Inside Ukraine's Azov Drone R&D: The Engineer Building AI Weapons 18 km From the Front Line | Alexander Palamarchuk

0:00The U.S. has said that we'll not develop fully autonomous lethal weapons for ethical reasons, but it seems on the battlefield, every side is trying to get an advantage. In two years, you'll be able to field fully autonomous drones. I mean, that's a frightening world. I would guess that that technology is going to be on the battlefield very soon if it's not there already. Dreads on with AI is on a particular ground. So now, there is some part of AI that good drugglasses in targets, but it footage recognizes, example, right target, wrong. It's very good drugglasses, but it is not good enough to know is it enemy or not.

0:45And that's the problem. The NATO countries are trying to get the right weapon. I mean, right model, they are trying to study this model, right? chief paradise is the call sign of a man named alexander a ukrainian drone specialist serving with the azov corps of the ukrainian national guard azov began in 2014 as a volunteer battalion it drew international notoriety for its links to far-right movements and was later incorporated into Ukraine's National Guard, where it has grown in one of the country's most prominent combat formations. Paradise joined in 2023 and now works in an R &D unit developing drone technology near the front.

1:36I spoke with him remotely about jamming, autonomous navigation, AI target recognition, and how drones are rapidly reshaping the battlefield and the future of warfare. First of all, I wanted to ask you about the history of Azov when you joined, when Azov started using technology. I joined Azov in 2023 and worked only with drones. It is different kinds of drones. If we are talking about Azov, as you know, the fast usage of drones in Azov was like in 2017.

2:25They were used to reconnaissance firstly and then to strike some bombs. To strike drones as anti-drone warfare? Is that what you mean? They were used to hit different kinds of targets. As a fast forward, it was used as reconnaissance drones, so you can just be in some kind of touch and use them when you can even see your enemy with your own eyes. But it gives you a good angle to view what is going on on the enemy side. Right. But the war didn't start, the invasion didn't happen until 2022. So in 2017, where were you using this? Actually, it started in 2014 for us. Actually, even on YouTube, you can find a video with Russian regular army in Donetsk region.

3:31So it was used with the same enemy from 2014. I see, yeah, in Donutsk. That was started with own-made drones. Then when some kind of DJI fandom survived, guys tried to use them. And as I know, they were glad that it gives good possibility for reconnaissance. Okay, and when you joined, they were still using DJI or home-built or when you joined the effort? Even for now, we are using some Mavic and Matrix drones for reconnaissance for some kind of bombers with light bombs. mainly we use them as bomber and now on this here we use them mostly for reconnaissance because for now we have more possibilities and to take a bigger bombs or fly Wanda stay in the air for longer actually they they have good cameras now and we are on the way to change the Chinese drones for our made drones.

5:00I mean, Ukraine made drones, our partners made drones. We are on the way for this. One of the big problems with drones on both sides is jamming. And I've seen all of these different jamming systems, but also systems to evade Russian jamming. I mean, how do you handle that? I know there's the fiber optic line that unspools up to 40 kilometers or something. but I've also seen drones with Starlinks mounted on them. So there's satellite communications, which I understand is unjammable. Can you talk about the jamming issue on both sides? Yeah, I can. I will divide it to three parts, like radio, like fiber optic, and like internet connection.

6:05Let's just start with radio, this is a classic drone switch I started to use in 2023. First we use its common frequencies, the same as LTE, for example, or some 2.4 GHz. It was regular, then we meet some jamming from enemy side. So we are trying to use the possibilities which we have had with ELRS. And we can move from 700 MHz up to 1000 MHz. And in this frequency range, we are trying to find the possibilities for us to fly over the enemy site. And actually until I guess the mid of 2024, it was more than enough. Then we moved to our LRS, maybe, you know, kind of LRS, which are made in Ukraine.

7:13This is a part of, this is based on common LRS, but it's made for our needs. to have a stronger communication with the drone and ground station, which was helpful to fight the jamming. And for 2025 year, we moved to not regular LRN receivers, transmitters. For now, we can use almost from 100 MHz up to 3000 MHz. We are trying to find some frequency ranges where can we work freely. And it mostly depends on where you're trying to fly. Just for example, we couldn't fly to Crimea region on common frequencies, but a guy who stands five kilometers next to us could freely fly on these frequencies on their point of interest and good for them.

