Intelligent Virtual Agents : 17th International Conference, IVA 2017, Stockholm, Sweden, August 27-30, 2017, Proceedings 978-3-319-67401-8, 3319674013, 978-3-319-67400-1

Table of contents :
Front Matter ....Pages I-XVI
Pedagogical Agents to Support Embodied, Discovery-Based Learning (Ahsan Abdullah, Mohammad Adil, Leah Rosenbaum, Miranda Clemmons, Mansi Shah, Dor Abrahamson et al.)....Pages 1-14
WalkNet: A Neural-Network-Based Interactive Walking Controller (Omid Alemi, Philippe Pasquier)....Pages 15-24
A Virtual Poster Presenter Using Mixed Reality (Vanya Avramova, Fangkai Yang, Chengjie Li, Christopher Peters, Gabriel Skantze)....Pages 25-28
Multiparty Interactions for Coordination in a Mixed Human-Agent Teamwork (Mukesh Barange, Julien Saunier, Alexandre Pauchet)....Pages 29-42
A Dynamic Speech Breathing System for Virtual Characters (Ulysses Bernardet, Sin-hwa Kang, Andrew Feng, Steve DiPaola, Ari Shapiro)....Pages 43-52
To Plan or Not to Plan: Lessons Learned from Building Large Scale Social Simulations (Anton Bogdanovych, Tomas Trescak)....Pages 53-62
Giving Emotional Contagion Ability to Virtual Agents in Crowds (Amyr Borges Fortes Neto, Catherine Pelachaud, Soraia Raupp Musse)....Pages 63-72
Selecting and Expressing Communicative Functions in a SAIBA-Compliant Agent Framework (Angelo Cafaro, Merijn Bruijnes, Jelte van Waterschoot, Catherine Pelachaud, Mariët Theune, Dirk Heylen)....Pages 73-82
Racing Heart and Sweaty Palms (Mathieu Chollet, Talie Massachi, Stefan Scherer)....Pages 83-86
Effects of Social Priming on Social Presence with Intelligent Virtual Agents (Salam Daher, Kangsoo Kim, Myungho Lee, Ryan Schubert, Gerd Bruder, Jeremy Bailenson et al.)....Pages 87-100
Predicting Future Crowd Motion Including Event Treatment (Cliceres Mack Dal Bianco, Soraia Raupp Musse, Adriana Braun, Rodrigo Poli Caetani, Claudio Jung, Norman Badler)....Pages 101-104
The Intelligent Coaching Space: A Demonstration (Iwan de Kok, Felix Hülsmann, Thomas Waltemate, Cornelia Frank, Julian Hough, Thies Pfeiffer et al.)....Pages 105-108
Get One or Create One: the Impact of Graded Involvement in a Selection Procedure for a Virtual Agent on Satisfaction and Suitability Ratings (Charlotte Diehl, Birte Schiffhauer, Friederike Eyssel, Jascha Achenbach, Sören Klett, Mario Botsch et al.)....Pages 109-118
Virtual Reality Negotiation Training System with Virtual Cognitions (Ding Ding, Franziska Burger, Willem-Paul Brinkman, Mark A. Neerincx)....Pages 119-128
Do We Need Emotionally Intelligent Artificial Agents? First Results of Human Perceptions of Emotional Intelligence in Humans Compared to Robots (Lisa Fan, Matthias Scheutz, Monika Lohani, Marissa McCoy, Charlene Stokes)....Pages 129-141
Pragmatic Multimodality: Effects of Nonverbal Cues of Focus and Certainty in a Virtual Human (Farina Freigang, Sören Klett, Stefan Kopp)....Pages 142-155
Simulating Listener Gaze and Evaluating Its Effect on Human Speakers (Laura Frädrich, Fabrizio Nunnari, Maria Staudte, Alexis Heloir)....Pages 156-159
Predicting Head Pose in Dyadic Conversation (David Greenwood, Stephen Laycock, Iain Matthews)....Pages 160-169
Negative Feedback In Your Face: Examining the Effects of Proxemics and Gender on Learning (David C. Jeong, Dan Feng, Nicole C. Krämer, Lynn C. Miller, Stacy Marsella)....Pages 170-183
A Psychotherapy Training Environment with Virtual Patients Implemented Using the Furhat Robot Platform (Robert Johansson, Gabriel Skantze, Arne Jönsson)....Pages 184-187
Crowd-Powered Design of Virtual Attentive Listeners (Patrik Jonell, Catharine Oertel, Dimosthenis Kontogiorgos, Jonas Beskow, Joakim Gustafson)....Pages 188-191
Learning and Reusing Dialog for Repeated Interactions with a Situated Social Agent (James Kennedy, Iolanda Leite, André Pereira, Ming Sun, Boyang Li, Rishub Jain et al.)....Pages 192-204
Moveable Facial Features in a Social Mediator (Muhammad Sikandar Lal Khan, Shafiq ur Réhman, Yongcui Mi, Usman Naeem, Jonas Beskow, Haibo Li)....Pages 205-208
Recipe Hunt: Engaging with Cultural Food Knowledge Using Multiple Embodied Conversational Agents (Sabiha Khan, Adriana Camacho, David Novick)....Pages 209-212
Development and Perception Evaluation of Culture-Specific Gaze Behaviors of Virtual Agents (Tomoko Koda, Taku Hirano, Takuto Ishioh)....Pages 213-222
A Demonstration of the ASAP Realizer-Unity3D Bridge for Virtual and Mixed Reality Applications (Jan Kolkmeier, Merijn Bruijnes, Dennis Reidsma)....Pages 223-226
An ASAP Realizer-Unity3D Bridge for Virtual and Mixed Reality Applications (Jan Kolkmeier, Merijn Bruijnes, Dennis Reidsma, Dirk Heylen)....Pages 227-230
Moral Conflicts in VR: Addressing Grade Disputes with a Virtual Trainer (Jan Kolkmeier, Minha Lee, Dirk Heylen)....Pages 231-234
Evaluated by a Machine. Effects of Negative Feedback by a Computer or Human Boss (Nicole C. Krämer, Lilly-Marie Leiße, Andrea Hollingshead, Jonathan Gratch)....Pages 235-238
A Web-Based Platform for Annotating Sentiment-Related Phenomena in Human-Agent Conversations (Caroline Langlet, Guillaume Dubuisson Duplessis, Chloé Clavel)....Pages 239-242
Does a Robot Tutee Increase Children’s Engagement in a Learning-by-Teaching Situation? (Markus Lindberg, Kristian Månsson, Birger Johansson, Agneta Gulz, Christian Balkenius)....Pages 243-246
The Expression of Mental States in a Humanoid Robot (Markus Lindberg, Hannes Sandberg, Marcus Liljenberg, Max Eriksson, Birger Johansson, Christian Balkenius)....Pages 247-250
You Can Leave Your Head on (Jeroen Linssen, Meike Berkhoff, Max Bode, Eduard Rens, Mariët Theune, Daan Wiltenburg)....Pages 251-254
Say Hi to Eliza (Gerard Llorach, Josep Blat)....Pages 255-258
A Computational Model of Power in Collaborative Negotiation Dialogues (Lydia Ould Ouali, Nicolas Sabouret, Charles Rich)....Pages 259-272
Prestige Questions, Online Agents, and Gender-Driven Differences in Disclosure (Johnathan Mell, Gale Lucas, Jonathan Gratch)....Pages 273-282
To Tell the Truth: Virtual Agents and Morning Morality (Sharon Mozgai, Gale Lucas, Jonathan Gratch)....Pages 283-286
Fixed-pie Lie in Action (Zahra Nazari, Gale Lucas, Jonathan Gratch)....Pages 287-300
Generation of Virtual Characters from Personality Traits (Fabrizio Nunnari, Alexis Heloir)....Pages 301-314
Effect of Visual Feedback Caused by Changing Mental States of the Avatar Based on the Operator’s Mental States Using Physiological Indices (Yoshimasa Ohmoto, Seiji Takeda, Toyoaki Nishida)....Pages 315-324
That’s a Rap (Stefan Olafsson, Everlyne Kimani, Reza Asadi, Timothy Bickmore)....Pages 325-334
Design of an Emotion Elicitation Tool Using VR for Human-Avatar Interaction Studies (P-H. Orefice, M. Ammi, M. Hafez, A. Tapus)....Pages 335-338
Toward an Automatic Classification of Negotiation Styles Using Natural Language Processing (Daniela Pacella, Elena Dell’Aquila, Davide Marocco, Steven Furnell)....Pages 339-342
Interactive Narration with a Child: Avatar versus Human in Video-Conference (Alexandre Pauchet, Ovidiu Şerban, Mélodie Ruinet, Adeline Richard, Émilie Chanoni, Mukesh Barange)....Pages 343-346
Who, Me? How Virtual Agents Can Shape Conversational Footing in Virtual Reality (Tomislav Pejsa, Michael Gleicher, Bilge Mutlu)....Pages 347-359
Cubus: Autonomous Embodied Characters to Stimulate Creative Idea Generation in Groups of Children (André Pires, Patrícia Alves-Oliveira, Patrícia Arriaga, Carlos Martinho)....Pages 360-373
Interacting with a Semantic Affective ECA (Joaquín Pérez, Yanet Sánchez, Francisco J. Serón, Eva Cerezo)....Pages 374-384
Towards Believable Interactions Between Synthetic Characters (Ricardo Rodrigues, Carlos Martinho)....Pages 385-388
Joint Learning of Speech-Driven Facial Motion with Bidirectional Long-Short Term Memory (Najmeh Sadoughi, Carlos Busso)....Pages 389-402
Integration of Multi-modal Cues in Synthetic Attention Processes to Drive Virtual Agent Behavior (Sven Seele, Tobias Haubrich, Tim Metzler, Jonas Schild, Rainer Herpers, Marcin Grzegorzek)....Pages 403-412
A Categorization of Virtual Agent Appearances and a Qualitative Study on Age-Related User Preferences (Carolin Straßmann, Nicole C. Krämer)....Pages 413-422
Towards Reasoned Modality Selection in an Embodied Conversation Agent (Carla Ten-Ventura, Roberto Carlini, Stamatia Dasiopoulou, Gerard Llorach Tó, Leo Wanner)....Pages 423-432
Lay Causal Explanations of Human vs. Humanoid Behavior (Sam Thellman, Annika Silvervarg, Tom Ziemke)....Pages 433-436
Generating Situation-Based Motivational Feedback in a PTSD E-health System (Myrthe Tielman, Mark Neerincx, Willem-Paul Brinkman)....Pages 437-440
Talk About Death: End of Life Planning with a Virtual Agent (Dina Utami, Timothy Bickmore, Asimina Nikolopoulou, Michael Paasche-Orlow)....Pages 441-450
Social Gaze Model for an Interactive Virtual Character (Bram van den Brink, Christyowidiasmoro, Zerrin Yumak)....Pages 451-454
Studying Gender Bias and Social Backlash via Simulated Negotiations with Virtual Agents (L. M. van der Lubbe, T. Bosse)....Pages 455-458
The Dynamics of Human-Agent Trust with POMDP-Generated Explanations (Ning Wang, David V. Pynadath, Susan G. Hill, Chirag Merchant)....Pages 459-462
Virtual Role-Play with Rapid Avatars (Ning Wang, Ari Shapiro, David Schwartz, Gabrielle Lewine, Andrew Wei-Wen Feng)....Pages 463-466
Motion Capture Synthesis with Adversarial Learning (Qi Wang, Thierry Artières)....Pages 467-470
Back Matter ....Pages 471-473

