L
Object localization
Vision measures where a part is in the workspace so the robot can approach or grasp without fixed fixtures for every pose.
Robotics Perception Engineering Guide
Machine Vision for Robotics: Cameras, Optics, Lighting and System Selection
An engineering guide to robot vision systems, cameras, lenses, illumination, calibration and vision-guided motion
Machine vision gives robots visual information about objects, surfaces and workspace geometry.
This guide explains how cameras, optics, lighting, calibration and processing interact as a system, and what engineers should evaluate before selecting a vision architecture for robotic applications.
What Is Machine Vision?
Machine vision is an engineered sensing system that captures images of a scene and extracts usable measurements or decisions for an industrial or robotic process.
In robotics, machine vision typically combines a camera, lens, lighting, acquisition interface, processing hardware or software, and calibration so that visual information can be transformed into object location, pose, identity, inspection results or guidance commands for the robot.
A camera alone is not a complete machine-vision system. Optics, illumination, timing, geometric calibration and the connection to the robot controller determine whether the image becomes actionable robot data.
In summary
- The camera and sensor capture light from the scene.
- The lens forms an optical image with a defined field of view and working distance.
- Lighting controls contrast, glare and feature visibility.
- Acquisition electronics transfer image data under defined timing.
- Software extracts measurements, detections or poses.
- Calibration relates image coordinates to robot or world coordinates.
- The robot controller uses the result to plan or correct motion.
How Does Machine Vision Work in Robotics?
- A trigger or continuous acquisition starts image capture.
- Lighting illuminates the object or workspace under controlled conditions.
- The lens projects the scene onto the image sensor.
- The camera digitizes the optical image and transfers it through a vision interface.
- Processing software detects features, measures dimensions, estimates pose or classifies objects.
- Calibration transforms image results into robot, tool or world coordinates.
- The robot controller or cell PLC receives the perception output.
- The robot executes guided motion, inspection response or process correction.
Vision-to-motion chain
- Camera + optics + lighting
- Image acquisition
- Calibration
- Perception output
- Robot controller
- Physical motion
Robot vision is therefore a system-integration problem involving optics, timing, geometry, software and robot control—not only a camera-purchasing decision.
Machine Vision vs Computer Vision vs Robot Vision
| Term | Primary meaning | Typical focus |
|---|---|---|
| Machine vision | Engineered industrial imaging system for measurement, inspection or guidance | Hardware, optics, lighting, timing, repeatability and integration |
| Computer vision | Algorithms and models that extract information from images or video | Detection, recognition, reconstruction and learning methods |
| Robot vision | Vision used to support robot perception, localization or guided motion | Camera-to-robot geometry, latency, hand-eye calibration and control interface |
The terms overlap. A robotic system may use computer-vision algorithms inside a machine-vision architecture to produce robot-vision results. This guide focuses on the complete vision system used with robots, not on generic AI model tutorials.
What Components Make Up a Machine Vision System?
| Component | Function | Important parameters |
|---|---|---|
| Camera | Captures the optical image | Sensor size, resolution, shutter type, frame rate, interface |
| Image sensor | Converts light into electrical signals | Pixel size, quantum efficiency, noise, global or rolling shutter |
| Lens | Forms the image and sets geometry | Focal length, aperture, working distance, field of view, distortion |
| Lighting | Controls contrast and feature visibility | Geometry, wavelength, intensity, strobe, uniformity |
| Triggering | Starts capture at the correct instant | Hardware trigger, encoder sync, exposure timing |
| Processing hardware | Runs acquisition and algorithms | CPU/GPU/FPGA, memory, I/O, thermal limits |
| Vision software | Extracts measurements and decisions | Tools, SDK, models, latency, licensing |
| Calibration | Relates image data to physical coordinates | Intrinsic, extrinsic and hand-eye transforms |
Related sensing outside cameras—encoders, force sensors, LiDAR and IMUs—is covered in the robot sensors guide. How perception results enter the motion stack is covered in robot controllers.
