Exploiting Imperfections in Perception-Action Learning