History of Machine Vision: From Slow Alignment to AI-Guided Robots

By A Niu | GIAI Photonics

About twenty years ago, I encountered machine vision for the first time. The equipment had to locate a small tab on a product. It was roughly the size of a fingernail, easy enough for a person to spot, yet the machine needed several seconds to find it. Even then, the result was often wrong.

Watching it work reminded me of someone trying to thread a fine embroidery needle. The thread moved back and forth, almost aligned with the eye, without quite finding its way through. Anyone watching wanted to help, while the machine continued searching at its own pace.

That was one machine, in one application. It did not represent the state of the whole industry. Nevertheless, for a first encounter, it left a strong impression.

Two decades later, machine vision can contribute to tasks ranging from defect inspection to guiding a robot through a bin of randomly oriented parts. How did it move from locating a small feature to supporting such complex operations? The answer involves several technologies developing together.

1. Early Machine Vision History: Turning Pictures into Data

Machine vision has no single universally accepted birthday. Digital image processing, pattern recognition, optics, computing and industrial automation developed along overlapping paths. Here, machine vision refers primarily to engineered visual systems used to inspect, measure, identify or guide a process; the research history also draws on the broader field of computer vision.

One useful early milestone is a photograph of a baby.

In 1957, Russell Kirsch and colleagues at the US National Bureau of Standards, the predecessor of today’s National Institute of Standards and Technology, scanned a photograph of Kirsch’s son into a computer. The image measured just 176 × 176 pixels. Its importance lay in making a picture available as numerical data that a computer could process. ^[1]

Receiving pixel values, however, does not mean understanding the subject. People readily separate a face from its background. A visual algorithm needs methods for finding relevant changes in intensity, grouping boundaries and relating image features to objects.

In 1963, Lawrence G. Roberts completed his MIT doctoral thesis, Machine Perception of Three-Dimensional Solids. His work on interpreting geometrically defined objects from images became an important early contribution to computer vision. Simple solids offered a more manageable setting for studying edges, surfaces and spatial relationships. ^[2]

MIT’s 1966 Summer Vision Project proposed dividing images into regions, describing those regions and identifying objects. These tasks remain recognizable parts of visual processing today, even though the available methods and computing resources have changed substantially. ^[3]

The challenge was already clear: a photograph had to become useful information about the physical world.

2. Moving from Research into Industrial Production

Factories gave visual systems a practical advantage: the task could be constrained.

A production system might need to check whether a component was missing, read a marked identifier or locate the center of a hole. Engineers could fix the camera position, control the lighting and present parts using fixtures. A clearly defined problem gave a visual system a better chance of producing consistent results.

Imaging hardware was developing alongside these applications. In 1969, Willard Boyle and George Smith invented the charge-coupled device, or CCD, at Bell Laboratories. The subsequent development of CCD image sensors provided an important foundation for electronic imaging. ^[4]

By 1982, Cognex had introduced its DataMan industrial optical character recognition system, designed to read letters, numbers and symbols on product surfaces. This is one documented example of industrial vision commercialization well before the 1990s. ^[5]

The same year saw the publication of David Marr’s Vision. Marr developed a computational framework for understanding how visual information could lead to representations of objects and their environment. The original book was published by W. H. Freeman in 1982; MIT Press republished it in 2010. ^[6]

Research and industrial engineering addressed related questions from different directions. Researchers investigated perception; manufacturers needed a dependable result within a production cycle.

Reading one identifier or detecting one assembly error could justify a system if it performed that task consistently. Industrial usefulness did not require solving every problem in visual understanding.

Selected Milestones in Machine Vision History

These are selected milestones, not a sequence in which each technology replaces the previous one.

3. Why a Small Locating Tab Could Still Be Difficult

Returning to the equipment I first encountered, its limitations make more sense in that wider context. By the mid-2000s, machine vision already had many established applications. Integrating the technology into a particular machine remained an engineering task with several interacting variables.

For a locating operation, the system usually needs a description of the target, a search region, permitted changes in orientation and acceptance criteria for a match.

Traditional approaches include intensity-based segmentation, edge extraction and template matching. Geometric matching methods use features such as contours and spatial relationships to improve tolerance to specified changes in rotation, scale or appearance. Cognex records the introduction of its PatMax geometric matching technology in 1997. ^[7]

These tools remain useful. A well-lit part with a clear outline and a stable background can be located quickly using an appropriate conventional method, with results that are relatively straightforward to interpret.

