Fitness Shaping on SLIP Locomotion Optimization

Abstract

Walking robots have been a thriving topic for years, and their impact in our future is undeniable. Through different walking techniques machines match their control parameters to environmental conditions, and in both simulation and real-world this control optimization always requires many iterations to find the best parameter. We benchmark four optimization methods and two fitness shaping methods to assess how fast a locomotion model, with two control parameters, can converge to stability. We find that a best overall solution does not exist, with inference-based methods such as Bayesian Optimization in some cases being as inefficient as Random Search. Fitness Shaping using additional information provided by the simulation after termination is shown to improve optimization speed in the presence of running gaits. Additionally, our results validate Bayesian optimization as the fastest optimization method for walking gaits, and present Neural Networks as the fastest for running gaits and. In the presence of so many methods and models, this comparative study aims to clarify the potential gains for optimization methods in bipedal locomotion.