Robots aren’t the fix if the infeed lies

We retrofitted a delta pick-and-place on a pouch line to chase 120 ppm. In FAT it looked great. On the floor, tiny shifts in pouch spacing and wrinkled seals turned into mis-picks and cascade faults. The robot caught the blame, but the real culprit was upstream variation and sloppy timing signals from a legacy encoder and photoeye.

What actually solved it: a short metering belt with encoder-based spacing, a simple mechanical singulator to kill doubles, and tighter seal jaw timing. We moved the camera, went to strobed lighting tied to conveyor position, and shielded the trigger cabling to stop noise. Added 300 mm of accumulation, a clean reject path, and a one-cycle retry before faulting. PackML state handling made recovery predictable.

Lesson for me: the robot was fine; the boring interfaces weren’t. Before you drop a robot on an existing line, what are your must-have prechecks? Do you insist on a reference encoder, metering, and defined rejects? Any go-to tricks for rock-solid vision and timing in washdown food plants?