The thesis begins with a relatively fundamental problem: how to dynamically plan vehicle routes when travel time is uncertain. The approach needs to continuously adjust routes based on the latest traffic information during the journey, rather than fixing the route at the beginning. Next, this thesis considers a more realistic scenario where orders are also dynamically generated. In this scenario, the vehicle must not only decide how to go, but also when to depart and which customers to serve first, thereby maximizing overall efficiency in an uncertain environment. The core difficulty is the travel time between the orders is hard to learn, because the orders are highly dynamic. Finally, this thesis extends the problem to a more practical scenario: online food delivery. In this scenario, the AI needs to simultaneously decide when to pick up the meal from the restaurant and how to deliver it to the customer and needs to rationally combine multiple orders to improve delivery efficiency. The designed solution is tested on data based on actual road networks and actual orders. Compared to common methods, the designed AI approach can not only serve more customers but also reduce the average delivery delay from 17 minutes to 7 minutes. This thesis primarily demonstrates that AI can effectively handle vehicle routing problems under uncertainty and the results show that designed AI are more effective at solving such problems than general AI. This lays the foundation for future use of AI to solve more complex problems with multi-source uncertainty.