 |
Josh Bongard |
|
Associate Professor |
|
329 Votey Hall, Burlington, VT 05405 |
|
Tel: (802) 656-4665; Fax: (802) 656-0696 |
|
Curriculum Vitae: PDF |
|
| New | Research | Teaching | Media | Publications | The Zoo | PhD Thesis | MSc Thesis | Misc |
| PhD Thesis, University of Zurich |
|
[PS] [PDF] |
Incremental Approaches to the Combined
Evolution of a Robot's Body and Brain The employment of evolutionary algorithms for the design of robots, known as evolutionary robotics, is becoming increasingly popular. In parallel, the importance of embodied robotics has come to the fore: that is, the realization that not just neural control, but rather the brain, body, and environment of the robot, as well as the interactions between all three systems, lead to interesting and useful behaviour. This thesis combines these approaches by evolving virtual robots in a physical, simulated environment. In this way the robots can exploit the physical dynamics of their environment to generate behaviour. We begin with a set of standard evolutionary robotics experiments, in which the robot body and neural controller are fixed, and only some of the parameters of the controller are optimized using simulated evolution. The following experiments then demonstrate the subjugation of increasingly more aspects of the robots' bodies and brains to evolutionary control. The results make clear many previously unknown interdependencies between robot brains and bodies, as well as generating testable hypotheses as to which combinations of controllers and body plans are best suited for particular tasks. In the final sections, a model of artificial development, based on genetic regulatory networks (GRNs), is introduced to evolve both the neural controller and body plans of robots. It is shown that evolutionary runs with high evolvability arise from GRNs with particular properties, which suggests how biological GRNs arose in response to natural, as opposed to artificial selection pressure. |