Deep Reinforcement Learning （DRL）-based vehicle routing planning methods have garnered significant attention for their rapid solving speed and end-to-end processing capabilities. However， most existing methods are limited to scenarios with uniformly distributed nodes and fixed node numbers， demonstrating performance degradation when handling unevenly distributed nodes or varying numbers of nodes. To address this issue， a meta-learning framework based on improved Model-Agnostic Meta-Learning （MAML） and Graph Variational AutoEncoder （GVAE） was proposed to obtain a well-initialized model through meta-training， and perform quick fine-tuning for out-of-distribution tasks， improving the model's generalization performance. Besides， a GVAE was employed for initializing parameters of the meta-learning framework to further enhance the effect of meta-learning. Experimental results show that the proposed method can handle Vehicle Routing Problems （VRPs） with different node distributions， performs well when dealing with varying numbers of nodes. The average gap across five tasks reduced by 0.45 percentage points compared to the method that does not use meta-learning. It can be seen that the proposed meta-learning framework enhances the effect of reinforcement learning， achieves comparable solution quality to state-of-the-art solvers while significantly shortening computation time.