Simultaneous Localization and Mapping (SLAM) has been of great assistance to explore perceptually-degraded underground environments, such as human-made tunnel, mine tunnel and cave. However, the recurring sensor failures and spurious loop closures in these scenes brings great challenges to the application of SLAM. In this paper, an architecture for underground localization and semantic mapping (ULSM) is proposed, that promotes the robustness and real-time character of odometry estimation and map-building. In this architecture, a two-stage robust motion compensation with sensor fusion is proposed to adapt sensor-failure situations. The proposed salient region loop closure detection contributes to avoid spurious loop closures. Meanwhile, 3D pose as the initial value for point cloud registration is estimated without additional input. We also design a multi-robot cooperative mapping scheme based on descriptors of salient region. Extensive experiments are conducted on datasets that are collected in Tunnel Circuit of DARPA Subterranean Challenge.