Step Selection in Data-Driven Geometric Gait Optimization

Geometric gait optimization has opened a frontier for the improvement of robotic gaits and analysis of animal behavior. Borrowing tools from the data-driven modeling of oscillators, we have developed a streamlined method for producing a data-driven geometric model of the dynamics in the neighborhood of a gait. This model informs a gradient ascent optimizer, for which the user can specify a variety of locomotive goal functions. We have found that step size selection has a strong effect on convergence and the quality of selected gaits. In our work, a new gait is constrained geometrically such that it lies within the neighborhood of trajectories collected in the experimental data of the previous gait. This prevents the system from venturing into unknown volumes of the gait design space. We select a step size by performing a line search along the gradient and within the sampled data. For a nine-link Purcell swimmer with injected noise, this step size selector allows optimization of gaits under a variety of system noise regimes. We anticipate that this algorithm's ability to handle noise and step safely will enable its application to hardware in the loop optimization on a broad class of mechanisms.