Simulations of simulations in evolutionary robotics Edgar Bermudez Contreras and Anil K. Seth Department of Informatics, University of Sussex, Falmer, Brighton, BN1 9QJ, UK. Email: {e.j.bermudez-contreras, a.k.seth}@sussex.ac.uk Abstract. In recent years simulation tools for agent-environment in- teractions have included increasingly complex and physically realistic conditions. These simulations pose challenges for researchers interested in evolutionary robotics because the computational expense of running multiple evaluations can be very high. Here, we address this issue by applying evolutionary techniques to a simplified simulation of a simu- lation itself. We show this approach to be successful when transferring controllers evolved for example visual tasks from a simplified simulation to a comparatively rich visual simulation. 1 Introduction For more than a decade, evolutionary robotics (ER) has struggled with the chal- lenge of producing controllers that function in real world environments. The approach of evolving in the real world itself is prohibitively time consuming in all but the simplest of cases [1],[3]. A popular alternative has been to evolve con- trollers in simulations, but simulations are often poor abstractions of the com- plexities of real world environments. This situation is changing. Recent years have witnessed enormous growth in the sophistication of simulation tools for modelling agent-environment interactions. Highly detailed physics-based simu- lations are now readily available ‘off-the-shelf’ which simulate not only complex morphologies but also rich streams of sensory input and motor output signals [5],[6]. While impressively realistic, these simulations can be highly computa- tionally expensive and as a result can pose challenges similar to those posed by evolution in the real world. This is not to say that evolution in a rich simulation is as problematic as evolving in the real world. Even a very rich simulation can likely be executed more rapidly (and with less chance of hardware failure) than a corresponding real world condition. If this is not true at present for a particular simulation, future increases in computational power will undoubtedly compensate. In addi- tion, rich simulations offer the possibility of exploring detailed but non-physical agent-environment interactions, which may shed light on adaptive behavior by providing alternative comparison conditions to agent-environment interactions in real-world situations.