Small AI Models Enable Drones to Identify and Attack Battlefield Targets

Summary

Scaleout Systems, a Swedish startup backed by NATO’s Defence Innovator Accelerator for the North Atlantic, is adapting small machine-learning models for drones, pilot tablets, and forward command posts. The models are designed to perform computer-vision inference on hardware ranging from embedded devices to edge workstations, avoiding continuous reliance on large data-center models. Through its Federated Aerial Intelligence for Recon project, the company’s system shares selective model updates between local computing nodes instead of transmitting sensitive raw battlefield data. Headquarters nodes can retrain models on data collected from several sources and send revised versions back to edge devices when communications are available. The approach is intended to address changing environments, such as models trained in deserts performing poorly in cities, as well as electronic warfare and jamming. In a January 2026 Swedish demonstration of the BAE Systems Bofors-led ALMA low-cost loitering-munition project, an onboard model detected, identified, and geolocated potential threats, selected an armored engineering vehicle as the highest-value target under the mission rules, and guided the drone to drop an explosive. A human operator could still control the aircraft, but the demonstrated mission ran without direct human commands. In June, a Swedish Air Force base test showed that a forward node could continue inference and active learning while disconnected from Scaleout’s central node, then synchronize its updates after reconnection. Scaleout says the decentralized learning strategy could eventually support collaboration across NATO member states, although the article describes this as a future possibility rather than an established deployment.