Citation preview

LNAI 10498

Jonas Beskow · Christopher Peters Ginevra Castellano · Carol O’Sullivan Iolanda Leite · Stefan Kopp (Eds.)

Intelligent Virtual Agents 17th International Conference, IVA 2017 Stockholm, Sweden, August 27–30, 2017 Proceedings

123

Lecture Notes in Artificial Intelligence Subseries of Lecture Notes in Computer Science

LNAI Series Editors Randy Goebel University of Alberta, Edmonton, Canada Yuzuru Tanaka Hokkaido University, Sapporo, Japan Wolfgang Wahlster DFKI and Saarland University, Saarbrücken, Germany

LNAI Founding Series Editor Joerg Siekmann DFKI and Saarland University, Saarbrücken, Germany

10498

More information about this series at http://www.springer.com/series/1244

Jonas Beskow Christopher Peters Ginevra Castellano Carol O’Sullivan Iolanda Leite Stefan Kopp (Eds.) •

•

•

Intelligent Virtual Agents 17th International Conference, IVA 2017 Stockholm, Sweden, August 27–30, 2017 Proceedings

123

Editors Jonas Beskow KTH Royal Institute of Technology Stockholm Sweden

Carol O’Sullivan Trinity College Dublin Ireland

Christopher Peters KTH Royal Institute of Technology Stockholm Sweden

Iolanda Leite KTH Royal Institute of Technology Stockholm Sweden

Ginevra Castellano Uppsala University Uppsala Sweden

Stefan Kopp University of Bielefeld Bielefeld Germany

ISSN 0302-9743 ISSN 1611-3349 (electronic) Lecture Notes in Artificial Intelligence ISBN 978-3-319-67400-1 ISBN 978-3-319-67401-8 (eBook) DOI 10.1007/978-3-319-67401-8 Library of Congress Control Number: 2017952717 LNCS Sublibrary: SL7 – Artificial Intelligence © Springer International Publishing AG 2017 This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, express or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Printed on acid-free paper This Springer imprint is published by Springer Nature The registered company is Springer International Publishing AG The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

Preface

This volume presents the proceedings of the 17th International Conference on Intelligent Virtual Agents (IVA 2017). The annual IVA conference represents the main interdisciplinary scientific forum for presenting research on modeling, developing, and evaluating intelligent virtual agents (IVAs) with a focus on communicative abilities and social behavior. IVAs are intelligent digital interactive characters that can communicate with humans and other agents using natural human modalities such as facial expressions, speech, gestures, and movement. They are capable of real-time perception, cognition, emotion, and action that allow them to participate in dynamic social environments. In addition to exploring theoretical issues, the conference showcases working applications. Constructing and studying IVAs requires knowledge, theories, methods, and tools from a wide range of fields such as computer science, psychology, cognitive science, communication, linguistics, interactive media, human–computer interaction, and artificial intelligence. The IVA conference was started in 1998 as a Workshop on Intelligent Virtual Environments at the European Conference on Artificial Intelligence in Brighton, UK, and was followed by a similar one in 1999 in Salford, Manchester, UK. Subsequently, dedicated stand-alone IVA conferences took place in Madrid, Spain, in 2001; Irsee, Germany, in 2003; and on Kos, Greece, in 2005. In 2006 IVA became a full-fledged annual international event, first held in Marina del Rey, California, followed by Paris in 2007, Tokyo in 2008, Amsterdam in 2009, Philadelphia in 2010, Reykjavik in 2011, Santa Cruz, California in 2012, Edinburgh in 2013, Boston in 2014, Delft in 2015, and Los Angeles in 2016. IVA 2017 was held in Stockholm, Sweden, at the Swedish National Museum of Science and Technology (Tekniska Museet) and KTH Royal Institute of Technology (Kungliga Tekniska Högskolan). IVA 2017’s special topic was “Situated Intelligent Agents”, that is, agents that have awareness of and/or make use of their environment (physical or virtual). The theme addresses the synergies between agents with different embodiments, from embodied virtual characters to social robots. Advances in both domains require the development of computational capabilities that allow robots and virtual characters to engage in those direct, unstructured, and dynamically evolving social interactions that characterize humans. We particularly welcomed contributions that addressed the cross-fertilization of state-of-the-art insights and methods from the domains of embodied virtual characters, computer games, social robotics, and social sciences in order to support the development of skills necessary to enable the vision of designing better machines capable of achieving better action, better awareness, and better interaction to engage in intuitive, lifelike, sustained encounters with individuals and groups.

