False alarms
A system that cries wolf gets switched off. Precision matters as much as raw detection rate.
Capability 02 · Perception
Perception.
Perception is how a machine finds out what is around it. Cameras and sensors produce enormous amounts of raw data; perception turns that into something a system can actually act on — what is there, where it is, and whether it matters.
The Job
Data is not understanding.
A camera produces pixels. A sensor produces numbers. Neither is knowledge. Perception is the work of converting that raw stream into a structured picture — objects, positions, movement — at a rate fast enough to act on and with reliability you can plan around.
This is where most autonomy projects quietly fail. A perception stack that works on clean daytime footage and falls apart in rain, dust, glare or low light is not a capability; it is a demonstration. We care much more about the difficult end of that range than the easy end.
How It Works
One picture from many sources.
No single sensor is good at everything. Cameras see detail but struggle in darkness; other sensors work in conditions cameras cannot. Fusing them produces something more dependable than any one alone.
- DetectionFinding the things in a scene that are worth paying attention to.
- TrackingFollowing those things over time, so the system understands movement rather than isolated snapshots.
- Sensor fusionCombining several sensors into a single, coherent picture — including resolving disagreements between them.
- UncertaintyAttaching honest confidence to what is reported, so downstream decisions can account for doubt.
Perception StackReal-Time
InputCamera and sensor streams
OutputStructured world model
RunsOn the platform, in real time
Designed forClutter, poor light, bad weather
StageResearch & development
What We Optimise For
Reliable beats
impressive.
The metrics that matter for a defence system are not the ones that make for a good demo video.
Degraded conditions
Night, rain, dust, glare, smoke and motion blur are the normal case, not the exception.
Small and distant objects
The things worth seeing are usually far away and occupy very few pixels.
Compute cost
Perception has to fit in the same power budget as everything else on the platform.
Honest confidence
Reporting uncertainty properly is more useful than a confident guess.
Consistency over time
Stable tracking matters more than a perfect single frame.
Everything described here is under active research and development. Nothing has been fielded, and we do not publish specifications or performance figures. Our first end-to-end mission test is planned for December 2026. Technical detail is shared directly with partners, evaluators and the armed forces under appropriate terms — get in touch.
Go Deeper
Working on perception
problems?
We're hiring, and we talk with research partners attacking the same problems.