Evolving Humanoid Behaviours for Language Games Frank Pasemann, Christian Rempis, Arndt von Twickel Institute of Cognitive Science, University of Osnabrueck, Germany {frank.pasemann | christian.rempis | arndt.von.twickel} @uni-osnabrueck.de, http://ikw.uni-osnabrueck.de/~neurokybernetik/ Abstract. Evolutionary techniques are applied to develop the neural control of humanoid robots. These robots were designed to act as agents in language games played in the context of the EU-project ALEAR. The basic ingredients needed to bring forth the desired behaviours are described: an appropriate physical simulator of the robots, an interactive evolution environment and various analysis tools. A modular approach to neural control is taken and is supported by a corresponding evolutionary algorithm, such that complete neural control networks are composed of specific functional units, the so called neuro-modules. Examples of such modules are described and their use is demonstrated by means of two developed networks for a walking and a gesture behaviour. Keywords: evolutionary robotics, humanoid robots, recurrent neural networks, neurocybernetics 1 Introduction The control of humanoid robots is one of the challenging problems of autonomous robotics research. Apart from comparably simple humanoid robots used in com- petitions like the famous Robocup challenge [1] – where the robots are specif- ically designed to be easy to control – there is a growing interest in complex, larger and anatomically more human-like robots equipped with many differ- ent sensors, more exotic actuators and body proportions close to humans. Such robots are of special interest because they may some day interact with humans and their corresponding environments in a more natural way and without the need of adapting the existing environments specifically for such robots. With bodies and behavioural abilities being comparable to humans, such robots can be integrated much easier to most environments that are already optimized for human(oid) beings. For language games, as performed in the ALEAR project, human-like be- haviours are desirable because language depends on the behavioural and percep- tive abilities of the users of that language. Thus, for perception camera sensors should be included to talk about colours and spatial object relations. Further- more, the robot should be able to manipulate objects to communicate about