Why Do Robots Use Machine Vision?
G
Motion guidance
Vision-guided robotics corrects robot trajectories based on measured part position, orientation or seam geometry.
I
Inspection
Robots and vision systems can inspect dimensions, presence, defects or assembly state during or after process steps.
P
Pose estimation
2D or 3D vision can estimate object pose for pick-and-place, assembly and bin picking.
A
Adaptive processes
Vision allows welding, dispensing, grinding or palletizing paths to adapt to real part variation.
S
Scene awareness
Cameras can detect obstacles, empty bins, fixture states or human presence as part of a broader sensing architecture.
2D Vision vs 3D Vision for Robotics
| Dimension | 2D machine vision | 3D machine vision |
|---|---|---|
| Primary measurement | Image-plane features and 2D coordinates | Depth, surface geometry or 3D pose |
| Typical strength | High-speed inspection, presence, barcode, planar guidance | Height variation, unstructured piles, complex pose |
| Main limitation | Ambiguous height and out-of-plane pose | Higher cost, denser data, more calibration and compute |
| Common technologies | Area-scan cameras with controlled lighting | Stereo, structured light, time-of-flight, laser triangulation |
| Best fit | Stable presentation, known height, planar parts | Bin picking, deformable parts, unknown orientation, height-critical tasks |
2D vision is not obsolete. Many robotic cells remain 2D when part presentation is controlled and height is known. 3D vision becomes necessary when depth, surface shape or unconstrained pose determines grasp or path success.
What Is Vision-Guided Robotics?
Vision-guided robotics uses camera-derived measurements to locate parts or features and adjust robot motion accordingly. The vision system may run once before a pick, continuously during a process, or in a closed loop with the robot controller.
Typical applications
- Pick and place with variable part pose
- Bin picking from unstructured piles
- Assembly and insertion guidance
- Seam tracking for welding or dispensing
- Palletizing and depalletizing
- In-process inspection with corrective motion
Required system elements
- Stable image acquisition under process conditions
- Perception output in a known coordinate frame
- Hand-eye or camera-to-robot calibration
- Latency low enough for the motion cycle
- Defined failure behaviour when detection is uncertain
How a robot locates an object with vision depends on whether the camera is fixed, mounted on the robot, or combined with structured lighting or 3D sensors. In every case, the output must be expressed in coordinates the robot can execute.
What Machine Vision Camera Types Are Used in Robotics?
Industrial and robot-vision cameras are selected for interface, shutter behaviour, resolution, spectral response and mechanical packaging. Consumer cameras and CCTV products are usually unsuitable where timing, calibration stability and industrial lifecycle matter.
Common categories
- Area-scan industrial cameras
- Line-scan cameras
- Monochrome and colour cameras
- Near-infrared or specialised spectral cameras
- Stereo and structured-light 3D cameras
- Smart cameras with onboard processing
Selection drivers
- Required field of view and working distance
- Object speed and motion blur limits
- Colour versus monochrome contrast
- Interface bandwidth and cable length
- Mounting mass and vibration environment
provides an objective basis for characterizing industrial image sensors and cameras when comparable sensitivity, noise and dynamic-range data are required.
Global Shutter vs Rolling Shutter
| Shutter type | Behaviour | Robotics implication |
|---|---|---|
| Global shutter | All pixels expose at the same time | Preferred for moving robots, moving parts and accurate geometric measurement |
| Rolling shutter | Rows expose sequentially | Can distort fast-moving objects; often unsuitable for precision guidance without motion constraints |
Global shutter is commonly preferred in robotics because robot motion and part motion both create geometric distortion under rolling-shutter readout. Rolling shutter may still appear in lower-cost depth cameras or inspection systems where relative motion is controlled.
