Using structured prediction, we can simultaneously learn what actions to suggest and what objects those actions should act on. In experiments on three tasks in cluttered table-top environments, this ...