The Role of Pleiotropy in Body-Brain Evolution of Simulated Robotic Agents

Abstract # ↰

Body-brain optimization of embodied agents is known to be a difficult task, as the two components may interact in non-trivial ways. This holds particularly in the case of modular robots, as bodies can be as complex as brains. When optimizing the agents through evolutionary computation (EC), this difficulty impacts on the design of the genetic representation which has to allow for mutually compatible body and brains. While the common approach is to “split” the representation in two distinct parts, one for the body, one for the brain, previous studies and biology suggest that pleiotropy may be beneficial. In brief, pleiotropic representations are those where one gene can affect more than one trait, possibly belonging to different components (body vs. brain). In this work, we experimentally study the effects of pleiotropy in genetic representations for body-brain evolvable robotic agents. We design a generic representation allowing for pleiotropy and instantiate it for two cases of different complexity: navigation of simulated differential-drive robots and locomotion of voxel-based soft robots (VSRs). Our results suggest that there is little interplay between evolutionary pressure, i.e., fitness, and pleiotropy. Pleiotropy tends to settle to some “natural” level mostly dictated by the representation.