Towards Real-World Reinforcement Learning: Constraints and Sample Efficiency