Evolving L-systems to generate virtual creatures

Abstract

Virtual creatures play an increasingly important role in computer graphics as special effects and background characters. The artificial evolution of such creatures potentially offers some relief from the difficult and time consuming task of specifying morphologies and behaviors. But, while artificial life techniques have been used to create a variety of virtual creatures, previous work has not scaled beyond creatures with 50 components and the most recent work has generated creatures that are unnatural looking. Here we describe a system that uses Lindenmayer systems (L-systems) as the encoding of an evolutionary algorithm (EA) for creating virtual creatures. Creatures evolved by this system have hundreds of parts, and the use of an L-system as the encoding results in creatures with a more natural look.

Introduction

As computers become more powerful it is increasingly our design ability, rather than computing power, that limits the richness of virtual worlds. Evolutionary algorithms (EAs), a technique inspired by biological evolution, have shown much promise in automating the process of producing creatures for virtual environments, yet the most recent work in this area [1], [2] has produced ungainly creatures with less than 50 components. The asymmetries of these creatures is a result of using a direct encoding, an explicit encoding with a one-to-one mapping from genotypic encoding to creature-part. As direct encodings have no re-use, symmetries and regularities do not occur, except by chance, and evolved structures tend to be unnatural looking.

In this we return to the spirit of [3], in which a graph structure was used as a generative encoding for the creatures. A generative encoding is a developmental method for producing a structure using a set of grammatical re-writing rules or a procedural process, not unlike a computer program with re-usable sub-procedures. Designs produced by a generative encoding have fractal-like self-similarities, giving them an organic look, and have been shown to have better scaling properties than direct encodings [4], [5]. Here we use Lindenmayer systems (L-systems) [6] as a more powerful, and general purpose, generative encoding than that of [3] to achieve moving creatures with hundreds of parts whose structure is more natural looking than [1], [2].

More common than co-evolving morphology and controller has been work evolving controllers for pre-specified morphologies. Control systems for these works has included stimulus-response rules [7], [8], neural controllers [9], and genetic programs [10]. The controllers of the creatures in this work are oscillator circuits; each actuated joint is controlled by an oscillator with its own frequency and relative phase offset. To achieve controllers that are reactive to the environment joints can be controlled by recurrent neural networks, as with [1], [2], [3], by including neural-network construction commands into the encoding language [11].

In the following sections, we first outline the design space and describe the components of our generative design system, then we present our results and finally close with a discussion and conclusion of our work.