● Autonomous edge-AI detection

Detect.
Machines that see before anyone else does.

Sunset R&D builds lightweight autonomous drones that perceive, classify and track in real time at the edge — engineered for civilian and non-lethal training environments. Sensing, never striking.

See the platforms How it works

Detection only · Non-lethal by design · No payload mount

13TOPS

Hailo-8L on board

~28fps

Edge inference

240mm

4Quad wheelbase

4.6.3

ArduPilot Copter

Offensive payloads

Test #1 4Quad printing
Build package 50 printable parts
Flight params verified against source
Classes human · animal · vehicle
Payload none sensors only
Printer Bambu A1 no supports
Ops FAA Part 107 daylight VLOS

A drone that streams video to the cloud is only as good as its link. Ours classify every frame on board, report only what they see, and carry nothing that could hurt anyone.

01 Perception

Perception that lives on the aircraft, not in the cloud.

Every frame is classified on board. Nothing leaves the airframe but a compact event: what was seen, where, and how confident the model was.

STAGE 01

Capture

Camera frames into the companion computer.

Raspberry Pi 5

STAGE 02

Motion gate

Background subtraction flags change before any inference is spent.

MOG2 gate

STAGE 03

Classify

Neural classifier on a dedicated accelerator.

Hailo-8L · 13 TOPS

STAGE 04

Event

Class, confidence and position. No video leaves the aircraft.

~28 fps

STAGE 05

Report

Authenticated telemetry to the operator console.

Bearer-token API

// POST /api/v1/telemetry — schema example
{
  "status": "airborne",
  "mode": "LOITER",
  "hailo_fps": 27.9,
  "detections": [
    { "class": "human", "confidence": 0.94 }
  ]
}

Classes: human · animal · vehicle
Inference: on board, offline-capable
Video uplink: none
Flight link: MAVLink · Guided mode
Privacy: no imagery stored off-craft

02 Capability

Three jobs, done on the edge.

01 / RECOGNISE

Object recognition

Human, animal and vehicle classes inferred at the edge at ~28 fps, with per-detection confidence and position streamed to the fleet monitor.

02 / DETECT

Motion detection

A background-subtraction gate flags movement across a fixed field of view before the classifier spends a single inference cycle on it.

03 / NAVIGATE

Autonomous flight

Pixhawk 6C running ArduPilot with gate-verified parameters. Guided-mode navigation from the companion computer; a hard geofence and failsafes underneath it.

Proof of work

Programme record · Aug–Sep 2026Verified

131

Automated tests passing

Across the perception, telemetry-protocol and flight-support codebase.

511

Cited sources

Across eleven domain briefs on regulation, flight stack, perception and materials.

Printable parts in Test #1

Each one gated manifold, flat-bottom and bed-fit before it reaches the printer.

Bench stages before air

E-stop, radio and ground-station failsafes proven on the bench first.

03 Test #1

The first airframe is on the print bed.

The 4Quad build package: printable parts, a KiCad schematic set, ArduPilot parameters verified against the flight-controller source, and a nine-stage bench plan.