Abstract: A coal mine rescue robot that performs a rescue task needs to obtain the environment model and plan a collision-free path from the current location to the target location with the built-in algorithm upon receiving a task instruction. In order to reduce the moving time of the rescue robot, the path is usually required to be time-optimal. However, the path planned by the conventional A* algorithm that is widely used at present under the grid map environment has many problems, such as many path redundancy points and large turning angle, which leads to the fact that the path is suboptimal for the rescue robot that could move flexibly in any direction. To solve these problems, an improved A* algorithm was proposed based on the conventional A* algorithm. Firstly, the improved algorithm has the number of adjacency points of the current expanding node increased on the basis of the conventional A* algorithm to obtain the initial path through quick search. Then, the redundant points of the initial path were removed by setting the distance threshold and reconnecting the path points. Thus, the set of path points with smaller spacing could be obtained by dividing the path according to the step size, and then the redundant points were further removed. Finally, in order to further smooth the turning point of the obtained path, an optimized path with fewer path points, lower path cost and smaller cumulative turning angle was obtained by fitting with the quintic B-spline curve. Besides, the improved A* algorithm was experimentally simulated in five grid map environments with different sizes and 20% obstacle coverage by MATLAB, and the simulation results of the improved A* algorithm were compared with that of the conventional A* algorithm. The result shows that the improved A* algorithm effectively improves the problems of the conventional A* algorithm, such as many path redundancy points and large path turning angle, by expanding the adjacency points, removing the redundancy points and path smoothing. In addition, the improved A* algorithm could decrease the number of expanding nodes during the generation of the initial path to a certain extent, and thus reduce the occupation of system memory.