INTERACTIVE EVOLUTIONARY DESIGN OF MOTION VARIANTS

Jonathan Eisenmann

Computer Science & Engineering, The Ohio State University, 2015 Neil Ave, Columbus, OH, U.S.A.

Matthew Lewis

Advanced Computing Center for the Arts & Design, The Ohio State University, 1224 Kinnear Rd, Columbus, OH, U.S.A.

Bryan Cline

Computer Science & Engineering, The Ohio State University, 2015 Neil Ave, Columbus, OH, U.S.A.

Keywords:

Animation, Evolutionary Graphics, Interaction Techniques.

Abstract:

This paper presents an intuitive method for novice users to interactively design custom populations of stylized,

heterogeneous motion, from one input motion clip, thus allowing the user to amplify an existing database of

motions. We allow the user to set up lattice deformers which are used by a genetic algorithm to manipulate

the animation channels of the input motion and create new motion variations. Our interactive evolutionary

design environment allows the user to traverse the available space of possible motions, presents the user with

populations of motion, and gradually converges to a satisfactory set of solutions. Each generated motion

sequence can undergo a motion ﬁltering process subject to user-speciﬁed, high-level metrics to produce a

result crafted to ﬁt the designer’s interest.

1 INTRODUCTION

The human visual system possesses acute pattern

recognition abilities which can expose unnatural qual-

ities in synthetic crowd animation if insufﬁcient mo-

tion variation is present. Often times, crowd anima-

tions will also need to display particular meaningful

expressive qualities in order to set the mood for a

scene. The motions of crowd agents must not only be

diverse and expressive, but they also often need to ap-

pear visually coherent. Therefore a certain amount of

similarity between some of the individuals in a crowd

is desirable.

A common method for achieving variation within

a crowd consists of providing agents with a broad

library of motions to choose from, using behavioral

rules and blending between the motions as the agents

transition from one action to another. Behavioral se-

lection of motion clips from a library should achieve

good results if the library is large enough to provide

satisfactory variation. However, this method is not al-

ways feasible for a small studio without access to an

extensive library of motion clips or for a crowd ani-

mation with novel motion that cannot be motion cap-

tured. Furthermore, it does not provide an easy way

to tune the crowd motion to ﬁt a particular style.

We have developed an interactive evolutionary de-

sign system that will help designers create a diverse

set of crowd motions that belong to a few families of

expressive motion and can be tailored to ﬁt the de-

signer’s preferences. Our system generates a range

of motion variants using a single input motion. The

input can be in the form of keyframe animation se-

quences and motion capture data should work as well.

Therefore our method can not only be easily used to

amplify a small database of motion clips for crowd

animation but also as an aid in the ideation process in

motion development for a character.

The system alters the input animation via user-

deﬁned free form deformations which reshape the an-

imation channels specifying the motion. Attributes

that control deformations act as genes within our de-

sign environment’s genetic algorithm. The resulting

population of motions can then be ﬁltered using sev-

eral techniques described in this paper that enforce

a set of constraints. Physical constraints can be uti-

lized to deﬁne the physical properties of each motion.

The designer views these ﬁltered motions at each gen-

eration to interactively determine the ﬁtness of the

crowd members. Through continued interaction, the

127

Eisenmann J., Lewis M. and Cline B. (2009).

INTERACTIVE EVOLUTIONARY DESIGN OF MOTION VARIANTS.

In Proceedings of the International Joint Conference on Computational Intelligence, pages 127-134

DOI: 10.5220/0002320301270134

 SciTePress

designer can guide the population to one or more ar-

eas of the space of solutions which portray an inter-

esting and desirable set of emotions or expressions.

Interesting motions can be saved in a library for use

later in the design process or as part of the ﬁnal set of

motions.

Our system also has the beneﬁt of providing the

designer with not just static samples from a space

of possibilities, but also a neighborhood of similarity

around each motion that can be used to tweak a sin-

gle motion if desired. This ﬂexibility is an improve-

ment over a static library of motion capture clips in

the same way that images produced by a procedural

shader are more ﬂexible than a set of scanned images.