VI

Preface

The interdisciplinary character of IVA 2017 and its special topic are underlined by the conference’s three renowned keynote speakers: – Bilge Mutlu, University of Wisconsin–Madison, USA – Petra Wagner, Bielefeld University, Germany – Iain Matthews, Oculus Research, USA IVA 2017 received 78 submissions. Out of the 50 long paper submissions, only 13 were accepted for the long papers track. Furthermore, there were 17 short papers selected for the single-track paper session, while 22 poster papers and 9 interactive demos were on display. This year’s IVA also included three workshops that took place before the main conference: – “Interaction with Agents and Robots: Different Embodiments, Common Challenges”, organized by Mathieu Chollet, Ayan Ghosh, Hagen Lehmann, and Yukiko Nakano – “Workshop on Conversational Interruptions in Human-Agent Interactions (CIHAI)”, organized by Angelo Cafaro, Eduardo Coutinho, Patrick Gebhard, and Blaise Portard – “Persuasive Embodied Agents for Behavior Change”, organized by Femke Beute, Robbert Jan Beun, Timothy Bickmore, Tibor Bosse, Willem-Paul Brinkman, Joost Broekens, Franziska Burger, John-Jules Ch. Meyer, Mark Neerincx, Rifca Peters, Albert “Skip” Rizzo, Roelof de Vries, and Khiet Truong We would like to express thanks to the rest of the conference’s Organizing Committee, listed herein. We would also like to thank the Senior Program Committee and the Program Committee for helping shape this excellent conference program and for their time, effort, and constructive feedback to the authors. Additionally, we want to thank our keynote speakers for sharing their outstanding work and insights with the community. Further, we would like to thank our sponsors, including Springer, Disney Research, and Furhat Robotics, and the organizers of IVA 2016 and the IVA Steering Committee. August 2017

Jonas Beskow Christopher Peters Ginevra Castellano Carol O’Sullivan Iolanda Leite Stefan Kopp

Organization

Conference Chairs Jonas Beskow Christopher Peters Ginevra Castellano

KTH Royal Institute of Technology KTH Royal Institute of Technology Uppsala University

Program Chairs Carol O’Sullivan Iolanda Leite Stefan Kopp

Trinity College Dublin KTH Royal Institute of Technology University of Bielefeld

Workshop Chairs Candace Sidner Björn Thuresson

Worcester Polytechnic Institute KTH Royal Institute of Technology

Sponsorship Chair André Pereira

Furhat Robotics

Posters and Demo Chair Catharine Oertel

KTH Royal Institute of Technology

Gala Chair Jens Edlund

KTH Royal Institute of Technology

Publication Chairs Maurizio Mancini Giovanna Varni

Università degli Studi di Genova Université Pierre et Marie Curie

Publicity Chair Maike Paetzel

Uppsala University

Webmaster Fangkai Yang

KTH Royal Institute of Technology

VIII

Organization

Senior Program Committee Elisabeth André Ruth Aylett Timothy Bickmore Joost Broekens Dirk Heylen Jill Lehman James Lester Stacy Marsella Yukiko Nakano Michael Neff Ana Paiva Catherine Pelachaud Laurel Riek Stefan Scherer Gabriel Skantze Hannes Vilhjálmsson Michael Young

University of Augsburg Heriot-Watt University Northeastern University Delft University of Technology University of Twente Disney Research Pittsburgh North Carolina State University Northeastern University Seikei University UC Davis INESC-ID and University of Lisbon CNRS - ISIR, Université Pierre et Marie Curie UC San Diego University of Southern California KTH Royal Institute of Technology Reykjavík University University of Utah

Program Committee Sean Andrist Kirsten Bergmann Elisabetta Bevacqua Johan Boye Hendrik Buschmeier Ronald Böck Angelo Cafaro Liz Carter Mathieu Chollet Luísa Coheur Iwan de Kok Etienne de Sevin Joao Dias Damien Dupr Kevin El Haddad Benjamin Files Samantha Finkelstein Farina Freigang Patrick Gebhard David Gerritsen Emer Gilmartin

Microsoft Research Bielefeld University Lab-STICC, Ecole Nationale d’Ingénieurs de Brest (ENIB) KTH Royal Institute of Technology Bielefeld University Otto von Guericke University Magdeburg CNRS-ISIR Université Pierre et Marie Curie Disney Research USC Institute for Creative Technologies INESC-ID and Instituto Superior Técnico, Technical University of Lisbon Bielefeld University University of Bordeaux INESC-ID and Instituto Superior Técnico, Technical University of Lisbon Queen’s University Belfast UMONS US Army Research Laboratory Carnegie Mellon University Bielefeld University DFKI GmbH Carnegie Mellon University Trinity College Dublin

Organization

Elena Corina Grigore Jacqueline Hemminghaus Laura Hoffmann W. Lewis Johnson Patrik Jonell Markus Kächele James Kennedy Peter Khooshabeh Dimosthenis Kontogiorgos Kangsoo Kim Mei Yii Lim Benjamin Lok José David Lopes Samuel Mascarenhas Maryam Moosaei Radoslaw Niewiadomski Aline Normoyle Magalie Ochs Jean-Marc Odobez Catharine Oertel Andrew Olney Slim Ouni Alexandros Papangelis Florian Pecune Andre Pereira Eli Pincus Ronald Poppe Aditi Ramachandran Tiago Ribeiro Lazlo Ring Justus Robertson Astrid Rosenthal-Von der Pütten Nicolas Sabouret Najmeh Sadoughi Samira Sheikhi Malte Schilling Ari Shapiro Mei Si Kalin Stefanov Stefan Sutterlin Reid Swanson Ha Trinh Daniel Ullman

IX

Yale University AG SCS, CITEC, Bielefeld University Bielefeld University Alelo Inc. KTH Royal Institute of Technology University of Ulm Disney Research US Army Research Laboratory KTH Royal Institute of Technology University of Central Florida Heriot-Watt University University of Florida KTH Royal Institute of Technology INESC-ID and Instituto Superior Técnico, Technical University of Lisbon University of Notre Dame University of Genoa University of Pennsylvania LSIS IDIAP KTH Royal Institute of Technology University of Memphis LORIA - Université de Lorraine Toshiba Research Europe CNRS - LTCI Furhat Robotics USC Institute for Creative Technologies Utrecht University Yale University INESC-ID and Instituto Superior Técnico, Technical University of Lisbon Northeastern University North Carolina State University University of Duisburg-Essen LIMSI-CNRS University of Texas at Dallas University of Chicago ICSI Berkeley USC Institute for Creative Technologies Rensselaer Polytechnic Institute KTH Royal Institute of Technology Oslo University Hospital, Norway Independent Northeastern University Brown University

X

Volkan Ustun Leo Wanner Fangkai Yang

Organization

USC Institute for Creative Technologies ICREA and University Pompeu Fabra KTH Royal Institute of Technology

Sponsors Springer

ABC Disney Research

Furhat Robotics

KTH School of Computer Science and Communication

Contents

Pedagogical Agents to Support Embodied, Discovery-Based Learning . . . . . . Ahsan Abdullah, Mohammad Adil, Leah Rosenbaum, Miranda Clemmons, Mansi Shah, Dor Abrahamson, and Michael Neff

1

WalkNet: A Neural-Network-Based Interactive Walking Controller . . . . . . . . Omid Alemi and Philippe Pasquier

15

A Virtual Poster Presenter Using Mixed Reality . . . . . . . . . . . . . . . . . . . . . Vanya Avramova, Fangkai Yang, Chengjie Li, Christopher Peters, and Gabriel Skantze

25

Multiparty Interactions for Coordination in a Mixed Human-Agent Teamwork . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Mukesh Barange, Julien Saunier, and Alexandre Pauchet A Dynamic Speech Breathing System for Virtual Characters . . . . . . . . . . . . Ulysses Bernardet, Sin-hwa Kang, Andrew Feng, Steve DiPaola, and Ari Shapiro To Plan or Not to Plan: Lessons Learned from Building Large Scale Social Simulations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Anton Bogdanovych and Tomas Trescak Giving Emotional Contagion Ability to Virtual Agents in Crowds. . . . . . . . . Amyr Borges Fortes Neto, Catherine Pelachaud, and Soraia Raupp Musse Selecting and Expressing Communicative Functions in a SAIBA-Compliant Agent Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Angelo Cafaro, Merijn Bruijnes, Jelte van Waterschoot, Catherine Pelachaud, Mariët Theune, and Dirk Heylen Racing Heart and Sweaty Palms: What Influences Users’ Self-Assessments and Physiological Signals When Interacting with Virtual Audiences?. . . . . . . Mathieu Chollet, Talie Massachi, and Stefan Scherer

