This thesis presents a framework for developing a library of gaits for highly non-linear, hyper-redundant, potentially compliant robotic systems. Examples of such systems that motivate this work include tensegrity robots, which combine both soft and rigid elements, and snake-like salamander robots. A library of gaits for such complex robots can be integrated with a search-based method so as to achieve more efficient exploration of the underlying state space when solving trajectory planning problems.

The first component of the work corresponds to the definition of a Central Pattern Generator (CPG) for a spherical tensegrity robot inspired by similar solutions in the domain of modular, bio-inspired snake robots. The CPG provides a reparametrization of the underlying system, which can easily result in the generation of rhythmic gaits. The second component is a novel framework for simultaneously discovering effective gaits along different directions of motion by searching the space of CPG parameters. The framework defines multiple objectives, which are maximized though a parallel Bayesian Optimization (BO) process. The samples, which correspond to different gait parameters, are biased towards areas of previously observed high reward using a set of binary kNN classifiers with on-line updates. This integrated method is shown to be more efficient than Monte Carlo sampling of gait parameters or BO without classification or the classification only approach. The evaluation is performed in simulation using a high-dimensional spherical tensegrity robot.