Optimal navigation for AGVs: A soft actor-critic-based reinforcement learning approach with composite auxiliary rewards.