Luxonis unveiled OAK 4 FFC on October 27 — a vision board that takes up to five cameras in any layout, with synchronization, stereo depth and AI processing handled on board. The developer kit costs $849 and is shipping now.
Key takeaways
- Up to five camera modules on a 70 × 53 mm carrier board
- $849 developer kit available immediately, a $730 barebone version planned for Q1 2027
- RVC4-based OAK 4 compute module: six-core ARM CPU, 8 GB RAM, 128 GB storage
- Dynamic Calibration aligns cameras from scene features, without a checkerboard or ChArUco target
- Camera modules sold separately — 48 MP IMX586, 5 MP OG05 and 1 MP OV9282
Why would a humanoid need five cameras?
A builder is assembling a humanoid head. It needs a forward-facing pair to compute distance from stereo, plus wide-angle cameras on each side and one at the back so the robot can see who is approaching.
Until now that meant assembling the perception system from separate parts: cameras, a unit synchronizing their frames and a computer calculating depth. Each bought separately, each with its own power and its own driver.
OAK 4 FFC folds that into a single 70 × 53 mm board. Five cameras plug in over flexible cables, and synchronization, depth and computer vision run on board.
But a question follows: if the cameras do not sit in a factory enclosure, how does the board know how they are positioned relative to each other?
That is not a detail. Whether the distance measurement means anything at all depends on it.
How does target-free calibration work?
Stereo depth derives distance from the difference between two camera images. Shifting a mount by a fraction of a degree breaks the geometry that calculation rests on — and the measurement starts lying.
The classic answer is target-based calibration: hold a checkerboard?Calibration target: A flat board of known geometry — a checkerboard or a ChArUco pattern — from which the algorithm derives the cameras' relative position and orientation. of known dimensions in front of the cameras and derive the lenses' relative positions from its image. Luxonis ships factory-calibrated modules and establishes their relationship inside a given rig through Dynamic Calibration.
The builder enters approximate camera positions in the OAK Viewer tool, then runs calibration against features visible in the scene. The documentation also covers monitoring alignment and recalibrating after thermal changes, mechanical shocks or a remounted camera.
The diagram shows where the risk sits. The builder performs the first two steps, the third depends entirely on the environment — and that is what decides whether target-free calibration works at all.
The method carries conditions. Luxonis recommends textured objects at varied depths and warns against dark scenes, blank walls, reflective and transparent surfaces and repetitive patterns. Dropping the target simplifies one setup step but does not remove the duty to check that the environment supplies enough information.
What do these names mean?
The board's spec sheet runs on abbreviations that decide what the hardware is good for.
- FFC — the flat flexible cable linking a camera module to the board. It lets the camera sit where the design requires rather than where a rigid enclosure allows.
- Global shutter — a sensor exposing the whole frame at once. Under motion it avoids the skew of sensors read line by line, which matters on a walking robot.
- Stereo baseline — the distance between two cameras facing the same way. Wider means better accuracy at range, worse up close.
- TOPS — trillions of operations per second, a measure of raw AI-accelerator throughput. It does not say how many frames per second a given model will process.
- PoE+ — power over the network cable. One wire instead of two, which simplifies wiring inside a cramped chassis.
Together these explain why the board ships without cameras: picking the sensor, lens and spacing is a design decision, not a matter of taste.
The three supported camera modules, sold separately:
Why not just use a finished stereo camera?
A simpler route suggests itself: buy a sealed stereo camera with a fixed baseline and skip calibration entirely.
The problem is that in a humanoid, camera placement is dictated by the structure, not by a catalogue. Head width sets the stereo baseline. A rear-facing camera has to be at the back. The side wide-angles have to fit whatever space the neck actuators left.
A finished enclosure imposes one geometry for every application. With five cameras in different roles — two for depth, two for field of view, one looking back — a single rigid geometry simply does not fit.
The second reason is computational. Running depth, detections and tracking on the board offloads the robot's main computer. That is an integration option, not a claim the board replaces the compute every humanoid AI model needs — the company positions it as the main computer for simpler or stationary machines, or as a dedicated perception module working alongside a Jetson or another host. Either way it shifts part of the work to edge AI.
Is it live yet?
Yes, but only partly. The $849 developer kit is available immediately. The $730 barebone version and volume availability are planned for Q1 2027. Both require separately purchased camera modules.
Specs for the RVC4-based OAK 4 compute module: a six-core ARM CPU, the Linux-based Luxonis OS, 8 GB of RAM and 128 GB of storage. The company markets it at 52 TOPS, while the detailed specification splits that into a 48-TOPS INT8 AI accelerator and a 4-TOPS FP16 GPU — figures that should not be read as a benchmark for a particular robot workload.
Connectivity covers 2.5 GbE, USB-C and CAN, with a 6–32 V DC input and optional PoE+. The barebone assembly weighs 36 grams, the developer kit with passive heatsink, PoE+ module and connectors weighs 215 grams.
| Parameter | Value |
|---|---|
| Carrier board | 70 × 53 mm |
| Camera count | up to 5 modules |
| Compute module | RVC4-based OAK 4 |
| CPU | six-core ARM |
| AI accelerator | 48 TOPS (INT8) |
| GPU | 4 TOPS (FP16) |
| RAM / storage | 8 GB / 128 GB |
| OS | Luxonis OS (Linux) |
| Connectivity | 2.5 GbE, USB-C, CAN, optional PoE+ |
| Power input | 6–32 V DC |
| Weight | 36 g (barebone) / 215 g (developer kit) |
| Price | $849 now · $730 barebone in Q1 2027 |
Why does this matter?
Perception is the part of a humanoid hardest to buy whole, because it depends on the shape of the specific machine. Moving synchronization and calibration onto a finished board shortens a stage every team previously worked through alone. At 36 grams and a price in the hundreds of dollars, that changes the maths for smaller builders, not just large manufacturers.
What's next?
- The $730 barebone version and volume availability are announced for Q1 2027
- Camera modules remain a separate purchase — supported are the 48 MP IMX586, 5 MP OG05 and 1 MP OV9282
- Dynamic Calibration's effectiveness depends on the scene, so the real test will be low-contrast halls with repetitive patterns, which the manufacturer itself warns about
Sources
- Humanoids Daily — Luxonis Launches OAK 4 FFC to Put Five Robot Cameras on One Vision Computer
- Luxonis — OAK 4 FFC
- Luxonis Documentation — Dynamic Calibration
