Evolving a novel bio-inspired controller in reconfigurable robots J¨ urgen Stradner 1 , Heiko Hamann 1 , Thomas Schmickl 1 , Ronald Thenius 1 , and Karl Crailsheim 1 Artificial Life Laboratory of the Department of Zoology Karl-Franzens University Graz, Universit¨atsplatz 2, A-8010 Graz, Austria, {juergen.stradner, heiko.hamann, thomas.schmickl, ronald.thenius, karl.crailsheim}@uni-graz.at Abstract. Evolutionary robotics uses evolutionary computation to op- timize physically embodied agents. We present here a framework for per- forming off-line evolution of a pluripotent robot controller that manages to form multicellular robotic organisms from a swarm of autonomously moving small robot modules. We describe our evolutionary framework, show first results and discuss the advantages and disadvantages of our off-line evolution approach. In detail, we explain the single parts of the framework and a novel homeostatic hormone-based controller, which is shaped by artificial evolution to control both, the non-aggregated single robotic modules and the joined high-level robotic organisms. As a first step we present results of this evolutionary shaped controller showing the potential for different motion behaviours. 1 Introduction Recently evolutionary robotics (ER) has become a fascinating field that ex- ploits evolutionary computation (EC) to optimize physically embodied agents. In some studies robot controllers were adapted by EC in simulated worlds [1]. In contrast to that, other studies [2] showed evolution of single robots and small robot groups in a process of on-line evolution, where a sort of genetic algorithm (GA) was optimizing artificial neural networks (ANN) during runtime of the real robot(s). The advantages and disadvantages of both approaches are clear: On-line evolution profits from the fact that real hardware is used in real-world environments but suffers from lower number of generations. Off-line evolution profits from computational speed (parallel processing, grids) but suffers from dif- ferences between models and reality [3]. As an intersection of these approaches, the swarm-bots project [4] used a set of simulation tools of varying levels of detail (physics, robot model) to perform off-line evolution of ANNs, which were finally tested on real robotic hardware. In former studies we used an evolutionary strategy [5] to shape algorithms that aggregated robots autonomously in various group sizes at target areas [6]. The 1 Supported by: EU-IST-FET project ‘SYMBRION’, no. 216342; EU-ICT project ‘REPLICATOR’, no. 216240.