Luxonis M8 Controller Box Adds Relays and CAN to OAK4

Closing the Loop Between Seeing and Doing

Luxonis has introduced a prototype of the M8 Controller Box, an industrial I/O expansion module that plugs into the M8 connector on its OAK4 edge AI cameras. The OAK4 line runs on Qualcomm's DragonWing QCS8550 and does its inference on-camera; what it has lacked is a clean way to act on that inference without a separate controller in the loop.

• 16 GPIOs, four relays rated 16A, CAN 2.0, RS232, and two USB 2.0 Type-A ports in one enclosure
• Powered directly from the OAK4's M8 connector — no separate supply, no second power run to the mounting point
• Isolated strobe output, onboard audio, three buttons and three status LEDs, with DIN rail and VESA mounting
• $249 for the PR1 prototype, including the M8 cable, DIN rail clip, and plug-in screw terminals

Why This Removes an Entire Box From the Cabinet

The standard machine-vision deployment has three parts: a camera that sees, a controller that decides, and an actuator layer that does. Even when the camera runs its own model — as the OAK4 does — the output usually has to travel to a PLC or industrial PC before anything physically happens. That intermediate box costs money, occupies rail space, adds a network hop, and becomes another thing to configure and maintain.

Hanging relays and CAN directly off the camera collapses that chain. Detection to actuation becomes a local decision, and the deployment turns into a camera and a terminal box instead of a camera, a box, and a controller. For a fixed-function station — reject a part, trigger a diverter, sound an alert — that is a genuinely simpler architecture with fewer failure points.

What Do 16A Relays Actually Let You Switch?

Real equipment. Four contacts at 16A is well past signaling territory and into pumps, solenoids, contactors, conveyor motors, and lighting rigs. That capacity is what separates a controller you can deploy from an evaluation board you demo with.

The supporting interfaces matter just as much. CAN 2.0 reaches motor drivers and vehicle-derived subsystems; RS232 covers the long tail of instruments and legacy machinery that never got an Ethernet port; the isolated strobe output is aimed squarely at machine vision lighting, where a camera firing its own illumination in sync with capture is standard practice. Readers who followed our mini computer coverage of the ESP32-C5 relay board with RS-485 will recognize the same design philosophy applied one tier up the compute stack.

Is a $249 Prototype Worth Buying?

For the right buyer, yes, and Luxonis is refreshingly clear about what "PR1 prototype" means. CAN, USB-A, UART/RS232, LEDs, buttons, relays and GPIOs are listed as fully supported in this hardware revision; the strobe output is partially supported. That is exactly the disclosure an integrator needs to decide whether to prototype now or wait.

The audience is people building a proof of concept for an automated inspection or sorting cell who want to demonstrate end-to-end behavior — camera sees defect, arm diverts part — without standing up a full control stack first. At $249 against the cost of an industrial controller plus integration time, the arithmetic works even if a later revision supersedes it.

Where Edge AI Vision Is Heading

The broader pattern is inference moving to where the sensor already is, and the peripherals following it there. A camera with an NPU, a relay bank, and a fieldbus is functionally a small industrial computer that happens to have a lens — much the same convergence we noted in the credit card-sized TI edge AI boards covered earlier this week.

DIN rail and VESA mounting are the tell that this is meant for cabinets and machine frames rather than benches. Small detail, but it is the kind of thing that decides whether a device gets deployed or stays in the lab.

Sources: CNX Software — July 30, 2026; Luxonis shop — July 2026.