Difficulties arise when production conditions change. A thin oil film may alter reflections. A part may arrive at a different angle. Background structures may resemble the target. Normal variations between batches may change features that appeared stable during development.

The system can then fail to find the target or select the wrong feature. Expanding the search space or adding checks may increase computation, depending on the implementation.

There is not enough information to diagnose the particular machine in my recollection. Imaging, processing hardware, algorithm settings, communication and mechanical coordination could all have influenced its behavior.

Progress over the following years involved improvements throughout that chain. Interfaces also became more standardized. GigE Vision was formally introduced in 2006, while GenICam provided a generic programming interface across different camera interface technologies. These developments reduced parts of the integration burden and helped make components easier to combine. ^[8]

4. Deep Learning Changed How Visual Features Were Developed

Some inspection tasks are difficult to express as a short set of rules.

Scratches vary in shape, brightness and direction. Acceptable material texture can resemble damage. A partially printed character may remain readable to a person while failing a strict geometric comparison. Adding exceptions to a rule set can make maintenance increasingly complicated.

In 2012, AlexNet demonstrated a major advance in ImageNet image classification using a deep convolutional neural network and GPU computation. It was an influential milestone in the wider adoption of deep learning for vision, rather than the beginning of machine vision itself. ^[9]

Learning-based approaches gradually expanded the available industrial tools. Engineers could collect representative images, define labels and train models to learn task-relevant features. This changed part of the development process: some feature design could be learned from examples instead of being specified explicitly.

People still had to define the inspection task, establish reliable labels and decide what constituted an acceptable product.

Industrial inspection also presents a particular data problem. Good products may be plentiful, while examples of a rare defect can be difficult to obtain. Anomaly detection offers an alternative approach: learn the appearance of normal products, then identify departures from that appearance.

Introduced in 2019, the MVTec AD dataset became an important benchmark for industrial anomaly detection. Its training images are defect-free, while the test sets include both normal and anomalous examples. ^[10]

An unusual appearance is not automatically a reject condition. Acceptable texture variation may trigger a false alarm, and a consequential defect may produce only a weak visual change. Thresholds and model outputs therefore need to be evaluated against the actual quality criteria.

Practical systems can combine methods. Geometric matching can establish the inspection region, a learning-based model can analyze appearance, and dimensional or process rules can provide additional checks. The engineering objective determines the combination.

5. From 2D Localization to 3D Robot Guidance

Two-dimensional imaging can solve many tasks when parts lie on a known plane and arrive in a controlled orientation. A bin of overlapping, randomly oriented components introduces additional questions: how high is the target, how is it tilted, and where can a gripper approach it?

Three-dimensional vision provides spatial measurements to support those decisions. Its research history is much older than recent AI developments. Improved sensors, computing and integration have made a wider range of industrial 3D applications practical.

Different methods obtain depth information in different ways. Stereo vision uses correspondences between views. Structured light projects a known pattern and analyzes its appearance on the object. Laser triangulation measures surface profiles using the geometry between projected light and an observing camera. The resulting surface information is often represented as a point cloud: a set of points with spatial coordinates. ^[11]

A point cloud is an input to robot guidance. The system must still identify the target, estimate its pose, transform locations between coordinate systems, select a grasp and plan motion that avoids collisions. Calibration establishes the necessary relationship between the vision system and the robot.

Mechanical execution matters too. A correct pose estimate cannot compensate for an unsuitable gripper or an unstable grasp.

Optical 3D methods also have limits. Occluded surfaces cannot be measured directly from a view that does not see them. Transparent materials and highly reflective surfaces may produce incomplete or misleading depth information. A known CAD model can support inference about an object, but that inference should not be confused with a direct measurement of its hidden surfaces. ^[11]

When a robot successfully picks a part from a bin, the result reflects the operation of this whole chain, from acquiring useful measurements to executing the grasp.

6. What Machine Vision Can Do, and What Still Needs Validation

As of October 2026, machine vision encompasses established industrial tools and active research into broader visual understanding. A single claim about being a certain number of times faster cannot describe that range.

