Hindsight States: Blending Sim and Real Task Elements for Efficient Reinforcement Learning