Coupled inverted pendulums

Abstract

The challenging scientific field of self-reconfiguring modular robotics (i.e., decentrally controlled 'super-robots' based on autonomous, interacting robot modules with variable morphologies) calls for novel paradigms of designing robot controllers. One option is the approach of evolutionary robotics. In this approach, the challenge is to achieve high evaluation numbers with the available resources which may even affect the feasibility of this approach. Simulations are usually applied at least in a preliminary stage of research to support controller design. However, even simulations are computationally expensive which gets even more burdensome once comprehensive studies and comparisons between different controller designs and approaches have to be done. Hence, a benchmark with low computational cost is needed that still contains the typical challenges of decentral control, is comparable, and easily manageable. We propose such a benchmark and report an empirical study of its characteristics including the transition from the single-robot setting to the multi-robot setting, typical local optima, and properties of adaptive walks through the fitness landscape.