A NATO-backed startup put the targeting decision inside the drone
A NATO-backed startup has put the whole targeting decision inside a small drone — detection, ranking, and the attack run — with a human able to steer but no longer required to. Scaleout Systems, a Swedish company spun out of Uppsala University in 2018, showed the capability under the ALMA project led by BAE Systems Bofors: an onboard model identified an armored engineering vehicle at 89 percent confidence, decided it was the highest-value target for the mission, and flew to it. A remote operator could still take over. The drone did not wait for one.
The interesting part is what is not in the loop. Scaleout is not running frontier models from OpenAI or Anthropic at the edge — it packs lean computer-vision models onto drone hardware, pilot tablets and forward-deployed workstations, where power, size and the absence of a reliable connection all bind. Where processing happens matters when jamming is routine and data centers have been physically destroyed in this war.
The training loop is the real product. Models improve through federated learning: each drone and forward node keeps its raw sensor data local and sends only selective model updates back to a computing node at platoon or company level, which retrains on the aggregate and pushes an updated model out when there is an opportunity to sync. "Models might have been trained in a desert environment, and if we try to deploy them in an urban environment, they're not going to perform well," cofounder and CEO Andreas Hellander said. In June the company demonstrated the failure case deliberately at a Swedish Air Force base in Uppsala — the local node kept running inference and active learning after losing its link to the lab, then reconciled once the connection returned. That is the design that lets the same approach scale across allied countries without shipping classified footage anywhere.
Scaleout joined NATO's DIANA accelerator in 2025 and works on the FEDAIR project under it; the Swedish military already licenses the core platform. The company is pairing up with France's AI Verse, whose GAIA platform generates physically accurate synthetic battlefield imagery — infrared signatures, lens distortion, motion blur at specific shutter speeds — to solve the cold-start problem of training counter-drone models before enough real footage exists.
What this test actually proves is narrower than the headline. A single demo with a purpose-built target and a human able to intervene is not a deployed autonomous weapon. But the engineering direction is unambiguous: the intelligence is moving onto the airframe, the human is becoming optional rather than structural, and the legal review that assumes a person pulls the trigger is being outrun by the hardware. NATO now has demonstration footage of an autonomous kill chain it did not have to write a policy for.
We covered the commercial side of this data problem in September — Ukraine is selling its battlefield drone footage to AI companies.
What to watch: whether a NATO member fields an autonomous engagement outside a demonstration, and whether any procurement contract writes a per-shot human sign-off into the requirement.
Should an AI be allowed to choose who it kills, or should that decision stay with a human even when the link is jammed? Tell us in the comments.
