Machine vision for robots: 2D, 3D and inspection
What vision solves, when 2D is enough and when 3D is needed, how it is calibrated and when it is not needed.
Vision is not an accessory that improves any cell: it is a way to solve uncertainty. If the part always arrives in the same position, good feeding can be cheaper and more robust than a camera. If position varies, you must measure, light and calibrate.
What vision solves
Locating a part, orienting it, verifying presence, sorting, inspecting or correcting a path. It also enables conveyor tracking. It does not solve a feeding problem that prevents separating or presenting the part.
2D, 3D and bin picking
The table organizes the decision. 2D is often enough for planar position; 3D adds height and orientation; bin picking needs separating stacked parts and a feasible grip.
| Criterion | What it measures | Best for | Limit |
|---|---|---|---|
| 2D vision | Planar position and presence. | Flat parts, conveyors and known positions. | Does not solve height or spatial orientation. |
| 3D vision | Height, volume and orientation. | Stacked, curved or tall parts. | More calibration, data and cost. |
| Bin picking | Part inside a container. | Separable parts and good visibility. | Fails with tangles, transparency or low contrast. |
Camera, lighting and calibration
A camera does not compensate for bad lighting. Contrast, shadow and reflection define whether the part can be seen. Robot-camera calibration maps pixels to robot coordinates; without it, vision locates the part but the robot does not reach it.
Conveyor tracking and inspection
If the part moves, the system needs an encoder, synchronization and pick logic. Inspection can be presence verification, measurement or defect control; each requires defining the criterion and validating it with samples.
Integration with PLC and robot
Vision provides coordinates or a decision to the robot and PLC. You must define who commands, what happens if the part is not found, how retries work and how the result is logged. Safety does not depend on the camera.
When vision is not needed
When the part arrives in a fixed, repeatable position, when fixturing always presents it the same way or when a better-designed tray removes variation. In those cases, vision adds cost and maintenance without solving anything.
How it is validated
With real samples, site lighting, good and bad parts, and an acceptance criterion. The trial must include the variations to be tolerated in production, not only the perfect demo part.
Data needed to assess vision
Part and material, position variation, speed, lighting, camera space, tolerance, defects to detect, rate and systems to integrate. That determines 2D, 3D, bin picking or none.
Frequently asked questions
When is machine vision needed?
When position varies, sorting or inspection is needed, or the part moves. If presentation is fixed and repeatable, good feeding may be enough.
2D or 3D?
2D for planar position and presence; 3D for height, volume or spatial orientation. The decision is validated with real samples.
When is bin picking needed?
When parts arrive stacked or random in a container and can be separated and gripped. It does not solve tangles or transparency without feeding design.
How is the camera calibrated to the robot?
With a procedure that maps pixels to robot coordinates, using a calibration part or pattern and checking error at several points.
What lighting is needed?
Lighting that creates stable contrast for the part, without shadows or reflections that confuse it. It is tested on site, not only in the lab.
How does it integrate with conveyor tracking?
With an encoder, synchronization and logic that provides position to the robot. You must define what happens if the part is not found or is lost.
When is vision not needed?
When the part is always presented the same way or a better-designed tray removes variation. Vision adds cost and maintenance.
How is a vision solution validated?
With good and bad parts, site lighting, real variations and an acceptance criterion. A perfect demo does not replace production testing.
