Real Robot Reproduction: Towards Evolving Robotic Ecosystems

The field of evolutionary robotics is continuously evolving. Recent developments in material science and progress in manufacturing techniques are making it possible to move current research from simulated robots into real world robotics. But the differences between simulated environments and the real world are still large and it is currently very difficult to bridge the gap between the two. The research I present in this thesis is the culmination of years of work and has been aimed at narrowing this gap. In this thesis I present the techniques I have used and the results I bring to the field. In the search for improvements, I identify and address two different sub-problems that contribute to the gap between simulation and reality. These are: (1) the opaque approach taken in respect of new and more complex robot genetic encodings, which empower more interesting evolutionary robotics at the price of explainability; and (2) the prevalent use of static environments for the robots, which helps with repeatability but hinders evolution. To address the first point, I borrow from the field of biology. The use of increasingly complex genetic codes empowers robots to evolve more interesting shapes, at the price of transparency on the relationship between genes and the robot phenotype (shape and behaviour of the robot). In an attempt to shed some light on this relationship, we have applied a simple but powerful analysis tool commonly used in biology: heritability. Heritability is widely used in biology when one wants to study the relationship between phenotypic traits, genes, and the environment. In our research we measured the heritability of evolving robots in a fixed and simple environment to study the relationship between phenotypic traits and genes across two very different genotype encodings. As a proof of concept, we compared a direct genotype versus an indirect genotype. What we discovered is that the heritability of a trait is highly dependant on the genotype used. We measured how our indirect representation has lower overall heritability compared to the direct one. Lower heritability values are caused by bigger epistasis effects that cause bigger jumps in the evolutionary space and a more “exploratory” evolutionary algorithm. In addition we observed how heritability of several phenotypic traits can vary over the course of a few generations, and how the change correlates to shifts in population diversity. To address the second point, in this thesis I present several publications where me and my fellow researchers study robot evolution under very different environments. Given how in nature the environment usually defines the “task” of an individual, we observed how evaluating robots on different tasks already has a profound impact on the evolution. We also studied different approaches on learning during the robot lifetime and how it impacts evolution, in order to evolve robots that can solve complex tasks in future research. To study a more dynamic environment where sexual selection is influenced by the environment itself, we studied and measured how isolating sexual selection to smaller groups can have a profound impact on evolution. We achieved this isolation by means of both a virtual separation based on phenotypic traits and a more natural separation based on physical distance. The latter also expanded the dynamicity of the environment by the fact that the robots shared the same physical simulation and could actively push each other. To conclude, I present a proof of concept of robot evolution in the real world, and a reflection on the ethics of doing robot evolution in the real world at the scale necessary for space exploration.