Up to 12 British companies will receive access to Ukrainian battlefield sensor data under a competition intended to accelerate development of AI models for collaborative drones and autonomous systems.
The Ministry of Defence has opened the first competition under the UK-Ukraine AI Partnership, giving selected businesses access to the Avengers AI Labs platform. The database contains imagery and video collected in Ukraine using daylight and thermal sensors and is used to develop, train and evaluate artificial intelligence models.
More than six million detected objects are represented in the database, including tanks, artillery, air defence systems, infantry and aerial targets. Companies will use the material to develop systems intended to operate under conditions that are difficult to recreate in conventional development environments.
The competition is being run by the MOD’s Rapid AI Delivery Taskforce, TF RAID. Proposals can address autonomous target recognition, distributed decision making, adaptive mission execution and collaborative sensing and information fusion, with proposed applications required to serve both British and Ukrainian defence requirements.
Defence Secretary Wes Streeting said: “This world-first access to real battlefield data will allow British companies to continue pushing the frontier of defence innovation.” The MOD plans to select participating companies within weeks and expects exploitation of the data to begin soon afterwards.
The engineering requirement goes beyond teaching a vision model to recognise an object. Collaborative drones have to remain useful when satellite navigation is unavailable, communications deteriorate or individual vehicles cannot maintain constant contact with a central controller.
Those operating conditions push more decision making towards onboard and distributed software. A system relying on uninterrupted communications with a remote operator can lose much of its capability when links are jammed or degraded, whereas distributed autonomy allows aircraft to alter routes, priorities and behaviour within defined rules.
Scaling the number of vehicles also changes the workload for operators. A fleet in which every aircraft requires continuous individual control becomes increasingly manpower intensive. Collaborative architectures are intended to let smaller numbers of personnel supervise multiple systems while software handles parts of navigation, sensing, task allocation and coordination.
The British Army is already examining a related operating model through a separate software-defined swarming demonstrator. An eight-UAV test bed is being used for continuing swarm trials, including work on collective control, open architecture and future acquisition options.
The Avengers competition addresses the training and model-development side of that broader capability. Operational sensor material gives engineering teams examples of environmental clutter, target variation and degraded conditions that are difficult to reproduce convincingly with clean laboratory datasets.
A model trained predominantly on ideal imagery can perform differently when weather, thermal clutter, obscuration, damage or unfamiliar target configurations alter what its sensors see. Access to battlefield data broadens the range of conditions available for training and evaluation, although it does not remove the need for independent testing before software is trusted in operational systems.
Hardware constraints will shape the models that can actually be deployed. Smaller drones have limited processing power, electrical capacity, payload mass and cooling, so software developers cannot assume the aircraft will carry the same computing resources available in a ground-based data centre.
Communications architecture presents a similar constraint. Collaborative sensing can distribute information across several platforms, but the system has to decide what data is worth transmitting when bandwidth is restricted and what decisions have to remain local when links disappear altogether.
Open interfaces become increasingly useful as processors, sensors, radios and aircraft designs change. A model that can only function on one tightly integrated proprietary platform is harder to update or transfer than software built around interfaces that allow components to be replaced as technology develops.
The Ministry of Defence has increasingly emphasised software-defined equipment for precisely that reason. Algorithms and electronics can evolve much faster than complete military platforms, forcing procurement programmes to account for software updates, data access and processor refreshes as continuing parts of the equipment life cycle.
Selected companies will initially receive access to data rather than a production order. The competition nevertheless creates a direct route from Ukrainian operational experience into British development teams, followed by testing intended to determine whether the resulting models improve collaborative systems under realistic constraints.
The first commercial and technical filter will be the selection of the companies allowed into the programme. The more demanding test begins afterwards, when systems trained on the Avengers material have to show useful performance despite degraded communications, contested navigation and the unpredictable sensor conditions represented in the data itself.
