Autonomy and onboard AI

Autonomy that needs a radio link, a satellite constellation or a ground station is not autonomy. We build vehicles that keep deciding correctly when all three are gone.

Autonomy & onboard AI

A control loop cannot wait for the ground

Inference on the vehicle is not a fashion. It is the only way to keep a control loop closed when the link drops, and the only way to hold latency inside a budget a fast airframe can tolerate. Sending frames down and waiting for an answer is a design that demonstrates well and fails in service.

So the perception and decision pipelines run entirely on the vehicle, sized to the thermal and power envelope the airframe actually has — not the one a development kit implies.

Commercial compute, made to fit the airframe

The accelerators are commercial, NVIDIA and Qualcomm class. The carrier boards under them are ours, twelve layers up from the schematic, because the work is making a module fit an airframe's mass, power and thermal budget and still hold the loop rate.

  • Reference-free navigation — vehicles that localise with no GNSS, no beacons and no prepared environment, on onboard processing alone.
  • Flight control optimisation for manoeuvres outside the envelope stock controllers are tuned for, where standard gains lose authority.
  • Perception pipelines on embedded accelerators, from detection and tracking through obstacle avoidance to terrain assessment, built against the platform's constraints rather than against a benchmark's.
  • Bounded autonomy — explicit limits on what the vehicle may decide alone, with defined escalation and a deterministic fallback when confidence drops below threshold.

We care more about what it does when it is unsure

A model that scores well on a curated dataset and a model that behaves safely at dusk, in rain, over water, against a low sun are different artefacts. We validate against the conditions the vehicle will actually meet.

And we spend longer on the low-confidence behaviour than on the high-confidence accuracy, because the first is what decides whether the aircraft comes home. A detector that is right ninety-nine times and confidently wrong once has not been tested. It has been averaged.

What you get

  • Autonomy that survives loss of link, GNSS and ground infrastructure
  • Latency and power inside the airframe's budget, not a kit's
  • A defined, testable behaviour for the moment the system is unsure

Related work

Tell us the constraint you cannot get past

The mass budget, the latency, the environment, the certification route. We will tell you honestly whether it is something we should be working on — and if it is not, who should.