A drone over a Swedish test range in January spotted four objects, ranked an armored engineering vehicle as the one worth hitting, flew to it and dropped an explosive on it. A human had set the mission. No human picked that vehicle.

The AI came from Scaleout Systems, an Uppsala University spin-off working on a low-cost loitering munition project run by BAE Systems Bofors. Scaleout sits inside NATO’s innovation accelerator, which took it on in 2025. Every step of the target work ran on the aircraft’s own computing, with nothing sent back to a server.

Six years covering this beat and the sequence has held: a human designates, then the AI flies the last stretch. Sweden moved the designating onto the drone.

Scaleout’s Onboard AI Ran The Entire Target Chain

Scaleout supplied the onboard AI for ALMA, the Affordable Loitering Modular Ammunition project headed by BAE Systems Bofors. At Winter Demo 2026 in Sweden this January, the drone detected and geolocated threats, then ranked an armored engineering vehicle highest. It flew to that vehicle and struck without direct human commands.

Jeremy Hsu reported the demonstration for Ars Technica on September 17. BAE Systems Bofors had announced Winter Demo 2026 in January, showing ALMA alongside roughly twenty startups and technology companies over three days in Karlskoga, with Swedish defence minister Pål Jonson joining a panel. Scaleout’s own presentation describes the aircraft detecting, identifying and geolocating threats using AI, with all of it handled by dedicated onboard computing and none of it processed externally.

An operator could still take the controls. Nobody needed to.

In June, Scaleout ran a second demonstration at a Swedish Air Force base in Uppsala, where the military already licenses its main software platform. A forward computing node kept running inference and active learning after losing contact with the company’s lab, then synced its updates on reconnect.

Electronic warfare and jamming routinely break communication links on a modern battlefield, which is why a growing number of Ukrainian drones now fly with onboard AI. The Ukrainian answer has been the same as Scaleout’s: push the intelligence onto the airframe. DroneXL covered Kolibri, the Ukrainian FPV that flies 25 miles and finishes the strike alone, in July, and Mykhailo Fedorov’s open-source battlefield training dataset in March. Scaleout is chasing the same problem with a formal architecture and a NATO badge.

CEO Andreas Hellander framed the payoff as an alliance-wide one. “In principle, you can unlock collaboration between NATO member states,” he told Ars.

Small Models Beat Frontier AI On Hardware This Size

Scaleout does not run frontier models from OpenAI or Anthropic. It trains smaller computer-vision models sized for what a drone or a forward workstation can carry. Federated learning ties them together: devices share model updates rather than raw sensor footage, keeping reconnaissance data on local hardware.

The specific failure it targets is model drift. A detector trained on desert imagery performs badly over a city, and a war does not pause while somebody retrains it in a data center. Under the federated scheme, headquarters nodes retrain on aggregated battlefield data from several units and push revised models back out to the edge whenever a link opens.

The work runs under a project called FEDAIR, short for Federated Aerial Intelligence for Recon, inside NATO’s Defence Innovation Accelerator for the North Atlantic. Uppsala University says Scaleout was selected in February 2025 alongside five other Swedish companies. It fits the wider European buildout DroneXL has been tracking, from Sweden’s $440 million commitment to unmanned military systems to Helsing and Stark taking €600 million in German loitering munition contracts away from Rheinmetall.

Geneva Settled For A Diluted Text Two Weeks Before This Landed

States party to the Convention on Certain Conventional Weapons agreed a non-binding autonomy text in Geneva on September 5, after the United States and Russia pushed late changes and Washington sought flexibility on human-judgment language. The Seventh Review Conference in November decides whether treaty negotiations start.

Reuters reported during the session that the US position had hardened over the previous six months, with Washington favoring guidelines over a binding treaty and formally approaching at least six countries to back it. Consensus arrived overnight after talks ran long, and Dutch foreign minister Tom Berendsen announced the deal on the Saturday. The agreed text has not been published. Nicole van Rooijen of Stop Killer Robots told Reuters three years of work had been “substantially diluted in the last hours.”

The US does have a domestic rule. DoD Directive 3000.09, reissued in January 2023, requires that autonomous weapon systems be designed so commanders can exercise “appropriate levels of human judgment over the use of force,” and sends systems outside its exemptions through senior review before development and again before fielding. DroneXL covered the Pentagon’s own AI-powered kamikaze drone push in January 2025. The directive governs American programs. It has nothing to say about a Swedish demo.

DroneXL’s Take

The strike is the headline. The learning loop is the story. A drone that picks its own target is not new, and DroneXL has covered several of them. What changed in June is a system that keeps getting better while nobody can reach it.

DroneXL reported in January on the New York Times investigation into Eric Schmidt’s AI attack drones in Ukraine, where autonomous terminal guidance was landing above a 70% hit rate and a human still designated the target first. June is where Scaleout added the other half. The way I read that test, a node that keeps improving while cut off and hands its improvements back on reconnect turns the models from a fixed capability into a compounding one, and compounding capability is what nobody at the CCW has a definition for.

The Swedish case is strong on the merits. Jamming breaks links constantly in Ukraine, and an aircraft that goes stupid the second its link drops is an aircraft you have already lost. Ukraine reached that conclusion from the bottom up, without a NATO accelerator behind it. Designing for a contested link is sound engineering. It still does not answer who is accountable when the model ranks wrong, and that question is the one Geneva spent September sanding down.

Washington spent the past year regulating what a Chinese camera drone might send back to Shenzhen, and spent the first week of September asking for flexibility on whether a machine may choose a target. Same government, same year. Only one of those involves an explosive. DroneXL has argued from the start that the FCC’s foreign-drone ban is protectionism wearing a security jacket, and this is the cost of that misallocation in plain view: the enforcement energy went to the thing that photographs a roof, and the thing that selects an armored vehicle got a footnote.

Scaleout’s DIANA announcement, its Scaleout Vision and defense pages, and BAE’s Winter Demo release all describe what the system does on its own. None of them says where operator veto lives once the link is gone. The whole design premise is that the link will be gone. Whether an abort still reaches the aircraft in that state is the question the November text will be judged on, and neither company has answered it in public.

November 16 to 20, the Review Conference meets in Geneva to decide whether the September text becomes a negotiating mandate, another rolled-over discussion, or nothing. Watch whether the United States runs the same six-country lobbying effort Reuters documented this month. If it does, every rule written after November will be describing something that already flew in January.

Sources: Ars Technica, Scaleout Systems, BAE Systems Bofors, Uppsala University, Reuters, AFP, US Department of Defense, Reaching Critical Will

DroneXL uses automated tools to support research and source retrieval. All reporting and editorial perspectives are by Haye Kesteloo. Read our editorial standards.