Ukraine's AI‑powered drones regain edge in Russia conflict

Currently, Ukrainians are deploying hundreds of thousands of drones for various tasks, with the most notable footage capturing FPV drone attacks. These drones are used to attack equipment like BMP-3 infantry fighting vehicles, various tanks, including the latest T-90M, and even individual soldiers or helicopters.

After the Ukrainian drones' initial effectiveness, the Russians began deploying control signals or GPS signal jammers extensively. This resulted in a classic shield-and-sword race with alternating periods of drone dominance and impotence, during which experiments with control signal frequencies were conducted.

To address these challenges, solutions akin to professional military drones were sought, utilizing artificial intelligence algorithms that enable autonomous attacks on designated targets. After months of trials, Ukrainians, sometimes with assistance from Western companies, achieved satisfactory results. Below, you can see a crashed FPV drone supported by a Google Coral AI module.

The key is to provide a module with enough computing power to create computer vision capable of tracking and following a selected object. Due to the limitations of vehicle-mounted jammers, their effective range is a maximum of 1,300-1,600 feet, necessitating fully autonomous flight to cover such distances.

Today, there are many commercial AI solutions proficient in object recognition. The difference is that, instead of recognizing humans or, for example, cute kittens, the algorithm must identify a T-72 family tank and pursue it. Some solutions, such as Skynode-S modules, even offer navigation based on triangulating the drone’s position by comparing the terrain viewed by the drone’s camera with preloaded satellite maps.

The development of machine learning algorithms presents a challenge, along with ensuring systems have sufficient performance, compact dimensions, and low power consumption. Consequently, a drone is created that can only be stopped by shooting it down.

Professional solutions include Israeli SMASH modules from Smart Shooter, which have gained significant interest in recent years from armed forces and allow drones to be shot down with handheld firearms even from several hundred meters (the effective range varies with the caliber of the weapon it’s mounted on).

Meanwhile, 12-gauge smoothbore shotguns prove effective for improvised solutions. With appropriate buckshot ammunition, they can be effective even at distances of around 160 feet. Utilizing new ALDA ammunition from Beretta Holding, drones can be targeted at distances of about 260-390 feet.

Photos reveal that a part of the drone includes a G950-04742-01 development board, characterized by a performance of 4 trillion floating-point operations per second (TOPS) while consuming only 2W. Google claims that this system can execute the MobileNet v2 SSD neural network architecture instruction in 14 ms, whereas a 3 GHz Intel Xeon Gold 6154 processor requires 106 ms.

The Google board features NXP’s iMX 8M system with four ARM Cortex-A53 cores, an ARM Cortex-M4 core, integrated Vivante GC7000Lite graphics, Google’s proprietary Edge TPU coprocessor, and a cryptographic coprocessor. It is equipped with 8 GB eMMC memory, 1 GB LPDDR4 memory, and necessary I/O ports.

Such boards are freely available for purchase, and Ukrainians, unlike Russians, have unhindered access to them. However, as the sanctions reality shows, some are likely to be available in Russia as well.