Learning-based intelligent algorithms have demonstrated remarkable effectiveness in navigation tasks. However, the significant gap between diverse scenarios and tasks poses a challenge for efficient and convenient transferability of navigation algorithms. This paper proposes a modular transferable navigation network for visual navigation in unstructured environments, aiming to achieve streamlined and effective navigation transfer through module transfer and reuse. The design incorporates Domain Randomization to enhance the network’s adaptability across different scenes and employs Model-Agnostic Meta-Learning (MAML) for efficient transfer between tasks. To validate the effectiveness of the proposed network, experiments are conducted in the Habitat simulation platform. Comparisons are made against several baseline models using navigation metrics such as Success Rate (SR), Success weighted by Path Length (SPL), and Distance to Success (DTS). The results indicate that the transferable navigation network outperforms the best-performing baseline model by 12.4%, 0.4%, and 1.7% in SR, SPL, and DTS, respectively, while exhibiting higher interpretability.