Planning for potential: efficient safe reinforcement learning