29 43

53 63

73

83

Effects of Social Priming on Social Presence with Intelligent Virtual Agents . . . Salam Daher, Kangsoo Kim, Myungho Lee, Ryan Schubert, Gerd Bruder, Jeremy Bailenson, and Greg Welch

87

Predicting Future Crowd Motion Including Event Treatment. . . . . . . . . . . . . Cliceres Mack Dal Bianco, Soraia Raupp Musse, Adriana Braun, Rodrigo Poli Caetani, Claudio Jung, and Norman Badler

101

XII

Contents

The Intelligent Coaching Space: A Demonstration. . . . . . . . . . . . . . . . . . . . Iwan de Kok, Felix Hülsmann, Thomas Waltemate, Cornelia Frank, Julian Hough, Thies Pfeiffer, David Schlangen, Thomas Schack, Mario Botsch, and Stefan Kopp Get One or Create One: the Impact of Graded Involvement in a Selection Procedure for a Virtual Agent on Satisfaction and Suitability Ratings . . . . . . Charlotte Diehl, Birte Schiffhauer, Friederike Eyssel, Jascha Achenbach, Sören Klett, Mario Botsch, and Stefan Kopp Virtual Reality Negotiation Training System with Virtual Cognitions. . . . . . . Ding Ding, Franziska Burger, Willem-Paul Brinkman, and Mark A. Neerincx

105

109

119

Do We Need Emotionally Intelligent Artificial Agents? First Results of Human Perceptions of Emotional Intelligence in Humans Compared to Robots. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Lisa Fan, Matthias Scheutz, Monika Lohani, Marissa McCoy, and Charlene Stokes

129

Pragmatic Multimodality: Effects of Nonverbal Cues of Focus and Certainty in a Virtual Human . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Farina Freigang, Sören Klett, and Stefan Kopp

142

Simulating Listener Gaze and Evaluating Its Effect on Human Speakers . . . . Laura Frädrich, Fabrizio Nunnari, Maria Staudte, and Alexis Heloir

156

Predicting Head Pose in Dyadic Conversation. . . . . . . . . . . . . . . . . . . . . . . David Greenwood, Stephen Laycock, and Iain Matthews

160

Negative Feedback In Your Face: Examining the Effects of Proxemics and Gender on Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . David C. Jeong, Dan Feng, Nicole C. Krämer, Lynn C. Miller, and Stacy Marsella A Psychotherapy Training Environment with Virtual Patients Implemented Using the Furhat Robot Platform. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Robert Johansson, Gabriel Skantze, and Arne Jönsson Crowd-Powered Design of Virtual Attentive Listeners . . . . . . . . . . . . . . . . . Patrik Jonell, Catharine Oertel, Dimosthenis Kontogiorgos, Jonas Beskow, and Joakim Gustafson Learning and Reusing Dialog for Repeated Interactions with a Situated Social Agent. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . James Kennedy, Iolanda Leite, André Pereira, Ming Sun, Boyang Li, Rishub Jain, Ricson Cheng, Eli Pincus, Elizabeth J. Carter, and Jill Fain Lehman

170

184 188

192

Contents

XIII

Moveable Facial Features in a Social Mediator . . . . . . . . . . . . . . . . . . . . . . Muhammad Sikandar Lal Khan, Shafiq ur Réhman, Yongcui Mi, Usman Naeem, Jonas Beskow, and Haibo Li

205

Recipe Hunt: Engaging with Cultural Food Knowledge Using Multiple Embodied Conversational Agents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sabiha Khan, Adriana Camacho, and David Novick

209

Development and Perception Evaluation of Culture-Specific Gaze Behaviors of Virtual Agents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tomoko Koda, Taku Hirano, and Takuto Ishioh

213

A Demonstration of the ASAP Realizer-Unity3D Bridge for Virtual and Mixed Reality Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jan Kolkmeier, Merijn Bruijnes, and Dennis Reidsma

223

An ASAP Realizer-Unity3D Bridge for Virtual and Mixed Reality Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Jan Kolkmeier, Merijn Bruijnes, Dennis Reidsma, and Dirk Heylen

227

Moral Conflicts in VR: Addressing Grade Disputes with a Virtual Trainer . . . Jan Kolkmeier, Minha Lee, and Dirk Heylen Evaluated by a Machine. Effects of Negative Feedback by a Computer or Human Boss. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Nicole C. Krämer, Lilly-Marie Leiße, Andrea Hollingshead, and Jonathan Gratch A Web-Based Platform for Annotating Sentiment-Related Phenomena in Human-Agent Conversations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Caroline Langlet, Guillaume Dubuisson Duplessis, and Chloé Clavel Does a Robot Tutee Increase Children’s Engagement in a Learning-byTeaching Situation? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Markus Lindberg, Kristian Månsson, Birger Johansson, Agneta Gulz, and Christian Balkenius The Expression of Mental States in a Humanoid Robot . . . . . . . . . . . . . . . . Markus Lindberg, Hannes Sandberg, Marcus Liljenberg, Max Eriksson, Birger Johansson, and Christian Balkenius You Can Leave Your Head on: Attention Management and Turn-Taking in Multi-party Interaction with a Virtual Human/Robot Duo . . . . . . . . . . . . . . . Jeroen Linssen, Meike Berkhoff, Max Bode, Eduard Rens, Mariët Theune, and Daan Wiltenburg Say Hi to Eliza: An Embodied Conversational Agent on the Web . . . . . . . . . Gerard Llorach and Josep Blat

231

235

239

243

247

251

255

XIV

Contents

A Computational Model of Power in Collaborative Negotiation Dialogues . . . Lydia Ould Ouali, Nicolas Sabouret, and Charles Rich Prestige Questions, Online Agents, and Gender-Driven Differences in Disclosure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Johnathan Mell, Gale Lucas, and Jonathan Gratch

259

273

To Tell the Truth: Virtual Agents and Morning Morality . . . . . . . . . . . . . . . Sharon Mozgai, Gale Lucas, and Jonathan Gratch

283

Fixed-pie Lie in Action . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Zahra Nazari, Gale Lucas, and Jonathan Gratch

287

Generation of Virtual Characters from Personality Traits . . . . . . . . . . . . . . . Fabrizio Nunnari and Alexis Heloir

301

Effect of Visual Feedback Caused by Changing Mental States of the Avatar Based on the Operator’s Mental States Using Physiological Indices. . . . . . . . Yoshimasa Ohmoto, Seiji Takeda, and Toyoaki Nishida

315

That’s a Rap: Increasing Engagement with Rap Music Performance by Virtual Agents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Stefan Olafsson, Everlyne Kimani, Reza Asadi, and Timothy Bickmore

325

Design of an Emotion Elicitation Tool Using VR for Human-Avatar Interaction Studies. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . P-H. Orefice, M. Ammi, M. Hafez, and A. Tapus

335

Toward an Automatic Classification of Negotiation Styles Using Natural Language Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Daniela Pacella, Elena Dell’Aquila, Davide Marocco, and Steven Furnell Interactive Narration with a Child: Avatar versus Human in Video-Conference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Alexandre Pauchet, Ovidiu Şerban, Mélodie Ruinet, Adeline Richard, Émilie Chanoni, and Mukesh Barange Who, Me? How Virtual Agents Can Shape Conversational Footing in Virtual Reality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Tomislav Pejsa, Michael Gleicher, and Bilge Mutlu Cubus: Autonomous Embodied Characters to Stimulate Creative Idea Generation in Groups of Children . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . André Pires, Patrícia Alves-Oliveira, Patrícia Arriaga, and Carlos Martinho

339

343

347

360

Contents

XV

Interacting with a Semantic Affective ECA . . . . . . . . . . . . . . . . . . . . . . . . Joaquín Pérez, Yanet Sánchez, Francisco J. Serón, and Eva Cerezo

374

Towards Believable Interactions Between Synthetic Characters . . . . . . . . . . . Ricardo Rodrigues and Carlos Martinho

385

Joint Learning of Speech-Driven Facial Motion with Bidirectional Long-Short Term Memory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Najmeh Sadoughi and Carlos Busso Integration of Multi-modal Cues in Synthetic Attention Processes to Drive Virtual Agent Behavior . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Sven Seele, Tobias Haubrich, Tim Metzler, Jonas Schild, Rainer Herpers, and Marcin Grzegorzek A Categorization of Virtual Agent Appearances and a Qualitative Study on Age-Related User Preferences . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Carolin Straßmann and Nicole C. Krämer

