26 Comments

Many thanks for posting this!

Expand full comment

UA drones production is well diversified: top10 producers have around 20-25k monthly production. It is like 30% of total production of 45-50k monthly. And even those 25k monthly are produced on different locations and facilities.

Next Schwerpunkt is FPVs with AI to minimize operator influence on “last mile”-75-100m to target

Expand full comment

This developer believed AI might be functional this year. It will certainly change the equation between drones and EW. It will also probably be bad news for Russian airfields.

Expand full comment

I would say mass production of FPVs with AI should definitely start till the EoY 2024. Russkies are working hard on it either. the one who starts first, will have huge advantage

Expand full comment

Yes, and the quality of the AI will improve over time making drones increasingly deadly for both sides. I'm also interested in the counter measures to AI. If initial AI homes in on shapes, how realistic does a decoy have to be? Could the same decoy be hit multiple times?

Also, EW will still have an impact on AI drones if their video feeds can be jammed, denying the enemy reconnaissance.

If AI on the front lines is equally destructive to both sides but Russian aircraft are destroyed on the air bases then I think Ukraine comes out ahead. But a counter to that point defense target is the machine gun/cannon teams that Ukraine already deploys as a mobile defense.

Expand full comment

During forthcoming months we will see some elements of AI solutions only. Full functional AI to be expected en masse maybe in a year.

I have no idea about ru-solution, but UA side works hard on it to avoid decoys to be hit and own vehicles as well

Expand full comment

Yeah, that "own vehicles" part seems important.

Expand full comment

I think "AI" is a bit of a misnomer here, or at least it encompasses many rather different things. The simplest form of terminal guidance for drones is the ability to automatically stay aimed at an operator-selected object in the video stream. This is what Javelin missile does, for example, and this is hardly AI, but rather "conventional" image processing, although some AI elements may enhance the algorithms.

Automatically classifying objects in the video stream, and picking targets among them - like soldiers, or cars, or planes on the ground - would better fit AI definition, but may be a bit further away.

In any case though - either approach needs high-performance processing chips - heavily-parallel GPUs, FPGAs, etc., and it's vitally important to cut off any access to them for the Russians. And I think in this specific field Western sanctions are relatively efficient.

Expand full comment

Training a model is what requires the high performance chips. Running a trained model is vastly less computationally expensive. The hardware for running a CV model on a drone *may* not be too sensitive to lose. However, protecting the trained model from extraction would be a difficult problem.

Expand full comment

Yes, running a trained model is vastly less computationally expensive than training it, but training takes place offline, on virtually-unlimited HW and power resources. Running must be implemented in real time, off a drone battery - you can't do it on an R-Pi or something like that. And I'd say developing a model that really works in real-life conditions is a far more difficult problem than protecting it from extraction by the enemy - if that's what you meant.

Expand full comment

It can be done but I have no idea about power consumption impact. https://pytorch.org/tutorials/intermediate/realtime_rpi.html

Expand full comment

From what I read, Ukraine was working on object recognition, differentiation and prioritization. I don't know of the quality of chips that are coming from the west, but over a billion dollars worth and 2500 components (which includes transistors, not just chips) made it to Russia last year.

Expand full comment

Ukraine is certainly working on it, and the whole world is working on it, but I'm not aware of any stark breakthrough yet. I certainly wish Ukrainians will be the first to make one.

As to the chips, I think a difference is between what the Western governments perceive as of military importance, and what the Russians can "press" into their military production, but is also used for civilian applications. The later are much more numerous, cheaper, and more difficult to control.

Having said that, I do not know if Russia manages to smuggle in even the most advanced chips. She might.

Expand full comment

Perfect article, Tom&Don! Many thanks!

Expand full comment

In Ukraine there is a society of volunteers printing parts for the military called DrukArmy https://drukarmy.org.ua/ua

Expand full comment

Thank you for such a great support!

Expand full comment

"uses infrared and is much cheaper than thermal optics" - Google reports that all thermal cameras are infrared cameras.

Expand full comment

Not really: IFR are just day/night mode; teplovision cams could show termal objects , all other could be hardly identified. Ideally to set up irr+thermal cam, but it makes a FPV too expensive

Expand full comment

Thermal cameras use longer-waves IR - typically around 10 um. The light-sensing pixels in such cameras are microbolometers, or exotic semiconductors, like HgCdTe, and they are expensive.

On the other hand, conventional cameras with silicon sensors typically have peak sensitivity in near-IR - around 0.8....0.9 um, and can still produce discernable images in very low ambient light. Additionally, in case of total darkness, a few IR LEDs can sufficiently illuminate the scene up to tens of meters, perhaps, even hundreds of meters in a narrow beam.

Expand full comment

Thanks for this inside tour on a very interesting subject. Ukraine is truly a nation in arms.

Expand full comment

Truly, and they can use all the help they can get.

Expand full comment

Thanks Tom I found this to be very interesting

Expand full comment

Could you pls tell us something about these new Russian ZAL-16 reconnaissance drones, which supposedly have a control and video range of about 40-50 km, and the transmission is very difficult to disrupt? Sources among volunteers dealing with drones and drone jamming systems indicate that this is a serious problem. This drones are intended to replace Orlans soon

Expand full comment

Sorry, ZALA Z-16, more precisely.

Expand full comment

And they work with Ukranian company SWD (Shock Wave Dynamics) for new FPV adapted warheads

Both anti personel and anti armor

https://twitter.com/wilendhornets/status/1773095867123163486

https://pbs.twimg.com/media/GJvD2SQXQAA2dlF?format=jpg&name=large

Expand full comment

Many thanks to Don and Wild Hornets' Dima for this great article.

It's important to highlight that Wild Hornets are a 100% donation-funded charity. Their crowdfunded drones have destroyed or damaged an estimated $500 million of enemy equipment since starting operations a year ago.

Wild Hornets' high effectiveness was why they were chosen as the only FPV drone to get dedicated fundraising support by the well-known Ukrainian news site Censor.net, run by prominent military reporter and anti-corruption activist Yuri Butusov. A Censor.net editor sits on the Wild Hornets board.

Find out more on the Wild Hornets website: dykishershni.com/en.html

Expand full comment