Emotionally Responsive Robotic Avatars as Characters in Virtual Worlds David Burden Daden Limited Stuart Slater University of Wolverhampton

Evolving UI Paradigms Teletype Character VDU Windows Audio-Visual Virtual Worlds

Non-Player Characters (NPCs)‏ Task focussed Usually no life beyond the user Privileged access to information Common in MMORPGs, less common in Virtual Worlds

Task focussed

Usually no life beyond the user

Privileged access to information

Common in MMORPGs, less common in Virtual Worlds

Virtual Worlds Socially focussed Can be persistent Access the world on the same basis as a human Visually can be identical to a human user's avatar Roles include receptionists/greeters, salesmen, actors, tutors

Socially focussed

Can be persistent

Access the world on the same basis as a human

Visually can be identical to a human user's avatar

Roles include receptionists/greeters, salesmen, actors, tutors

Emotional Chatbots Given that computer controlled avatars (robotars) can access all of the gesture and emotional expressions of a human controlled avatar, how important is the use of such emotions with computer controlled avatars?

Given that computer controlled avatars (robotars) can access all of the gesture and emotional expressions of a human controlled avatar, how important is the use of such emotions with computer controlled avatars?

Altair Robotar Architecture libsecondlife altair #2 Second Life Servers ASML/ AAML via web service Perl API Bus Human User BotIF altair #1 Other Engines Emotion Engine Navigation Engine Discourse AIML Chatbot Engine Web Services Server (on web) (Perl)‏ PC or Server (on web) (C#)‏ SL Interface Discourse RDF Engine

E-AI Architecture Developed by University of Wolverhampton Followed assessment of CogAff7, Emile8,SOAR9 and Tok10 Grounded in psychology and psychology research and models Intended for developers who already have a broader bot architecture

Developed by University of Wolverhampton

Followed assessment of CogAff7, Emile8,SOAR9 and Tok10

Grounded in psychology and psychology research and models

Intended for developers who already have a broader bot architecture

Emotions Modelled Happy Sad Fear Disgust Anger Surprise/Startle

Happy

Sad

Fear

Disgust

Anger

Surprise/Startle

E-AI Architecture

Previous Implementations Quake-3 combat bots Over 300 combat situations Got around 20% less kills, won a below average (almost half) number of matches Results clearly supported the assumption that the emotionally enhanced bots, performed less well in combat situations. Lack of combat effectiveness was partly attributed to the e-Bot running away from combat situations when experiencing fear, and often froze for a second (startle response) when an attacking bot came round a corner.

Quake-3 combat bots

Over 300 combat situations

Got around 20% less kills, won a below average (almost half) number of matches

Results clearly supported the assumption that the emotionally enhanced bots, performed less well in combat situations.

Lack of combat effectiveness was partly attributed to the e-Bot running away from combat situations when experiencing fear, and often froze for a second (startle response) when an attacking bot came round a corner.

E-AI Architecture with Altair Object and avatar appearance/ disappearance (ASML)‏ 1. Surprise & Startle vs EAD 2. Explicit vs AIML 3. Implicit vs EAD/Chat Lookup => AEML Avatar expressions/ gesture/movement/chat (AAML)‏ Not yet modelled Not yet modelled Fuzzy mixing of emotional states => AEML Reason for last state stored, and accessible from AIML

Additional Features High Road vs Low Road Low-road (“gut”) response implemented close to avatar in C#/BotIF High-road (“considered”) response implemented in AIML, other side of a web service Moods Each new emotion effects the on-going mood Moods have half-lives and decay over time Habituation With repeated exposure the response to an emotional stimulus can decrease/increase Each EAD trigger has an habituation factor

High Road vs Low Road

Low-road (“gut”) response implemented close to avatar in C#/BotIF

High-road (“considered”) response implemented in AIML, other side of a web service

Moods

Each new emotion effects the on-going mood

Moods have half-lives and decay over time

Habituation

With repeated exposure the response to an emotional stimulus can decrease/increase

Each EAD trigger has an habituation factor

Learnt Behaviour Manual EAD Creation Time consuming No guarantee objects will have the right name Not scalable to “true” AI Bot Learning Programme only “fundamental” responses, eg: “ Be fearful of things that hurt you” “ Like things that give you money” Bot then builds EAD as it experiences things in the world

Manual EAD Creation

Time consuming

No guarantee objects will have the right name

Not scalable to “true” AI

Bot Learning

Programme only “fundamental” responses, eg:

“ Be fearful of things that hurt you”

“ Like things that give you money”

Bot then builds EAD as it experiences things in the world

Demonstration Video

User Evaluation - Provisional 2 sessions of 10 – 16 students (Computer and Games courses)‏ Session 1: Evaluate robotar with/without emotional gestures/expressions, but same chat responses Most noted some difference Majority preferred “emotional” bot Session 2: Evaluate robotar with emotional expression and chat responses and without expression/chat responses Almost all noted some difference Clear majority preferred “emotional” bot

2 sessions of 10 – 16 students (Computer and Games courses)‏

Session 1:

Evaluate robotar with/without emotional gestures/expressions, but same chat responses

Most noted some difference

Majority preferred “emotional” bot

Session 2:

Evaluate robotar with emotional expression and chat responses and without expression/chat responses

Almost all noted some difference

Clear majority preferred “emotional” bot

Conclusions Virtual world robotars provide a greater challenge (and opportunity) than game NPCs The E-AI architecture is a useful way of modelling emotions in avatars E-AI can be used to support habituation and learnt behaviour Initial user evaluation suggests that users prefer interacting with a more “emotional” avatar

Virtual world robotars provide a greater challenge (and opportunity) than game NPCs

The E-AI architecture is a useful way of modelling emotions in avatars

E-AI can be used to support habituation and learnt behaviour

Initial user evaluation suggests that users prefer interacting with a more “emotional” avatar

Emotionally Responsive Robotic Avatars as Characters in Virtual Worlds David Burden Daden Limited Stuart Slater University of Wolverhampton [email_address] [email_address]