Developers have developed a method to generate traversability maps for autonomous ground robots using GPU-accelerated processing. The approach evolved from 3D voxel maps to 2D height maps, addressing performance limitations of CPU-based solutions that struggled with real-time data processing.
The system integrates data from RGB-D cameras and LiDAR sensors. While 3D voxel maps provide high detail, they are computationally expensive, as approximately 90% of the volume consists of empty space. The new method compresses 3D data into a 2.5D height map, which is then converted into a 2D traversability map.
The mapping process utilizes a sliding window technique on a 400x400x100 voxel grid, centered on the robot. By prioritizing layers based on the robot's height and surface elevation, the algorithm collapses 3D data into a surface representation. This allows the robot to navigate both indoor and outdoor environments effectively.
To determine traversability, the system employs the least squares method to calculate local and global slopes, as well as height variance within a window. These metrics are used to assign cost values to map cells, classifying them as free, lethal, or unknown, which enables efficient path planning.
