Motion Preference Learning

Abstract

In order to control systems to meet subjective criteria, one would like to construct objective functions that accurately represent human preferences. To do this, we develop robust estimators based on convex optimization that, given empirical, pairwise comparisons between motions, produce both (1) objective functions that are compatible with the expressed preferences, and (2) global optimizers (i.e., “best motions”) for these functions. The approach is demonstrated with an example in which human and synthetic motions are compared.