Autonomous navigation in complex environments is a crucial task in\ntime-sensitive scenarios such as disaster response or search and rescue.\nHowever, complex environments pose significant challenges for autonomous\nplatforms to navigate due to their challenging properties: constrained narrow\npassages, unstable pathway with debris and obstacles, or irregular geological\nstructures and poor lighting conditions. In this work, we propose a multimodal\nfusion approach to address the problem of autonomous navigation in complex\nenvironments such as collapsed cites, or natural caves. We first simulate the\ncomplex environments in a physics-based simulation engine and collect a\nlarge-scale dataset for training. We then propose a Navigation Multimodal\nFusion Network (NMFNet) which has three branches to effectively handle three\nvisual modalities: laser, RGB images, and point cloud data. The extensively\nexperimental results show that our NMFNet outperforms recent state of the art\nby a fair margin while achieving real-time performance. We further show that\nthe use of multiple modalities is essential for autonomous navigation in\ncomplex environments. Finally, we successfully deploy our network to both\nsimulated and real mobile robots.\n