Area Scan vs Line Scan
| Camera type | How it captures | Best fit |
|---|---|---|
| Area scan | Captures a full 2D frame in one exposure | Most robot guidance, pick-and-place, inspection of discrete parts |
| Line scan | Captures one line at a time; builds an image through motion | Continuous webs, long products, high-resolution inspection with encoder sync |
Line-scan systems require precise motion synchronization. Area-scan cameras dominate most robot-arm vision cells unless the application is built around continuous product movement.
How Should a Machine Vision Lens Be Selected?
The lens determines field of view, working distance, depth of field, magnification and much of the optical resolution. Camera megapixels do not compensate for an unsuitable lens.
Approximate field of view: FOV ≈ (sensor dimension × working distance) / focal length
Use consistent units. Exact results depend on lens design, object-side geometry and whether the system is telecentric.
Key optics parameters
- Focal length and aperture
- Working distance and field of view
- Depth of field and resolution across the image
- Distortion, chief-ray angle and mounting standard
- Telecentricity where measurement geometry requires it
A telecentric lens keeps magnification nearly constant with object distance over a defined range, which helps dimensional measurement. It is not automatically required for every robot-guidance task.
Why Lighting Often Matters More Than Camera Resolution
Lighting controls whether the feature the algorithm needs is visible. Increasing camera resolution cannot recover contrast that illumination failed to create.
Common lighting geometries
- Ring light for compact local illumination
- Backlight for silhouettes and edges
- Dome light for reducing specular glare
- Bar or directional light for texture and height cues
- Structured light for 3D shape
- Strobe lighting for freezing motion
Engineering questions
- What feature must be made visible?
- Is the surface specular, matte or translucent?
- Must ambient light be rejected?
- Is strobe synchronized to exposure?
- Does lighting create heat or mechanical interference?
Reduce glare by changing illumination angle, using diffuse or polarized light, controlling wavelength, or switching to backlight or dome geometries rather than simply increasing exposure.
Intrinsic, Extrinsic and Hand-Eye Calibration
Calibration establishes the geometric relationship between image measurements and the physical world the robot occupies. Without it, detections remain in pixel space and cannot reliably drive robot motion.
Intrinsic calibration
Models the camera’s internal geometry: focal length, principal point and lens distortion. It explains how 3D rays project onto the image sensor.
Extrinsic calibration
Defines the camera’s pose relative to another coordinate frame such as a fixture, conveyor or world origin.
Hand-eye calibration
Relates the camera frame to the robot flange or tool frame so measured object poses can be converted into robot motion commands.
Installed validation
Confirms that calibration remains accurate after mounting, cable forces, temperature change and vibration in the real cell.
A factory camera calibration certificate does not replace hand-eye calibration after installation. Camera-to-robot geometry must be established and verified in the actual mounting configuration.
supports measurement-management requirements when calibration systems must demonstrate confidence in measurement validity and reliability.
Resolution, Accuracy and Repeatability
| Term | Meaning |
|---|---|
| Pixel resolution | Number of pixels on the sensor |
| Optical resolution | Smallest spatial detail the optics and sensor can usefully resolve |
| Measurement accuracy | Closeness of the vision result to a reference measurement |
| Repeatability | Variation when the same condition is measured repeatedly |
| Robot positioning accuracy | Ability of the robot to reach the commanded pose after vision guidance |
Pixel resolution ≠ optical resolution ≠ measurement accuracy ≠ robot positioning accuracy. Higher megapixels do not automatically improve machine-vision accuracy if optics, lighting, calibration, depth uncertainty or robot mechanics dominate the error chain.
Approximate spatial sampling can be estimated from field of view divided by pixel count along that axis, but useful accuracy also depends on edge quality, distortion correction, temperature drift and algorithm behaviour.
Frame Rate, Exposure, Bandwidth and Latency
| Parameter | Question answered |
|---|---|
| Frame rate | How many images can be acquired per second under the configured mode? |
| Exposure time | How long does the sensor collect light for each frame? |
| Bandwidth | How much image data can the interface and host sustain? |
| Latency | How old is the scene information when the robot uses it? |
| Jitter | How much does acquisition or result timing vary? |
Latency matters because the robot and the part may move while the image is acquired, transferred and processed. A high frame rate does not prove low end-to-end latency if software, network or host scheduling dominate delay.
