Mobile manipulator demo: when a detector hallucinates a graspable object, make the grab outcome end the pursuit fast
Setting: a small arm robot collects objects on a wooden table with a YOLO model fine-tuned on photos of the parts on carpet. In the first table run it took a flat white dot on the table top (screw cap / sticker) for a 'cup' at 0.85-0.88. The grab closed its jaws on nothing; the retry logic then backed up, searched a whole turn for that class only (40 s) and tried again 'not seen' - ~2.5 min lost while a real object was in view.
What did not work as a filter: - Rejecting boxes whose content looks like the learned table colour: the grab's close-up camera view had a different white balance than the search view, so the table model did not match there at all. - Rejecting by estimated real width (flat-floor geometry): fine for one calibrated view, but photos from other arm poses gave nonsense widths and real small parts (knobs) came out as narrow as the phantom.
What works: treat the grasp result as evidence. If the target has no distinctive feature (no colour) and the jaws closed freely (the gripper's own miss check - here a battery-voltage sag test), drop that object class immediately instead of retrying. After any miss, re-acquire only straight ahead (the robot just backed up along its line; arm pans allowed, no body turns) and give up the object if it is not there. The cost of a phantom drops to one grab (~45 s).
General point: a detector trained on one floor will hallucinate on a new surface; on site, the cheapest guard is to make 'grab failed' strongly informative rather than to tune the detector live.