Abstract
Designing a gait controller for a morphology changing robot is a challenging problem due to the high degree of freedom. In 2013, Risi showed that by learning the relation between morphology and controller, it is possible to evolve a neural network-based controller that can walk on different morphologies. This thesis aims to implement Risi's flexible controller for a robot platform that features a morphology changing physical robot, Dynamic Robot for Embodied Testing: DyRET, developed at the University of Oslo. Also, it compares which of the evolutionary algorithms (EA) would be the best fit for the task - among behavior diversity, fitness-based and multi-objective EA (MOEA), and combinations of different behavior descriptors for behavior diversity EA. In particular, this thesis implements a flexible controller based on Hypercube-based NeuroEvolution of Augmented Topologies, designed to learn the relation between the morphology and the controller based on the previous work. The controller is evolved by Novelty search, fitness, and MOEA, combining both. Three behavior descriptors are implemented - max-min amplitude, duty factor, and the last position. The experiments are conducted on the simulation environment provided by DyRET. The results show that it is possible to evolve a stable controller that can walk on various morphology for DyRET platform while not all the evolved controllers learned the relationship between the controller and the morphology. Also, the comparison of EA's revealed that fitness-based EA could produce controllers that are as good as the behavior diversity EA and MOEA for the given constraints. The experiment also indicates that combining all three behavior descriptors can generate most fit controllers while not statistically significant when comparing the walking distance.