In this paper, we propose a FastSLAM method that can generate a large-scale voxel map. FastSLAM is easily implementable since adopting random sampling to generation of maps. However, FastSLAM is unable to generate a consistent map because of huge memory usage. When the map consists of voxels. In resampling step, FastSLAM requires making many copies of a map. In the case of a voxel map, the copy procedure requires huge memory usage and computational cost. In order to resolve the problem, we propose a method that divides a map to sub-maps and re-aligns the maps by sampling. Our method only generates one set of sub-maps, it doesn't require making many copies of a map, and it only realigns sub-maps by position sampling. Additionally, in this paper, we propose a scalable voxel map system in order to reduce memory usage. Generally, voxel maps are implemented as an array on memory to randomly access each voxel. For the implementation, a voxel map has to fix the size before mapping. However, the requisite minimum size depends on environments and it cannot fix before mapping. In order to resolve this problem, we propose a scalable map system applying spatial indexing. Our system can reduce memory usage since the map can be scaled as necessary. Experimental results indicate our method can generate large-scale voxel maps whose memory usage disables applying original FastSLAM.