Robotic assembly
Vision guidance and verification on an assembly cell
Part guidance and inline verification on a mixed-model manual assembly station, inside takt.
- -74%assembly error escapes
- ~110 msverification per placement
- 100%units with an image record
Stack
- Industrial area-scan cameras
- OpenCV + lightweight CNN
- Pick-to-light integration
- PLC handshake
- Edge GPU inference
The problem
An electronics manufacturer ran a station that assembled several product variants on the same line. Operators picked from shared bins, and the occasional wrong or misoriented part only surfaced at final test, expensive to trace back. They wanted guidance and a verification gate that didn't add to takt.
Approach
- 01 Mounted an overhead camera over the bins and a wrist-height camera at the fixture.
- 02 Classified part presence and orientation per variant, lighting a pick-to-light prompt for the operator.
- 03 Verified each placement against the variant recipe before the fastening step, with a soft Andon on mismatch.
- 04 Logged a per-unit image record so any later failure could be traced to the build.
Outcome
Wrong-part and orientation escapes to final test dropped sharply, and the rework they caused with them. Verification ran well within the existing takt, so throughput was unchanged. Operators leaned on the prompts during variant changeovers, where most of the errors had been.
