The creation of object models for computer graphics applications, such as interior design or the generation of animations is a labour-intensive process. Today's computer aided design (CAD) programs address the problem of creating geometric object models quite well. But almost all users find common tasks, such as quickly furnishing a room, hard to accomplish. One of the basic reasons is that manipulation of objects often does not yield the expected results.
This paper presents a new system that exploits knowledge about natural behavior of objects to provide simple and intuitive interaction techniques for object manipulation. Semantic constraints are introduced, which encapsulate such `common knowledge' about objects. Furthermore, we present a new way to automatically infer a scene hierarchy by dynamically grouping objects according to their constraint relationships.