Accelerated Sim-to-Real Deep Reinforcement Learning: Learning Collision Avoidance from Human Player