On the Utility of Model Learning in HRI