Vision-guided picking stands or falls on calibration — the mapping between what the camera sees and where the robot moves. The good news: done methodically, it is a routine procedure, not an art.
Three frames, one mapping
The camera has a pixel frame, the robot has a base frame, and the part lives in both. Calibration builds the bridge: the camera observes a target at known robot positions, and the software solves the transformation between the two. Every guided pick afterwards rides on that bridge.
The target plate does the heavy lifting
A flat plate with a regular pattern — dots or checkerboard — gives the solver many precise reference points in one view. Keep the plate clean and unbent, fill a good share of the field of view, and capture positions spread across the working area rather than clustered in the middle.
Height changes everything in 2D
A 2D camera assumes the part sits on a known plane. If the pick surface moves — stacked layers, warped trays — either re-reference the plane, add a height sensor or step up to a 3D sensor. Most mystery offsets in 2D guidance trace back to a part that was not where the plane said it was.
Verify with a round trip, then monitor
After calibrating, command the robot to touch a feature the camera located — the residual error is your working accuracy, measured end to end. Log it, repeat the check at maintenance intervals, and recalibrate when the trend drifts rather than when the picks start missing.
