DATK: counter-drone synthetic datasets

A tool that fabricates labelled data — trajectories, radar returns, swarm telemetry — for drones that do not exist, so that counter-drone models have something to learn from.

Read first

Everything here is fictional — invented drones, specifications and records; no real platform, programme or munition is modelled.
The capabilities are deliberately speculative; several do not exist and may never. That is the point, explained below.
Only a slice is shown — a few parameters, scenarios and runs, not the full threat or parameter space.
It is not operational: it plans nothing, targets nothing, and holds no guidance, coordinates or recommendations.

Fig. 1Two runs at speed. An EFP swarm services a moving armoured column through an urban canyon; then an interceptor swarm meets an incoming hostile swarm over a defended asset. Alongside the arena: a PPI radar scope, live telemetry, swarm state, and an event log — each a stream the run writes to the dataset.

DATK generates synthetic datasets for battle drones that have not been built. Given a description of a platform, a swarm, a payload, and the scene it operates in, it runs a deterministic top-down simulation and records what a sensor network would have seen: where each unit went, what the radar returned, how the swarm coordinated, and what happened to the targets. The output is a labelled dataset for a threat that does not yet exist.

Why the question is being asked now

The war between Russia and Ukraine led many defence agencies to reconsider the future of drones, and with it the future of counter-drone interception. Cheap uncrewed systems, used in large numbers and in coordination, have changed what a contested space looks like; the response — detecting, classifying, tracking, and intercepting those systems, increasingly as swarms rather than as single aircraft — has become a research priority across several agencies. The threat that drives the planning is rarely the drone fielded today. It is the one expected next: faster, quieter, more autonomous, harder to see, and acting as a group.

Counter-drone models need data that does not exist yet

A detection-and-tracking model learns from labelled examples: this radar return is a drone at this range and bearing; this track is a striker and that one a decoy; this swarm is converging and that one is dispersing. For threats already in the field, such examples can be gathered — slowly, at cost. For threats still on the horizon, the silent and cloaked and self-coordinating swarm, there is nothing to gather, because the thing has not been built. A counter-drone system trained only on what exists is, by construction, blind to what comes after it.

Synthetic data is the established way out of that bind. If the trajectories, radar signatures, and swarm telemetry of a hypothetical threat can be produced with their labels already attached, a model can be trained and stress-tested against that threat before it is ever met in the field. The value is not in reproducing today’s drones exactly; it is in generating large, varied, fully-labelled datasets for drones that are plausible but not yet real, and in being able to move the threat as engineering moves.

What DATK does

The tool has two stages. The first is a dense parameter cockpit: the airframe and its skin, the means of propulsion and its energy, the flight profile, the swarm’s size and its method of coordination, the payload, the environment, and what the sensors record. Some parameters are numbers and switches; others — the swarm’s objective, its rules of engagement, its target priority — are entered as plain language, the way an analyst would state intent rather than encode it.

The second stage is the run. The configured swarm attacks the configured scene, drawn from above and symbolically, while the engine records, at every step, a labelled dataset: the trajectory of each unit (position, heading, speed, energy, status), synthetic radar frames in the manner of a plan-position indicator (range, bearing, cross-section, doppler), swarm telemetry, the comms graph as it heals and breaks, and a discrete event log of launches, detections, hits, losses, and link drops. The dataset exports as CSV and JSON. Each run is one small, reproducible, fully-labelled corpus.

Deliberately unreal, on purpose

The drones in DATK are not realistic, and are not meant to be. The parameters reach past current engineering by design — near-silent ionic propulsion, in-flight recharge from beamed power, self-healing and dissolving airframes, metamaterial cloaking, stigmergic and distributed-brain coordination, mid-flight fission and fusion, directed-energy and cyber effectors. None of these is claimed to exist. They are present because a counter-drone system should be exercised against capabilities it has not yet encountered.

This is the argument for putting unreal capabilities in a generator rather than keeping them out. A dataset confined to what exists trains a model to recognise yesterday’s threat. A dataset in which the signature can be made quieter, the turn tighter, the link denied while the swarm keeps coordinating, trains a model against the envelope of plausible threats rather than a single point inside it. Speculative capability is a property of the tool, deliberately exposed and adjustable, not an error to be corrected.

What the AI does, and what the mathematics does

The simulation itself is mathematics. Motion is integrated from kinematic limits. Swarm behaviour is computed from a few rules — separation, alignment, and cohesion for flocking; a decaying pheromone field for stigmergy; an auction for distributing targets across the swarm. Radar returns follow from geometry, with a cross-section modulated by aspect, cloaking, and fog. Lethality is resolved through a fixed table of effect by target class. A seeded random-number generator makes the entire run reproducible. None of this is learned; all of it is arithmetic, and that is what makes the exported labels trustworthy — every recorded number has a known cause.

The role of AI sits on either side of that mathematical core. At the front are the parameters written in plain language — objective, rules of engagement, target priority — where intent is expressed as description, and where a language model is what turns description into the weightings and behaviours the mathematics then executes. (In this demonstrator those fields shape the run directly; the language layer is sketched rather than built out.) At the back, in the tool’s intended use, the synthetic dataset is what an AI counter-drone model is trained and evaluated against. The mathematics makes the world consistent and labelled; the AI reads intent into it at one end, and learns from what comes out at the other.

Fig. 2A strike into a denied zone. As links drop under jamming, the swarm keeps coordinating through a stigmergic field and short-range relays; the event log fills with link drops and the radar view fills with ghost returns from spoofing. The point of the run is the labelled record it leaves behind — exactly the kind of denied-environment data a counter-drone model would otherwise never see.

A synthetic dataset is only ever as useful as the realism and the breadth of the generator that made it, and a model trained on one inherits the generator’s assumptions — including the speculative ones. DATK does not resolve that limit; what it offers is to make those assumptions explicit and adjustable rather than hidden, so that the threat a counter-drone system is trained against can be stated, varied, and reproduced. The demonstrator shown here exposes a small part of that, on a deterministic engine, so its behaviour is legible and repeatable.

Everything shown is fictional by design.

Recordings are AAIC’s own, captured from the DATK demonstrator.