NATO-backed startup Scaleout Systems is putting compact computer-vision models on drones and field computers so they can identify targets without a continuous connection to a central data center. The Swedish company’s system can also share selected model updates with local military nodes rather than transmitting raw battlefield sensor data.
Those local nodes retrain models on observations from multiple devices and distribute updates when communications become available. This federated-learning approach is designed for environments where jamming or attacks on infrastructure make centralized processing unreliable. It may also help models adjust when terrain changes, such as moving from desert training data to an urban operation.
Scaleout has demonstrated the technology in surveillance and strike settings. In a January test of the ALMA loitering-munition project, a drone used onboard processing to detect and rank possible threats, selected an armored engineering vehicle and flew to drop an explosive. A human could take control, but the aircraft completed that mission without direct commands. A separate Swedish Air Force test showed a base node continuing inference and active learning after losing its central connection, then synchronizing later. The work shows that autonomy is moving closer to deployed weapons, making the exact role of human authorization and safeguards a consequential issue beyond the technical benefit of operating offline.