To allow a highly natural interaction between the robot and the operator we consider a visual interface to be mandatory. The idea is that the robot is controlled by the operator using his gaze point to to define objects in the environment following or earlier commands refer to. The commands themselfes are given by the person by performing simple facial gestures such as nodding or shaking the head. With the current repertoire 12 different gestures can be recognised by the system. To achieve this goal the system has to be able to localise the operator in its environment and to track the operators face in real-time.