Motions generated by the system can be used to in-

spire animators with new movement ideas or stored

in a database for generating crowd animation.

2 RELATED WORK

Our approach draws from previous work in areas

ranging from genetic algorithms to motion variation

techniques and crowd design methods.

Interactive evolutionary design, unlike many ge-

netic algorithm applications, lets the user interac-

tively determine the ﬁtness of the evolved solutions.

Sims applied this technique to a wide array of ap-

plication areas, including procedural models and tex-

tures (Sims, 1993). In the application area of char-

acter animation Lim and Thalmann presented an in-

tuitive interface for searching through a design space

by selecting one from a pair of options in “tourna-

ment” style (Lim and Thalmann, 2000). Ventrella

introduced a physically-based animation and creature

design system that allows users to interactively deter-

mine ﬁtness of individuals explicitly by altering ﬁt-

ness functions as well as implicitly by selecting indi-

viduals of high ﬁtness. His motion model consisted

of single DOF joints activated by parameterized sinu-

soidal functions for the sake of fast interactivity (Ven-

trella, 1995). We extend his novel work in this area

by introducing a ﬂexible framework for generating

motion which allows for a wide range of expressive

motion.

There are a wide variety of ways to generate mo-

tion variation. Sung proposes a method for synthe-

sizing motion clips for crowd animation using motion

graphs (Sung, 2007). Amaya et al. use signal pro-

cessing techniques to embed varying emotions into a

neutral motion clip (Amaya et al., 1996). Chi et al.

introduce a system for modifying human motion us-

ing a Laban inspired effort and shape parameteriza-

tion (Chi et al., 2000). Neff and Fiume introduce a

method that uses both low-level and composite prop-

erties to edit character motion iteratively and interac-

tively (Neff and Fiume, 2005). Wang et al. present a

motion signal ﬁlter method for making a motion more

animated or cartoon-like (Wang et al., 2006). Gle-

icher has provided an extensive survey of constraint-

based motion editing techniques with particular atten-

tion paid to “per-frame inverse kinematics plus ﬁlter-

ing” techniques (Gleicher, 2001).

Crowd design has typically centered on the prob-

lems of navigation and behavioral patterns. For ex-

ample, Kwon et al. present a graph-based approach

for intelligently deforming group motion trajecto-

ries (Kwon et al., 2008). In addition, Treuille et al. in-

troduce a particle-based solution for crowd navigation

without the use of agent-based dynamics (Treuille

et al., 2006). Li and Wang have used interactive evo-

lutionary design to tune the parameters in a virtual

force based system (Li and Wang, 2007). Musse and

Thalmann created a real-time interactive system with

three methods for editing crowd behaviors (Musse

and Thalmann, 2001). Sung et al. presented an efﬁ-

cient statistics-based scalable model for goal-directed

crowd behavior that can satisfy duration, orientation,

position, and pose constraints for individuals within

the crowd (Sung et al., 2005). Similarly, the crowd

simulation middleware in use today focuses mainly

on navigation and behavioral rules while relying on

extensive motion libraries to provide variation of mo-

tion (Massive Software, 2009). Time warping and

blending between actions selected by behavior mod-

els are a popular means of generating motion variation

in crowds.

Our method seeks to complement prior work in

this area by addressing the need that crowd design-

ers may have to craft the expressive motion of crowd

members. We choose to focus exclusively on gener-

ating the variety of expressive motion portrayed by

the crowd. We encourage use of our system in con-

junction with the existing methods for navigation and

behavior.

3 INTERACTIVE

EVOLUTIONARY DESIGN

There are often two types of designers involved in the

evolutionary design paradigm: the meta-designer and

the designer. The meta-designer deﬁnes the search

space by creating the original prototype to be altered

by the genetic algorithm. In our case this is a mo-

tion clip with parameterized deformers applied to it.

