The Watchmaker's guide to Artificial Life: On the Role of Death, Modularity and Physicality in Evolutionary Robotics

Optimizing robots through the implementation of evolutionary processes
plays a key role in evolutionary robotics and artificial life. Challenges
that arise in the evolutionary optimization of robots are usually related to
an algorithm’s compromise between trying new solutions and improving
previously found solutions (the exploration vs. exploitation trade-off),
and whether to express genomic information directly or generatively
(genotype to phenotype mapping). An additional challenge is that there
is a discrepancy between simulation and reality (reality gap) where robots
‘evolved’ in simulation environments function differently when transferred
to the real world.
The exploration vs. exploitation trade-off is addressed through
describing and experimenting with the inclusion of biologically inspired
intrinsic mortality and how this affects the evolvability of populations.
The results contribute to our understanding of the relationship between
intrinsic mortality and mutation rate. The results further indicate how it
can be utilized to develop algorithms that can outperform state-of-the-art
algorithms.
This thesis continues by addressing the challenge of mapping the
genotype to the phenotype through investigating the influence of generative
encodings on the evolution of simulated modular robots. It is investigated
how an L-System as a generative encoding can lead to the formation
of plant-inspired virtual creatures and shows that movement for tracking a
moving light source is not an emerging phenotypic trait. Afterwards, this
generative encoding is shown to better evolve modular robots for locomotion
compared to a direct encoding. The addition of real-world solar panel modules demonstrates how modular robots can be evolved toward energy
autonomy.
The final part of the thesis describes the evolution of embodiment and
control of physical robots. As part of this, an automated process for
assembly and disassembly of modular robots is demonstrated, which can
be used to evaluate evolved individuals in the real world. A viable method
for implementing evolution directly is demonstrated through evolving
the behavior of a knifefish-inspired physical soft robot. Both approaches
represent strategies for addressing the reality gap.
The experimental results of this thesis contribute to the understanding
of biological phenomena and elucidate how improvements can be made
to existing methods in evolutionary robotics. It shows that we can utilize
concepts from evolutionary biology to advance our understanding of
evolutionary dynamics, encodings and physical designs that are valuable
for the automated design of robots.