Bayesian optimization of gaits on a bipedal SLIP model

Abstract

The Spring-Loaded Inverted Pendulum (SLIP) model has been in play for many years as the most plausible explanation for both walking and running gaits. Although the current knowledge on this model has advanced significantly over the years, a more in-depth analysis of the vast parameter space was always hindered by inefficient searching tools. Beyond finding a stable representation for different velocities, an ideal search method also needs to consider different gait alternatives and quickly optimize its velocity to the maximum possible. In this work we utilize a bipedal SLIP model and present Bayesian Optimization (BO) as a fast alternative to explore gaits within a large parametric search space. Our results show that BO can efficiently be used to find stable parameters for different gait patterns, where the best behavior for four different gaits was usually found with less than 40 iterations. These results provide a powerful tool to tune a robot's behavior on the fly, and future robots can be designed to easily adapt their gaits to new environmental conditions.