System Agent

Conceptual Farm

Shuen-Huei Guan Sheng-Yao Cho Yu-Te Shen

Ming Ouhyoung*

Rung-Huei Liang Bing-Yu Chen

{drake,ven,edwards,liang}@cmlab.csie.ntu.edu.tw, abpnm3@r.postjobfree.com, *abpnm3@r.postjobfree.com

Communication and Multimedia Laboratory,

*

Dept. of Computer Science and Information Engineering / Dept. of Information Management,

National Taiwan University

1. Converting simple descriptions into complex be-

haviors Complex autonomous behaviors are cre-

Abstract

ated using compact table-based descriptions (Fig-

ure 3) and flexible scripts instead of complicated

Conceptual Farm is a virtual reality platform for

codes.

generating and observing behaviors of different

2. Semi-interactive editing The behavior is per-

autonomous characters. By providing 1) descriptions

formed in real time as users adjust characters

for characters behaviors and 2) 3D animations and

properties, except when there is a need to regener-

sound, life-like characters in a realistic habitat can be

ate the animation clips.

created, modified, and interact both with users and

3. Extensibility Conceptual Farm is a platform that

other characters in real time. The flexible, manageable

can be used to easily realize every type of animal

and scalable nature of Conceptual Farm leads to its

including insects, mammals, fish, and can even be

desirability in zoological research, general education,

extended to autonomous characters of every type

game and film production, and even decorative arts.

such as aircraft or soccer players.

4. Adoptability for AFX According to the former

1. Introduction features, the design philosophy of Conceptual

Farm can be applied to the higher levels in Anima-

Imagine you have a paintbrush in your hand. You tion Framework eXtension in MPEG-4, which are

draw some pigeons with gray feathers in a square, and not yet clearly defined.

teach them to walk, to eat, to fly, and so on. Then, the The first two features, which are not easily per-

pigeons start wandering around the ground, eating the formed by previously mentioned approaches, are our

feed you spread, and flying up from time to time. An main contributions, since our system reflects user s

idea occurs to you, Why not add kids to play with the input with characters behaviors directly, and users can

pigeons? So you draw three children in the square, focus on characters behaviors without annoying pro-

running amongst the pigeons for fun as soon as run- gramming problems.

ning is taught. As new ideas continue to strike you, the

picture becomes more and more enriched

2. Overview

Interactive artificial life as imaged above has been

presented so far [2, 3, 5, 6, 12, 13]. Based on biologi-

cal theories, these systems simulate animals behaviors

realistically, but, on the other hand, too complicated to

be integrated with a compact and systematic interface

such that users can create artificial lives easily. Some

languages are proposed for describing cognitive be-

haviors [4, 5, 7], but they are not intuitive enough.

To simplify the process of creating artificial lives,

we designed a system, Conceptual Farm, for easy

creation, modification, observation of and interaction

with virtual autonomous characters. Four important

features of Conceptual Farm are:

Figure 1: Flocking pigeons in Conceptual Farm

Based on Conceptual Farm, Dove project creates but also internal factors such as current action, degree

pigeons in a 3D world that can act autonomously in of hunger, etc. Following the rules given by users,

namely, Plans, Decision Maker will select the action

Figure 1. These autonomous behaviors come from a

with the highest utility to send to Pilot and Performer.

table-based system, in which the three resources of

simulated characters (Plans, Navigating Styles and Plans describe 1) the relationship (occurrence prob-

Appearances) are provided by users. abilities) between actions and percepts, 2) the mapping

of each action to its corresponding Navigating Styles

This paper describes the mechanism which enables

and Appearances, and 3) the scope of each percept.

an efficient approach to simulate artificial lives as in

Dove, the implementation issues, and the impact on The following is a sample of (1) and (2) of a pigeon s

Plans.

AFX.

Wander Eating Pursuit

3. Behavioral Model

See(feed) 0.2 0.0 0.8

Destroyed(feed) 0.7 0.0 0.3

In Conceptual Farm, all simulated characters can be Hungry 0.0 0.5 0.5

viewed as autonomous agents, which repeatedly per- Wander 0.7 0.0 0.3

ceive information from World and perform certain re- Eating(feed) 0.2 0.8 0.0

NS_2DWander NS_2DStill NS_2DPursuit

actions, as illustrated in Figure 2. Each agent consists

Animation wander.asf eat.asf pursuit.asf

of Decision Maker, Pilot, and Performer, and has its

Sound wander.wav eat.wav pursuit.wav

own resources provided by users. During each simulat-

ing process, Decision Maker selects a Navigating Style Action Percept

and an Appearance according to the information ob-

Navigating Style Appearances

tained from World, and passes them to Pilot and Per-

former, respectively. Pilot determines new position,

Figure 3: Plans

velocity, and orientation for the next instant, and Per-

former outputs proper Appearances to World. In the Decision Maker consists of 2 agents: Percept Agent

implementation, World is the union of all the other and Action Agent. We refer to the c4 architecture [2]

characters. of the MIT Media Lab.