8:29So we move to, we are trying to resolve our problems and try to find the solution for this by ourselves in the unit. For now we are using not ELRES for radio communication, but some kind of Ukrainian stuff or our partner stuff. Okay, and then what about internet communication via satellite? Actually, maybe you know some names from drones that use it even from 2022 or 2023, like MESIS or something like that. This is a common internet connection, actually via Starlink or some satellites like this. Um, that's all actually what is, there is nothing, uh, there is no actually secret, uh, from this, you can use a malware link, uh, which can be routed to via internet or via some local, uh, network or something like this.

9:46And you have your, uh, control and you carry videos from Android. So it gives you possibilities to fly where the something else wouldn't fly. We will have actually no problems to fly over some kind of buildings or something like that even far from ground stage. Is there a possibility to jam the Starlink connection? I guess that is possible. So we know actually what happened in Iran. So it is possible. I haven't made it on Battleground for now, but I think that is possible somehow. This is not the thing where you could find easily, but you need some kind of very expensive stuff which could easily be hit by something else.

10:52Are the Russians also using something like Starlink? Yeah, yeah, yeah. You can find even videos where they're using Starlink. Right. And so at this point, neither side has an advantage, is that right, with technology? This is the kind of tradition that we could find the solution in our own unit by one person. And when you are trying to solve some problem like UI jammed or something else, you can find actually in three parts of the internet very, very, very many and very much information about what you can do with it. And this is good when we have some kind of engineers who can... sources. So we are good at making some very interesting projects, very interesting solutions to some kind of jamming, if you're talking about jamming.

12:06But I guess the Russians are good to do it more as a part of the system. They are good at producing many, many, many drones using this kind of communication, for example. Are you guys using any artificial intelligence in your... We are using it as early programmers to make our lives with code. It's much, much easier. For example, actually using when we are working with some data, for example, it is very helpful too. Yeah. It's a common, I guess, the way to use AI, mainly using it in actually the same way. Yeah. And that's, are you using cloud code or what model are you using for... Actually, when they have something we should code, we are trying to use it.

13:27in this model or this model or this model, and which would get us better results. We use that. So I couldn't tell that this one model is great for us, but they use different models, which works well. Okay. We use it. Yeah. And are you using AI in drone, to help the drone navigate or to automate? Yes, of course. Yeah, again, how you're using... There are two kinds of ultra-con recognition for drones, like where you track some pixels, or when you recognize some kind of UAVs or something like that. Or not UAVs, but some kind of cars. What's this? When you're tracking the pixels, You move some little, for example, forest and you just lose your target and the drone couldn't recognize this far.

14:34This is armored vehicle or this is civilian cars. It's not great, but because you always, the operator should track what the drone is doing. You could, you should, you must actually, you couldn't use drone as like move to this area and hit everything you see. Cause you need to recognize this civilian car or this already destroyed car or some, all of this. And we, if we are talking about AI models for auto recognition, they are much clever. They are much more clever, they can recognize by the drone. Is this a correct target for hit or this isn't. But the operator should be present and he has a phone control just in the case that this should be enemy.

15:40So you're saying that they're models for recognizing objects. And what about navigation? I mean, how autonomous are the drones today? You know, there's FPV, of course, where the operator is watching through a screen, either goggles or on a screen, and guiding the drone that way. But if you have object recognition in the drone, can you make it autonomous so that once it locks on an object, you no longer have to guide it? There's a case where the LIS help, because it could crack the drone correctly. Exactly. If you are talking about navigation, I guess you know that GNSS is almost everywhere is jammed and near the bottom ground.

16:49So we have two ways to resolve this problem that the drone couldn't navigate by itself because of GNSS jamming. This one is visual positioning. and the drone should have a camera on the bottom of it. Let's see, it's a check what is above this and compares to maps which preloaded to the drone and the result also, I'm here. So it can move to a position with some kind of accuracy. There's some problems with accuracy up to 50 meters, for example. Different position gives different results.

17:55And other way to solve these problems is to use beacons, like if one communicates with ground state, with one, two, three beacons, and it gives you a point where it is. It depends on the time which it communicates with beacon, with ground stations, it compares the time it needs to communicate with ground station. Beacons, it gives different time. So, you know, you are closer to ground station and far from beacon. You know, this time needs to communicate. For example, you need to 10 milliseconds for one kilometer, for example. If you have 60 milliseconds, you are far as about six kilometers. And beacons help it to engage where you are.

19:08I see. Yeah. Yeah. And have you guys experimented with fully autonomous drones? Yep. Yep. Can you tell me how that works? Actually, you just place it on the ground, then you make some kind of fusion for it. It's takeoff by itself, so that calibrate, for example, compost or maps or finding the radio link with beacons, some kind of this, and then move to the place or area where it should be. That is the case where AI is needed. Because the problem is that you should mark this correct target. Just for now, it is impossible to hit only enemy targets. So it could be civilian, so you couldn't hit this kind of target.