The meta-designer also deﬁnes which attributes are

adjustable and gives ranges to these attribute values.

IJCCI 2009 - International Joint Conference on Computational Intelligence

128

Figure 1: Interactive evolutionary design interface - the phe-

notypes are arranged in a grid for easy viewing and selection

with the mouse.

There must be enough meaningful attributes to pro-

vide a sufﬁciently rich search space for the algorithm

to traverse. The designer who directs the evolution

process will be able to explore solutions that exist in

the space deﬁned by the meta-designer.

The technique described in this paper assumes that

a parametric model with a rich search space has al-

ready been designed. Problems in the ﬁeld of para-

metric modeling are both challenging and interesting,

but they are beyond the scope of this paper. Further-

more, the solutions to parametric modelling problems

are typically domain speciﬁc. Bezirtzis, et al. have

given attention to some of the general concerns of

parametric search space design in their industrial de-

sign case study (Bezirtzis et al., 2007) with emphasis

on the fact that the design of a parametric space is

an iterative process that can only be veriﬁed empiri-

cally. In general, the areas of high ﬁtness in a para-

metric search space for interactive evolutionary de-

sign should be much larger than those in typical non-

interactive evolutionary design applications. This is

due to the small number of individuals that can be

viewed in one population by the human visual sys-

tem and the low number of generations required for

convergence in a reasonable interactive environment.

Also, discontinuities or sharp features in the search

space should be avoided as much as possible. Other-

wise, the designer will ﬁnd that slight changes in gene

values can result in dramatically different phenotypic

expression.

In our system the list of attributes belonging to the

parametric model corresponds to a genotype which is

represented by a ﬁxed-length array of ﬂoating point

numbers (gene values). The phenotype in our sys-

tem is the motion produced when these gene values

are mapped to the model’s parameters. The designer

chooses a set of parents that will participate in the re-

production process for the next generation. Our ge-

netic algorithm chooses two distinct, random parents

from this set each time it produces a new offspring. It

copies the genes from one parent and then switches to

copying the other parent’s genes with a user-deﬁned

crossover probability. The genes are then mutated

given a user-deﬁned probability by adding a random

value between -1 and 1 scaled by a user-deﬁned mu-

tation amount.

There are many variants on reproduction algo-

rithms for genetic search, and we are not married to

this particular algorithm. However, the ability to let

the user choose more than just two parents for the next

generation provides a very nice property for the evo-

lutionary design of crowd motions: The designer is

able to simultaneously engineer distinct subspaces of

motion for an entire crowd. Alternately, the designer

can save interesting motions from exploration of one

area of the search space into a motion library and can

reintroduce these motions later, when exploring other

areas of the search space. In addition, we encour-

age designers to explore different areas of the space

by using a functionality in our interface that allows

designers to backstep to previous generations and ex-

plore characteristics found in previously ignored parts

of the population. It is important to note that our sys-

tem cannot provide enough variation from one input

motion to create all the types of action that might be

required in a crowd scene. For example, it cannot turn

a walking motion into a sitting motion or vice versa.

Instead it provides varieties of walking motions that

can be placed in a motion clip library along with va-

rieties of sitting motions that were generated from a

separate input motion.

A designer cannot design just any particular pre-

conceived motion using our system. It will only be

possible to generate motions that exist in the avail-

able search space. Furthermore, if a designer has a

speciﬁc visual expressive quality in mind, it can often

be modelled more quickly using conventional meth-

ods. This system is more useful for discovering novel

motions and expressive qualities during exploration

of the search space. Interactive evolutionary design

has typically been aimed at aiding designers in the

ideation phase of development of a character or idea,

and the same applies here. Our system can be used

to craft a novel motion for an individual character or

a set of characters, but it may also serve as inspira-

INTERACTIVE EVOLUTIONARY DESIGN OF MOTION VARIANTS

129

tion in the planning stages of new character motion

development.

