Unmanned heavy transport vehicles have been widely adopted in various industry scenarios, prompting a shift from individual to fleet intelligence. However, managing a large fleet of unmanned vehicles presents significant challenges, including road congestion, vehicle conflicts, and deadlock situations. This study introduces a novel two-stage online algorithm to address these issues, specifically focusing on automated guided vehicles in container terminals. The algorithm redefines the traditional static vehicle routing problem into two manageable sub-problems. In the initial stage, an improved A* algorithm is designed to identify the most efficient route for each vehicle. This improvement not only aims at optimizing traffic flow but also takes into account the dynamic nature of network workload and the likelihood of conflicts between vehicles. The subsequent stage introduces a dynamic vehicle grouping algorithm followed by a deadlock-free path planning algorithm. These algorithms are crucial for categorizing vehicles into clusters and streamlining the vehicles' movement within their respective clusters. The simulation demonstrates that the new algorithm surpasses traditional static vehicle routing methods by 20% in terms of efficiency and achieves an 80% reduction in computation time. Remarkably, it supports the real-time operation of 100 to 500 vehicles without any incidents of deadlock. The implications of these findings are significant, offering valuable insights for the future implementation of large-scale unmanned vehicle systems in complex industrial environments.