Ukrainian FPV Operator Reviews AI-Enabled ‘Lock On Target’ Drones

Small drones, especially FPVs, dominate the battlefield in Ukraine, destroying roughly twice as much enemy equipment as every other weapon put together. Their main issue is jamming, which anecdotally brings down at least half of FPVs before they can reach the target.

Jamming can be countered with a “lock on to target” function: the operator selects the target, the drone flies the rest of the way, so no control signal is required though the jamming zone. FPVs with this feature have been tested on the battlefield since last year. I talked to a frontline user in Ukraine about their experience with this technology.

From Walleye To ActiveTrack

Back in the early days of guided weapons, a video camera able to lock on to a target was an expensive capability. The AGM-62 Walleye, an unpowered glide bomb with a TV camera in its nose was first used in 1967, scoring a direct hit on the main power plant in Hanoi the first time it was used in action. The 1,100-pound Walleye cost around $160,000 in modern money, but the lock-on-target needed good lighting and contrast to work reliably. Even in the clear conditions of Desert Storm in 1991, Walleyes only scored a hit rate of about 60%, far less than laser-guided bombs.

Computers are now millions of times more powerful, and many consumer quadcopters, down to the humble $199 DJI Neo, now boast an ActiveTrack function. This allows the operator to draw a box around an object and have the drone follow it. Hands-free operation means the drone can follow you skateboarding, skiing or cycling. It does not take much imagination to see how similar technology could lock a drone on to a target and crash into it.

The added price is a key issue FPVs cost around $400 each. Expensive AI hardware is unaffordable; the AI systems developed and used by Saker for reusable bomber drones are in another league to FPVs. Ukrainian FPV developers seem to use the lowest-cost hardware they can get. While there are some highly capable Auterion Skynode-S units available, Ukrainian developers are more likely to use a cheap Raspberry Pi Zero – costing perhaps $100 – or Google Coral AI dev board, with software developed using the popular YOLO vision family.

A year ago we started seeing the first results, rough-and-ready AI enabled drones which locked on and flew through jamming for the last few hundred meters. But how well did they actually work?

Inconclusive Results

‘Michael,’ Commander of the Typhoon drone unit of the National Guard of Ukraine, told me about his experience with early AI-enabled drones.

“We’ve tested this solution a couple of times, and the results were quite inconclusive,” says Michael. “While it was able to lock on target, the target needed to be highly distinguishable from the background.” [The same problem as the Walleye in the 1960s] “When dealing with well-camouflaged targets, it became very challenging to accurately detect the target at all, or the exact point where the strike should occur.”

Developers claim that AI drones, which have fast reflexes and never panic or get distracted, can achieve a hit rate of 80% which they say surpasses human operators. So far, Michael says, humans are keeping up.

“A skilled pilot with a good technical setup and properly configured stack settings can achieve this level of success,” says Michael.

He also notes that the hit rate of AI drones depends very much on the type of target. Vehicles moving in the open are easy for machine vision to lock on to and follow. It is a different story with smaller, less obvious targets moving behind cover – like dismounted infantry.

“Nowadays, Russian forces do not rely heavily on armored vehicles during assault operations,” says Michael. “Instead, they primarily use small infantry teams. Even if the system is able to recognize a target in the forest, I’m not sure it will be able to navigate around all the obstacles on the way to the target. Target acquisition is especially difficult in summer conditions.” [i.e. when there is dense foliage]

Michael sent a video from a recent operation which shows just how difficult it is to track soldiers moving through woodland even in winter. From a machine vision point of view, they keep changing size and shape and disappearing behind cover. It is easy to see why the automated lock on has difficulty following on a target.

The Vision Challenge

Others in Ukraine have noted similar problems.

“Machine vision isn’t developing as fast as we’d expected”, Oleksii Babenko, CEO of Ukrainian drone manufacturer Vyriy Drone, told Ekonomichna Pravda recently. He notes that manufacturers sent drones to the front with underpowered guidance algorithms which are now being upgraded.

Russia Lancet – a high-end, military-grade loitering munition using U.S.-made NVIDIA processors – also experienced problems with its lock-on target function in 2024. These now seem to have been solved, and the function is used in about a third of attacks.

In February 2024, Breaking Defense described AI drones as a “revolution that wasn’t” but added “just yet”.

But while they may not be suitable for everything, AI-enabled drones appear effective against some targets.

“This feature could be particularly useful for engaging unarmored vehicles such as cars and motorcycles, as well as small infantry groups on roads,” says Michael.

This suggests that such drones would be effective in the ambush role, tactic in which FPVs are landed beside roads to wait for likely targets when cued by a scout drone. The technology could even automate the ambush process, turning the drone into a long-range anti-tank mine.

Michael notes most FPVs tend to have cheap, low-resolution cameras. Historically, FPVs used an analog data link so anything better would be wasted, but a better, high-resolution camera could feed a better image to the AI and significantly improve the system’s ability to stay locked on.

Michael also says the AI drones they tested were not as effective at taking out heavily armored targets because they simply flew directly at them. A skilled operator knows to fly around to hit the target’s most vulnerable spot, like the thinly-armored turret rear of Russian tanks, where a strike from an FPV can destroy the vehicle instantly. AI would need to match this.

“The key question is whether an AI-driven ‘lock on target’ system would be able to recognize these specific weak points,” says Michael.

Newer AI-enabled drones are claimed to be more capable. In November the Ukrainian government announced that it had ordered the first batch of 3,000 drones with machine vision. Mykhailo Fedorov, Deputy Prime Minister and Minister and driving force behind Ukraine’s drone effort said a month later they planned to “significantly increase” the proportion with autonomous targeting.

Low-cost AI drones have not yet swept the battlefield. But in the last 18 months they have moved from laboratory experiments to viable if limited weapons. As Deepseek recently reminded us, AI technology is improving fast and may make sudden rapid advances, even without new hardware.

This will definitely be a space to watch in 2025.