Abstract: Simultaneous Localization and Mapping (SLAM) is a key technology for achieving autonomous navigation of mining robots. However, due to sparse and highly repetitive environmental features in underground coal mines, significant cumulative localization errors and prolonged relocalization times occur. To address this problem, a multi-sensor mapping and localization method for underground coal mines based on semantic landmarks was proposed. The method integrated observation information from visual odometry, inertial odometry, and LiDAR odometry to construct a tightly coupled multi-sensor fusion odometry system, thereby improving localization robustness in feature-deficient environments. Semantic landmarks suitable for underground environments were defined. By establishing a mapping relationship between roadway structural features and landmark encoding information, a fused semantic landmark map containing spatial geometric features and customized semantic labels was constructed to solve low relocalization efficiency and feature mismatches caused by the high repetitiveness of roadway features. Semantic landmarks were used to correct cumulative odometry errors in real time to achieve dynamic pose correction of the robot. Experiments were conducted using a dust suppression robot platform in surface tunnel environments and in underground industrial tests. The results showed that the average mapping error in the ground tunnel was 0.020 m, the maximum static localization error was 0.035 m, the maximum absolute pose error in dynamic localization was 0.153 m, and the average relocalization time was 3.3 s. In underground roadways, a global map covering 2 400 m was constructed, with an average error of 0.038 m per 100 m, and autonomous navigation of the robot was achieved.