389

403

413

Towards Reasoned Modality Selection in an Embodied Conversation Agent . . . Carla Ten-Ventura, Roberto Carlini, Stamatia Dasiopoulou, Gerard Llorach Tó, and Leo Wanner

423

Lay Causal Explanations of Human vs. Humanoid Behavior. . . . . . . . . . . . . Sam Thellman, Annika Silvervarg, and Tom Ziemke

433

Generating Situation-Based Motivational Feedback in a PTSD E-health System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Myrthe Tielman, Mark Neerincx, and Willem-Paul Brinkman

437

Talk About Death: End of Life Planning with a Virtual Agent . . . . . . . . . . . Dina Utami, Timothy Bickmore, Asimina Nikolopoulou, and Michael Paasche-Orlow

441

Social Gaze Model for an Interactive Virtual Character . . . . . . . . . . . . . . . . Bram van den Brink, Christyowidiasmoro, and Zerrin Yumak

451

Studying Gender Bias and Social Backlash via Simulated Negotiations with Virtual Agents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . L.M. van der Lubbe and T. Bosse

455

The Dynamics of Human-Agent Trust with POMDP-Generated Explanations. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ning Wang, David V. Pynadath, Susan G. Hill, and Chirag Merchant

459

XVI

Contents

Virtual Role-Play with Rapid Avatars . . . . . . . . . . . . . . . . . . . . . . . . . . . . Ning Wang, Ari Shapiro, David Schwartz, Gabrielle Lewine, and Andrew Wei-Wen Feng

463

Motion Capture Synthesis with Adversarial Learning. . . . . . . . . . . . . . . . . . Qi Wang and Thierry Artières

467

Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

471

Pedagogical Agents to Support Embodied, Discovery-based Learning 1 Ahsan Abdullah1 , Mohammad Adil1 , Leah Rosenbaum2 , Miranda 2 2 2 Clemmons , Mansi Shah , Dor Abrahamson , and Michael Neff 1

University of California, Davis [email protected], [email protected], [email protected] 2 University of California, Berkeley [email protected], [email protected], [email protected], [email protected]

Abstract. This paper presents a pedagogical agent designed to support students in an embodied, discovery-based learning environment. Discovery-based learning guides students through a set of activities designed to foster particular insights. In this case, the animated agent explains how to use the Mathematical Imagery Trainer for Proportionality, provides performance feedback, leads students to have different experiences and provides remedial instruction when required. It is a challenging task for agent technology as the amount of concrete feedback from the learner is very limited, here restricted to the location of two markers on the screen. A Dynamic Decision Network is used to automatically determine agent behavior, based on a deep understanding of the tutorial protocol. A pilot evaluation showed that all participants developed movement schemes supporting proto-proportional reasoning. They were able to provide verbal proto-proportional expressions for one of the taught strategies, but not the other. Keywords: pedagogical agents, discovery-based learning, dynamic decision networks

1

Introduction

Discovery-based learning is an educational activity paradigm whereby students are led through well-specified experiences that are designed to foster particular insights relevant to curricular objectives. It differs from many of the applications in which pedagogical agents have traditionally been used, in that the knowledge desired for the student is never explicitly stated in the experience. Rather, the child discovers it on her or his own. Testing also differs, as the goal is a deeper conceptual understanding, not easily measured by right or wrong answers. Although discovery-based learning has been a major approach in reform-oriented pedagogy for over a century in classrooms, only recently has begun to be incorporated into interactive technology. The broadest objective of the current paper 

The first two authors made equal contributions to technical aspects of the research.

© Springer International Publishing AG 2017 J. Beskow et al. (Eds.): IVA 2017, LNAI 10498, pp. 1-14, 2017. DOI 10.1007/978-3-319-67401-8_1

1



A. Abdullah et al.

is to highlight an approach, along with challenges and responses, for building autonomous pedagogical agents for discovery-based interactive learning. This work explores the application of pedagogical agents to the experiential goal of discovery-based learning. In particular, we add a pedagogical agent to an embodied math learning environment, MITp, designed to teach children proportion. MITp is described in detail in Sec. 2. The basic idea is that a child is encouraged to move two markers on a screen with her fingers. As she does this, the screen changes color. If the height of the two markers is in a particular ratio, say 1:2, the screen will go green, and as the ratio varies away from this it will go to yellow, and then red. The child is never told anything about ratios or proportion or how to make the screen green. Rather she is guided to discover different ways to create a green screen, and by doing so, begins to build an understanding of proportion. This system has been used extensively in learning research with a human tutor guiding students. In this work, we seek to understand what is required to make an animated pedagogical agent effective in this tutoring role. MITp is an embodied learning experience where the child learns through performing physical movements. Embodied pedagogical agents are particularly useful in this setting because of the engagement they engender and, most importantly, because they can enact virtual actions and gestures they wish the learner to perform. This type of learning application creates unique computational challenges. Chief among them, it is very Fig. 1: A child listens to the pedagogical difficult to measure the student’s agent explain concepts within the MITp progress when it is not possible to ask learning environment. questions with right or wrong answers that are easy for a computer to grade. Our design process employed a deep analysis of the process used by human tutors, including reviewing many hours of video recorded interactions. We identified the key stages in the tutorial process, the types of actions tutors took and when they took them. From this analysis, we identified the following activity types the agent must engage in: instructing the child what to do; valorizing success; waiting so the child can explore on her own; providing remedial training when the child is blocked and advancing the child through the tutorial process. While the human tutor sits beside the learner, we placed our animated agent on the other side of the screen from the child, with access to the same touch surface as the learner (Figure 1). Dubbed Maria, our agent can execute a sequence of action blocks. Each block consists of any subset of spoken audio, facial animation, lip syncing and body animation, including arm and finger gestures. There are well over 100 actions that the agent can perform. We use Dynamic Decision Networks to decide when the agent should perform an action block and which action to perform.

Pedagogical Agents to Support Embodied, Discovery-based Learning

3

We have conducted a pilot evaluation of the system. It showed that students were able to effectively explore the screen to find greens and enact a particular search strategy taught by the system. Students provided proto-mathematical descriptions of one solution strategy, although not the other. The results demonstrate good progress and also illuminate potential directions for future work.

2

MITp Learning Environment and Tutorial Protocol

Fig. 2: The MITp environment. The screen is green when the hands’ heights match a pre-programmed ratio.

The current study builds upon an earlier educational-research effort to design and evaluate embodied-interaction technology for mathematics instruction. Specifically, the pedagogical agent described herein was integrated into an existing activity design architecture called the Mathematical Imagery Trainer for Proportionality (MITp) [1, 2, 17], which we now present. Proportional reasoning is important yet difficult for many students. It involves understanding multiplicative part-whole relations between rational quantities; a change in one quantity is always accompanied by a change in the other, and these changes are related by a constant multiplier [6, 26, 31]. Our MITp approach to support students in developing multiplicative understanding of proportions draws on embodiment theory, which views the mind as extending dynamically through the body into the natural-cultural ecology. Thus human reasoning emerges through, and is expressed as, situated sensorimotor interactions [3, 25]. Educational researchers informed by these theories have created technologies to foster content learning through embodied interaction (e.g., [5, 11]). The MITp system (Figure 2) poses the physical challenge of moving two hands on a touch screen to make it green, a result which occurs when the ratio of hand heights matches the pre-programmed ratio of 1:2. Through this process, students can develop pre-symbolic mathematical understanding by engaging in this embodied activity and building particular movement schemes related to proportions. By introducing specific tools into the environment, here a grid and numbers, students are given progressively more mathematical tools with which to express those strategies. The MITp system has been extensively tested for its educational effectiveness. Using qualitative analyses, the researchers demonstrated the variety of manipulation strategies students developed as their means of accomplishing the task objective of moving their hands while keeping the screen green [17]. Moreover, it was shown that students engaged in deep mathematical reflection as they were guided to compare across the strategies [2]. The studies have presented empirical



A. Abdullah et al.

data of students shifting from naive manipulation to mathematical reasoning as they engage the frames of reference introduced into the problem space and the tutor’s critical role in facilitating this shift [1]. 2.1

Interview Protocol

Maria is programmed to lead students through a series of activities on the MITp touchscreen, each supporting the development of particular movement strategies deemed relevant to proportional reasoning. Broadly, the two main phases are exploration, targeting the strategy “Higher-Bigger,” and “a-per-b.” In “HigherBigger,” participants meet Maria on a screen with a red background, which is later overlaid with a grid. Participants are instructed to move the cursors up and down to make the screen green. At each green, Maria valorizes their work and asks them to make another green, either higher, lower, or elsewhere. Other than moving the cursors up and down, participants receive little guidance on particular movement strategies. With time, the grid is overlaid on the screen. The goal of this stage is for students to notice that when they make a green higher on the screen, the gap between their hands is bigger (“Higher-Bigger”). Next, in the “a-per-b” phase, participants are given instructions to start at the bottom of the screen, move their left hand up one grid unit, and then place their right hand to make green. Finally, the grid is supplemented with numerals. Participants are periodically asked to reflect on their rule for making green. Though Maria does not (yet) recognize speech, these reflections promote verbal description through which the developers can assess the participants’ protoproportional understanding. Participants took about 20 minutes on average to complete the task. While participants interact with Maria, the presiding human interviewers try to minimize human-to-human interaction, albeit occasionally they respond to participant queries, confusion, or frustration.

