Robotics Vision Camera: How to Choose the Right One
A robot arm reaches for a part on a moving conveyor and closes on empty air. The controller was right, the trajectory was right, and the part position it was given was wrong by eleven millimetres. The camera reported where the part looked like it was, not where it was.
That failure has a specific cause, and it is the reason shutter type is the first specification on any robotics vision camera datasheet you should read. This post covers what vision-guided robotics demands from a camera, why global shutter is a hard filter rather than a preference, how interface choice shapes your architecture, and which Vadzo modules fit each case.
What Vision-Guided Robotics Actually Requires from a Camera
Vision-guided robotics means the camera output drives motion. That is the whole difference from inspection.
In an inspection setup, a wrong frame produces a wrong verdict. In vision-guided robotics, a wrong frame produces a wrong movement, and a wrong movement has physical consequences for the part, the gripper, and the cell.
So a robot vision system carries requirements a general-purpose camera does not. Geometric accuracy under motion. Deterministic latency. Repeatability across thousands of cycles and across a temperature range.
A robotics vision camera is therefore selected on constraints, not on image quality alone. A frame that looks good and arrives 40 ms late is useless to a controller running a 100 Hz loop.
This is why a machine vision camera for robots is specified differently from a webcam or a general industrial camera. An embedded camera for robotics has to hold its accuracy while the scene, the arm, or the whole chassis is in motion.
2D Vision Guided Robotics vs 3D Vision Guided Robotics
2D vision-guided robotics works on a known plane. The part sits on a conveyor or a table at a known height, and the camera supplies X, Y, and rotation. Most pick-and-place vision system deployments are 2D, and 2D is where a single well-chosen camera does the entire job.
3D vision-guided robotics adds depth, and it is a different architecture rather than a different camera. Depth comes from stereo pairs, structured light, or time-of-flight, so a bin-picking camera setup means multiple synchronized cameras or a dedicated depth sensor, not a single module.
Choose 2D vision-guided robotics wherever your fixture can guarantee the plane, because it is faster to deploy, cheaper to calibrate, and lighter on compute. Most cells can guarantee it with a fixture change that costs far less than moving to 3D vision-guided robotics.
Robot Guidance Camera vs Robot Vision Inspection Camera
A robot guidance camera feeds coordinates to a motion controller. Timing is part of the specification, because a stale coordinate is a wrong coordinate.
A robot vision inspection camera feeds a pass or fail decision to a logger or a reject gate. It can afford more latency, and it usually wants more resolution, because defect size sets the pixel budget.
Many cells need both. A vision-guided robot arm may use a fast global shutter module for guidance and a high-resolution module for the robot vision inspection camera role, because one sensor rarely optimises for both.
Global Shutter vs Rolling Shutter: The First Filter in Any Robot Vision System
A global shutter camera for robotics exposes every pixel at the same instant. A rolling shutter sensor exposes row by row, top to bottom.
For a static scene, the difference does not show. For anything moving, rolling shutter records different rows at different moments, so a straight edge crossing the frame is captured as a slanted one.
That slant is not a cosmetic artifact. It is a measurement error, and your vision algorithm has no way to know the edge was ever straight.
Why Motion Skew Breaks Pick and Place Vision System Accuracy
Consider a pick-and-place vision system watching a conveyor. The part moves during the readout interval, so the top of the part is recorded at one position and the bottom at another.
Your algorithm finds a centroid that corresponds to no real instant. The offset scales with belt speed, which means the error changes whenever line speed changes, and a calibration that worked at one speed fails at another.
A global shutter camera for robotics removes the failure mode rather than compensating for it. Every row shares one exposure window, so the geometry in the frame is the geometry that existed at that moment.
The same logic applies to a moving camera. On a vision-guided robot arm with a wrist-mounted module, the camera is what moves, and rolling shutter skews the whole scene during arm motion.
Do Robots Need a Global Shutter Camera?
The honest answer to do robots need a global shutter camera is: not all of them, but every one that measures during motion does.
If your cell indexes and stops, and the camera triggers only while everything is stationary, rolling shutter is adequate and usually cheaper. That is a real design, and it is common in low-cycle-rate cells.
