In the global race for battlefield AI, Ukraine has one thing others don’t: its gigantic trove of organized battlefield data, which can be used to train weapons. Now, Kyiv is making that data available to eligible allies, so they could train their weapons as well.
Ukraine has launched a new cooperation initiative that allows the government, Ukrainian companies and international partners to work together training AI models used in semi-autonomous weapons, newly-appointed Defense Minister Mykhailo Fedorov announced last week.
“Today, Ukraine has a unique array of battlefield data that is not found anywhere else in the world. These are millions of annotated frames collected during tens of thousands of combat flights,” Fedorov wrote in his statement.
“This data is already used to train neural networks that automatically detect ground and air targets in the Delta system,” Ukraine’s widest-known battle management network.
According to Fedorov, the data sharing platform will allow partners to safely train models without direct access to sensitive information, and work with a large array of labeled photo and video materials, which is constantly updated.
How well this data is collected in one place and annotated will be critical to the initiative’s success. A centralized, secure database with controlled permission access significantly improves development efficiency, an AI drone developer for Ukraine’s National Guard, who goes by ‘Jack,’ pointed out.
“Ukraine does have an advantage because they do have systems like Delta that are integrated so there's a platform that AI can learn from,” said Kateryna Stepanenko, a Russia analyst with the Institute for the Study of War.
Another AI expert who develops drones for the National Guard, who goes by ‘Walrus,’ also emphasized the importance of database security.
“Security is the primary concern because the video or images used for model training and evaluation could contain highly sensitive information about area of operations, tactics and targeting,” he said.
Terabytes per diem
Over 9,000 Ukrainian drones are in the air, collecting many terabytes of footage every day. In terms of runtime, that’s the entire Game of Thrones series multiple times over.
That data includes extensive pics and videos of terrain and Russian forces and equipment. Ukraine also collects massive amounts of information from acoustic sensors, EW, and other systems.
“Ukraine does have the largest databases of all of those things,” said Deborah Fairlamb, the founder of Green Flag Ventures and a prominent defense tech investor in Ukraine. “They definitely have far more up-to-date information on EW than anybody in the West even understands is happening.”
Western countries, especially the US, are also working to develop this tech, using satellite images, terrain footage, and mockup targets. However, the Ukrainians are constantly updating their data with new battlefield information of the ever-shifting battle-scarred landscape and Russia's evolving technology.
“It's the constant updating of data and imaging that the Ukrainians have this advantage,” Fairlamb said.
Some foreign companies are already getting in on the game. Merlin Hipp, the founder of the German AI company Lancelot Systems, which is working in Ukraine, praised Ukraine’s multifaceted data collection from the battlefield.
“Generally I believe that Ukraine is doing an amazing job and recognizing that its data is extremely valuable for the next generation of warfare,” he told Euromaidan Press. “Not just for Ukraine but for the entire Western world.”
“Ukraine is giving back to partners,” he added. “I think a lot of the advancements in military artificial intelligence would not have been possible without the efforts of, for example, Brave One,” Ukraine’s battle tech incubator program.
Quality vs quantity
It’s not enough just to have a lot of raw data. In interviews with over a dozen AI developers, dronemakers and military insiders, a consistent message emerged: battlefield data often has very inconsistent quality, and thus, usefulness.
“The challenge is not necessarily having enough raw data but rather having enough data that is properly labeled and cleaned,” Walrus said.
“This is always the most expensive and time consuming part of developing high quality AI models along with training the model. In my experience, the quality of the data and the data labels is far more important than the quality of the model itself.”
Trending Now
However, to Hipp, founder of Merlin, quantity has a quality of its own — Lancelot’s AI has been trained on “millions” of videos. The German developer credits his system’s effectiveness to having “a lot of data and consistent high-quality training.”
Ukraine leads Russia in data organization
Data is also often scattered across multiple databases and units, making it difficult to assemble. While Ukraine also has this issue to contend with, it is likely ahead of Russia in organizing it into unified, searchable databases.
“The Russians don't have that system just yet,” Stepanenko said. “They're trying to develop it but they're behind because Ukraine has been having those systems since at least the start of the full scale invasion.”
Samuel Bendett, a Russia adviser at CNA, concurred: “When it comes to Delta, obviously they're ahead of the Russians. When it comes to other applications, they are ahead of the Russians.”
The Russians are also behind in making data available to their troops. While Russia has some systems, they exist as a “patchwork… on a very limited scale, mostly within the higher command echelon,” rather than with regular soldiers who would benefit from these systems, he added.
Ukraine also has better access to cloud data storage. Russians heavily relied on Western servers, which impeded their ability to save a lot of the footage from every single drone across the battlefield.
“They now not only have to develop the system like a Delta, they also have to develop their own domestic cloud or get that from one of their partners,” Stepanenko said. “Ukraine has cloud systems that are based outside the country. That is also an important technical component.”
On the other hand, if Russia wants cloud storage, Beijing is likely to offer a hand, provided they get something out of it — be it access to useful data, or just cold, hard cash.
Data only as useful as its application
Having data is good, but it is not a silver bullet to develop effective battlefield AI. Multiple sources who spoke to Euromaidan Press on this topic said almost the same thing, verbatim: “it’s not about the data, it’s what problem you are trying to solve.”
For example, “Detecting a Shahed-type UAV at long range is a different challenge than tracking a quadcopter at 200 meters,” Jack said. “Night thermal tracking is different from daylight optical tracking. Each scenario requires focused training and evaluation.”
Brian Streem, the CEO of Vermeer, which produces drones that can navigate by terrain image matching, had an even blunter assessment — the data is only as good as the model, the engineering, and how well the two work together to produce an effective weapon.
“If someone had this magical data, and if you actually gave it to like a team of AI engineers, they'd be like, what the fuck is this?” he said. “What are you trying to do with it?”
In other words, not every AI model, or attempt at a solution will be created equal. Many might fail where a few succeed.
Fedorov said that Ukraine is “ready to work with partners on joint analytics, model training, and creation of new technological solutions.”
