A Swedish startup has published video of a strike drone selecting and engaging a target without a human deciding which target, and without any radio link to the outside world. The processing runs on an Nvidia Jetson Orin Nano, a developer board that costs a few hundred dollars and fits in a hand.
Scaleout's demo, titled Onboard Edge Intelligence for Autonomous UAV Missions, shows the aircraft flying a mission against an armoured engineering vehicle inside parameters a human set beforehand. After the button press that starts the system, further input is optional. A pilot stayed on the controls as a failsafe. The drone spent roughly 200 seconds on reconnaissance and finished the whole mission in under 320 seconds, with no communication in or out. That last detail is the point of the exercise: a system that never transmits cannot be jammed.
Loitering munition is the accurate term
Coverage of the video reached for kamikaze drone. The programme's own label, loitering munition, describes the behaviour better. Almost the entire flight is spent searching, ranking candidate targets, and waiting. The strike is the short part.
The software stack is less exotic than the capability suggests. In an arctic demonstration at BTC Karlskoga in February, flown at minus 18 degrees Celsius, an Airolit S1 airframe ran YOLOv8 Nano on the Jetson. Scaleout reported 30 frames per second at around 20 metres per second of airspeed, with a latency target of 30 milliseconds and roughly that sustained in the field. Distance estimation came from a pinhole camera model and known object dimensions rather than a depth sensor, which removes a part and a failure mode.
YOLO is a general-purpose object detector that thousands of hobbyists have run on the same board. The engineering here is not the model. It is making a small model good enough on a cold airframe at speed, and building the network around it.
The interesting part is what happens when the link dies
Scaleout calls its architecture a Tactical Computer Vision Network, combining edge devices that train locally with federated learning so model updates propagate without central collection. The company published a June test with the Swedish Air Force using two ground nodes under a cloud control plane. The node at the air base had a stable link. The one in a Uppsala lab had its connection degraded and then cut entirely.
The disconnected node kept running inference and active learning at full frame rate while offline, logged detections locally, and backfilled on reconnect in a set priority order: heartbeat, critical alerts, drift, model updates, telemetry. That queue is the actual product. Anyone can run a detector on a drone. Running a fleet that keeps learning through sustained electronic warfare and reconciles cleanly afterwards is a harder problem, and it is the one that decides whether a demonstration becomes a system.
The programme, FEDAIR, sits inside NATO's DIANA accelerator with 100,000 euros of development funding, training, and test access. As Tom's Hardware laid out in its write-up, Scaleout Edge is already deployed under an active licence.
The rules are still being written in Geneva
None of this is a procurement order and none of it has been used in combat. A United Nations group of experts has been working on lethal autonomous weapons and recently concluded that human judgment and control are required for compliance with the laws of war. Its recommendations go to Geneva in November.
The gap between that timetable and the video is the story. The expert group is debating whether meaningful human control can be defined; a startup has already flown a mission where the human contribution was a button and a set of parameters agreed in advance. Scaleout would argue those parameters are the control. Whether that satisfies a treaty is not a question engineers get to answer.
It is the same pattern showing up everywhere capability outruns oversight, from agents being deployed faster than anyone can review them to federal agencies quietly scaling AI ahead of any public accounting. The difference here is the payload. A drone that decides on its own, on hardware anyone can buy, with no link to cut, is not a future problem for the November session to contemplate. It flew in February.