The paper introduces a novel meta-learning method for domain generalization (DG), which aims to improve the ability of models to generalize to novel testing domains with different statistics. Unlike previous DG methods that design specific models robust to domain shift, this method proposes a model-agnostic training procedure. The algorithm simulates train/test domain shifts by synthesizing virtual testing domains within each mini-batch during training. The meta-optimization objective ensures that improvements in training domain performance also lead to better testing domain performance. This approach trains models with good generalization ability to novel domains. The method is evaluated on a recent cross-domain image classification benchmark and demonstrates promising results on two classic reinforcement learning tasks. The key contributions include a gradient-based meta-learning algorithm that can train any type of base network and apply to both supervised and reinforcement learning settings. The method achieves state-of-the-art results on the PACS benchmark and shows promising performance in reinforcement learning tasks like Cart-Pole and Mountain Car.The paper introduces a novel meta-learning method for domain generalization (DG), which aims to improve the ability of models to generalize to novel testing domains with different statistics. Unlike previous DG methods that design specific models robust to domain shift, this method proposes a model-agnostic training procedure. The algorithm simulates train/test domain shifts by synthesizing virtual testing domains within each mini-batch during training. The meta-optimization objective ensures that improvements in training domain performance also lead to better testing domain performance. This approach trains models with good generalization ability to novel domains. The method is evaluated on a recent cross-domain image classification benchmark and demonstrates promising results on two classic reinforcement learning tasks. The key contributions include a gradient-based meta-learning algorithm that can train any type of base network and apply to both supervised and reinforcement learning settings. The method achieves state-of-the-art results on the PACS benchmark and shows promising performance in reinforcement learning tasks like Cart-Pole and Mountain Car.