4 MOTION GENERATION

Our method requires interaction from the designer at

each iteration of the algorithm (see ﬁgure 2). We be-

gin by preparing the input motion for modiﬁcation by

the genetic algorithm which, at the discretion of the

designer,sends the resulting motion through a number

of constraint-based ﬁlters and a dispersion algorithm.

Figure 2: Pipeline Overview.

4.1 Input

The underlying representation of our motion evolu-

tion system consists of the set of animation channels

taken from the input motion clip. We do not use

a simpliﬁed motion model based on sinusoidal sig-

nals or any other similar technique. Therefore our

system is not limited to cyclical motions, but can be

used to generate variants on any type of motion that

can be keyframed. Each animation channel is con-

verted to piecewise Bezier spline geometry based on

the data frames and their tangents. The meta-designer

sets up deformers (Sederberg and Parry, 1986) to al-

ter these splines and also deﬁnes which attributes of

the deformers can be changed by gene values from

the genetic algorithm. The splines are then altered by

these deformers during the evolutionary design pro-

cess. Once the spline geometry has been changed, the

information is then transferred back to the animation

channels of the individual being evolved. The result-

ing motion is displayed to the designer. We set the

system up this way so that any 2D deformation that

does not violate the one-to-one correspondence of the

function represented by the animation channels may

be applied, giving the meta-designer complete ﬂexi-

bility in creating and controlling the space of possible

motions. See ﬁgure 3 for an example of how lattice

deformers can be applied to an animation channel.

Ideally, the parametric models produced via the

application of lattice deformers will provide a wide

range of variation with few problems. However, some

parameters may ﬁght each other and occasionally lead

to unrealistic or undesirable results. We address this

by adding a layer of ﬁlters to the pipeline that auto-

matically detect problematic results above a tolerance

threshold and generate individuals to replace these

unacceptable phenotypes. Though it would be ideal to

correct these phenotypes, we choose to replace them

in the interest of faster interactivity. This will be dis-

cussed further below. These ﬁlters, though general in

some regards, should speciﬁcally apply to the type of

motion being generated. In our case, any number of

high-level qualities of locomotive animation can be

addressed here, but we choose to focus on two partic-

ular properties of motion where we have observed the

abuses of large variation: balance and self-collision.

4.2 Balance

In order to measure the level of balance in individual

phenotypes and decide if they are acceptable we have

adopted the zero moment point (ZMP) algorithm of

Tak et al. ZMP in dynamic motion analysis is sim-

ilar to the center of gravity in the static case. It is

deﬁned as the point on the ground plane under a char-

acter where there is zero moment. In other words, if

this point were modeled as a joint between the char-

acter and the ground, there would be no actuation at

this joint. As a character moves, the ZMP will create

a trajectory over the ground plane.

Our balance ﬁlter ensures that the ZMP is always

within the character’s support area. The support area

is the convex hull of the contact area between the

feet and the ﬂoor. This deﬁnition encapsulates both

the single and double stance phases of bipedal mo-

tion and allows for seamless calculation between the

two phases. In some cases, the designer may want the

motion of individuals in the crowd to appear exag-

gerated or cartoon-like. In a case such as this, a pre-

cise balance constraint might cause more harm than

good. Therefore we have implemented the ﬁlter in

such a way that the designer has control over the al-

lowable balance error. We deﬁne balance error as the

IJCCI 2009 - International Joint Conference on Computational Intelligence

130

Figure 3: A variety of motions achieved using free form deformations on the rotation channels for the arm. Only the elbow

joint values are represented by the graphs, but the same deformations are being applied to the shoulder in the animation

sequence to the right. The X-axis of each lattice controls the passage of time (labeled in frames), and the Y-axis values control

the rotation angle of the joint at any point in time. The ﬁrst curve and accompanying animation sketch represent the original

motion. Notice the lattice is in its initial, undeformed state.

distance from the ZMP to the closest point on the sup-

port area, and we sum this error over the entire motion

sequence.