Under suitable conditions, systems can locate features, read codes, check assemblies, inspect surfaces and measure dimensions. Appropriate magnification, lens performance, lighting, sampling and calibration are required for observing and measuring small features. Camera pixel count alone does not establish measurement accuracy. ^[12]

Systems needing depth information can use 3D measurements for surface inspection, pose estimation and validated robot guidance. Two-dimensional appearance and 3D geometry can also be used together.

Research is extending the connection between images, language and actions. RT-2, published in 2023, is a representative vision-language-action model. It combines visual-language pretraining with robot data to generate actions from observations and instructions, exploring generalization beyond individually programmed tasks. ^[13]

Such results need to be distinguished from an industrial workcell that has met its production acceptance criteria. A research demonstration does not, by itself, establish the required cycle time, reliability or recovery behavior on a particular line.

Inspection remains sensitive to changes in operating conditions as well. The MVTec AD 2 research includes difficult industrial scenarios involving transparent or overlapping objects, very small defects and changing illumination. Its journal version was published in the International Journal of Computer Vision in 2026. ^[14]

Two distinctions are particularly useful when evaluating a system:

These distinctions make performance claims more useful to an engineer making an integration decision.

7. Why Better Algorithms Still Need Good Optics

The optical system determines which information reaches the algorithm.

If glare obscures a defect, or if the target has little contrast against its background, analysis becomes more difficult. Once an image region is saturated, the recorded data no longer preserve the original intensity differences in that region. A model may make an inference, but it cannot be assumed to recover a trustworthy measurement of the lost detail.

People often change their viewing angle naturally to find a feature. A fixed inspection camera generally relies on the imaging conditions designed for it. Lighting direction, wavelength, lens selection and exposure all influence the result.

Match spectral filtering to the illumination and detector

With controlled illumination, a suitable bandpass filter can attenuate unwanted wavelengths outside its passband while transmitting the useful band. Its effectiveness depends on the source spectrum, background spectrum, detector response and optical geometry. Background light within the passband may still reach the sensor. ^[15]

For some reflective surfaces, a coordinated polarizer and analyzer arrangement can reduce troublesome specular reflections. The improvement depends on the material, viewing geometry and polarization behavior, and must be balanced against the loss of useful light. ^[15]

GIAI’s guide to choosing an optical filter for machine vision discusses the specification inputs involved in matching a filter to an imaging system.

Evaluate the filter at its intended angles

For typical interference bandpass filters, increasing incidence angle from normal shifts the passband toward shorter wavelengths. A cone of incident rays can also alter the effective passband shape and transmission. The actual response depends on the coating design, angular distribution and polarization. A normal-incidence curve alone may not describe performance in the installed system. ^[15]

Coordinate exposure with motion

Global shutter operation helps avoid distortions associated with capturing different image rows at different times. It does not eliminate motion blur caused by an object moving during the exposure. Exposure duration, illumination and motion must still be coordinated. ^[12]

Improvements in algorithms therefore complement the work done in illumination and image formation. A well-designed acquisition stage makes the subsequent analysis more reliable.

8. What to Ask When Evaluating a Machine Vision System

When I visit a machine vision exhibition now, I still want to see faster cameras, clearer images and well-executed robot demonstrations. I also want to know what happens when the conditions change.

Does the inspection continue to work when a part rotates? What happens when reflections change? How much setup is needed for a new batch or product? How does the machine detect a failed grasp, and what does it do next?

These details help determine whether a demonstration can become a dependable production tool.

A useful evaluation starts with real samples and an agreed task definition. Include acceptable product variations as well as representative defects. Establish the required throughput, what counts as an error, and how the result will be verified against a reference or quality decision.

For robot guidance, test difficult orientations, occlusion, parts near bin walls and the final few parts remaining in the container. For optical inspection, assess changes in surface finish, illumination and environmental light within the intended operating conditions.

Optical component review at GIAI Photonics

GIAI Photonics supports optical filters, lenses, windows, optical coatings and other custom optical component projects. These components are reviewed against the drawing, sample, optical requirements, geometry and inspection criteria before a manufacturing route is defined.

