Bandwidth
A detection is bytes; the video it came from is megabits per second. Deciding onboard is what makes the link budget work.
Edge AI
Inference where the pixels are
Sending every frame to the cloud to ask whether anything is in it wastes the one resource a flying robot never has enough of. Edge AI puts the model on the aircraft and sends the answer instead.
Why the edge
Three reasons the model belongs onboard
Latency
When a detection should change what the aircraft does next, a cloud round trip is too slow to be part of the loop.
Privacy
Imagery that never leaves the vehicle cannot leak in transit — which matters where the footage covers people or protected sites.
Simulation · a thermal hand-off from a car to a fire truck in CARLA, detections boxed, visible and simulated 640×512 thermal side by side
Interchangeable models
One aircraft, a different model per mission
Swap the model, not the hardware
The mission decides what matters — people, corrosion, vegetation, vehicles. Models are interchangeable so the aircraft does not have to be.- Detection and segmentationObject detection, anomaly detection and mobile segmentation models of the SAM Mobile and SAM Lite class run on the same device.
- OAK-D Lite as referenceThe reference edge device today, with the architecture kept open for other accelerators as they earn their place.
- No vehicle downtimeChanging the model is a deployment step, not a maintenance window — the aircraft stays as it is.
One aircraft, many models
Integration
Detections that go somewhere
An inference nobody acts on is an expensive way to heat a circuit board.
Onto the live map
Detections appear on the shared operational map as they happen, visible to everyone watching the mission.
Into alarms
A detection can raise an alarm, and an alarm can dispatch an aircraft — the loop closes without an operator relaying it.
Into reports
Findings carry into the mission report with their position and timestamp, so the deliverable assembles itself.
Where this stands today
On the platform first, on the aircraft next
Today the models run on the platform: an AI pipeline takes a live stream from any aircraft or camera, runs the model and puts the detections back on the video wall; the DJI Dock 3's on-board recognition is driven from the same console.
Interchangeable models on the aircraft's own compute, with the OAK-D Lite as the reference device, are in development on the Skyclient side. Detections into alarms and reports follow.
Bring us your detection problem
Tell us what you need to find and we will talk through what runs onboard today.