∑

i=1

kZMP

− closestPoint(ZMP

, K · SA

)k (1)

F is the number of frames in the animation, ZMP

contains the zero moment point trajectory over all the

frames, and SA contains the support area hull over all

the frames. We also allow the user to enforce a lesser

or greater degree of balance on the entire sequence

by adding a custom support area scale factor K. The

value of K can be adjusted through the system’s inter-

face thus allowing the user to shrink the support area

to constrain the ZMP error more towards the center of

support if more stable motion is required.

4.3 Self-Collision

Since the variation of arbitrary animation channel at-

tributes may introduce self-collisions, we have devel-

oped a set of tests to ﬁlter out these self-colliding mo-

tions. We employ a simple bounding box method,

testing for overlap of world-aligned bounding boxes

and then calculate the volume of object aligned

bounding boxes if two links are found to be intersect-

ing. We then sum the volumes of all the intersections.

∑

j =1

∑

k =j+1

{volume(O

∩O

) | A

j,k

= 0, W

∩W

0} (2)

L is the number of links or bones in the char-

acter, O contains the object-aligned bounding boxes

of the links, W contains the world-aligned bounding

boxes, and A is an adjacency matrix which describes

the spatial adjacency of the links. The above summa-

tion is accumulated over every frame of the motion

sequence. By summing the total volume we avoid

over-penalizing quick, grazing collisions while still

penalizing quick, high-volume collisions as well as

slow, grazing collisions. We ignore collisions that are

design artifacts of the given model’s geometry as well

as collisions between adjacent links. Since we do not

test at the polygon-polygon level for intersection in

order to reduce running time, this ﬁlter only serves as

a heuristic and not an exact measure of self-collision.

Nevertheless, in practice, it gives a good indication of

the level of self-collision for a given phenotype, espe-

cially since we want to keep computation to a mini-

mum in the interest of higher levels of interaction.

INTERACTIVE EVOLUTIONARY DESIGN OF MOTION VARIANTS

131

4.4 Replacement Method

It would be preferable to ﬁx unacceptable phenotypes

and bring them back into the allowable space of mo-

tion deﬁned by our balance and self-collision con-

straints. In fact, the method of Tak et al. optimally

transforms unbalanced animation sequences into bal-

anced motions (Tak et al., 2002). There are also meth-

ods for correcting self-colliding animation via inverse

kinematics optimization (M¨uller, 2004). However,we

do not attempt to ﬁx unacceptable motions because

this would require an optimization process. Optimiza-

tion would introduce longer wait times between gen-

erations, reducing the interactivity of our evolutionary

design environment. Instead, we generate a new indi-

vidual motion to replace the old one and resubmit it

to the ﬁlters for evaluation. We are able to proceed in

this way because our space of possible motions was

created from a balanced original motion that was free

from self-collisions and so the rich search space of

variants will provide a viable, balanced replacement

motion within a few iterations. Furthermore, the re-

production process uses the user-selected motions as

input for any newly generated replacement motion so

the offspring will most likely have similar balance

and self-collision error to the parents’ errors. Even

so, generating replacement individuals with our re-

production algorithm may not always result in an ac-

ceptable individual, especially if the user’s constraints

are too restrictive or if the search space is not rich

enough. If this occurs, we let the replacement process

run for a user-deﬁned maximum number of iterations

and then force it to move on, replacing the unbalanced

individual with the most balanced option found so far.

4.5 Diversity

Although duplicate motions in a crowd are harder to

spot than duplicate appearances, they can be spotted

just as easily whether or not each character has a dif-

ferent appearance. Moreover, varied motion between

two visually equivalent individuals can help to ob-

scure the fact that they havea similar appearance(Mc-

donnell et al., 2008). It follows that diversity of mo-

tion provides desirable qualities for crowd animation.

