Luxonis Puts Five Cameras on One Board for Robots That Don't Fit a Single Lens

Luxonis began taking orders Tuesday for the OAK 4 FFC, a board-level vision computer that connects up to five of the company's camera modules to a single processor delivering 52 trillion operations per second of compute, aimed squarely at humanoid robots, drones and other machines whose shape makes a single integrated camera module impractical. A Developer Kit version ships in November at US$849; a lighter Barebone version built for custom enclosures ships in the first quarter of 2027 at US$729, both prices temporarily carrying a US$200 surcharge the company attributes to memory and storage costs.

Luxonis is a Denver, Colorado-based machine-vision hardware company founded in 2019 that built its reputation on the OAK line, stereo-depth cameras built around the open-source OpenCV computer-vision library, and raised a US$14 million Series A round in July 2026 explicitly framed around powering what the company calls "the perception layer behind physical AI." The OAK 4 FFC is the clearest expression yet of that framing: rather than selling a self-contained camera a robot builder bolts on in one place, Luxonis is selling a vision brain that a robot's distributed cameras all plug into.

Why Robots Run Out of Camera Ports

A wheeled delivery robot or a fixed warehouse camera can usually get by with one or two cameras mounted in an obvious, unobstructed spot. A humanoid robot or a small drone cannot. A humanoid needs vision at roughly head height for navigation, in its wrists or fingertips for manipulation, and often at its feet or hips for balance and obstacle detection, and each of those locations has different space, weight and cabling constraints that a single integrated camera module, which bundles its own lens, sensor and compute in one housing, cannot satisfy everywhere at once. Luxonis's answer is to separate the sensing from the compute entirely: the camera modules themselves become small, cheap, interchangeable sensors, starting at US$34.99 each, while the OAK 4 FFC board, mounted wherever convenient, does the actual synchronization, calibration, stereo-depth processing and AI inference for all five feeds at once.

The module lineup at launch spans three sensor types across nine individual products: a 48-megapixel color sensor, and two global-shutter sensors at one and five megapixels, each available in standard and wide fields of view. Luxonis says thermal, time-of-flight and additional global-shutter modules are coming, which would extend the same one-board, many-camera architecture to applications like low-light navigation and precise distance measurement without requiring a second compute board.

The Calibration Problem Luxonis Says It Solved

Running five independent camera modules through one stereo-vision pipeline only works if those cameras stay precisely calibrated relative to each other, and vibration, thermal expansion and ordinary wear on a moving robot will knock a factory calibration out of alignment over time. Luxonis ships every module factory-calibrated, then layers on what it calls Dynamic Calibration, a system that re-aligns the cameras using the texture already visible in the scene rather than requiring an operator to hold up a physical checkerboard pattern, the conventional method for recalibrating stereo cameras.

That detail matters more to a robotics integrator than it might first appear. A checkerboard recalibration routine means pulling a robot out of service, finding a technician who knows the procedure, and running it in a controlled setting, which is a real operational cost for a fleet of machines deployed across multiple sites. A system that can quietly recalibrate itself using whatever a robot's cameras happen to be looking at removes a maintenance step that would otherwise scale with fleet size, provided the self-calibration proves as accurate in the field as a checkerboard-based routine, a claim Luxonis has not yet had independent third-party testing to confirm.

Two Ways to Use the Same Board

Luxonis designed the OAK 4 FFC to serve two different roles depending on how a robot builder wants to architect their system. It can run as the robot's main onboard computer, using Luxonis's own Linux-based operating system to handle navigation and control logic directly, which suits a simpler robot that does not need a separate, more powerful general-purpose computer. Or it can run as a dedicated perception co-processor, handling all camera synchronization, depth computation and object detection, then sending only the processed results, depth maps, detections and tracked objects, to a separate main computer running the robot's higher-level autonomy stack.

That second mode is the more significant option for established robotics builders, because it means adding five-camera perception to an existing robot design without replacing the computer that already runs navigation, planning and manipulation control. A company with a working robot architecture can bolt on expanded vision as an add-on perception layer rather than re-architecting its entire compute stack around a new main processor, a meaningfully lower-risk integration path than most "replace your robot's brain" vision-hardware pitches require.

