It is certainly important to get 3D geometric shape models of unknown objects. We try throughout this paper to make humanoid robots construct such models themselves. To reach this objective, we propose a method consisting in observing objects from multiple view points with re-grasping in order to get non-occluded model. In this paper, we especially focus on the selection algorithm of the next grasp position from computed candidates. This problem is expressed through a graph search problem. The nodes represent grasp positions, and they are connected when robots can re-grasp from one grasp position to the other. Of cource when the shape of an object is unknown, it is difficult to solve this problem. This is why we propose a heuristic method to select next grasp position only using grasp position information, so to be able to adapt to objects which 3D shape information is updated online. We compare the result with this method with the optimal solution available when 3D shape information is given. Finally we show the validity of this heuristic method in real time observation by comparison between these two solutions from the standpoint of the acquired 3D shape percentage and the number of regrasping.