20:19this civilian car, even if enemy uses civilian cars as its own. So you couldn't hit this kind of target. So you need to be sure that this enemy won't harm the system. This operator needed.

20:39If it is not coming out, it's wrong. It is possible to make a solution. of a solution, such as move to this area, watch this, watch this, watch this, and then it returns back. Are the Russians using fully autonomous from what you can see? I guess we know about the heads. It's using GNSS, some kind of CRPA antennas, and it even needs many jammers to jam CRPA antenna. So up to, I guess they have some problems with this for now. Cause we saw the CUP antennas with 16 elements. So you need 16 jammers, 3 jamses. They are moved to Starlink. For now they are using it. So if you saw the one with the last, they hit the train, civilian train, which or active host.

21:50I guess, actually. Even now, you could find some drones fully automated. They just take over this and move to this area, hit this. And if it would be jammed, it hits anywhere. It would. I don't know. if the drone thinks he is there, who will hit this? Or even if in general. Yeah. I mean, because the Russians don't seem worried about civilian casualties on the Ukrainian side. So I'm wondering if they're going to move to fully autonomous that will target any vehicle or any person in its range. Naturally, it's much easier for them. They don't care about any civilians or something like that. If you could hit any target that can move faster, for example, than 20 kilometers, And the car with your car.

23:04This is a target. It's much easier than if you can just move to this area, hit everything that is moving. So it's much easier than hitting, but too much, too many civilians was hit like this. Actually, they even have telegram channels where they show that, oh, this was a woman with, for example, a bicycle here. Even with Mavic in manual control, they saw what they did. It's okay for them. So it's much easier for them to make some absolutely autonomous missions because they shouldn't care about the type of term.

24:05I mean, do you analyze the remains to see what kind of, I don't know if there would be anything left, but to see how autonomous the Russian drones have become? I mean, do you periodically capture one before it's detonated and look at how it's being guided? Yes, of course. Of course, we are trying to get something interesting from their engineers. Absolutely, I'm confident they do something like we do. They are trying to reproduce our good ideas. We are trying to reproduce their good ideas. And that's the way that it should work. Because if you don't want to use good idea of your enemy, you will use some kind of good idea.

25:11It wasn't some good ideas. The inherent this is very efficient actually. They can produce a lot and this is this kind of drones isn't so accurate for now but they are using some mesh system some star links to make it to have a control of it or get a video from them to get some kind of telemetry from them. Actually, everyone wanted to have a super autonomous drone. Just press six to win. Just press six, 10 ,000 drones fly, hit every enemy. But for now, I see that both sides of this conflict of our war, trying to get as much control drone as it was. that everyone's trying to use in some network systems like MASH or Starlink or any kind of internet connection to the drone.

26:29Yeah. I mean, what you said about launching, you know, 10 ,000 drones at once, do you think if the war continues, it'll get to that point that Russia could launch 10 ,000 autonomous drones. So the drones kill everything or hit everything that's moving. I mean, that's pretty frightening. I don't think that Russia is closer to this than we are, actually. Yeah, except that, as you said, you're fighting over Ukrainian territory and you don't want to hit Ukrainian civilians. Actually, you know, we have the Ukrainian army, we have an invasion in Russia and the same thing. The civilians, you couldn't hit it.

27:31But we saw this battalion commander here of Russian army. No, we couldn't. We couldn't. Just civilians. And the partners look at what we are doing. They try not to hit civilians as possible. So that's one of the problems to use Celical Autonomous Kamikaze John Foss. But it is possible to make some kind of autonomous

28:09missions when you couldn't hit someone. Is it, do we are close to the live 10 ,000 drones, for example, 10 ,000 drones? Actually, yes, I guess the two, three years, I mean, we have the situation where we could launch many, many, many drones. The one problem is, you said correctly, the navigation of the drone. If you have some kind of systems that could recognize the enemy correctly without hitting any wrong targets, it's possible to use many, many autonomous drones. The one problem with 10 ,000 was you could have so much place to launch there. So it takes much time to prepare for launch. Two, three years, it will be possible.

29:20I am really with launching on five with one ground station drones at one time. But I think it is possible to make, if it is fully autonomous drone, if you don't need to control drones, that it's a bit easier. If you need to control any of drones, for example, this 1 ,100 drones you need to control this in this time, it's much harder than it's fully autonomous. Yeah, so you say in two years it's possible that you would be able to launch 10 ,000 autonomous drones. I mean, you mean the tech will be there? Yes, so actually it could be even in this year, not about 10 ,000, but even if you could launch the 100 drones.