Synchronize cameras with robots using hardware triggers, encoder pulses, timestamps and verified clock behaviour. Soft polling alone is often insufficient for high-speed guidance.
Camera Interfaces, Bandwidth and Synchronization
Common interfaces
- GigE Vision / industrial Ethernet
- USB3 Vision
- Camera Link
- CoaXPress
- Vendor-specific or embedded interfaces
Evaluation fields
- Cable length and industrial robustness
- Payload bandwidth and multi-camera load
- Hardware trigger and I/O support
- GenICam compatibility and SDK maturity
- Determinism under host and network load
GigE Vision and USB3 Vision are common industrial camera transport standards. GenICam provides a generic programming interface model across many cameras. Protocol support alone does not guarantee interoperability with every host, cable plant or timing configuration.
Can a Normal Camera Be Used for Robot Safety?
Not automatically. A standard machine-vision camera provides application data. A safety-rated vision system must form part of a defined safety-related control architecture and meet the diagnostic, fault-response and performance requirements for that application.
Standard vision camera ≠ safety-rated vision system. Navigation or guidance cameras should not be assumed to provide protective safety functions.
and address industrial-robot and robot-cell safety scopes. Sensor selection alone does not establish cell safety.
How Should a Machine Vision System Be Selected for a Robot?
Selection should start from the robot decision that depends on vision, then work through optics, lighting, timing, calibration and controller integration—not from camera resolution alone.
1
Define the robot decision
- Locate, inspect, guide, classify, track or verify?
- What happens if the vision result is uncertain?
2
Define the scene
- Part geometry, material, colour, reflectivity
- Presentation, fixtures, clutter and ambient light
3
Choose 2D or 3D
- Is depth or unconstrained pose required?
- Can presentation keep the problem planar?
4
Define geometry
- Working distance, field of view, depth of field
- Camera mounting: fixed, robot-mounted or hybrid
5
Select optics
- Focal length, distortion, telecentricity needs
- Resolution across the usable image area
6
Design lighting
- Contrast method, wavelength, strobe, glare control
- Mechanical clearance and heat
7
Select camera and shutter
- Global vs rolling shutter, mono vs colour
- Sensor size, interface and industrial packaging
8
Define timing
- Exposure, frame rate, trigger source
- End-to-end latency and jitter budget
9
Define calibration
- Intrinsic, extrinsic and hand-eye requirements
- Recalibration interval and validation method
10
Define processing
- On-camera, edge PC, GPU or robot-controller host
- Algorithm class, model lifecycle and licensing
11
Define robot interface
- Coordinate frames, units, pose format
- Communication to controller, PLC or ROS
12
Validate in the cell
- Real parts, real lighting, real motion and vibration
- False-positive and false-negative behaviour
How Should a Robot Vision System Be Tested?
Optical and imaging
- Field of view, focus, distortion, uniformity
- Contrast under representative lighting
- Motion blur at production speeds
Measurement performance
- Accuracy against reference artefacts
- Repeatability across cycles
- Depth accuracy for 3D systems
Timing
- Trigger latency, exposure consistency
- End-to-end result latency and jitter
- Behaviour under multi-camera load
Calibration
- Intrinsic residual, hand-eye residual
- Temperature and remounting sensitivity
- Recalibration repeatability
Integration
- Robot pose conversion and units
- Communication loss and recovery
- Controller handshake and fault handling
Application robustness
- Part variation, dirty surfaces, ambient light
- False detects, misses and confidence thresholds
- Long-duration production behaviour
Every performance claim should identify camera and lens model, lighting configuration, working distance, exposure, trigger method, calibration status, processing settings, reference method and environmental conditions.
