Abstract: The visual semantic navigation of unmanned systems in complex environments relies on high-quality training data; however, such data are often difficult to be obtained in large quantities. Consequently, data augmentation techniques have become an important research direction in this field. Nevertheless, current data augmentation methods frequently suffer from issues of blind application, focusing solely on generating more data while neglecting the effectiveness of the augmented data for model training. This can lead to the generation of data that is redundant or misaligned with actual navigation requirements. To address this issue, an adaptive data augmentation method based on model performance is proposed. This method dynamically evaluates the performance of model in navigation tasks and leverages the powerful generative and reasoning capabilities of large models to construct high-value augmented data that is related to the semantics of the navigation scene and performance of the model, thereby optimizing the model adaptively. The experimental results demonstrate that the proposed method significantly enhances the visual semantic understanding and navigation performance of unmanned systems, providing new insights for addressing data challenges.