Swedish attack drone chooses its target and strikes alone, with an Nvidia Jetson Orin Nano card

The start-up Scaleout Systems runs a scale vision model aboard a prowling munition drone, capable of prioritizing targets and striking without a radio link or human intervention.

Scaleout Systems, swarming from the University of Uppsala, develops the embedded intelligence brick of the ALMA (Affordable Loitering Modular Ammunition) project led by BAE Systems Bofors. During a winter demonstration in Sweden in January 2026, a drone detected and geolocated several objects, classified an armoured engineering vehicle as a priority target, and then struck it without an operator ordering it.

The technical choice goes down from race to large models: the calculation is based on a Nvidia Jetson Orin Nano card, with small vision networks sized for drone electronics. A test conducted in June on the basis of the Swedish Air Force verified that the aircraft continued to operate after the loss of connection with the outside servers. Training is based on federated learning: the devices exchange model updates without bringing out recognition images.

The initiative, named FEDAIR, was selected in February 2025 by NATO's DIANA accelerator. The diplomatic calendar is moving more slowly: a non-binding text on arms autonomy was adopted in Geneva on 5 September under the Convention on Certain Conventional Weapons, with the United States and Russia having obtained a relaxation of human judgement formulations prior to the November Review Conference.