Abstract: With the continuous advancement of intelligent mine construction, the application of underground inspection robots plays a significant role in improving mine safety and ensuring operational efficiency. It addresses the challenges of autonomous exploration and safe navigation in complex and unknown underground environments, with a focus on effectively identifying the boundary between unknown and known areas and strictly adhering to underground traffic and safety regulations during navigation. A system framework is constructed, consisting of a mapping module, an autonomous exploration module, a rule-based navigation module, and a local coordination module. Firstly, a global map construction method based on Fast-LIO is employed, which fuses LiDAR and IMU data through an error-state Kalman filter to build a dense 3D map of the underground environment in real time, providing accurate localization information for subsequent path planning. Secondly, considering the complexity and spatial limitations of underground environments, an innovative boundary expansion algorithm based on information gain is proposed. The algorithm generates candidate target points at the frontier between known and unknown regions, calculates their information gain based on the number of unknown voxels, and selects high-value targets under safety constraints to guide the robot in continuously expanding the known area. Subsequently, in accordance with fixed underground traffic regulations, a constraint navigation method based on simplified topological abstraction is used to model the environment as a topological graph of nodes and edges. This reduces the search space and defines passable areas through rule constraints, ensuring the robot follows safety requirements such as “keep to the right” during path planning. Furthermore, to address dynamic obstacles, a local path cluster sampling strategy is designed. By smoothing local paths and evaluating multiple cost metrics, the optimal trajectory is selected to ensure the robot can make timely adjustments and navigate safely in dynamic environments. In simulation experiments, compared to RRT-based viewpoint sampling and traditional frontier-based exploration methods, the proposed exploration algorithm reduces trajectory length by approximately 42.8% and 26.8%, and decreases exploration time by 34.8% and 18.6%, respectively. Meanwhile, the simplified topological A* algorithm achieves nearly a 98% reduction in planning time and significantly reduces map memory usage, all while complying with traffic rules. Future work will further explore the safe navigation tasks of inspection robots operating in rugged and complex mining face terrains.