Autonomous robotic systems are becoming prevalent in our daily lives. Many robots are still restricted to manufacturing settings where precision and repetition are paramount. However, autonomous devices are increasingly being designed for applications such as search and rescue, remote sensing, and tasks considered too dangerous for people. In these cases, it is crucial to continue operation even when some unforeseen adversity decreases performance levels---a robot with diminished performance is still successful if it is able to deal with uncertainty, which includes any unexpected change due to unmodeled dynamics, changing control strategies, or changes in functionality resulting from damage or aging.The research presented in this dissertation seeks to improve such autonomous systems through three evolution-based techniques. First, robots are optimized offline so that they best exploit available material characteristics, for instance flexible materials, with respect to multiple objectives (e.g., speed and efficiency). Second, adaptive controllers are evolved, which enable robots to better respond to unforeseen changes to themselves and their environments. Finally, adaptation limits are discovered using a proposed mode discovery algorithm. Once the boundaries of adaptation are known, self-modeling is applied online to determine the current operating mode and select/generate an appropriate controller.These three techniques work together to create a holistic method, which will enable autonomous robotic systems to automatically handle uncertainty. The proposed methods are evaluated using robotic fish as a test platform. Such systems can benefit in multiple ways from the integration of flexible materials. Moreover, robotic fish operate in complex, nonlinear environments, enabling thorough testing of the proposed methods.
Read