For a machine vision optical project, provide the source wavelength or spectrum, detector information, intended angle of incidence, angular range, dimensions, clear aperture, coating requirements, operating conditions and expected quantity where available. The manufacturing capabilities page provides additional context, and drawings or specifications can be submitted through Contact GIAI Photonics.

I still remember the machine that needed several seconds to locate that small tab. The progress since then involves much more than a quicker answer. Visual systems now address a wider range of variation, provide more spatial information and connect more closely with machine actions.

On a production line, the practical question remains familiar: can this system keep making the right decision under the conditions that matter?

Frequently Asked Questions

When did machine vision begin?

There is no single agreed starting date. Digital imaging experiments in the 1950s and computer vision research in the 1960s provided important foundations. Industrial vision later developed into commercial systems for specific production tasks. The 1957 Kirsch image is a milestone in digitizing pictures, rather than the invention date of every form of machine vision.

How is machine vision different from computer vision?

Usage varies, but computer vision generally describes the broader methods used to extract information from images. Industrial machine vision combines such methods with cameras, optics, lighting, interfaces and control systems to perform a defined task. Consequently, a successful algorithm is only one part of a functioning production system.

Why can a conventional vision method still be the better choice?

A stable, well-defined task may be solved efficiently with geometric matching, edge tools or dimensional rules. Development effort, interpretability, processing requirements and maintenance all matter. The choice should be based on representative validation, including the expected variations, rather than on whether a product is described as AI-enabled.

What should an AI inspection accuracy claim include?

Ask how defects and acceptable variations were defined, whether evaluation images were independent of training data, and whether they represent production conditions. Request separate false-accept and false-reject results, the decision threshold, complete cycle time and performance after relevant batch or lighting changes. An aggregate score without this context is difficult to apply to a real line.

References

1. NIST. First Digital Image; Fiftieth Anniversary of First Digital Image Marked. Digital image history · Background on the baby photograph.

2. Roberts, L. G. Machine Perception of Three-Dimensional Solids. MIT doctoral thesis, 1963. MIT DSpace.

3. Papert, S. A. The Summer Vision Project. MIT AI Memo AIM-100, 1966. MIT DSpace.

4. Royal Swedish Academy of Sciences. The Nobel Prize in Physics 2009. CCD invention background.

5. Cognex. 2015 Annual Report Timeline. DataMan and early commercial systems.

6. Marr, D. Vision: A Computational Investigation into the Human Representation and Processing of Visual Information. W. H. Freeman, 1982; MIT Press reissue, 2010. MIT Press.

7. Cognex. Advances in Search and Location Algorithms. Geometric matching history.

8. AIA / A3. GigE Vision Interface Standard Premieres at Vision East, 2006; EMVA. GenICam. GigE Vision release · GenICam.

9. Krizhevsky, A., Sutskever, I., Hinton, G. E. ImageNet Classification with Deep Convolutional Neural Networks. NIPS, 2012. Paper.

10. Bergmann, P., Fauser, M., Sattlegger, D., Steger, C. MVTec AD — A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection. CVPR, 2019. Paper · Dataset.

11. LMI Technologies. How 3D Works; Pickit. How Does the Pickit 3D Camera Work? 3D measurement · Acquisition principles and limitations.

12. Basler. Fundamentals of Image Processing Systems; Electronic Shutter Types. System design · Shutter behavior.

13. Zitkovich, B. et al. RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control. CoRL, PMLR 229:2165–2183, 2023. Conference paper.

14. Heckler-Kram, L. et al. The MVTec AD 2 Dataset: Advanced Scenarios for Unsupervised Anomaly Detection. International Journal of Computer Vision, 134, article 175, 2026. DOI: 10.1007/s11263-026-02743-0. Journal paper.

15. Edmund Optics. Machine Vision Filter Technology; The Importance of Optical Filter Orientation, Angle of Incidence, and Cone Half Angle. Filter and polarization principles · Angular effects.

16. NEPCON ASIA official website. VisionChina Shenzhen Exhibition and Conference, October 27–29, 2026, Shenzhen World Exhibition & Convention Center, Bao’an. Event information.

Editorial context: This article was prepared to coincide with VisionChina Shenzhen 2026, held October 27–29 at Shenzhen World Exhibition & Convention Center in Bao’an. The technical discussion is intended for use beyond the exhibition dates. ^[16]