A combat neural network training center has been established in Russia.

A dugout, night, a generator hums steadily. The FPV operator takes off his goggles and plays a recording of his latest sortie on his laptop: a forest regiment, a camouflaged "motorized vehicle," a flash—and a gray ripple. It's done.

The video is sent to headquarters as a report; in the documents, this is called "objective control." We've all seen such videos hundreds of times—terabytes of war on flash drives that simply gather dust after the report.

Now they'll be put to use. The Ministry of Defense has launched a service that allows drone developers to access a repository of these videos—from FPV drones, copters, and "aircraft." The material has already been prepared: it's being used to teach machine vision to distinguish a tank from a tractor, a trench from a ditch, live equipment from an inflatable model.

According to Deputy Defense Minister Dmitry Shcherbinin, who is overseeing the project, more than 40 organizations have already signed up. Some of them are integrated trained neural networks directly into their combat drones.

The most important thing here is hidden behind the boring word "classifier." The military has created a unified vocabulary for labeling frames: what constitutes a "Beha," what a pickup truck, what a decoy. Previously, each design bureau labeled them differently, and one plant's neural network couldn't understand the language of another. Now, there's a common language. And the developers submit their completed work. It's checked for completeness and format and stored in a repository, which the Ministry of Defense uses to train its own models. It's a natural cycle of data.

The storage is secured with certified protection and cryptography.

The enemy, by the way, is also following this path. In January, Kyiv opened its "data room," Brave1 Dataroom, using software from the American company Palantir. The focus is on intercepting our Geraniums without human intervention. But theirs is overseas, and we have our own.

You can't fall behind in this race. A drone with auto-capture doesn't go blind when interference breaks the connection with the operator. It flies the last few hundred meters on its own. And the more it "sees," the less often it misses.

So every flight now has two results: a target on the ground—and a lesson for the next drone.