1 1. representing and parameterizing agent behaviors jan allbeck and norm badler 연세대학교...

17
1 1. Representing and 1. Representing and Parameterizing Parameterizing Agent Behaviors Agent Behaviors Jan Allbeck and Norm Badler 연연연연연 연연연연연연 연연 연연 연연 2004 2 연연 10410898 연 연 연

Upload: alan-mclaughlin

Post on 05-Jan-2016

214 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: 1 1. Representing and Parameterizing Agent Behaviors Jan Allbeck and Norm Badler 연세대학교 컴퓨터과학과 로봇 공학 특강 2004 2 학기 10410898 유 지 오

1

1. Representing and Parameterizing1. Representing and ParameterizingAgent BehaviorsAgent Behaviors

Jan Allbeck and Norm Badler연세대학교 컴퓨터과학과

로봇 공학 특강2004 2 학기

10410898 유 지 오

Page 2: 1 1. Representing and Parameterizing Agent Behaviors Jan Allbeck and Norm Badler 연세대학교 컴퓨터과학과 로봇 공학 특강 2004 2 학기 10410898 유 지 오

2

AgendaAgenda

sub-title

• Introduction• Control vs. Autonomy• AI-Level Representation• Network Simulation• Parameterized Action Representation

– PAR Architecture– Action Representation– Object Representation

• PAR for Agent Modeling– Personality and Emotions– EMOTE for Displaying Affect

• Interfaces to Representations• Conclusions and Future Research

Page 3: 1 1. Representing and Parameterizing Agent Behaviors Jan Allbeck and Norm Badler 연세대학교 컴퓨터과학과 로봇 공학 특강 2004 2 학기 10410898 유 지 오

3

IntroductionIntroduction• The world is complex difficult to represent…• In order to create an interactive world that meets natural

expectations substantial amount of computer S/W Engineering is required– Graphical depictions, motion models or generators, collision

detection and avoidance, communication or synchronization channels, planning and navigation, cognitive modeling, psychosocial and physiological modeling …An action representation is IMPORTANT!!

• In this chapter…– Outline some thing to consider when adopting an action

representation– Present a representation, Parameterized Action

Representation (PAR)

Page 4: 1 1. Representing and Parameterizing Agent Behaviors Jan Allbeck and Norm Badler 연세대학교 컴퓨터과학과 로봇 공학 특강 2004 2 학기 10410898 유 지 오

4

Control vs. AutonomyControl vs. Autonomy• Control

– Key-frame animation– Detailed control over the movement of the characters– A time consuming process, required a large storage, specific

to a character– Cannot be altered to context Difficult to…

Interact with objects and other agentsCreate transitions between motionsAlter the expression of the motion to new context

• Autonomy– Decrease the data, enable context-sensitive actions– Use Inverse kinematics– Motion capture– Example) Jack, DI-Guy (Human Simulation) …

Low-level motion representations

Page 5: 1 1. Representing and Parameterizing Agent Behaviors Jan Allbeck and Norm Badler 연세대학교 컴퓨터과학과 로봇 공학 특강 2004 2 학기 10410898 유 지 오

5

AI-Level RepresentationAI-Level Representation• High-level representations

– Can vary in their purpose and their semantics

• Communicative or conversational Agents– Mechanisms to synchronize facial expressions with speech– Extract semantic information from text

• Perform autonomously in a virtual world– Concentrate on an agent’s interactions and autonomy

• Planning for characters in virtual environments– Require representations of the state of the environment

(dynamic) Object must also be represented

• Cognitive and social modeling– Emotional states, goals, motivations, and more…

Page 6: 1 1. Representing and Parameterizing Agent Behaviors Jan Allbeck and Norm Badler 연세대학교 컴퓨터과학과 로봇 공학 특강 2004 2 학기 10410898 유 지 오

6

Network SimulationsNetwork Simulations• Design dimensions for distributed or networked simulations

– Bandwidth, synchronization, agent autonomy, agent control, latency, visualization, interfaces…

– Trade offEx) Minimize bandwidth vs. maximize control

• Packets describing agent actions must be formulated, sent, received, and interpreted

• Increasing the autonomy decreasing in necessary bandwidth– Frame-by-frame joint angle vs. string “enter the building”

• “enter the building + carefully + through the blue door”– Modification the detailed joint or motion capture data is

IMPOSSSIBLE!!– If the actions are suitably parameterized POSSIBLE!!

