Swedish startup demos fully autonomous NATO attack drone with onboard AI
The loitering munition detects, selects and strikes targets offline, raising fresh questions over human control in armed conflict.
Swedish startup Scaleout Systems demonstrated an autonomous NATO loitering munition that uses onboard AI to detect, rank, and strike targets without external communications, as part of BAE Systems Bofors’ ALMA program at the Winter Demo 2026 event. The system runs small, non-frontier computer-vision models fully on the drone, with human involvement limited to starting the mission and acting as a failsafe while the AI executes a 320‑second end‑to‑end engagement. In a separate February “arctic strike” test, Scaleout flew an Airolit S1 airframe at -18°C running the YOLOv8 Nano model on an Nvidia Jetson Orin Nano, achieving around 30 fps at roughly 20 m/s with ~30 ms latency using only a pinhole camera model for ranging. Scaleout’s architecture combines its Scaleout Edge platform and federated learning into a “Tactical Computer Vision Network” that keeps edge nodes operating and learning through network degradation, then backfills model and telemetry updates when links are restored. The project, called FEDAIR, is part of NATO’s DIANA accelerator, which has provided €100,000 in funding plus training and test access, even as UN experts reiterate that human control is required for lawful use of lethal autonomous weapons.
Why it matters
By showing a low-cost loitering munition that can independently find, choose and hit a target in a few minutes without any data link, Scaleout Systems and BAE Systems Bofors highlight how lethal decisions can move from remote human operators to inexpensive hardware running compact models. This pushes autonomy closer to front-line use just as UN experts insist that human judgment and control are necessary for lethal autonomous weapons to comply with the laws of war, sharpening a live policy and ethics conflict.