In this paper, we describe an algorithm for acquiring occupancy grid maps with mobile robots. The standard occupancy grid mapping developed by Elfes and Moravec in the mid-eighties decomposes the high-dimensional mapping problem into many one-dimensional estimation problems which are then tackled independently. Because of the independencies between neighboring grid cells, it often generates maps that are inconsistent with the sensor data. To overcome it, we propose the concept of the cluster which is a set of cells. The cells in the clusters are tackled dependently with another occupancy grid mapping with EM algorithm. The occupancy grid mapping with EM algorithm yields more consistent maps in the cluster, but it find out the maximized map for the whole environments. Hence, it takes more time and gets the map to be inaccurate partly. Instead of this, our approach use the EM algorithm only in the clusters. As we use mapping algorithm adaptively with clusters according to the sensor measurements and we emphasize not the cells but the clusters , the maps generated by our approach have less calculation time and are more accurate than the previous mapping algorithms.