Eye-in-Hand Cameras Reach Points a Fixed Station Cannot
Inspection points on automotive assembly parts are spread over several faces and locations, and parts arrive on the conveyor with good units released and exceptions blocked. Missing machining, missing assembly and wrongly assembled parts all have to be caught inside 60 seconds, and every point needs a camera that can physically reach it. Put a ring of fixed cameras at every station and one layout change means rebuilding brackets.
Scattered points and a changing layout make fixed stations expensive
The scheme mounts 2D and 3D cameras on the robot end effector, an eye-in-hand arrangement: the camera travels to any inspection point with the arm. Compared with a fixed station, the change is that reach stops being a matter of layout and becomes a matter of path. Adding or removing a point means editing the path program, not moving brackets or rebuilding structure. On an assembly part with scattered points and frequent layout changes, that often decides whether the project works, more than any accuracy figure.
Shooting while moving: blur and error are two problems
Capturing while the arm moves runs into two separate things. Motion blur, because the camera is still moving during exposure and the image smears; and positioning error, because the robot stops somewhere near the taught position and the captured region shifts. Blur is handled by coordinating exposure with motion timing, positioning error by repeatability and trigger synchronisation, and here the motion and trigger are designed together.
2D decides presence, 3D decides correctness
Missing machining and missing assembly are mostly presence questions; wrong assembly is a correctness question. The 2D camera handles planar features such as surface marks and hole presence; the 3D camera handles spatial features such as assembly height and part position. Each is compared with the standard model and fused by rule.
The split is not arbitrary: presence can be answered from a flat image, while a correctness error usually sits in depth, where only 3D can see it. One sensor for both means missing depth errors or paying 3D cost for a flat question.
The path is the cycle budget
The cycle is 58 seconds, inside the 60-second target, subject to site acceptance results. Most of it is the robot travelling between points, so ordering the path by inspection item and cutting empty travel is the main way the cycle is compressed. More points and longer paths tighten the budget; where there is slack, it is better spent on more stable imaging conditions than on forcing in extra points.
The comparison step loads the standard model for the assembly part, so changeover switches model and path program with no hardware change. That lowers mixed-production changeover cost, and means model maintenance has to move with product changes. If the model changes and the standard does not, the system judges confidently against an old standard, and that surfaces only after a batch of complaints.
When a fixed multi-camera station is still right
Where the points are concentrated, flat judgment is enough and the product range is single, a fixed 2D station is cheaper and structurally simpler. The value of eye-in-hand is flexibility and coverage, not single-point image quality, so where points never move and the cycle is relaxed, the premium does not pay back.
Accuracy and the record
Recognition accuracy is above 99.9 percent and inspection efficiency rises several times over, subject to site acceptance results. Verdicts for the three POKA-YOKE classes are recorded against the assembly part serial number, and exception images and conclusions are archived to support process investigation and standard refinement. Robot repeatability and trigger synchronisation decide whether it works, and both should be assessed separately rather than inferred from camera specifications.
When to change the configuration
Can eye-in-hand and fixed multi-camera be mixed?
Yes. Fixed cameras on the main faces, the robot on scattered points needing spatial judgment, with coordinate systems unified.
Does moving capture cost accuracy?
A little, for coverage and flexibility. Precision-sensitive features are captured with the robot stopped; loose ones can be captured in motion.
How many points can 58 seconds take?
It depends on per-point capture and judgment time and path length. Calculate path time first, then decide between more points and a second robot.