Connectivity Built for a Robot, Not a Desktop

The board's connectivity options read like a checklist built from robotics integration complaints rather than desktop-computing convention: 2.5 gigabit Ethernet and Power over Ethernet Plus for a single-cable data-and-power run, a wide 6-to-32-volt DC input that tolerates the voltage swings common on battery-powered mobile robots, USB-C, CAN FD for the vehicle-style networking many industrial robots already use internally, and hardware frame sync to keep all five cameras capturing at precisely the same instant, which matters for any application fusing multiple camera views into a single 3D reconstruction.

Luxonis is also releasing the board's reference design as open source, a decision that invites custom hardware variants and board-level modifications from robotics manufacturers building their own products around the chip, rather than locking integrators into Luxonis's own Developer Kit or Barebone form factors as the only options.

What the Software Stack Determines

A vision-compute board is only as useful as the software layer that turns raw camera frames into something a robot's control software can act on, and this is where Luxonis is leaning on an existing asset rather than building from zero. DepthAI, the company's open-source perception software, already runs on its earlier OAK camera generations and has accumulated several years of community use, example pipelines and integrations with common robotics middleware. Extending that same software base to a five-camera board rather than introducing a new, unproven stack alongside new hardware lowers the risk a robotics team takes on when adopting the OAK 4 FFC, since engineers evaluating the board can test much of its claimed functionality against DepthAI's existing documentation and community forum before committing engineering time to a full integration.

The six-core RVC4 processor at the center of the board is doing several jobs simultaneously that a less specialized chip would need separate silicon for: synchronizing five camera feeds to a common clock, computing stereo depth across whichever pairs of sensors a given robot design uses, and running the AI inference models that turn raw depth and image data into object detections and tracked identities. Collapsing all three functions onto one chip is what lets the board stay small and power-efficient enough to mount directly on a robot rather than requiring a separate server-class computer nearby, which is the design constraint that most distinguishes a robotics vision processor from a data-center AI accelerator repurposed for the same math.

Budgeting for a Pilot Fleet, Not Just a Prototype

The pricing structure says as much about who Luxonis expects to buy the OAK 4 FFC as the specs do. A single Developer Kit at US$849 is cheap enough for an individual engineer or a small robotics startup to buy on a procurement card for evaluation, without the multi-week enterprise sales cycle that typically accompanies higher-end vision and compute hardware. But the real test of the pricing comes at fleet scale: a robotics company outfitting ten or twenty prototype units with the Barebone version at US$729 each, plus five camera modules per unit at a minimum of roughly US$175 before accounting for the pricier sensor options, is looking at a vision bill in the tens of thousands of dollars before any of the surrounding robot hardware, software integration labor or the robot's own chassis and actuators are counted. Whether that line item is reasonable depends entirely on what budget a given robotics program has already allocated to perception, and Luxonis's temporary US$200 memory and storage surcharge on both board variants is a reminder that components pricing, not just finished-board pricing, can move the total faster than a published spec sheet suggests.

Pricing Against a Crowded Field

The OAK 4 FFC is entering a vision-compute market that already includes Nvidia's Jetson line and Intel's RealSense cameras, both widely used in robotics, alongside a growing set of smaller specialist vision-processor startups chasing the same physical-AI buildout. Luxonis's pitch is not raw compute power, Nvidia's larger Jetson modules outpace the OAK 4 FFC's 52 TOPS on paper, but the combination of multi-camera flexibility, self-calibration and robotics-specific connectivity in a single board, sold at a price a small robotics team can put on a procurement card rather than negotiate through an enterprise sales process.

Whether that combination holds up outside Luxonis's own demonstrations will depend on developers who buy the Developer Kit in November and try to build real multi-camera perception pipelines on top of it, the group whose feedback over the next few months will say more about the OAK 4 FFC's actual reliability than any spec sheet Luxonis has published at launch.

This account draws on Luxonis's own product and pricing materials for the OAK 4 FFC.

This article is provided for general information purposes only and does not constitute investment, procurement, legal, or engineering advice. Product specifications reflect public company information available at the time of writing and may change without notice.