Page 7: 1 1. Representing and Parameterizing Agent Behaviors Jan Allbeck and Norm Badler 연세대학교 컴퓨터과학과 로봇 공학 특강 2004 2 학기 10410898 유 지 오

7

Parameterized Action RepresentationParameterized Action Representation• PAR allows an agent to act, plan, and reason

• A knowledge base and intermediary between natural language and animation

• Specify (parameterize) the agent– Any relevant objects, information about paths, locations,

manners, and purposes

PAR

Page 8: 1 1. Representing and Parameterizing Agent Behaviors Jan Allbeck and Norm Badler 연세대학교 컴퓨터과학과 로봇 공학 특강 2004 2 학기 10410898 유 지 오

8

PAR ArchitecturePAR Architecture

• Actionary stores uninstantiated PARs (UPARs)• Agent Process create instantiated PARs (IPARs)

– Consider emotion, personality factors, current state of the world

• Motion Generators simply replay stored joint angle data or alter this data for context or affect

PAR

Page 9: 1 1. Representing and Parameterizing Agent Behaviors Jan Allbeck and Norm Badler 연세대학교 컴퓨터과학과 로봇 공학 특강 2004 2 학기 10410898 유 지 오

9

Action RepresentationAction Representation• Include fields for low-level animation

concepts– Kinematics, dynamics, …

• Participants– Object or other agents involved in

the action or can be affected by it

• Applicability conditions– True can perform the action

• Preparatory specifications– A list of <CONDITION, action>

statements

• Termination conditions– A list of conditions which when

satisfied indicate the completion of the action

PAR

Page 10: 1 1. Representing and Parameterizing Agent Behaviors Jan Allbeck and Norm Badler 연세대학교 컴퓨터과학과 로봇 공학 특강 2004 2 학기 10410898 유 지 오

10

Object RepresentationObject Representation

• Stored Actionary• Virtual world created retrieve object from the actionary

instantiated placed updated throughout the simulation• Associated with a graphical model in a scene graph• Many of the fields can be filled in as the simulation begins

– Ex) bounding volume

• Help orient actions that involve objects

PAR

Page 11: 1 1. Representing and Parameterizing Agent Behaviors Jan Allbeck and Norm Badler 연세대학교 컴퓨터과학과 로봇 공학 특강 2004 2 학기 10410898 유 지 오

11

Funge et al[19],hierarchy of computer graphics modeling

PAR for Agent ModelingPAR for Agent Modeling• PAR and PARSYS enable each level

– Geometric PAR representsand PARSYS automatically recognizes

– Kinematics and dynamics (physical) explicitly represented in PAR

– Behavioral component World model + agent processes+ motion generators in PARSYS

– Cognitive modeling PARSYS contains mechanismsfor planning and also filtering and prioritizing the actions Individualizing the agentUse conditions (Actionary)

Page 12: 1 1. Representing and Parameterizing Agent Behaviors Jan Allbeck and Norm Badler 연세대학교 컴퓨터과학과 로봇 공학 특강 2004 2 학기 10410898 유 지 오

12

Personality and EmotionsPersonality and Emotions• Personality OCEAN

– “Big Five” OpennessConscientiousnessExtroversionAgreeablenessNeuroticism

• Emotion OCC– Emotion are generated through the agent’s construal of and

reaction to the consequence of events, actions of agents, aspects of objects

PAR for Agent Modeling

Page 13: 1 1. Representing and Parameterizing Agent Behaviors Jan Allbeck and Norm Badler 연세대학교 컴퓨터과학과 로봇 공학 특강 2004 2 학기 10410898 유 지 오

13

EMOTE for Displaying AffectEMOTE for Displaying Affect• EMOTE system

– Based on movement observation science

– Laban Movement Analysis (LMA) Effort and Shape

PAR for Agent Modeling

Page 14: 1 1. Representing and Parameterizing Agent Behaviors Jan Allbeck and Norm Badler 연세대학교 컴퓨터과학과 로봇 공학 특강 2004 2 학기 10410898 유 지 오

14

EMOTE ExampleEMOTE Example• Hitting a balloon

– Differing EMOTE setting

PAR for Agent Modeling

Page 15: 1 1. Representing and Parameterizing Agent Behaviors Jan Allbeck and Norm Badler 연세대학교 컴퓨터과학과 로봇 공학 특강 2004 2 학기 10410898 유 지 오

15

EMOTE and OCEAN linkageEMOTE and OCEAN linkage

• Future work in EMOTE system and the motion quality recognizer– Train the system to correlate captured motions with actor affect,

behavior, mood, and intent

PAR for Agent Modeling

Page 16: 1 1. Representing and Parameterizing Agent Behaviors Jan Allbeck and Norm Badler 연세대학교 컴퓨터과학과 로봇 공학 특강 2004 2 학기 10410898 유 지 오

16

Interfaces to RepresentationsInterfaces to Representations• Basic scripting languages

– Create outline to perform …Specified actionSpecified time

• Drag-and-drop creation applications– For virtual environments

• Natural language

Page 17: 1 1. Representing and Parameterizing Agent Behaviors Jan Allbeck and Norm Badler 연세대학교 컴퓨터과학과 로봇 공학 특강 2004 2 학기 10410898 유 지 오

17

Conclusions and Future ResearchConclusions and Future Research• An action representation

– Autonomy and control– Minimize data storage– Provide semantic for planning

• Level of detail– Nearby action: Inverse kinematics– Further distance: replaying motion capture data– Cognitive representation for conveying action information between

agents• Flexible representation

– Different types of information• Trade-off

– Parameterization specificity vs. program complexity• Future work

– PAR to XML representation– EMOTE parameterization models of personality and emotion– Natural language interface