Solutions · 04 — Integrate

One Stack. Any Platform.

Standard interfaces, no proprietary protocol, no bespoke firmware.

The same sensor, compute and software integrate across unmanned platforms from multirotors at 50 m to pseudo-satellites at 25,000 m. The camera joins your MAVLink network as a standard camera component, so most of the integration is configuration rather than development — and where the airframe allows, several cameras combine into one system.

Request a briefing Integration documentation

MAVLinkThe camera is a camera component on your existing network — ArduPilot or PX4.
From 350 gSensor and mission optics.
One pod, many camerasCombine cameras for a wider swath or two ground resolutions — or change one camera’s mission by changing its lens.

Platform classes

From 50 Metres to the Stratosphere

Four classes of unmanned platform carry the same stack. What changes with altitude is the area each pass covers and the mission it suits — not the sensor, the interfaces or the edge compute.

10–25,000 mASL

HALE UAVs & High-Altitude Pseudo-Satellites (HAPS)

Persistent surveillance

3–5,000 mASL

Medium-Altitude Long-Endurance (MALE) UAVs

Persistent surveillanceTactical reconnaissance

300–1,500 mASL

Mini & Tactical UAVs

Threat detectionTactical reconnaissance

50–300 mASL

Single-Rotor & Multirotor UAVs

Threat detectionPersistent surveillance

Altitudes are the bands each class typically operates in, not limits of the sensor. Coverage per pass widens with altitude; ground resolution trades against it, set by the optic you fit.

What you integrate

Three Interfaces, All Open Industry Standard

Integration touches the stack in three places. Each one uses an interface your platform already has, so there is no proprietary bus to adopt and no Sapient-specific firmware to maintain.

01 — Sensor
Camera over MAVLink
ROOK or ECHO joins your MAVLink network as a standard camera component. Your flight controller supplies position and orientation; the camera requests the streams it needs at startup and reports back on the same link.

Protocol
MAVLink

Flight stack
ArduPilot or PX4

Gimbal
With or without
02 — Compute
FORGE processing module
The processing module carries the sensor and runs the intelligence layer onboard. It needs DC power and a telemetry link to the autopilot; Ethernet and USB-C are there for data, configuration and offload.

Supply
11–20 V DC

Power
30 W typical, 65 W peak

Compute
NVIDIA Jetson Orin NX
03 — Software
IGNITE:AI framework
Bring your own model in a container. Sapient handles everything either side of inference — ingest and tiling before, then consolidation, tracking, georeferencing and export after. Findings leave as Cursor-on-Target.

Model
Yours, unchanged

Runtime
Container on target

Output
CoT · JSON

Run your own model

The Integration Boundary

Inference is the only stage you own. Everything either side of it runs onboard, so integration means packaging a model against a fixed boundary — not building a pipeline.

Sapient · before inference

Ingest from the sensor
Debayer and correct
Resample and tile to your format

Your container

Your model, unchanged

An isolated Docker container holding your framework, your weights and your logic. It runs as built.

Framework: Any — agnostic
Language: C++17 or Python
Acceleration: TensorRT on target
Retraining: None
Your data: Never shared
New version: Swap the container

Sapient · after inference

Consolidate across tiles
Track across frames
Georeference and export

Three calls, and that is the whole surface

01
Configure
Declare the tile size, pixel format and target ground resolution your model was trained on.

You
02
Receive
Tiles arrive in GPU memory — debayered, corrected, scaled and overlapped. No file handling.

IGNITE delivers · you consume
03
Return
Hand back detections. Consolidation, tracking, georeferencing and export follow automatically.

You submit · IGNITE consolidates

IGNITE:AI framework →SDK & integration guide →

Scaling up

Combine Cameras, or Change the Mission

One airframe can carry more than one camera. Arrange their footprints side by side and you widen the swath covered per pass; nest a longer optic inside a wide one and you get two ground resolutions from the same pass. Where the airframe cannot carry a second camera, changing the optic gives the one it has a different mission.

Three camera apertures mounted flush on one airframe, with their footprints overlapped so coverage is continuous across the join. Three cameras cover roughly three times the width of one in the same pass — the alternative being a second aircraft, or three passes.

A wide optic and a longer optic on the same airframe. The wide one holds the whole scene; the longer one resolves fine detail inside it. Two ground resolutions from one pass, rather than choosing between them or flying twice.

×NSwath scales with camerasEach added camera adds its own footprint. Coverage per pass rises roughly in proportion; how many you fit is set by the airframe, not the architecture.
From 350 gPer cameraSensor and mission optics. “From” because the figure depends on the optic fitted.

One coordinate frame, one timebase

Each camera detects independently, but every one of them takes position, orientation and GPS time from the same autopilot. Findings therefore arrive already in a single world coordinate frame and on a common timebase — so what reaches your command system is one consolidated picture to act on, not several streams of imagery to reconcile first.

A different mission, a different lens

ECHO takes interchangeable optics behind a standard mount, so one camera moves between wide-area survey and long-range standoff without requalifying the system. Same interfaces, same edge compute, same integration — a different lens.

Size a Fit for Your Airframe

Bring the airframe, the autopilot and the mission profile. Detailed interface control documents and integration guides are available through the resource hub.

Request a briefing Integration documentation