3

Related Work on Pedagogical Agents

Our work finds its roots in previous work on Pedagogical agents and Intelligent Tutoring Systems. Intelligent tutoring systems are computer softwares designed to simulate a human tutor. Pedagogical agents aid the process by adding a human-like character to the learning process. Research over the past few decades [21] has validated the positive impact of having an embodied presence in virtual learning environment. They have been a success primarily because they add emotional and non-verbal feedback to the learning environment [22]. More expressive pedagogical agents tend to improve the learning experience [15]. Intelligent tutoring systems (ITS) have been developed for a wide range of topics. Cognitive Tutors [24] have been adapted to teach students mathematical and other scientific concepts like genetics. The Andes Physics tutor [36] focuses on helping students in introductory Physics courses at college level. Writing Pal [33] and iStart [19] help students in developing writing and reading strategies

Pedagogical Agents to Support Embodied, Discovery-based Learning

5

respectively. Decision theoretic tutoring systems have also been very successful and range from generic frameworks like DT Tutor [29] to domain specific systems such as Leibniz [13]. The feedback and learning mechanism behind all these activities revolves around the tutor provided instructions or some sort of rule specification, followed by student responses, given as either text or multiple choice selections, to posed challenge questions. Our discovery based learning methodology differs fundamentally as the system never describes how to achieve the desired goal, and the student response has to be gauged in real-time based solely on the touch screen coordinates. There are not questions that can be used to directly gauge progress. Pedagogical agents can interact with the student in various roles, such as interactive demonstrators, virtual teammates and teachers. Steve [20] is an early example of a demonstration based pedagogical agent to train people to operate ship engines. INOTS [7] teaches communication and leadership skills to naval officers. The agent is questioned about a case by officers during training, and their performance is evaluated by rest of the class watching this interaction. AutoTutor [14] is a modern system used to teach concepts in Science and Mathematics. The student agent works with the human student to solve the problems in different ways. Adele [34] and Herman the Bug [27] are two classic pedagogical agents designed to teach medicine and botanical anatomy respectively. Decision Networks have also previously been used in the development of adaptive pedagogical agents such as [8] and [32] and for narrative planning [28]. These decision theoretic agents work on concrete feedback from the user in form of biological signals or responses to questions. In fact, all the above mentioned agents use the standard teach and test framework in order to gauge student’s performance, allowing them to focus on specific concepts in the learning that the student is struggling with. Our pedagogical agent operates in a quite different, discoverybased learning paradigm. Some previous pedagogical agents have also targeted more open-ended learning environments. For example, a system designed for children with ASD allows children to interact with the system by telling stories, control the agent by selecting pre-defined responses or author new responses to create new stories [35]. Related work has sought to use agents not as instructors, but as virtual peers [23, 12].

4 4.1

System System Overview and Architecture

An overview of our MITp autonomous agent system architecture is shown in Figure 3. It consists of a control system and Unity3D front end. Students interact with our agent using a touch screen. The screen is virtually divided into left and right halves around the agent, designating two large tracks where the learner can move the markers. Maria is standing in the middle as shown in Figure 1 and can reach most of the screen. The Unity client sends the system state consisting of the two touch locations to the control system, which then instructs the agent to perform particular actions by specifying Action IDs.



A. Abdullah et al.

Action IDs

Character Rig

Animations & Audio

Screen Color

Decision Networks Action Blocks

Control System

System state

Unity3D

Touch Screen Touch coords

Fig. 3: System Architecture

The control system employs dynamic decision networks [10] to model the behavior of our pedagogical agent. The decision networks are updated based on the evidence received from Unity and history maintained throughout the interaction. They may decide to do nothing or have the agent perform one of many pre-designed actions, depending on what appears most efficacious in the current context. Triggered actions are sent to Unity as action block IDs. Each action block consists of an audio file, a facial animation and a body animation. Our database contains 115 different action blocks. Generating meaningfully labelled data for learning agent behavior is a challenge in a discovery-based setting. At this point, we do not make assumptions about student’s learning during interaction and instead rely on a robust tutorial process which can adapt and provide remedial instruction if the student is struggling. Modeling student’s learning on the fly requires mapping from patterns of finger movement to their mental state, which remains future work. Due to these issues, learning based techniques such as [18] are not a good fit for the current problem. DDNs allow us to leverage our strong understanding of the tutorial process by pre-encoding it into system parameters. 4.2

Decision Networks

Decision networks find their roots in Bayesian Networks [30], which are graphical models consisting of chance nodes representing a set of random variables. Random variables are events that could possibly occur in the given world. Chance nodes, drawn as ovals in the graph, can take any type of discrete value, such as a boolean or an integer. These values are generally finite and mutually exclusive. Arcs between these nodes capture conditional dependencies. Cycles are not allowed. Given the conditional probabilities, prior and evidence, inferences can be made about any random variable in the network. Decision networks [16] extend Bayesian Networks by using Bayesian inference to determine a decision that maximizes expected utility. Decision networks add decision and utility nodes, represented by a rectangle and diamond respectively. Decision nodes represent all choices the system can make while the utility node captures our preferences under different factors impacting the decision. This concept can be further extended to Dynamic Decision Networks (DDNs)

Pedagogical Agents to Support Embodied, Discovery-based Learning

7

[10]. DDNs consist of time varying attributes where there are conditional dependencies between nodes at different time steps. Decisions made in previous time steps can impact the probability distribution or state of the network in future steps. DDNs provide a useful way of modeling evolving beliefs about the world and changing user preferences. 4.3

Building Decision Networks for Experiential Learning

Discovery-based learning is challenging compared to other learning settings because the pedagogical agent must make decisions with very impoverished information as there is no continuous stream of concrete verbal or q&a based feedback that the agent can use to assess the student’s understanding. For example, in our case, we only have student’s touch coordinates on the screen as input. Our agent guides the learner through the discovery process using this input and a deep understanding of the tutorial process, encoded in the DDNs. Per the tutorial analysis, the agent must instruct, valorize, wait, provide remedial training and advance the child to the next stage. To decide if the agent should wait, the child must have a notion of the passage of time, which can be combined with the child’s activity pattern to determine if allowing time for exploration is appropriate. Remedial actions are triggered when the child is not executing desired movement patterns after repeated instructions and ample trial time. An example remedial strategy involves the agent displaying a marked location for the child’s one hand and then encouraging them to find green by only moving the other hand. Based on their performance, the agent may ask students to repeat activities to deepen their learning. The learning experience progresses through multiple activities tied to particular interaction strategies: “Higher-Bigger” exploration without and with the grid, “a-per-b” without and with numbers, and an optional “speed” activity. All the activities are modeled using dynamic decision networks. Space limitations preclude discussion of each, but we will explain our approach using the “Higher-Bigger” task as an exemplar. Exploring “Higher-Bigger” is the introductory activity students go through, as outlined in Sec. 2. The activity guides the student to find greens at several locations on the screen, with the goal of having the child realize that the separation between her hands is larger for greens higher on the screen than greens that are lower. Figure 4 and 5 show the decision network which governs the behavior of our pedagogical agent for this activity at two different points during an interaction. The decision network is updated multiple times each second to ensure real time responses. The network encodes our agent’s belief about the state of tutorial process and student’s interaction at each time step. Factors impacting the agent’s decision for the network shown in Fig. 4 and 5 are: Goal : Models agent’s temporally evolving expectations for the student. During this activity the student is first expected to find a couple of greens anywhere on



A. Abdullah et al. t+1

t

t+1

Introduction

0

0

Mid

FindGreen

0

0

Low

MoreGreen

1

0

HighGreen

0

1

LowGreen

0

0

High

Location

Timeout

Color

Location

Goal

Decision

Utility

.

