SafePicking: Learning Safe Object Extraction via Object-Level Mapping