Using Shapes for Function in Evolutionary Robotics

Using Shapes for Function in Evolutionary Robotics A senior thesis presented to the Faculty of the School of Arts and Sciences of Brandeis University, Waltham, Massachusetts by Peter Macko Most work in evolutionary robotics has been done on very simple models, such as pointand-stick models. When these robots are constructed or just displayed on the screen, they are built from components designed by human engineers. The lack of ability of the evolutionary algorithms to evolve their own shapes limits the kinds of robots that can evolve. On the other hand, the evolution in nature involves evolution of shapes. For example, wings of birds are effective in helping the animal to fly because of their shape. Consequently, no computer program can evolve a bird if it is not capable of evolution of shapes, and no human engineer built a wing as a primitive. Computers became faster, and physics engines became more available, which allowed us to start experimenting with evolution of shapes for functionality. We surveyed and experimented with several possible representations, including direct representations of polygons and bitmaps, and a rotational representation, which was found to be very suitable for evolution of gear-like shapes. We have designed a novel language-like representation, which describes both shapes and how the shapes are assembled. Although the work on this representation is still in progress, our current results are very promising.