Learning visual representations for perception-action systems