Common Mistakes When Selecting Robot Vision Systems
| Mistake | Likely consequence |
|---|---|
| Selecting by megapixels alone | Optics or lighting remain inadequate |
| Ignoring lighting design | Unstable detection and glare |
| Using rolling shutter for fast motion | Geometric distortion and guidance error |
| Treating pixel size as measurement accuracy | Overstated system capability |
| Skipping hand-eye calibration | Robot cannot use vision poses reliably |
| Ignoring end-to-end latency | Guidance lags moving parts |
| Choosing 3D when presentation can remain 2D | Unnecessary cost and complexity |
| Choosing 2D when depth is required | Grasp and path failures |
| Assuming protocol name equals interoperability | Integration delays |
| Using consumer cameras in industrial cells | Timing, lifecycle and support risk |
| Treating a guidance camera as a safety device | Inadequate safety architecture |
| Validating only with ideal demo parts | Field performance collapses |
| Ignoring cable, mount and vibration effects | Calibration drift and intermittent faults |
| Accepting accuracy claims without test conditions | Invalid supplier comparison |
What Evidence Should a Machine-Vision Supplier Provide?
Product definition
- Camera, sensor, lens and lighting specifications
- Mechanical drawings, connectors and pinouts
- Supported interfaces and trigger options
Imaging evidence
- Sensitivity, noise and dynamic-range data where available
- Frame-rate and exposure envelopes
- EMVA 1288 or equivalent characterization where claimed
3D and measurement evidence
- Depth range, accuracy and repeatability conditions
- Working distance and field-of-view maps
- Reference artefacts and test methods
Software and calibration evidence
- SDK, API, GenICam support and ROS drivers
- Calibration tools, hand-eye workflow and residuals
- Firmware ownership and update policy
Manufacturing and lifecycle evidence
- Sensor and optics sourcing, assembly and burn-in
- End-of-line testing and traceability
- Repair, recalibration, spare parts and end-of-life notice
Confirmed through primary documentation
Supplier-reported
Observed or independently tested
Not confirmed
Not disclosed
Sourcing Machine Vision for Robotics in China
China has substantial manufacturing capability across industrial cameras, lenses, LED lighting, structured-light projectors, 3D cameras, vision controllers, cables and integrated robot-vision packages.
The central sourcing question is not whether a supplier can ship a camera with an Ethernet port. The buyer must establish who owns the imaging design, which sensors and optics come from third parties, how calibration is performed, which interface and SDK stacks are controlled in-house, and how long firmware and spare-parts support will last.
Classify suppliers as camera technology owners, sensor-module manufacturers, optics manufacturers, lighting designers, 3D-vision system companies, vision-controller or software providers, robot OEMs with packaged vision, private-label assemblers, distributors or trading companies.
Evaluate legal identity, camera OEM versus reseller status, image-sensor source, lens and optics ownership, lighting design capability, calibration equipment and procedures, SDK and firmware ownership, GigE Vision / USB3 Vision / GenICam compliance evidence, multi-camera timing tests, English documentation, overseas support and lifecycle policy.
A supplier may manufacture housings and boards while depending on imported sensors, third-party lenses or external calibration software. The technology and responsibility map must be established below the catalogue-product level.
What Should Be Included in a Machine-Vision RFQ?
Robot and application context
- Robot category, application and vision function
- Fixed camera, robot-mounted camera or hybrid
- Development stage, destination market, pilot and annual volume
Scene and measurement
- Part materials, sizes, pose variation and presentation
- 2D or 3D requirement, accuracy and repeatability targets
- Working distance, field of view and depth of field
Camera, optics and lighting
- Resolution class, shutter type, mono or colour
- Lens constraints, telecentricity, distortion limits
- Lighting geometry, wavelength, strobe and ambient rejection
Timing, interface and software
- Frame rate, exposure, latency budget and trigger method
- GigE Vision, USB3 Vision, CoaXPress or other interface
- SDK, GenICam, ROS support, processing host and licensing
Calibration, safety and lifecycle
- Intrinsic, extrinsic and hand-eye calibration requirements
- Safety-related use or explicit non-safety application scope
- Repair, recalibration, firmware support and end-of-life policy
Required evidence
- Datasheets, optical drawings and lighting configuration
- Accuracy, repeatability and latency test conditions
- Calibration residuals, interface compliance and lifecycle policy
Frequently Asked Questions
What is machine vision?
