Gravar-mail: Learning and exploration in action-perception loops