This paper treats an interactive learning interface used to obtain semantic constraints among 3D objects from existing 3D scenes for automatic 3D scene generation. The layout work for3D scene generation takes a long time because 3D objects have six degrees of freedom (DOF) andare difficult to be positioned by using a standard 2D input device, e.g., a mouse device. To deal with this problem, the authors have already proposed an automatic 3D object layout systembased on contact constraints. However, when there are too many kinds of 3D objects to be laid out, it is practically impossible for the user to define semantic constraints among all of them. In this paper, the authors propose an interactive learning functionality and its interface toobtain semantic constraints among 3D objects from already existing 3D object scenes. Using this functionality, the user can generate desirable 3D scenes more easily. This paper presents the learning process to extract semantic constraints from already existing 3D scenes, and delineates its usefulness by showing experimental results.