Robot grasp retries must re-position first: after a miss the camera no longer sees the object
A small arm robot grabs with a camera on its arm: look pose finds the object ahead, the car creeps up until the object is within reach, then grips. A wrapper retried a failed grab by simply running the same approach again, up to 5 times. On the floor only try 1 was a real attempt. After the miss the car stood right over the object, so the look pose saw past it; tries 2-5 reported 'not found' or locked onto a different thing further ahead (a large box placed as an obstacle, whose glare the vision model accepted because the prompt was 'any object') and the car crept towards that instead. Fixes that follow directly: (1) before every new try back up by roughly the creep distance (about 18 cm here) so the object is back in the look-pose view; cap the total blind retreat when the failure says the car never moved; (2) when anything else is in view, ask the detector for the named object rather than 'any'; (3) keep the full output of each attempt on the robot, since a wrapper that keeps only the last line cannot tell 'missed' from 'looked at the wrong thing'. Separate finding from the same run: a classical floor/object mask (floor colour fitted over the image, object = different hue or brighter) failed under strongly coloured room light: the cube's shadow came out as object and its bright face as floor, so a compactness test called the cube 'not a block' and the gripper aimed at a 1 cm sliver. Re-running the mask offline on the saved photo showed it in one step (component filled 44 percent of its box, threshold 62). Passing the known object height to skip that shape guess is the workaround; not yet re-tested on the robot.