Abstract: To address the issues of low safety and poor smoothness in path planning for underground mining unmanned vehicles operating in complex tunnel environments, a global path planning algorithm, RA-A*, is proposed by integrating multi-gradient risk maps with an improved A* algorithm. The algorithm transforms discrete obstacle environments into continuous risk potential fields via multi-gradient risk maps. Collision risk and steering penalties are incorporated into the cost function, while a forward-looking risk estimation term toward the target direction is added to the heuristic function to guide the search toward nodes balancing safety and efficiency. Cubic smoothing B-spline curves combined with an adaptive collision detection mechanism are employed to optimize the path, thereby improving its continuity and feasibility. Simulation results show that RA-A* achieves superior safety in complex environments compared to A*, DW-A*, and BA*-MAPF, reducing average cumulative risk by 7.6%–33.2% with the fewest turns. Field experiments demonstrate that when navigating in underground tunnels with static equipment and dynamic pedestrians, RA-A* guides the vehicle toward low-risk areas, maintaining a safety distance over 0.75 m. Compared with A* and DW-A*, the cumulative risk is reduced by over 74% and the average obstacle avoidance distance is increased by over 68%, significantly enhancing the operational safety and stability of UMVs in underground environments.