even lower. We started using Raspberry Pi zero two. It's like a smaller, smaller thing, which is like much less like performance than Raspberry Pi four. So it still runs in real time.

• 10:06Craig S. Smith?

  On the navigation, so there are two parts, right? Navigation and target acquisition.

• 10:11Speaker 01

  Actually, it's mostly target acquisition and control. Navigation part in classical terms, when we know exact our location, is not introduced yet.

• 10:21Craig S. Smith?

  These are not loitering drones. That you're using right now. Because my understanding what's coming are drones that you can geofence an area.

• 10:31Speaker 01

  Yeah, yeah.

• 10:32Craig S. Smith?

  And you can say to the drone, you can give the drone instructions.

• 10:38Speaker 01

  Yeah.

• 10:39Craig S. Smith?

  Or the module instructions. Look for military vehicles in this.

• 10:45Speaker 01

  Yeah.

• 10:46Craig S. Smith?

  Uh, geofenced area. If you see a military target.

• 10:52Speaker 01

  Yeah.

• 10:53Craig S. Smith?

  Lock on it. And then what? The operator has to give permission for it to attack?

• 10:59Speaker 01

  Yeah, yeah, actually. We are working on such kind of. Uh, scenario, uh, actually very close to what you described. And yeah, uh, it is important that the operator approves the target. Like ba- basically, we. Hard code this uh, condition that operator needs to select with already.

• 11:22Craig S. Smith?

  That last uh, switch that the operator has to toggle to allow the drone to attack. I I mean, one of the things I'm interested in at some point. Either Russia or Ukraine or some other state. Will decide that the human in the loop happens when you launch the drone. And once the drone is launched. If it's restricted to a geofenced area and that area. You're confident is an enemy line or then you you could allow the drone to go ahead and and hit the target. Autonomously.

• 12:05Craig S. Smith?

  And I know there's uh, a lot of discussion about international human rights law. Uh, and right now. Everyone's very careful not to take that step, but at some point. Someone's gonna take that step. How close do you think? The technology is to to moving the human. Earlier in the process to the launch of the drone. Once the drones up. It searches, acquires and attacks.

• 12:36Speaker 01

  I think people are very close and technology is very close to that. I'm not sure i- if it's gonna be an efficient way to use the drone. But overall, the technology is very close.

• 12:52Craig S. Smith?

  And not efficient because uh, the models are not yet accurate enough.

• 12:57Speaker 01

  Yeah, it's mostly about accuracy, about accuracy of positioning of of how the. Navigation part will perform in in in different conditions. Another important thing, you need to have proof of uh. Uh, in in all military organizations, you need to prove that you had hit the target. Yes, otherwise you just used your resources, you used and drones for what?

• 13:22Craig S. Smith?

  Yeah.

• 13:23Speaker 01

  And in this case, you need to monitor what had happened with drone. Whether it tar- uh, whether it engaged in a proper target or it or it engaged in a false positive uh. Scenario, right? So in this case, you need to. Monitor them and if you can monitor them, you know, you can control them. And if you can control it, why you need to uh, complete and one in it? Yeah, so. Uh, in my opinion, technology is close, but. I don't think it is. Completely necessary to do to use it.

• 13:54Craig S. Smith?

  Right, although part of the problem. Uh, I mean, certainly there's a supply problem, but. Uh, one operator uh, can really only keep his eye on one or two drones. But if you had the autonomy, you could launch. A thousand drones in a in a geofenced area and the operator could could watch.

• 14:19Speaker 01

  Yeah, yeah, exactly. That is, that is why the drone can notify you when it sees a target. And then you can pull up the picture and approve it. So i- and you still can control.

• 14:32Craig S. Smith?

  I see.

• 14:32Speaker 01

  Lot lots of drones simultaneously, but uh, still be in a loop.

• 14:38Craig S. Smith?

  Yeah, and and so how many uh, drones can you launch at a time right now?

• 14:44Speaker 01

  It's not part of our system to to do like a swarming part.

• 14:49Craig S. Smith?

  Right.

• 14:50Speaker 01

  Uh, but from uh, it mostly depends on a on a ra- radio connection system used.

• 14:56Craig S. Smith?

  Right.

• 14:57Speaker 01

  One of one of systems viewers, uh, it allows a connection from one pilot to thirty two drones simultaneously.

• 15:05Craig S. Smith?

  To how many?

• 15:06Speaker 01

  Thirty two.

• 15:07Craig S. Smith?

