At the Edge, Every Second Counts
In today’s defence landscape, the most critical decisions happen at ‘the edge’—where frontline personnel must act swiftly under pressure, often with limited data, power, or connectivity. In a physical sense, the domain is characterised by uncertainty and chaos, as well as the paradox of sensory overload from too much information and insufficient good information for effective decision-making.
At the edge, there is often a sense that one more bit of information will be the missing piece of the puzzle and enable the best decision to be made, leading to hesitation and even decision paralysis. But in conflict, delay can be more dangerous than error. This is why generations of military leaders are told to make timely decisions, whether they be ‘right’ or ‘wrong’, rather than no decision at all.
The edge also has a series of technical constraints that make introducing new technologies more challenging. However, there is also an opportunity. Artificial Intelligence (AI) and Machine Learning (ML) can be used to reduce the overwhelming volume of information while increasing the quality and frequency of decision-support information, enabling better and more timely decision-making. Although this is hard to achieve, it is both possible and demonstrable.
Compute Power at the Edge
It’s critical to design these systems with deployment in mind, rather than for the lab. Developing a compute power-hungry model and then trying to strip it down so it can operate at the edge is impossible. These models have to be designed from first principles to operate under the constraints of the edge, and so one thing that must be kept in mind is that they are often low-power environments.
Even in environments where more power is available, building low-power models is often the best option. Sensor arrays in the air and sea consume significant power, so adding power-hungry models only compounds the burden. Instead, smaller, targeted models that tackle specific parts of the problem are far more effective at delivering the best outcomes. These smaller models require less compute power than a single universal model. These modular models are often more explainable when implemented effectively, as their outputs can be traced back to narrower, well-defined decision processes. In operational terms, this makes them easier to trust and integrate into decision-making cycles, giving operators powerful and practical tools.
Battery Power and Connectivity at the Edge
Soldiers can’t carry endless power supplies for all their equipment, and today’s infantry is already burdened with a significant battery load just to power their radios, night vision, and sensors. Batteries are now one of the heaviest contributors to a soldier’s load, often rivalling food and ammunition, making introducing any new, power-hungry system practically unfeasible.
However, efficient lightweight AI models can run continuously on limited energy. Importantly, there isn’t a linear relationship between model performance and power used. With a well-designed, modular, and flexible architecture, the trade-off between capability and power use can be tuned and optimised for each specific deployment.
Unsurprisingly, constant WiFi or 5G coverage at the edge is unlikely, but with the right AI and ML algorithms, it doesn’t need to be. Trying to send a constant stream of data over weak or blocked networks isn’t realistic, and in future conflicts, we shouldn’t expect our adversaries to let us freely use those networks anyway.
As the ongoing war in Ukraine has demonstrated, satellites can be jammed or intercepted, and cellular infrastructure destroyed. This means that AI systems built for the edge must be designed to operate in disconnected and intermittent communication environments. By processing as much information locally as possible and only transmitting what is absolutely essential, operators can continue to function effectively without relying on constant connectivity.
Sending large volumes of raw data isn’t just wasteful, but potentially dangerous since it creates a greater electronic signature for adversaries to detect and exploit. Therefore, only mission-relevant data should be shared and in as small a packet as possible to ensure resilience and prevent adversaries from gaining an advantage.
Normally, AI models are trained in central locations with lots of computing power, but with federated learning, it can take place locally at the edge to improve the efficiency and security of communications. This protects sensitive information and distributes resilience across the force, so that if one node is lost, the rest will continue to function. With data remaining local, there is also no central target for hackers to attack or a single point of failure. This transforms the edge from a vulnerability into a strength.
Human Centricity at the Edge
For AI to be useful at the edge, it must “speak the language” of its users, not the data scientists who made it. While F1 scores, precision, and recall are useful for engineers and scientists to measure performance and progress, this data doesn’t make it easier for non-technical users to understand the outputs.
When people are given information, they usually ask three questions: Can I trust this? Can I understand this? Do I need to do anything about this? Models at the edge must answer these questions and be what we like to call ‘humble’, meaning they declare their uncertainty and take their environment into account.
For example, a platoon commander who has repeatedly experienced false positives from a detection system may become desensitised to its warnings, even when they indicate a genuine threat. This risks leaving units vulnerable to real dangers. However, a system that communicates its own uncertainty (such as by stating ‘with 80% certainty, this is a hostile UAV’) allows the human to weigh that input alongside their own judgement before making a decision. Instead of treating the AI as absolute, this kind of “AI in the loop” approach is vital to enhance long-term trust in systems and reduce errors.
AI must also be able to account for the fact that its use cases may vary dramatically. If someone is detecting drones in real time to protect personnel, speed matters more than accurately identifying types of drones. They will happily accept some false alarms as long as all drones are detected promptly. However, if someone is analysing drone behaviour to inform long-term strategy, they will need detailed classification and tracking. Two similar tasks with very different human requirements and different trade-offs between performance, explainability, and context.
Integrate, Don’t Replace
This doesn’t mean you need to completely overhaul your setup to deploy AI at the edge. Effective integration means enhancing, not replacing, what is already there, as no single entity has all the answers or all the expertise. Ultimately, AI at the edge should be an enabler of resilience and adaptability. It must be designed to work alongside existing communications and operational infrastructure, empower human judgement rather than replace it, and evolve in line with the changing nature of conflict. Getting this balance right will not only reduce the burden on operators but also provide a decisive advantage in the most demanding environments.
The edge is not an abstract concept; it is where the mission lives and dies. It is where information is scarce and overwhelming, decisions must be made in seconds, and power, compute, and connectivity are all at a premium. But it is also where AI and ML will have the most significant impact if designed to fit the fight, not the lab.