Reinforcement Learning for Sparse-Reward Object-Interaction Tasks in First-person Simulated 3D Environments