Let us take an example of the whole process. There The Percept Agent senses all external events within

is a dove. At one moment, Decision Maker decides to the sensory scope for each percept. Internal self-

eat food, then Pilot steps a forward little, and finally awareness, such as the current action, is also sensed.

Performer plays head-lowering animation and cooing Users can dynamically add or remove customized per-

sound. cept functions as well as the built-in ones, and the be-

haviors will change immediately. Plans not only en-

able realistic simulation for sophisticated behaviors

with the uncertain nature of the probability values, but

Resources

Autonomous Agent

also, as a table, provide a compact way for easy ma-

Decision Plans

nipulation.

Maker

The Action Agent is a utility-based agent. It uses

World

Navigating

the percepts stored by Percept Agent and calculates the

Styles

Pilot

score for every candidate action using (1). If nothing is

sensed within the scope of a percept, the probability of

Appearances

Performer this percept will be set to zero. The resulting probabil-

ity of each action is proportional to its score.

Figure 2: An agent with its resources vis-a-vis. World. N

Score ( A j ) = Prob ( P i, A j ) ( P i ) (1)

i=0

3.1 Decision Maker

Aj : the jth action

Pi : the ith percept

Decision Maker, as a brain with sense organs, per-

N : the number of percepts.

ceives information from World and then chooses the

Prob (Pi, Aj) : the occurrence probability of ith per-

best responding action. Influences affecting decision-

cept when the jth action happens

making include not only events made by users with the (Pi ) : 1 if Pi occurs, 0 otherwise.

system UI or by other characters in the virtual world

Performer is responsible for perceptible outputs of

After selecting the action with the highest score, the

Action Agents will send the Navigating Style to the autonomous characters by animation, sound or any

Pilot, while the corresponding animation and sound expressive media provided by users. Our system han-

dles most kinds of sounds by FMOD3.

will be sent to Performer.

Several exporters were implemented to allow users

to make animations with their favorite tools such as

3.2 Pilot

Maya, 3D Studio Max.

For each character, Pilot determines its own path

and orientation according to the Navigating Style (ex. 4. Experimental Results

seek, pursuit and wander) from Decision Maker.

Pilot is built based on Reynolds s OpenSteer li- We used Conceptual Farm to build Dove which is a

brary 1 . Extending Reynolds 16 common steering square at the Chiang Kai-Shek Memorial Hall in

styles for autonomous agents [9], we provide 31 built- Taipei where pigeons gather to find food, clean their

in Navigating Styles. Although the built-in styles meet feathers, and where dogs wander around, and user can

the demands of most cases, users can also provide cus- spread crumbs and run amongst the pigeons in Figure

tomized Navigating Styles through scripts. We provide 6. Taking the pigeons resource as an example, there

a high-level script based on Small 2, which enhances are 13 actions and 8 percepts from Plans, 12 anima-

the flexibility of Conceptual Farm. tions and 4 sounds for Appearances, and 2 scripts for

Navigating Styles.

For example, when a bird is landing with a built-in

Navigating Style, NS_YParabolicUp, it will have the

unreal velocity-alignment problem as illustrated at the 5. AFX & Conceptual Farm

left of Figure 4. The problem can be solved by a script

as below: According to the characteristics mentioned above,

the concept of our system can be applied to enhance

current multimedia standards.

public doLanding (force, elapsedTime) {

As illustrated in Figure 5, MPEG-4 proposed AFX

applyForce (force, elapsedTime)

(Animation Framework eXtension) [1] in order to pro-

velocity[1] = 0

vide a standardized description for computer animation

calculateOrientation (velocity)

and interaction, similar to video and audio standards.

}

AFX is layered into six components, which are, in a

top-down order, cognitive, behavioral, biomechanical,

First, it calls applyForce to get the new position and

physics, modeling and geometry components. The last

velocity and then, it applies calculateOrientation with

four are specified clearly in details, while cognitive

the y-zeroed velocity to get the reasonable orientation

and behavioral components are not, since they are AI-

with the result as illustrated on the right side of Figure

intensive and difficult to formalize.

4.

Audio Audio

Audio CB

DB Decode

Composite

Render

Video Video

Video CB

DB Decode

DIMF

AFX AFX Decoded

Animation

DB Decode AFX

Framework

eXtension

Figure 4: Two landing styles.

