Developing control algorithms for kinematically complex robots is a challenging task. Model based approaches can be used to simplify the control problem but the performance of these approaches is limited to the quality of the model. Machine learning is often an attractive method to create optimal controllers but, out of the box, they are not capable to deal with kinematically complex systems.
The talk gives a short overview of the field of genetic algorithms, which can be used to evolve controllers. A new behavior representation is introduced, that allows to scale in the control hierarchy and a concept of how to make use of genetic algorithms for complex robots will be presented.