The flight controller keeps the aircraft upright; the companion computer makes it worth flying. Every capability your customers actually pay for — detection, tracking, mapping, smart inspection, live video with overlays — runs on this second brain. Here's how to size it, wire it and keep it from cooking itself.

The division of labour

The architecture rule that keeps aircraft safe: the flight controller must remain able to fly and land the aircraft if the companion computer dies mid-mission. The FC (see FC selection) owns rate loops, estimation and failsafes on a real-time MCU. The companion owns perception, planning and connectivity on Linux — powerful, and precisely as deterministic as a Linux box under load, which is why nothing safety-critical lives there.

Hardware classes

ClassExamplesComputePowerFits when
SBC (ARM)Raspberry Pi 5/CM5, Orange PiCPU only (+ light NPU)3–8 WMission logic, geotagging, streaming, cloud link
Edge-AI moduleJetson Orin Nano / NX classGPU/NPU, 20–100+ TOPS7–25 WOnboard inference: detection, tracking, depth
x86 moduleIntel N-series / Core mini-PCsStrong CPU, weak GPU10–35 WLegacy x86 software stacks, heavy single-thread
FPGA/ASIC add-onsHailo, Coral acceleratorsEfficient fixed-point inference2–5 WBoosting an SBC to inference duty at low watts

Choose by the heaviest workload with numbers: model, resolution, frame rate, latency target — the full workflow around this is in our AI drone integration guide. "Run YOLO-class detection at 1080p 30 fps under 50 ms" is a spec you can shop with — and the vendor benchmark you should reproduce before committing the airframe design.

Power and heat: the taxes nobody budgets

Every companion watt is battery mass or minutes (do the arithmetic with the endurance guide) — 20 W of compute on a small quad can be two minutes of flight. And those watts leave as heat from a module that throttles at ~85–95 °C junction, mounted inside an enclosure in direct sun with no fan-friendly dust environment. Working practices:

  • Duct real airflow — prop wash is free cooling if the mechanical design uses it; a sealed box is a slow cooker.
  • Structure as heatsink: thermally strap the module to an aluminium plate doing double duty.
  • Test the thermal corner: full inference load, zero airspeed (hover), maximum sun, then check for throttling — throttled perception at the moment of obstacle detection is a systems failure, not a component one.
  • Power it cleanly: a dedicated regulated rail with brown-out protection; companion reboots from voltage sag during aggressive manoeuvres are a classic "ghost bug".

Wiring and software patterns that age well

  • Link: Ethernet FC↔companion where hardware allows (bandwidth, electrical robustness); UART/MAVLink remains fine for command-and-telemetry-only roles.
  • Software interface: MAVSDK or ROS 2 (uXRCE-DDS on PX4; MAVLink/DDS on ArduPilot). Resist inventing a bespoke protocol — the ecosystem tooling you inherit is worth more than the elegance you imagine.
  • Deployment: containerise the autonomy stack, version it like firmware, and make the update path field-safe (A/B partitions or rollback). A fleet of aircraft with hand-configured companions is technical debt with propellers.
  • Watchdogs both ways: companion monitors its processes; FC monitors companion heartbeat and downgrades the mission gracefully on loss — a behaviour that belongs in your failsafe matrix and gets exercised in flight test.
Cameras count too

Perception compute is sized by its input. A global-shutter camera on a stable mount with decent optics reduces the model quality you need; a rolling-shutter camera on a vibrating mount manufactures artefacts no network fully unlearns. Spend on the sensor before the TOPS — and mind USB3's notorious RF interference with GNSS receivers when you route it.

Build vs buy, once more

Integrated FC+companion combos and vendor "autonomy kits" compress your first year usefully. The trade is lock-in exactly where your product differentiation lives. A defensible pattern for startups: standard FC + standard compute module + your own carrier board when volumes justify it — the carrier is a modest engineering project that buys you connectors you chose, power sequencing you trust, and a bill of materials you control. It's also a well-bounded first project to hand an external electronics partner if the skill isn't in-house yet.

Frequently asked questions

What does a companion computer do on a drone?

It runs everything too heavy for the flight controller: computer vision, AI inference, mapping, mission logic, video encoding and cloud connectivity. The flight controller keeps flying the aircraft; the companion computer decides what the aircraft should do.

Jetson or Raspberry Pi for a drone?

Raspberry Pi (or similar SBC) suffices for mission logic, geotagging and video streaming. The moment you run neural networks onboard — detection, tracking, obstacle perception — you want a Jetson-class module with a real GPU/NPU. Budget 10–25 W and the thermal design to match.

How does a companion computer talk to the flight controller?

Traditionally MAVLink over a UART; increasingly Ethernet. On PX4, ROS 2 via uXRCE-DDS is first-class; on ArduPilot, MAVLink libraries (pymavlink, MAVSDK) and the DDS bridge cover most needs. Keep the safety-critical loop inside the flight controller regardless.