Genetic algorithms by their very nature seek conver-

gence to a speciﬁc area of the search space. Because

we want to design coherent sets of motion and be-

cause proximity in the search space corresponds to

phenotypic similarity in our system, this convergence

is beneﬁcial. However, it can also be problematic

since we want to ﬁnd a diverse group of motions. We

do not want the algorithm to converge so far that the

resulting population becomes homogenous. The de-

signer in the interactive evolutionary design paradigm

exercises ultimate control over how far the popula-

tion converges. As mentioned earlier, if the designer

chooses a variety of individuals as parents in one gen-

eration, the chances of diversity in the next generation

are greater.

In order to encourage diversity in the population

and ensure sufﬁcient sampling of the local search

space surrounding the designer’s regions of interest,

we run the evolution dispersion algorithm introduced

by Marks et al. (Marks et al., 1997) at each genera-

tion. This algorithm alters the genes of individuals in

a way that will discourage similarity between pheno-

types. We measure the phenotypic difference between

two individuals by comparing their point clouds. Our

method is somewhat similar to the technique used by

Kovar, et al. in their paper on motion graphs (Kovar

et al., 2002). The point cloud for a character consists

of the positions of a subset of the character’s vertices

over time. We consider the distance between point

clouds to be the sum of the root mean squared error

between all the corresponding vertices of two point

clouds. We allow the designer to specify which parts

of the model to choose vertices from when making

this comparison. The designer can also specify the

level of dispersion required at each generation.

4.6 Using the Filters

Because the motion ﬁlters require extra computation,

using them will inevitably result in longer wait times

between generations. The balance and self-collision

ﬁlters are most useful during the ﬁrst couple gener-

ations of the design process when the algorithm is

sampling a wide area of the search space. These ﬁl-

ters make the designer’s job easier by avoiding the

areas of the search space where two or more param-

eters ﬁght each other and produce unwanted motion

artifacts. As the design process progresses and the

algorithm begins to converge, these ﬁlters should be

turned off to speed up the interactivity of the system.

In contrast, the diversity ﬁlter should only be used

near the end of the design process when the algorithm

is sampling a smaller area of the search space. Using

the above ﬁlters in this way will maximize the ﬁtness

of the options presented to the designer throughout

the process in a way that is customizable to ﬁt the de-

signer’s needs.

5 RESULTS

We implemented our interactive evolutionary design

motion generation method in a layer of MEL and

IJCCI 2009 - International Joint Conference on Computational Intelligence

132

Python code over Autodesk’s Maya environment. We

chose to use Maya because it is ﬂexible enough to

apply evolutionary design not only to animation, but

also to modeling, texturing, and special effects do-

mains. We also chose this software environment be-

cause most of the designers at our research facility are

familiar with its interface, and we would like them to

be able to create and evolve designs from their own

parametric models.

Table 1: Average time (sec) and average number of rejected

phenotypes while producing a generation of size 25.

Filter None Balance Collide Disperse

Time 10.0 76.0 47.8 28.0

Reject - 1.5 3.5 2

Our hardware consists of an Intel Xeon 2.66 GHz

cpu with 4 GB of RAM and an Nvidia Quadro FX

5600 graphics card. Performance metrics of our cur-

rent implementation under varying conditions can be

seen in Table 1. Each ﬁlter was set up with a max-

imum of 2 replacements per individual except for

the dispersion ﬁlter which had a maximum of two

replacements per population.The number of replace-

ments required at each generation depends on the

richness of the search space as well as on the de-

signer’s selections and the constraint thresholds set by

the designer. The time required to replace an individ-

ual that does not meet the constraints is approximately

equal to the time required to generate the original in-

dividual. The running time is dependent on the size

of the population so smaller populations will take less

time per individual.

For these tests, we created a parametric walk cy-

cle model with lattice deformers on 14 different an-

imation channels. Note, however, that any type of

deformer available in Maya may be utilized. The de-

formers’ transformations were controlled by 14 ﬂoat-

ing point numbers from the ﬁxed-length array that

formed the genotype of each individual. Although the

size of the population is adjustable by the user, we

generally use populations of size 25 because larger

populations are harder for the human visual system to

fully and easily comprehend. We are currently work-

ing to improvethe running times for the various ﬁlters

in the interest of better interactivity.

