In scenarios where multiple types of road surfaces coexist, conventional laser SLAM mapping methods such as cartographer and gmapping tend to broadly classify different surface types as ‘ground’. This oversimplification results in standard laser navigation systems being unable to differentiate between and avoid different types of road surfaces during path planning. This limitation can lead to issues such as damage to robots with low clearance chassis and harm to lawn surfaces. This paper proposes a road-adaptive navigation method based on a 2D grid semantic map. It utilizes real-time semantic segmentation with the Yolact algorithm to discern different road surfaces. Subsequently, the selected road surface pixel coordinates are mapped into the robot’s map coordinate system. By modifying the pixel values corresponding to road surface coordinates on the static map, different types of road surfaces are assigned varying pixel values, thus enabling the construction of a 2D semantic grid map. Ultimately, this approach enables robots to autonomously select safer and more stable road surfaces for path planning when faced with multiple coexisting road types. Simulation and experimental results demonstrate that this method allows for more rational navigation strategies based on different types of road surfaces.