To extend the field of application of robots in unstructured environments it is necessary to develop new techniques of environment perception and interpretation. These methods must give machines the capability to generate sufficient information, which enables them to fulfill their tasks with the aid of their sensors. Therefore it is required to extract local and task dependend invariant structure out of the unstructured environment. A perception-action loop based on depth data for robot positioning tasks is developed, implemented and evaluated. The proposed method is based on the assumption that a major part of these tasks include invariant geometric properties. It is attempted to find a representation of these properties, that allows describing it in an intuitive way and compare them to perceived features. The approach goes without machine learning methods to avoid the need of huge amounts of data. Handcrafted models are used instead. Difficulties of the problem are discussed and illustrated with experimental results.