Learning to Recognise Objects and Situations to Control a Robot End-Effector

View based representations have become very popular for recognition tasks. In this contribution, we argue that the potential of the approach is not yet fully tapped: Tasks need not to be “homogeneous”, i.e. there is no need to restrict a system e.g. to either “object classification” or “gesture recognition”. Instead, qualitatively different problems like gesture recognition and scene evaluation can be handled simultaneously by the same system. This feature makes the view based approach a well suited tool for robotics as will be demonstrated for the domain of an end-effector camera. In the described scenario, the task is threefold: Recognition of object types, judging the stability of grasps on objects and hand gesture classification. As this task leads to a large variety of views, a neural network–based recognition architecture specifically designed to represent very non-linear distributions of samples representing views will be described. Künstliche Intelligenz 2:24–29, 2003