Hierarchical development of physics-based animation controllers

Abstract

In this work, a developmental hierarchy is applied to the evolution of a relatively complex physics-based character animation controller. This means that the artificial neural network that makes up that controller is composed from a number of interdependent sub-networks; the control modules. It is hypothesized that evolving these modules one-by-one, with each of them dependent on its predecessors, will allow evolution to converge faster, and possibly to better results, than for a pair of baseline controllers. Both muscle-based actuation and joint torque-based actuation are tested, but only the latter succeeds. It is demonstrated that developmental hierarchies can lead to faster evolutionary convergence, while dealing with compound animations more adequately.