Towards Real-World Reinforcement Learning: Constraints And Sample Efficiency