Crankshaft Blanks: 3D Vision Poka-Yoke Stops Misloads
Each crankshaft blank arrives in a bin, oiled, lying at whatever attitude the last pick left it, heavy enough that lifting one is a two-handed job repeated all shift. The variants share close outlines and two ends that differ by features small enough to miss under oil. Take the wrong one, or set it backwards, and the line finds out several operations later.
The fixture is where a handling error turns into a machining error
Clamping is the moment the mistake stops being recoverable. The fixture closes on a datum that does not match the blank's real orientation, and every later cut references the wrong face. Inspection downstream records the loss; it cannot undo it.
So the checkpoint sits on the transfer, after the grasp and before the fixture closes. The machine tool never sees a rejected part.
Pose, model and orientation are three answers, not one
Picking needs a pose. Clamping safely needs two more answers: is this the variant the line is running, and does it face the fixture the right way. They are computed separately because they fail on different blanks.
| Output | Question | Prevents |
|---|---|---|
| Pose | Where the blank lies | A reach into empty space |
| Model | Which variant it is | A close forging on the wrong datum |
| Orientation | Which end faces the fixture | A reversed blank clamped wrongly |
Model verification matches the blank to the production variant's template on contour and datum features; orientation compares the two ends, where the difference is small but consistent.
Oil breaks the image before it breaks the grip
Rolling oil reflects hard, and 2D imaging over an uneven film reads the same blank differently with every change of light. Point cloud geometry does not change with the coating, so an uneven film still yields a usable contour.
The same oil lowers friction, so grasp points sit at pronounced contour features the fingers can close around, and the approach avoids the thickest film. Slip is addressed in the pose, before the gripper closes.
A bin holds one pick per scan
Stacking is uneven and attitudes vary, so a taught position goes stale whenever the bin is refilled differently. The system measures each visible blank, ranks candidates by grip quality, and starts from the top layer.
What planning skips: removing one blank moves those beneath it. A scan is valid for one pick, so positions refresh every cycle, which keeps the last blanks pickable.
The verification stop rides the transfer instead of adding a beat
The blank is grasped, held at an intermediate stop inside the cell, then released into the fixture or diverted. The check reuses the pre-grasp point cloud, so its computation overlaps the arm's motion and opens no gap in the cycle. Rejected blanks go to a designated area, leaving the pick order untouched.
Pick and verification together take about 35 seconds per blank, and one robot tends a group of machine tools running the same cycle, shifting the line to one operator for several machines, subject to site acceptance results.
What a misload check does not cover
The verdict covers model and direction only. Cracks, chipped edges and missing material need their own imaging stage; expecting it to find material defects is a planning error.
Three conditions sit under it: blanks must stay pickable in the bin, surface condition sets point cloud quality, and each variant needs its own program. On a line already loading with 3D vision, the stop can be added as a module.
Questions asked before the layout is fixed
Can 3D vision still locate a blank coated in rolling oil?
Yes. Position comes from point cloud geometry, far less sensitive to reflection than 2D imaging. Oil changes the image, not the shape.
Does verification need its own station?
No. It runs after the grasp and before clamping, on the robot's path, using the pre-grasp point cloud.
How does one robot serve a group of machines?
Pick and verification take about 35 seconds per blank, matched to a group of machine tools on the same cycle, subject to site acceptance results.
Does the check also catch cracks and chipped edges?
No. Model and orientation only; surface defects need a separate stage.