Development and optimization of a robot navigation algorithm combining deep reinforcement learning and state representation learning