The January ALMA demo turned Andreas Hellander and Salman Toor's federated-learning research into onboard targeting, with a pilot held in reserve.
By [RuntimeWire Staff](https://runtimewire.com/author/runtimewire-staff)
· Published
Primary source: Tom's Hardware
Why it matters #
Scaleout shows how edge-AI infrastructure can move from distributed training into autonomous weapons without relying on a cloud connection. The founders have demonstrated the technical chain from detection to strike, while procurement, combat validation and the acceptable level of human control remain unsettled.
In January 2026, Scaleout Systems, co-founded by Uppsala researchers Andreas Hellander and Salman Toor, supplied the onboard AI for a drone that ranked possible targets, chose an armored engineering vehicle and released an explosive during BAE Systems Bofors' ALMA demonstration in Sweden.
The system worked from human-set mission parameters. Once started, it detected and geolocated objects, assigned the engineering vehicle the highest score and completed the approach without direct commands or an external communications link. A designated pilot remained available to intervene as a failsafe.
The January test is about eight months old. It returned to view after Ars Technica reported on Scaleout's defense work on September 17th and Tom's Hardware detailed the demonstration on September 20th. The renewed attention captures the turn Hellander and Toor have made since Russia's full-scale invasion of Ukraine: software first developed to train models across trucks and other distributed machines is being recast as infrastructure for autonomous military systems.
An academic spinout finds its defense market
Hellander and Toor spent roughly a decade researching systems that divide computing and data storage between central clouds and machines closer to where data is produced. Uppsala University says their business idea began taking shape in 2017, when university investment helped finance a software prototype; Scaleout lists 2018 as its founding year.
Hellander is an associate professor whose work spans computational science, distributed systems and systems biology. Toor is an associate professor in scientific computing, a former senior cloud architect at Sweden's national computing infrastructure and one of the lead architects of FEDn, Scaleout's federated-learning framework.
Their original thesis concerned data that could not safely or practically be moved into one central repository. Federated learning sends a model to local machines, trains it against data held there and shares selected model updates rather than the underlying records. The approach can apply to hospitals, vehicle fleets and industrial equipment. A battlefield adds a harder constraint: the network itself may disappear.
Hellander told Ars that the war in Ukraine and a changing security environment pushed Scaleout to apply its work to defense. Scaleout joined NATO's DIANA accelerator in 2025 with FEDAIR, or Federated Aerial Intelligence for Recon, receiving 100,000 euros in contractual funding, training and access to test facilities. The project extends federated learning across drones, pilot tablets and field command posts.
Scaleout had already raised 35 million Swedish kronor in February 2025. Fairpoint Capital and Navigare Ventures co-led the equity round, joined by Almi Invest, Uppsala University Invest, the Beijer Foundation and several individual backers. Scaleout said the capital would support its cloud-to-edge software across industrial, automotive and defense deployments.
What the ALMA demonstration showed
BAE Systems Bofors' Affordable Loitering Modular Ammunition program is aimed at a lower-cost autonomous weapon that can spend much of a mission searching rather than flying directly to a preselected coordinate. During the Winter Demo, Scaleout's computer-vision stack reportedly spent about 200 seconds conducting reconnaissance. The full mission finished in less than 320 seconds.
The important technical step was the sequence connecting perception to action. The onboard system classified objects, estimated their physical locations and applied weighted mission logic to produce a ranked target list. It then passed the selected target into the navigation and approach system. Processing remained aboard the aircraft, allowing the mission to continue without sending video to a remote data center or waiting for instructions over a radio link.
That setup narrows the role of the human operator without removing people from the mission architecture. Humans defined the target criteria and operating parameters before launch, and the pilot retained the ability to take control. The demonstration therefore supports a specific claim: the drone performed target ranking and the final approach autonomously within a mission envelope established by people.
The Nvidia hardware detail belongs to a separate arctic test that Scaleout described on February 16th. In that demonstration, an Airolit S1 airframe used Nvidia's Jetson Orin Nano and a YOLOv8 Nano object-detection model at minus 18 degrees Celsius. Scaleout says the system processed about 30 frames per second while flying near 20 meters per second and sustained roughly 30 milliseconds of inference latency.
Scaleout also says that test estimated distance without a dedicated depth sensor by combining a pinhole-camera model with the known dimensions of recognized vehicles. Its software retained a target's last estimated location when visual contact dropped, navigated toward that point and resumed closed-loop pursuit after reacquiring the object.
Those measurements come from Scaleout's field work. They offer engineering detail, though they do not establish battlefield accuracy, false-positive rates or performance against camouflage, decoys and unfamiliar vehicles.
The larger product is the learning network
Selling a targeting module would place Scaleout in a crowded field of defense autonomy suppliers. Hellander and Toor are pursuing a broader position: the software layer that deploys, monitors and updates models across disconnected drones, vehicles and ground computers.
Scaleout calls that architecture a Tactical Computer Vision Network. A drone can continue running inference offline, save detections locally and synchronize selected information when a connection returns. Ground nodes can retrain models against local data, then contribute model updates to a shared version without exporting the raw imagery.
Scaleout demonstrated that workflow with the Swedish Air Force on June 30th. One node operated at the F16 air base in Uppsala while another ran at Scaleout's lab. Engineers degraded and then severed the lab node's connection. Scaleout says inference and active learning continued at full frame rate, with detections stored locally. After reconnection, the system prioritized health signals, critical alerts, model updates and telemetry rather than transmitting the backlog indiscriminately.
That model-management layer is where Scaleout's academic roots become commercially useful. A computer-vision model trained on clear desert imagery can degrade in snow, fog or a dense urban environment. Scaleout's pitch is that military units can retrain and redistribute models from current sensor data without pooling every sensitive image in one cloud system.
A demonstration, before procurement
The public tests establish that Scaleout has connected target detection, ranking and autonomous flight control on small edge hardware. They do not amount to operational validation. The ALMA work remains a demonstration, with no announced procurement order or confirmed combat deployment.
The unresolved policy questions are as consequential as the engineering ones. The United Nations' 2026 discussions on lethal autonomous weapons have centered on keeping responsibility with states and people, maintaining a human chain of command and reducing unintended engagements. Scaleout's retained pilot and human-defined mission parameters fit within that debate, while the autonomous selection and strike sequence shows how much decision-making can occur after launch.
For Hellander and Toor, the defense pivot gives years of distributed-systems research an immediate customer problem: models must keep working when communications fail and conditions change. The January flight showed that the same architecture can also move from identifying an object to acting against it. That step will determine how militaries, regulators and buyers judge Scaleout's technology.