The moment you want to measure on the fly, or you mount the module on the arm, do robots need a global shutter camera stop being a question. Every camera Vadzo recommends for robotics uses a global shutter sensor: the Onsemi AR0234 and the Sony IMX900.
The trade-off is real. Global shutter pixels are more complex, so at equal resolution a global shutter sensor typically costs more and offers lower resolution than a rolling shutter equivalent. The Falcon-234MGS delivers 2.3MP at 120fps, and that resolution is a deliberate exchange for shutter accuracy and frame rate.
Interface Decides Your Architecture: USB, MIPI CSI-2 and GigE
Shutter type decides whether the measurement is valid. Interface decides whether it arrives in time and how far it can travel.
A USB camera for robotics is the fastest route to a working cell. UVC support means no driver development, and a USB camera for robotics connects to any industrial PC or SBC you already have. Cable length is the constraint.
A MIPI CSI-2 camera for robotics goes straight into an SoC. There is no protocol stack between sensor and processor, which is what makes a MIPI CSI-2 camera for robotics the low-overhead option for an embedded camera for robotics build. Cable runs are short, typically well under a metre, so an embedded camera for robotics on this interface has to sit close to the compute.
A GigE camera for robotics solves distance. Standard Ethernet cabling runs tens of metres, which is what you need when the controller cabinet is nowhere near the cell. A GigE camera for robotics also gives you existing network infrastructure and PoE options.
Low Latency Camera for Robotics: What Actually Adds Delay
A low latency camera for robotics is not just a fast sensor. Total latency is exposure, readout, transport, host buffering, and your algorithm, and any one of those can dominate.
Interface bandwidth matters because it sets readout transport time. The Vajra series runs USB 3.2 Gen 2×2, which is the widest USB link in Vadzo’s robotics lineup, and that headroom is what lets the Vajra-2020MRS move 20MP frames without starving the link. On USB 3.2 Gen 1, the Falcon-234MGS sustains 120fps at 2.3MP.
Frame rate and latency are related but not the same. A 120fps sensor gives you a fresh frame every 8.3 ms, which caps how stale your newest coordinate can be, and for a low-latency camera for robotics that ceiling is often the number that matters to the controller.
NVIDIA Jetson Camera Robotics Builds and the ROS Camera Driver
Most embedded robotics stacks today land on Jetson. An NVIDIA Jetson camera robotics design usually favours MIPI CSI-2, because the SoC has ISP hardware and CSI lanes ready to use.
The Bolt-234CGS is the MIPI CSI-2 module in this lineup, a 2MP global shutter camera at 1080p, positioned for robotics, AGV navigation and UAV imaging.
USB stays viable on the same platform. An NVIDIA Jetson camera robotics build using a UVC module skips driver work entirely, which matters when your schedule is tighter than your latency budget.
Whichever you pick, confirm ROS camera driver support early. A ROS camera driver that already publishes to the topics your stack expects saves weeks, and discovering there isn’t one after the mechanical design is frozen is an expensive way to learn.
Mono vs Color, Resolution and NIR: Matching the Sensor to the Task
Mono is the default in a robot vision system, and it surprises people.
A mono sensor has no colour filter array, so every pixel collects full-spectrum light and no demosaicing is needed. You get better sensitivity and true per-pixel detail, which is what edge detection, fiducial location and blob analysis actually consume.
Choose colour when colour is the feature. Sorting by colour, reading colour-coded labels, or verifying a painted finish all need it, and that is why a machine vision camera for robots is offered in both. Vadzo pairs most sensors this way, so you can move between a mono and a colour machine vision camera for robots without changing interface or form factor.
Resolution follows the feature, not the frame. Work out the smallest feature you must resolve and how many pixels across it your algorithm needs, then divide by field of view. Specifying more resolution than that costs you frame rate and bandwidth for nothing.
NIR matters when your lighting is not visible. The Vajra-900MGS and Falcon-900MGS use the Sony IMX900 with NIR response, which lets you light a cell with infrared illumination that human operators do not see, keeping the scene consistent regardless of ambient light.