Figure 5: AFX in MPEG-4

3.3 Performer To make up for this limitation in current AFX, the

concept of our system provides a solution for editing

1

http://opensteer.sourceforge.net

2 3

http://www.compuphase.com/small.htm http://www.fmod.org

tional Joint Conference on Autonomous Agents and Multi-

and functioning4 data in the cognitive and behavioral

agent Systems, AAMAS, Bologna, Italy, 2002, pp.362-333.

layers, and connects them with multimedia in the other

components of AFX, and even other parts in MPEG-4. [4] L. Chen, K. Bechkoum, and G. Clapworthy, A logical

The formalized data used to describe characters be- approach to high-level agent control, Proc. of the Fifth In-

haviors in Conceptual Farm is also suitable for storage ternational Conference on Autonomous Agents, Montreal,

Quebec, Canada, May 2001, pp.1-8.

and transmission, which is the characteristic of stan-

dardized data. [5] J. Funge, X. Tu, D. Terzopoulos, Cognitive modeling:

knowledge, reasoning and planning for intelligent charac-

ters, Proc. of SIGGRAPH, ACM Press, Los Angeles, CA,

6. Future Work and Conclusions

August 11-13, 1999, pp.29-38.

Conceptual Farm provides an easy way to create [6] M.P. Johnson, A. Wilson, B. Blumberg, C. Kline, and A.

virtual lives. There are, however, some limitations. Bobick. Sympathetic Interfaces: Using a Plush Toy to Di-

rect Synthetic Characters, Proc. of CHI, Pittsburgh, ACM

Characters cannot adapt themselves to the environment

because they cannot modify their own Plans. In addi- Press, May 1999, pp. 152-158.

tion, hardcode cannot be substituted for with our table- [7] J.E. Laird, It Knows What You re Going to Do: Adding

based input mechanism to provide user-character inter- Anticipation to a Quakebot, Proc. of the Fifth International

action, which is simulated-character, environment, and Conference on Autonomous Agents, ACM Press, Montreal,

input-device dependent. Canada, 2001, pp. 385-392.

In summary, we demonstrate a novel approach to

[8] K. Perlin, A. Goldberg, Improv: A System for Scripting

create interactive artificial lives with our virtual reality Interactive Actors in Virtual Worlds, Proc. of SIGGRAPH,

system, Conceptual Farm. The system simulates char- ACM Press, New Orleans, August 1996, pp.205-216.

acter behaviors realistically with compact and formal-

[9] C. Reynolds, Steering Behaviors for Autonomous Char-

ized input descriptions and preserves its flexibility by

acters, Proc. of Game Developers Conference, San Jose,

introducing scripts. It also suggests a practical method

CA, 1999, pp. 763-782.

for standardizing and easily manipulating the dynamic

[10] K. Sims, "Evolving Virtual Creatures", Proc. of SIG-

and real-time properties within the cognitive and be-

GRAPH, ACM Press, New York, July 1994, pp.15-22.

havioral levels of current AFX in MPEG-4.

[11] D. Terzopoulos, Artificial life for computer graphics,

7. Acknowledgements Communications of the ACM, Vol. 42, No. 8, ACM Press,

August, 1999, pp.33-42.

This research was supported in part by National [12] X. Tu, D. Terzopoulos, Artificial Fishes: Physics, Lo-

Science Council 92-2622-E002-002. We are graceful comotion, Perception, Behavior, Proc. of SIGGRAPH,

to Kuei-Yuan Zheng, Ping-Chun Kuo, Wei-Chih Liao, ACM Press, New York, 1994, pp.43-49.

and Tien-Jung Huang (National Taiwan University of

[13] H.S. Yang, H.-J. Park, Y.-J. Cho, Interactive Artificial

Arts) for providing their technical helps. We also thank Life based on Behavior and Perception in a Virtual Environ-

Wan-Chun Ma for his comments on MPEG-4 AFX. ment, Proc. of International Conference on Multimedia &

Expo, IEEE, New York, 30 July - 2 August, 2000, pp. 207-

8. References 210.

[1] ISO/IEC 14496-16:2003(E), Information technology

Coding of audio-visual objects Part 16: Animation

Framework eXtension (AFX)

[2] R. Burke, D. Isla, M. Downie, Y. Ivanov, and B. Blum-

berg, Creature Smarts: The Art and Architecture of a Vir-

tual Brian, Proc. of Game Developers Conference, San Jose,

CA, 2001, pp. 147-166.

[3] R. Burke and B. Blumberg, Using an Ethologically-

Inspired Model to Learn Apparent Temporal Causality for

Planning in Synthetic Creatures, Proc. of the First Interna-

4

We use the word functioning instead of playing because artifi-

Figure 6: Simulating pigeons with Conceptual Farm

cial intelligence is concerned in addition to animation.

Contact this candidate