Figure 4 shows a sampling of the variety of mo-

tions that can be produced from a single parameter-

ized walk cycle. These motions are shown in an ani-

mation sketchbook style where every fourth frame is

drawn to show the change in form over time. We are

displaying the motions in animation sketchbook style

only for the purposes of the paper. In our evolutionary

design interface, the motions are actually displayed as

Figure 4: A variety of sixteen motions evolved from a single

walk cycle motion clip.

time-varying animation.

We have conducted a user study to determine if

our software truly provides an intuitive way for de-

signers to create varieties of motion. In this study we

asked three graduate students from the college of de-

sign who were already familiar with the Maya soft-

ware environment to participate. After a brief 10-

15 minute tutorial on how to use the system, each

designer explored the space of options provided by

the deformers in our parametric model of a walk cy-

cle motion and decided on a style of motion to in-

teractively develop. The entire process, including

the tutorial, search space exploration, and motion de-

sign took anywhere from an hour and a half to two

hours for each designer. The designers all felt that

the system was easy to learn and enjoyed experienc-

ing the interactive evolutionary design process. They

agreed that the motions produced would make for

interesting, varied background crowd character mo-

tions or as a fertile starting point for the develop-

ment of a character motion with an individualized

style. However, they said that the motion produced

would need to be reﬁned with a high-level of con-

trol over speciﬁc keyframes if it were to be used for

foreground or hero characters. The populations they

created (See the animations on the project web site

at http://accad.osu.edu/Projects/Evo/) including a set

of energetic dancers, a group of feminine characters,

and a mob of zombies demonstrate the wide variety of

motions that are acheivable from just one input mo-

tion.

6 FUTURE WORK

There are many more ways to ﬁlter the individual mo-

tions that could prove to be helpful to the crowd de-

signer. A few such ﬁlters that could be useful in-

clude: joint torque analysis based on limits from a

comfort model (Ko and Badler, 1996), various meth-

INTERACTIVE EVOLUTIONARY DESIGN OF MOTION VARIANTS

133

ods of psychological analysis (aggressiveness, energy

level, coordination, etc.), Laban movement analysis

(effort and shape), as well as gender or age analysis.

We are investigating methods for implementing the

balance and self-collision ﬁltering process in parallel

on subsets of the population using our quad-core pro-

cessors which should speed up computation times and

enhance the interactivity of the system.

7 CONCLUSIONS

Our method presents a novel approach to evolving

families of expressive motion, making it easier for a

crowd designer to quickly and intuitively ﬁnd a satis-

fying combination of motion variations for a speciﬁc

application. This method could prove especially use-

ful to those who do not have access to motion capture

facilities or cannot afford to spend time capturing a

wide range of motion clips. Our interaction model al-

lows the user to view and make decisions about entire

generations at once, and our reproduction algorithm

allows for evolution of multiple (even mutually ex-

clusive) styles of motion simultaneously. Our use of

user-deﬁned constraints plus the designer’s selections

as the determination of ﬁtness exempliﬁes a hybrid

system that seeks to maximize the designer’s time and

attention in the evaluation of populations by ﬁltering

out the individuals who do not meet the given criteria.

REFERENCES

Amaya, K., Bruderlin, A., and Calvert, T. (1996). Emotion

from motion. In Graphics Interface ’96, pages 222–

229.

Bezirtzis, B. G., Lewis, M., and Christeson, C. (2007). In-

teractive evolution for industrial design. In C&C ’07:

Proceedings of the 6th ACM SIGCHI Conference on

Creativity & Cognition, pages 183–192, New York,

NY, USA. ACM.

Chi, D., Costa, M., Zhao, L., and Badler, N. (2000). The

emote model for effort and shape. In SIGGRAPH ’00

Proceedings, pages 173–182, New York, NY, USA.

ACM Press/Addison-Wesley Publishing Co.