30:28It's very hard to deal with it. I mean, do you worry though that Russia is going that direction? Everyone, everyone is going in this direction. Russia, Ukraine, and anyone who wanted to have an up-to-date weapon. They are about fully autonomous weapons that can recognize the correct target. They are, that couldn't be jammed or something like that. Everyone wants this kind of weapon. Yeah. And the development, because I mean, there's the hardware production, but a lot of that is software, is the AI. Do you have... Yeah, of course. Absolutely. This is AI. Yeah. And do you have AI developers that are working on models?

31:26Because that kind of model exists today. Yep, yep, yep, yep. We have not only one, many of them. The NATO countries trying to get the right weapon. I mean, the right model for this, they are trying to learn, study this model correctly. USA, actually, anyone who wanted this kind of drone are on the way to have this. If they're trying to use AI, they're trying to place it on the drone and test it in our bottom ground. Yeah. I mean, the reason I ask so much about this is the U.S. has said that it will not develop fully autonomous lethal weapons for ethical reasons. But it seems on the battlefield, every side is trying to get an advantage, and that would be an obvious advantage.

32:48And you said everyone's working in that direction, and you said in two years you'll be able to field fully autonomous drones. I mean, that's a frightening world if two armies have huge numbers of fully autonomous drones that would kill every living thing in a certain zone. Is that where warfare is headed, do you think? Even in this century, we will see the conflict with, if it would lead enough conflict with some kind of big countries, which you have wanted to have very expensive weapons, it would be so. This is just my opinion, but you know, everyone knows it. Canis are trying to get this kind of drones, they are testing their software here.

34:02They send people, their people here to get some kind of experience.

34:16the rules of the war just to...

34:26just for example, the heavy armored vehicle like tanks. Russians, for example, in their assaults, don't use the tanks as attacks. This is armored vehicle, just armored vehicle. Nothing. and they take some infantry on them. And the one reason is if you have a big, big, big jammer on it and move as far as it can. So, so, and isn't a tank. Actually, the artillery is suffering from drones, infantry is suffering from drones, everything is suffering from drones. And as a war, we would have the rules. This war provides new rules of what we are doing, what we are afraid of. Is some kind of autonomous drone is present here?

35:25Of course, it's very good at the diagnosis, but it is not good enough to know is it enemy or not. That's the problem. And this is what is much easier for us. If you don't care about, is it enemy or just vehicle, you can just hit. And this is the point. They have drones with AI, just like U2V or something like that. But they don't have many of them. They have many stupid drones. You should have a ray of the moves, recognize the target, hit the target, but they don't have many of smart drones. A Bumblebee is a smart drone, but Russia don't have something like that. I would guess that that technology is going to be on the battlefield very soon if it's not there already.

36:36Is that fair to say? Yeah. As I saw later, the drone with AI is on a battleground for now. They are present here. but they couldn't recognize with 100 % accuracy that it is target what you should hit. And this is a problem because they could use just civilian cards, just sit in civilian cards and use it as your own. So would you hit it or wouldn't you hit it? The drone just has no information. They source a car, they recognize this as a car, but the operator should understand, is it enemy car? For example, if you have a rear, there is no civilians, absolutely no civilians. You can use autonomous mission, any car, just hit it.

37:41But if you know that there is some civilians could be there, or you couldn't use this kind of missions. Does Azov have just one R &D group or many R &D groups around the country?

37:59Actually, for now, many units have R &D groups with different results, I would say. This way, with different results, some of them are good, some of them are not so good. They are trying to find their ways, for example. You know, as I know, I am on the call and some other R &D group is present.

38:36I know some people from other R &D groups. I mean, they are doing their stuff. We are doing our stuff. Sometimes we are doing in communication, and as usually, we are different in deep parts. Yeah. And can I ask, in your group, is the primary focus navigation, jamming, or object recognition? I mean, which is most important? I mean, target recognition. I understand. Actually, we are focused on navigation. We have some good results in it. I would have to tell about this. We are focused on it. We are making some results. about some kind of recognition. This is an extra. When you can move to this repair, then you can use some kind of recognition.

39:50For now, we are focused on navigation. If we are talking about autonomy, we can just, first of all, navigation then, of course, navigation is enough for reconnaissance drones. And it could do some difference, even without recognition and then we just work with this thing. Yeah. And again, on the models, the AI models, you don't want to say where you're getting them. Is that right? And on the chips, where are you getting the semiconductors to run the models?