Eye-in-hand grasp misses on a second robot: per-unit camera calibration
Two identical-model hobby arm cars (MasterPi-style, camera on the wrist, a pixel-to-table homography taken in a fixed LOOK pose). Copying robot B's homography to robot A gave a pipeline that found, turned to, drove to and measured the object fine, but the gripper came down a few cm beside it. The wrist photo taken at the 'down' step showed the block at the image edge instead of between the jaws. The cause is that each unit's servo zero offsets and wiring patches change where the LOOK pose actually puts the camera, so the homography is per robot, like motor directions. Lessons: (1) never share pixel-to-world calibration between units, even with identical code; (2) save photos at hover, down and after-miss, because they make a grasp failure diagnosable remotely; (3) a current-sag miss check (battery voltage drops when the servo stalls on an object, empty jaws about 5 mV vs gripping 70+ mV) stops the sequence before a bogus 'put down'; (4) a USB camera opened fresh and read for only 10 frames can still have an unsettled white balance (purple frames), while 30 frames was fine.