Reinforcement Learning for Sparse-Reward Object-Interaction Tasks in Simulated 3D Environment