Machine vision is an engineered imaging system that captures scenes and extracts measurements or decisions for industrial or robotic processes. It typically includes camera, lens, lighting, acquisition, processing and calibration—not only a camera. See the definition section.
What is machine vision in robotics?
In robotics, machine vision provides visual information for localization, guidance, inspection or scene awareness, then transforms that information into coordinates or decisions the robot controller can use. See how machine vision works in robotics.
How does machine vision work?
Lighting illuminates the scene, the lens forms an image on the sensor, the camera digitizes and transfers the image, software extracts features or poses, calibration converts results into robot coordinates, and the controller executes motion or process actions. See the eight-step sequence.
What is a robot vision system?
A robot vision system is a machine-vision architecture integrated with a robot so that image-derived measurements guide or verify robot behaviour. It includes optics, timing, calibration and controller interfaces as well as cameras. See vision system components.
Why do robots use machine vision?
Robots use vision to locate parts, guide motion, inspect quality, estimate pose and adapt processes when fixtures alone cannot control every variation. See six engineering reasons.
What are the components of a machine vision system?
Typical components include camera, image sensor, lens, lighting, triggering, processing hardware, vision software and calibration. See the component table.
What is the difference between machine vision and computer vision?
Machine vision emphasises the engineered industrial system—optics, lighting, timing and integration. Computer vision emphasises algorithms that extract information from images. Robot vision is the application of vision to robot perception and guided motion. See the comparison.
What is the difference between 2D and 3D machine vision?
2D vision measures features in the image plane and is strong when height is known. 3D vision measures depth or surface geometry and is preferred for unstructured pose, bin picking and height-critical tasks. See 2D vs 3D.
How do robots locate objects with vision?
The vision system detects object features or estimates pose, then uses calibration to express that pose in robot coordinates. The controller commands motion to the measured location. See vision-guided robotics and hand-eye calibration.
What is hand-eye calibration?
Hand-eye calibration establishes the geometric relationship between the camera frame and the robot flange or tool frame so measured object poses can be converted into executable robot commands. See calibration section.
Why is lighting important in machine vision?
Lighting creates the contrast the algorithm needs. Higher camera resolution cannot recover features that illumination failed to make visible. See lighting for machine vision.
Does higher resolution improve machine vision accuracy?
Not automatically. Pixel resolution is only one factor. Optical resolution, lighting, calibration, depth uncertainty and robot mechanics often dominate the final error. See resolution vs accuracy.
Global shutter vs rolling shutter—which should robots use?
Global shutter is generally preferred when robots or parts move during exposure because all pixels capture the scene at the same instant. Rolling shutter can distort fast motion. See shutter comparison.
How do you synchronize a camera with a robot?
Use hardware triggers, encoder pulses, verified timestamps and a defined latency budget. Soft polling alone is often insufficient for high-speed guidance. See timing and latency and interfaces.
Can a normal camera be used for robot safety?
Not automatically. Safety-rated vision systems must meet defined safety performance and fault-response requirements as part of the complete safety architecture. See safety distinction.
How do I select a machine vision system for a robot?
Start from the robot decision that depends on vision, then define scene, 2D/3D need, optics, lighting, camera, timing, calibration, processing and controller interface. Validate with real parts and motion. See the 12-stage selection framework.
How do I source machine vision from China?
Define the complete vision requirement first, then establish which sensors, optics, lighting, calibration methods and software the supplier owns versus what comes from third parties. See the China sourcing section and component sourcing service.
Need Help Sourcing Machine Vision for Robots?
If you have defined the perception task, optics envelope, lighting constraints and robot interface requirements, Yana can help structure supplier research, vision-system comparison and validation planning for robotics components in China.