  Thirty two, wow. Uh, and is that being deployed?

• 15:12Speaker 01

  Uh, as far as I know, some some squads using.

• 15:15Craig S. Smith?

  The targeting module, uh, so you have uh, Raspberry Pi on device.

• 15:21Speaker 01

  Yeah.

• 15:22Craig S. Smith?

  You have uh, a small model on the Raspberry Pi. That's uh, has object recognition. Or s- wh- what uh, is the targeting?

• 15:36Speaker 01

  Actually, we we have different options. Uh, most the the one which is mostly used. Doesn't have any models at all. It runs uh, regular math algorithms on it, so no, no, no, nothing fancy as a model.

• 15:50Craig S. Smith?

  Right.

• 15:51Speaker 01

  Uh, we have, we have other and it it doesn't have any object recognition. It just uh, locks on a specific target. Uh, we have higher version, which has object recognition and it runs a model.

• 16:05Craig S. Smith?

  And the drones, uh, I'm seeing. Uh, that you've been showing me, uh, do those have object recognition or or you're just uh, the pilots doing the locking?

• 16:16Speaker 01

  No, uh, the one we've shown, they're just have a lock on a target. So pilot selects an object, a vehicle or a place in a field or.

• 16:24Craig S. Smith?

  Right.

• 16:25Speaker 01

  Wh- whatever pilot chooses and we just uh. Uh, targeting the drone there.

• 16:30Craig S. Smith?

  Right, and then once the target is acquired, whether it's through object recognition. Or by the pilot locking on a target. The module takes over. Control of the drone and and.

• 16:45Speaker 01

  Yes, example.

• 16:47Craig S. Smith?

  And so these these drones uh, currently have a range like the fixed wing you were telling me is thirty to forty kilometers, is that right?

• 16:55Speaker 01

  Uh, no, thirty to forty, it's is a with a with a higher payload.

• 16:59Craig S. Smith?

  That's right, yeah.

• 17:00Speaker 01

  I would say it's up to eighty kilometers of.

• 17:03Craig S. Smith?

  Yeah.

• 17:03Speaker 01

  This electrical fixed wings.

• 17:06Craig S. Smith?

  Yeah.

• 17:06Speaker 01

  Uh, up to eighty.

• 17:07Craig S. Smith?

  And how long is the uh, target? I mean, how far out can you lose? Radio uh, contact and have the drone take over?

• 17:18Speaker 01

  Uh, usually our distance of uh, terminal guidance is like up to one kilometer.

• 17:25Craig S. Smith?

  Yeah.

• 17:26Speaker 01

  It depends on the situation, uh, because like when you have fog or lower. Uh, lower clouds or. It just uh, the sun is lower or something like that, i- it it may vary.

• 17:40Craig S. Smith?

  Yeah.

• 17:40Speaker 01

  But overall, it's up to kilometers.

• 17:42Craig S. Smith?

  And is that getting longer or is that a pretty much a fixed parameter?

• 17:47Speaker 01

  No, i- i- we can do mu- we can do better.

• 17:51Craig S. Smith?

  Yeah.

• 17:52Speaker 01

  Yeah, it it's by but by using a different different optics, by using uh. Like gy- uh, gyro stabilized cameras, we can increase distance, I think up to two and a half kilometers, things like that.

• 18:06Craig S. Smith?

  And the the drones themselves, uh, is NORDA supplying the drones with the module built in or you?

• 18:14Speaker 01

  No, we we supply modules or we supply a software. Uh, either this or that, we don't do drones. Basically, we collaborate with other drone manufacturers in Ukraine. It's much more efficient because we can focus on uh, research and develop part of the stuff and. People on all manufacturing do their job.

• 18:36Craig S. Smith?

  And how many manufacturers are there today?

• 18:39Speaker 01

  Like we are collaborating with?

• 18:41Craig S. Smith?

  Or first of all, in in Ukraine?

• 18:44Speaker 01

  Oh, I know, hundreds.

• 18:46Craig S. Smith?

  Yeah.

• 18:47Speaker 01

  Hundreds of manufacturing manufacturers.

• 18:49Craig S. Smith?

  Yeah.

• 18:50Speaker 01

  And we are uh, in contact and collaborating in one way or another with like about forty.

• 18:57Craig S. Smith?

  A forty or a fourteen?

• 18:59Speaker 01

  Forty.

• 19:00Craig S. Smith?

  Forty, wow.

• 19:01Speaker 01

  Yeah.