HDR handles mixed lighting. The Vajra-2020MRS offers eHDR with dynamic ROI, and the Vajra-821CRS offers eHDR with dual endpoint streaming, which are the ones to look at when a cell has both a bright window and a shadowed fixture. The Vajra-900MGS adds Quad HDR on a global shutter sensor, which is the combination to check when the scene has mixed lighting and moves at the same time.
Calibration and Integration: Hand-Eye Calibration and What It Costs You
Buying the right module is half the job. The other half is telling the robot where the camera is.
Hand-eye calibration is the procedure that establishes the transform between the camera coordinate frame and the robot coordinate frame. Without it, the camera reports positions in its own frame and the robot has no way to act on them.
There are two mounting cases. Eye-in-hand puts the module on the wrist of a vision-guided robot arm, and the transform is fixed relative to the tool. Eye-to-hand mounts it in the cell, and the transform is fixed relative to the base.
Budget real time for hand-eye calibration in your schedule. It needs a target, a routine that moves the robot through many poses, and a validation pass, and it has to be repeated if the module is ever remounted.
Mechanical stability matters more than most teams expect. A camera bracket that flexes under vibration will quietly invalidate a hand-eye calibration that was correct on the day it was run.
The Vadzo Robotics Vision Camera Lineup
The Vajra series is Vadzo’s purpose-built robotics family. Everything else in this section supports a specific constraint the Vajra range does not cover.
Vajra series: USB 3.2 Gen 2×2, built for robotics
Camera | Sensor | MP | Color/Mono | Interface | Focus | Special Feature |
Sony IMX900 | 3.2MP | Mono | USB 3.2 Gen 2×2 | Fixed | Global Shutter, Quad HDR, NIR | |
Onsemi AR2020 | 20MP | Mono | USB 3.2 Gen 2×2 | Fixed | Dynamic ROI, eHDR, WOM | |
Onsemi AR0821 | 8.3MP | Color | USB 3.2 Gen 2×2 | Fixed | eHDR, Dual Endpoint Streaming |
Start here for a cobot vision camera or a guidance module on a fast cell. The Vajra-900MGS gives you a global shutter sensor with Quad HDR and NIR response on the widest USB link Vadzo offers, which covers demanding 2D vision-guided robotics work directly.
Falcon series: USB 3.2 Gen 1 global shutter
Camera | Sensor | MP | Color/Mono | Interface | Focus | Special Feature |
Sony IMX900 | 3.2MP | Mono | USB 3.2 Gen 1 | Fixed | Global Shutter, NIR | |
Onsemi AR0234 | 2.3MP | Color | USB 3.2 Gen 1 | Fixed | Global Shutter | |
Onsemi AR0234 | 2.3MP | Mono | USB 3.2 Gen 1 | Fixed | Global Shutter, 120fps, NIR |
Use the Falcon range when your host has USB 3.2 Gen 1 rather than Gen 2×2. Every module here is still a global shutter camera for robotics, so the measurement integrity is unchanged.
Bolt series: MIPI CSI-2
Camera | Sensor | MP | Color/Mono | Interface | Focus | Special Feature |
Onsemi AR0234 | 2MP | Color | MIPI CSI-2 | Fixed | Global Shutter, 1080p |
This is the MIPI CSI-2 camera for robotics, positioned for robotics, AGV navigation, and drone imaging. If you are building an AGV camera or an AMR camera into a Jetson-based chassis, the Bolt-234CGS connects directly to the SoC.
Innova series: GigE for long cable runs
Camera | Sensor | MP | Color/Mono | Interface | Focus | Special Feature |
Onsemi AR1335 | 13MP | Color | GigE | Auto | — | |
Sony IMX678 | 8.4MP | Color | GigE | Fixed | HDR, Starvis 2 | |
Onsemi AR0821 | 4K | Color | GigE | Fixed | HDR | |
Sony IMX900 | 3.2MP | Mono | GigE | Fixed | Global Shutter | |
Onsemi AR0234 | 2.3MP | Color | GigE | Fixed | Global Shutter | |
Onsemi AR0234 | 2MP | Mono | GigE | Fixed | Global Shutter |
A GigE camera for robotics is what you specify when the cell and the controller are far apart. The Innova-900MGS, Innova-234CGE, and Innova-234MGS keep the global shutter requirement intact over that distance, which matters for a warehouse robot camera network where a single controller may serve several cells.