. .

. . .

. . . .

. . . . .

. . . . .

Valorize

0

Goal

Utility

No

Green

Educate

0

MakeGreen

0

MakeGreen

HighGreen

20

HighGreen

LowGreen

0

Wait

0

Done

-5

Valorize Educate

LowGreen Wait

Decision

Color

Green

Yes

NotGreen

No

Timeout

Mid

More Green

Done

Fig. 4: Decision Network for the exploration task at time step t. This figure shows an example query and update mechanism of our decision network. The arcs labeled t+1 indicate that the state of Goal node at time t+1 depends on the state of Location, Color and Goal at time t. At time t, the agent’s goal is for the student to find MoreGreen on screen. The student finds green somewhere in the middle of the screen, and the evidence for nodes Location, Color and Timeout (bold words) is set. We now query the network to give us a decision with maximum utility given the circumstances represented by the network state. As shown in green, the agent decides to guide the student to the new task of finding green in the upper portion of the screen.

the screen, followed by greens in specific portions of the screen. Possible node states are shown in the network above. Timeout : Models the time elapsed since important events. It includes time since the last agent action, time since last touch by the student, time since the last green and time since last achieved goal. These factors, both individually and in combination, are critical in behavioral modeling when the student is struggling in finding patterns on the screen. Location : Captures the portion of the screen that the student is currently exploring, discretized into {High, Medium, Low, NA}. Screen Color : Models the current screen background color as a binary node that can be either be green or non-green. Decision : Decision node that contains all the high level decisions the agent can take. For this activity, the agent can choose to instruct the student about the current task, valorize them, prompt them to explore different areas for green, provide location specifications for finding a green or stay quiet to give the student time to think and explore. At each time step, the network is updated with available evidence and provides a decision with maximum expected utility. Directed arcs in the network above show conditional dependencies. Those labeled as ‘t+1’ show temporal dependencies, such that the state of node Goal at time t+1 depends on the goal, screen color and finger touch coordinates of the user at time t.

5

Evaluation

This section describes our first explorative experimental evaluation, methods, results and discusses implications for future pedagogical agent technology.

9

Pedagogical Agents to Support Embodied, Discovery-based Learning t+1

t

t+1

Introduction

0

0

Mid

FindGreen

0

0

Low

MoreGreen

0

0

HighGreen

1

1

LowGreen

0

0

Valorize Educate

High

Location

Utility

Decision

HighGreen

Wait Done

Color

Color

Location

Goal

Decision

Utility

.

. .

. . .

. . . .

. . . . .

. . . . .

Valorize

10

Goal

MakeGreen

LowGreen

Timeout

Green

Yes

NotGreen

No

Timeout

Yes

Not Green

Mid

High Green

Educate

10

MakeGreen

0

HighGreen

0

LowGreen

0

Wait

0

Done

-5

Fig. 5: At this point, the child is either actively exploring, but failing to satisfy the goal, or has stopped interacting with the system. The agent queries the network for the optimal decision to make after setting evidence for Timeout, screen Color and current exploration Location on the screen (bold words). The decision network suggests Valorization and Educating the child about the task as actions of equal and optimal utility (shown in blue). We choose either action randomly. Remedial activities override the decision network and are triggered if there have been multiple continuous timeouts and student hasn’t been able to achieve a goal for an extended time.

5.1

Experiment Description

The agent followed the protocol outlined in Sec. 2.1. Participants included 10 children (5 male, 5 female) aged 9 - 12 years old. Data Gathering: As participants worked with Maria, three sets of data were collected simultaneously. One researcher used an objective-observation instrument, detailed below, to note the occurrence and frequency of key participant movements and expressions. A second researcher took qualitative field notes of the participant’s interaction with Maria. Finally, the interview session was video and audio recorded. The qualitative field notes and video records were used to verify key moments indicated in the observation instruments. Note that these data sources are consistent with qualitative methods. As this system represents a new genre of agent-facilitated, discovery-based learning, we first seek to understand the types of interactions and insights that emerge as children work with the system. Such qualitative work is not consistent with statistical analyses or pre/post comparison, which pre-suppose which ideas will emerge as salient for participants and considers only those insights that can be quantitatively evaluated. With a qualitative understanding in hand, later work can focus the system on insights deemed most mathematically productive, at which stage a quantitative evaluation would be appropriate. This progression is well established within grounded theory [9] and design based research [4]. Observation Instrument: A one-sheet observation instrument was used during each interview to code, in real time, when each participant exhibited particular benchmark movements and expressions. These movements and expressions were determined in consultation with researchers familiar with analogous, human-run interviews. For example, a participant expression of “the higher I go on the screen, the bigger the distance must be between my hands to make green”



A. Abdullah et al.

is a proto-proportional expression that reflects the changing additive difference between numerator and denominator. Additionally, a movement scheme of raising the right hand 2 units for every 1 unit increase in the left hand reflects the constant multiplicative relationship between equivalent ratio (e.g., 4:8 and 5:10). These two strategies are termed “Higher-Bigger” and “a-per-b,” respectively. 5.2

Results

Subject 1 2 3 4 5 6 7 8 9 10

Greens Low 14 12 6 10 11 11 9 7 8 2

by Screen Region Middle High 4 5 7 7 9 3 16 7 4 6 15 7 11 5 5 5 15 8 7 10

“a-per-b” 1 8 4 1 7 9 4 5 5 4

“HigherBigger” 0 0 0 0 0 0 0 0 0 0

“a-per-b” 1 1 2 0 2 0 2 1 2 0

Other Insights 1 3 2 5 1 4 2 3 2 3

Table 1: Frequency of greens, by region, and Table 2: Participant expressions “a-per-b” performance. by category. The observation instrument was designed to measure study participants’ physical movements and/or verbal and gestural utterances that would imply they are engaging in “the higher, the bigger” or “a-per-b” strategies as their means of solving the bimanual manipulation problem. Performance of a particular movement pattern suggests that the participant is enacting a perceptuomotor strategy that could result in conceptual learning. Verbal description of those movements in proto-mathematical terms indicates further progress along that learning pathway per our instructional design. As indicated in Table 1, all participants produced green low down, in the middle, and high up on the screen. Interestingly, no participant described the changing distance between their hands at these various regions (Table 2). Additionally, all participants performed the “a-per-b” strategy (Table 1 indicates the number of times this was observed), and 7 participants verbally described it, such as “for every 1 my left goes, my right goes 2.” Notably, all participants developed other insights into the system’s functioning, which fell into 1 of 4 categories: observational (“Generally, my right hand has to be on top”), feedback-based (“If you see the red screen flash [to green], move back to where you were”), memorization (“4 and 7 or 8 will make green. 6 and 11 or 12 will make green. 2 and 3 makes green.”), and procedural (“Keep one hand in one spot and move your other hand around”). Unanticipated was that researchers were obliged to interact with participants on average twice per interview. In 7 of the 10 interviews, this interaction involved researchers restating the “a-per-b” instructions with similar phrasing to Maria’s.

