Multi-Objective Optimization for Speed and Stability of a Sony AIBO Gait

摘要

Locomotion is a fundamental facet of mobile robotics that many higher level aspects rely on. However, this is not a simple problem for legged robots with many degrees of freedom. For this reason, machine learning techniques have been applied to the domain. Although impressive results have been achieved, there remains a fundamental problem with using most machine learning methods. The learning algorithms usually require a large dataset which is prohibitively hard to collect on an actual robot. Further, learning in simulation has had limited success transitioning to the real world. Also, many learning algorithms optimize for a single fitness function, neglecting many of the effects on other parts of the system. As part of the RoboCup 4-legged league, many researchers have worked on increasing the walking/gait speed of Sony AIBO robots. Recently, the effort shifted from developing a quick gait, to developing a gait that also provides a stable sensing platform. However, to date, optimization of both velocity and camera stability has only occurred using a single fitness function that incorporates the two objectives with a weighting that defines the desired tradeoff between them. However, the true nature of this tradeoff is not understood because the pareto front has never been charted, so this a priori decision is uninformed. This project applies the Nondominated Sorting Genetic Algorithm-II (NSGA-II) to find a pareto set of fast, stable gait parameters. This allows a user to select the best tradeoff between balance and speed for a given application. Three fitness functions are defined: one speed measure and two stability measures. A plot of evolved gaits shows a pareto front that indicates speed and stability are indeed conflicting goals. Interestingly, the results also show that tradeoffs also exist between different measures of stability.