Automatically planning motion for robots or humans in a virtual environment is a complex task. The navigation cannot be done directly on the geometric scene. An internal representation of the environment is necessary. However, virtual environments complexity is growing at an important rate, and it is not unusual to work with millions of triangles and area as large as some square kilometers. In these cases, grid based methods require too much memory, and are too slow for A* algorithms. Polygons based methods are more memory and computation efficient, but they are difficult to generate and require complex preprocessing of the input mesh to ensure nice topological properties. In this paper, we propose an hybrid method, which uses 3D voxelization to generate a clean polygon mesh simple enough to perform fast A* search. Using this approach, we robustly generate Navigation Meshes for large environments with fine details, described by millions of triangles, without any assumption on the quality of the input mesh. The input mesh can contains holes, degenerated or intersecting triangles.