One robot, many cameras, one time base
A modern robot does not see with one webcam. The head watches the aisle and the human. The wrist watches the approach. A teaching gripper (UMI-style or similar) watches the grasp so human demonstrations transfer to the arm. Each viewpoint is a camera module. If those modules do not share an exposure time base, the stack is not a multi-camera robot. It is several unrelated video streams that software tries to glue together after the fact.
Camemake builds that stack as a manufacturer: OEM and ODM camera modules and CameVision multi-cam boards with in-house SMT, cleanroom optics, active alignment (AA), factory extrinsics, and serial tracking. Est. 2016, 20,000+ m² factory, about 30 million units per month capacity on public Camemake scale facts. Belgium engineering support, China production, Hong Kong HQ. Interfaces include USB, MIPI CSI-2, DVP/SPI, and GMSL when the dress pack leaves the short FPC zone.
The buyer question is not megapixels per site. It is whether head, wrist, and gripper exposures share one trigger domain so SLAM, teleop, and imitation learning see the same instant of the world.
Global shutter is the motion baseline
Rolling shutter shears edges when the neck turns, the arm swings, or the gripper snaps shut. Marker corners bend. Stereo disparity shifts mid-frame. Grasp datasets train on geometry that never existed in the cell. Global shutter integrates the whole frame in one window so fast motion stays geometrically clean.
That is why robotics modules for closed-loop work prefer global shutter at every site that moves: head stereo or context cams, wrist cams, and gripper spine cams. Rolling shutter remains a conscious exception with an ISP warp plan, not the default for a moving manipulator.
Sync is not a software afterthought
Software timestamps help logging. They do not replace a shared trigger. When the head fires a few milliseconds before the wrist, depth from stereo and pose from the gripper disagree. Teleoperators feel lag as mush. Teaching pipelines label the same grasp as two different scenes. Multi-view fusion then spends compute fighting time error instead of estimating pose.
Hardware sync means a shared FSYNC / PPS / trigger line (or equivalent multi-CSI time base) so every global-shutter module in the set opens and closes together. IMU samples that feed VIO sit in the camera frame they describe, with a known lever arm, not "somewhere on the chassis." CameVision double and higher boards on Camemake's robotic line are built around that idea: stereo and array cameras share one timed domain; arrays ship with factory extrinsics as one unit.
For long dress packs, serializers (GMSL / FPD-Link class) keep the same trigger discipline when raw CSI cannot run the full cable length. The sync story does not stop at the connector on the host.
Head, wrist, gripper: three jobs, one clock
Site | Job | What breaks without sync
Head | Navigation, HRI, teleop overview, stereo / multi-view SLAM | Neck motion plus delayed side cameras produce inconsistent maps and delayed collision cues
Wrist | Approach, tool alignment, close-range obstacle volume | Wrist frames that lag the head make hand-eye transforms jitter under motion
Gripper / UMI-style teacher | Grasp view for demonstration and transfer | Grasp frames that do not match head/wrist time poison imitation datasets
Pure-vision fleets that skip depth sensors still need this discipline. Adding a depth assist later does not fix a head and gripper that never shared an exposure. Depth must join the same time base or fusion invents ghost geometry.
What a module manufacturer must ship with the set
Camemake treats multi-site robotics as a module program, not a bag of independent UVC sticks:
- Shutter class locked per site (global shutter for motion sites unless the brief says otherwise).
- 2. Shared trigger plan across head, wrist, and gripper modules (GPIO / FSYNC / PPS language in the traveler).
- 3. Interface path (MIPI CSI-2 into embedded edge-compute hosts, USB for evaluation, GMSL when the run is long).
- 4. AA and calibration pack per serial, plus array extrinsics when boards share one mechanical datum.
- 5. Volume path from evaluation MOQs into production lots on one factory so field swaps stay in the same evidence class.
CameVision ODM covers single through 9-camera boards and finished ODM goods (headset, cap, helmet, backpack array, handheld teaching gripper). When the faceplate, baseline, or wrist envelope is not in that family, Camemake runs full OEM: PCB, optics, mechanics, firmware, then the same lines.
Exact sensor part numbers, FPS, and baselines stay quote-specific. This article does not name chip brands. The product is the sync architecture and the factory that can reproduce it.
Why this belongs on a robotics camera program
Teams that buy cameras as afterthought accessories discover sync in the field: sheared markers, mismatched teleop views, teaching logs that will not replay. Teams that brief sync first get modules designed as a set: same shutter class, shared trigger, calibration files that travel with the serial, and a manufacturer that can scale the set.
Send Camemake the robotics brief: camera roles (head / wrist / gripper), motion budget, host compute, cable length, and whether teaching (UMI-style) is in scope. We map it to a CameVision multi-cam ODM layout or a full OEM design, then move from AA samples into serial-tracked volume.
CameVision EGO (dual global-shutter AI stereo): https://www.camemake.eu/shop/cv-ego01-os-camevision-ego-dual-global-shutter-ai-stereo-camera-2414
CameVision robotic cameras: https://www.camemake.eu/robotic-solutions
MIPI CSI-2 OEM modules: https://www.camemake.eu/mipi-camera-modules
Services and certification path: https://www.camemake.eu/services-capabilities
Inside the Camemake factory: https://www.camemake.eu/inside-the-camemake-factory