Gleicher, M. (2001). Comparing constraint-based motion

editing methods. Graphical Models, 63(2):107–134.

Ko, H. and Badler, N. I. (1996). Animating human locomo-

tion with inverse dynamics. Computer Graphics and

Applications, IEEE, 16(2):50–59.

Kovar, L., Gleicher, M., and Pighin, F. (2002). Motion

graphs. In SIGGRAPH ’02 Proceedings, volume 21,

pages 473–482, New York, NY, USA. ACM Press.

Kwon, T., Lee, K. H., Lee, J., and Takahashi, S. (2008).

Group motion editing. In SIGGRAPH ’08 Proceed-

ings, pages 1–8, New York, NY, USA. ACM.

Li, T.-Y. and Wang, C.-C. (2007). An evolutionary ap-

proach to crowd simulation. In Autonomous Robots

and Agents, pages 119–126.

Lim, I. S. and Thalmann, D. (2000). Tournament selec-

tion for browsing temporal signals. In Symposium on

Applied Computing ’00 Proceedings, pages 570–573,

New York, NY, USA. ACM.

Marks, J. et al. (1997). Design galleries: a general ap-

proach to setting parameters for computer graphics

and animation. In SIGGRAPH ’97 Proceedings, pages

389–400, New York, NY, USA. ACM Press/Addison-

Wesley Publishing Co.

Massive Software (2009). Massive prime. Accessed online

www.massivesoftware.com/prime/.

Mcdonnell, R., Larkin, M., Dobbyn, S., Collins, S., and

O’Sullivan, C. (2008). Clone attack! perception of

crowd variety. In SIGGRAPH ’08 Proceedings, vol-

ume 27, pages 1–8, New York, NY, USA. ACM.

M¨uller, A. (2004). Collision avoiding continuation method

for the inverse kinematics of redundant manipulators.

In Robotics and Automation ’04 Proceedings, vol-

ume 2, pages 1593–1598 Vol.2.

Musse, S. R. and Thalmann, D. (2001). Hierarchical model

for real time simulation of virtual human crowds. Vi-

sualization and Computer Graphics, IEEE Transac-

tions, 7(2):152–164.

Neff, M. and Fiume, E. (2005). Aer: aesthetic exploration

and reﬁnement for expressive character animation. In

SCA ’05 Proceedings, pages 161–170, New York, NY,

USA. ACM Press.

Sederberg, T. W. and Parry, S. R. (1986). Free-form de-

formation of solid geometric models. SIGGRAPH ’86

Proceedings, 20(4):151–160.

Sims, K. (1993). Interactive evolution of equations for pro-

cedural models. The Visual Computer, 9(8):466–476.

Sung, M. (2007). Continuous motion graph for crowd sim-

ulation. In Technologies for E-Learning and Dig-

ital Entertainment, volume 4469, pages 202–213.

Springer Berlin / Heidelberg.

Sung, M., Kovar, L., and Gleicher, M. (2005). Fast and

accurate goal-directed motion synthesis for crowds. In

Symposium on Computer Animation ’05 Proceedings,

pages 291–300, New York, NY, USA. ACM Press.

Tak, S., Song, O.-Y., and Ko, H.-S. (2002). Spacetime

sweeping: An interactive dynamic constraints solver.

In Computer Animation ’02 Proceedings, pages 261–

271, Washington, DC, USA. IEEE Computer Society.

Treuille, A., Cooper, S., and Popovic, Z. (2006). Continuum

crowds. ACM Transactions on Graphics, 25(3):1160–

1168.

Ventrella, J. (1995). Disney meets darwin-the evolution of

funny animated ﬁgures. Computer Animation, 00.

Wang, J., Drucker, S. M., Agrawala, M., and Cohen, M. F.

(2006). The cartoon animation ﬁlter. ACM Transac-

tions on Graphics, 25(3):1169–1173.

IJCCI 2009 - International Joint Conference on Computational Intelligence

134