Conclusion: A Three-Point Decision Framework
Work through these in order, and the shortlist writes itself.
One: does anything move while you measure? If yes, a global shutter camera for robotics is a requirement, not an upgrade. If your cell fully indexes and stops, rolling shutter is a legitimate saving.
Two: where does the frame have to travel? Straight into an SoC means MIPI CSI-2. Onto an industrial PC means USB. Across the factory floor means GigE.
Three: what is your latency ceiling? Divide your control loop rate into your cycle time and see what staleness the controller tolerates, then pick frame rate and interface bandwidth to sit under it.
If you have those three answers, you can specify a camera for vision-guided robotics from the tables above. If you want to confirm before committing, request an evaluation module and test it in your own cell. The Vajra, Falcon, Bolt, and Innova ranges are all on the Vadzo products page, and our engineering team can shortlist against your platform, cable run, and cycle time.
Frequently Asked Questions
What camera is used in robotics?
The short answer to what camera is used in robotics is a global shutter industrial camera matched to the robot’s interface and cycle time. A robotics vision camera differs from a general-purpose camera because its output drives motion, so geometric accuracy under movement and predictable latency matter more than image aesthetics. Mono sensors are common because they give better sensitivity and per-pixel detail for edge and fiducial detection, with colour used only when colour is the feature being measured. Interface follows the architecture: USB for industrial PCs, MIPI CSI-2 for SoCs, GigE for long cable runs. Vadzo’s Vajra series is purpose-built for this, with global shutter modules such as the Vajra-900MGS running on USB 3.2 Gen 2×2, alongside the Falcon-234MGS delivering 2.3MP at 120fps over USB 3.2 Gen 1.
What is vision-guided robotics?
Asking what is vision guided robotics is really asking what changes when a camera drives motion instead of reporting on it. In vision guided robotics the camera locates a part and passes coordinates to the robot controller, which then plans a trajectory to it, so a wrong frame becomes a wrong movement rather than a wrong verdict. 2D vision guided robotics supplies X, Y and rotation on a known plane and covers most pick and place vision system work, while 3D vision guided robotics adds depth for stacked or tilted parts. The camera requirements are stricter than for inspection: repeatable geometry under motion, deterministic latency and mechanical stability. Vadzo builds the Vajra series specifically for this class of work, with global shutter sensors across the range.
Do robots need a global shutter camera?
Not every robot, but every robot that measures while something is moving. Rolling shutter exposes the sensor row by row, so a part crossing the frame is recorded at different positions in different rows, and your algorithm computes a centroid that never existed. That error scales with speed, which means a calibration valid at one line rate fails at another. If your cell indexes and stops completely before triggering, rolling shutter is adequate and cheaper, and that is a legitimate design choice. Every camera Vadzo recommends for robotics uses a global shutter sensor — the Onsemi AR0234 and the Sony IMX900, because a robot guidance camera usually cannot make that guarantee.
How do robots see?
The question how do robots see has a simple mechanical answer and a harder practical one. A camera captures a frame, software locates features in it, and a calibrated transform converts those image coordinates into robot coordinates the controller can move to. The hard part is that transform, established through hand eye calibration, which ties the camera frame to either the robot tool frame or the robot base frame depending on how the module is mounted. Everything downstream depends on that calibration holding, which is why bracket rigidity and mounting repeatability matter as much as sensor specification. Vadzo’s fixed-focus global shutter modules suit both eye-in-hand and eye-to-hand mounting on a vision guided robot arm.
How to choose a camera for robotics?
Work how to choose a camera for robotics as three filters in order. First, shutter type: if anything moves during measurement, a global shutter camera for robotics is mandatory. Second, interface: MIPI CSI-2 for a direct SoC connection, USB for an industrial PC or SBC, GigE when the controller sits far from the cell. Third, sensor: mono for edge and fiducial work, colour when colour is the feature, and only as much resolution as your smallest feature actually requires. Vadzo covers all three axes with global shutter options on every interface, from the Vajra series on USB 3.2 Gen 2×2 to the Bolt-234CGS on MIPI CSI-2 and the Innova-234MGS on GigE.
