Techniques in Evolutionary Robotics and Neurodynamics

Tutorial at KI 2005, Koblenz, Germany, September 11, 2005

Frank Pasemann, Martin Hülse, Steffen Wischmann and Keyan Zahedi

Topic

The tutorial provides basic facts about the dynamics of recurrent neural networks and its use for robot control applications. An introduction to evolutionary robotics is given, describing the aims of this approach as well as some basic examples of evolved robot controllers.

Furthermore, the software package ISEE (integrated structure evolution environment) is introduced that provides the implementation of the ENS^3 (evolution of neural systems by stochastic synthesis) algorithm, an interface to control physical simulations and real robots, a tool to analyze activities of recurrent neural networks, and a graphical user interface for online modifications of the evolutionary process. Participants are invited to run their own evolutionary experiments with simulated and real Khepera robots using ISEE.

Target Audience

Undergraduate and graduate students, and others interested in the basics of evolutionary robotics, are the intended audience for the tutorial. The tutorial is limited to a maximum of 14 participants.

Contents: Overview

First, there will be a general introduction (1 hr) comprising Neurodynamics, the Modular Neurodynamics Approach to Behavior Control, Concepts of Evolutionary Robotics, and Structure Evolution.

The main part (1.5 hrs) of the tutorial is devoted to an experiment in evolving recurrent neural networks for an obstacle avoidance behavior. For evolution a two-dimensional Khepera simulation is applied. The evolved networks will be tested on the real Kephera robots. A Kephera robot is a miniature robot of 55 millimeters diameter. Some space on the desk should be sufficient for the experiment. After this testing phase, control principles of one or more evolved controllers are investigated according to the sensorimotor signal flow and its corresponding robot behavior.

Time permitting, there will be a second experiment (30 min.). An already evolved RNN serving as obstacle avoidance module will be expanded via structure evolution to an RNN that is able to solve a light seeking task. Again, participants can either use their own evolved RNN or an other robust RNN, provided by the tutors. Before the experiment, basic mechanisms of this expansion technique are explained briefly. Depending on light conditions the experiment can probably only performed in simulation. Nevertheless, all the other topics of the previous session before can and should be realized in the same way.

Contents: Details

For details see the local Web page at the instructors' site.

Presenters Information

Prof. Dr. Frank Pasemann has studied Physics and Mathematics at Universities Marburg and Würzburg, and at the International Center for Theoretical Physics (ICTP), Trieste. Diploma (1971), Dr. rer.nat. (1977) and Habilitation (1985) in Theoretical Physics. Since 1985 Univ.-Professor, since 1992 Apl.Prof. for Theoretical Physics at Technical University Clausthal, since 2004 Hon.Prof., Institute of Cognitive Science, University Osnabrück. He headed research groups at the Research Centre Jülich, Jülich (1993 - 1996), the Max-Planck-Institute for Mathematics in the Sciences, Leipzig (1997 - 1999), the University Jena (2000 - 2001), and at the Fraunhofer Institute for Autonomous Intelligent Systems, Sankt Augustin (since 2002). His research interests include Dynamics of Recurrent Neural Networks, Cognitive Systems as Complex Systems, and Evolutionary Robotics.

Martin Hülse

Steffen Wischmann

Keyan Zahedi