Pedagogical Agents to Support Embodied, Discovery-based Learning

5.3

11

Discussion

We are encouraged by the widely adopted movement strategies, both for making green in all regions of the screen and in adopting the “a-per-b” strategy. Instructed by Maria, and largely independent of human intervention, all participants developed movement schemes supporting proto-proportional reasoning. This work surfaced a gap, however, between participants’ performed movements and their descriptions thereof. Participants adeptly made green in all regions of the screen and performed the “a-per-b” strategy, yet they did not develop proto-mathematical descriptions of “Higher-Bigger,” only of “a-per-b.” Comparing the conditions of “a-per-b” work with the conditions for “HigherBigger” work suggests sources of this disparity. In the “a-per-b” phase, participants received verbal instructions for the target movement strategy as well as a visual grid and numerals. The verbal instructions highlighted discrete, alternative hand movements (laying the “ -per- ” foundation) while the grid and numerals drew attention to unit quantities (supporting specifically “1 -per-2 ”). In contrast, instructions for the “Higher-Bigger” phase did not explicitly mention the distance between participants’ hands. Additionally, students were not instructed to follow a particular green progression, for example making green low down, in the middle, and up high, that would facilitate noticing a growing distance. Future efforts should focus on understanding how to embed in the pedagogical agent’s models and actions the nuances of human-tutor actions that have led students to attend to the interval between their hands. Patterns in researcher intervention suggest another area for design iteration. Researchers consistently interacted with participants by repeating the “a-per-b” instructions after participants worked unsuccessfully for 3 or more minutes. During this time, participants developed a host of less effective movement patterns - alternating left and right but moving each 1 unit or raising the left hand to the top of the screen then raising the right hand. Though Maria repeated fragments of her original instruction, the timing and particular fragment selected often did not correct the participant’s movement. And while researchers tried to mimic Maria’s exact instructional language, their choice of when to give those instructions and which words to emphasize gave more information than Maria is programmed to do. Further work is required to analyze these researcher interventions and convert their decisions into procedures for the autonomous agent. Overall, we find this work to tentatively support the added value of a virtual agent in discovery-based learning environments. The agent provides feedback on student work and suggests corrective or novel movement strategies that would likely not arise in agent-free work. In particular, the agent draws the student’s attention to multiple parametric regions of the problem space, such as particular spatial locations on the monitor, that had not occurred to the student in their free exploration, and the agent suggests new spatial-temporal interaction schemes, such as introducing a sequential bimanual manipulation regime where the student was trying only simultaneous actions. The agent also provides encouragement, validating the students’ efforts and encouraging them to explore in new ways. As none of the participants noticed the “Higher-Bigger” relationship,



A. Abdullah et al.

this first prototype was somewhat less successful than human tutors. However, this is not surprising. Human tutors perceive a wider range of student behaviors (posture, oral expressions, facial expressions) contemporaneous with their on-screen actions, giving more information upon which to determine the pacing and content of guidance. Additionally, human tutors enjoy the full range of their gestural and verbal vocabularies in responding to and guiding participants. Consequently, we did not expect that the virtual agent would perform to the same benchmark as human tutors. Nevertheless, we see the results of this work as a success, then, in that all participants performed, and almost all expressed, the proportional “a-per-b” relationship under the virtual agent’s guidance.

6

Conclusion

The MITp system presents a very challenging application for pedagogical agents as they must determine appropriate actions based on very little feedback from the learner, in this case, only the location of two markers on the screen. Such constraints are typical of discovery-based learning, where the asking of concrete questions is limited, and the learner is given freedom to explore. The system performed quite well on the task, leading to appropriate movement patterns in all cases and desired verbal expressions for one of the two movement strategies taught. This suggests that the potential for pedagogical agents in discovery-based learning is high and that DDNs represent an effective control strategy. The system was effective due to a very thorough understanding of the tutoring protocol that was then encoded in the DDNs. In our case, this was based on an analysis of human-led tutoring of the same task. Two significant shortcomings were noted in the study, students failed to verbally explain the “Higher-Bigger” pattern and some amount of human intervention was required, generally when students failed to progress later in the last task. Both of these suggest the need to further refine the protocol encoded in the DDNs. Future work should consider using verbal input from the learners. Acknowledgements Financial support for this work was provided by the National Science Foundation through grants IIS 1320029 and IIS 1321042. We gratefully acknowledge Huaguang (Chad) Song, Arman Kapbasov, Rhea Feng, Austin Berbereia, Quan Pham, Gregory Moore and Jessica Sheu for helping develop character content and Alyse Schneider, Virginia Flood and Seth Corrigan for valuable analysis and for providing voice and motion data. Finally, we thank the many children and their parents who participated in our IRB-approved studies.

References 1. Abrahamson, D., Guti´errez, J., Charoenying, T., Negrete, A., Bumbacher, E.: Fostering hooks and shifts: Tutorial tactics for guided mathematical discovery. Technology, Knowledge and Learning 17(1-2), 61–86 (2012)

Pedagogical Agents to Support Embodied, Discovery-based Learning

13

2. Abrahamson, D., Lee, R.G., Negrete, A.G., Guti´errez, J.F.: Coordinating visualizations of polysemous action: Values added for grounding proportion. ZDM 46(1), 79–93 (2014) 3. Anderson, M.L., Richardson, M.J., Chemero, A.: Eroding the boundaries of cognition: Implications of embodiment. Topics in Cognitive Science 4(4), 717–730 (2012) 4. Anderson, T., Shattuck, J.: Design-based research: A decade of progress in education research? Educational researcher 41(1), 16–25 (2012) 5. Antle, A.N.: Research opportunities: Embodied child–computer interaction. International Journal of Child-Computer Interaction 1(1), 30–36 (2013) 6. Boyer, T.W., Levine, S.C.: Prompting children to reason proportionally: Processing discrete units as continuous amounts. Developmental psychology 51(5), 615 (2015) 7. Campbell, J., Core, M., Artstein, R., Armstrong, L., Hartholt, A., Wilson, C., Georgila, K., Morbini, F., Haynes, E., Gomboc, D., et al.: Developing inots to support interpersonal skills practice. In: Aerospace Conference, 2011 IEEE. pp. 1–14. IEEE (2011) 8. Conati, C.: Probabilistic assessment of user’s emotions in educational games. Applied Artificial Intelligence 16(7-8), 555–575 (2002) 9. Creswell, J.W.: Qualitative inquiry and research design: Choosing among five approaches. Sage Publications (2012) 10. Dean, T., Kanazawa, K.: A model for reasoning about persistence and causation. Computational intelligence 5(2), 142–150 (1989) 11. Duijzer, C.A., Shayan, S., Bakker, A., Van der Schaaf, M.F., Abrahamson, D.: Touchscreen tablets: Coordinating action and perception for mathematical cognition. Frontiers in Psychology 8 (2017) 12. Finkelstein, S., Yarzebinski, E., Vaughn, C., Ogan, A., Cassell, J.: The effects of culturally congruent educational technologies on student achievement. In: International Conference on Artificial Intelligence in Education. pp. 493–502. Springer (2013) 13. Gluz, J.C., Cabral, T., Baggio, P., Livi, P.: Helping students of introductory calculus classes: the leibniz pedagogical agent 14. Graesser, A.C., Lu, S., Jackson, G.T., Mitchell, H.H., Ventura, M., Olney, A., Louwerse, M.M.: Autotutor: A tutor with dialogue in natural language. Behavior Research Methods 36(2), 180–192 (2004) 15. Hasegawa, D., Shirakawa, S., Shioiri, N., Hanawa, T., Sakuta, H., Ohara, K.: The Effect of Metaphoric Gestures on Schematic Understanding of Instruction Performed by a Pedagogical Conversational Agent, pp. 361–371. Springer International Publishing, Cham (2015) 16. Howard, R.A.: Readings on the principles and applications of decision analysis, vol. 1. Strategic Decisions Group (1983) 17. Howison, M., Trninic, D., Reinholz, D., Abrahamson, D.: The mathematical imagery trainer: From embodied interaction to conceptual learning. In: Proceedings of the SIGCHI Conference on Human Factors in Computing Systems. pp. 1989–1998. ACM (2011) 18. Jaakkola, T., Singh, S.P., Jordan, M.I.: Reinforcement learning algorithm for partially observable markov decision problems. In: Advances in neural information processing systems. pp. 345–352 (1995) 19. Jackson, G.T., Boonthum, C., McNAMARA, D.S.: istart-me: Situating extended learning within a game-based environment. In: Proceedings of the workshop on intelligent educational games at the 14th annual conference on artificial intelligence in education. pp. 59–68. AIED Brighton. UK (2009)



A. Abdullah et al.

20. Johnson, W.L., Rickel, J.: Steve: An animated pedagogical agent for procedural training in virtual environments. SIGART Bull. 8(1-4), 16–21 (Dec 1997) 21. Johnson, W.L., Lester, J.C.: Face-to-face interaction with pedagogical agents, twenty years later. International Journal of Artificial Intelligence in Education 26(1), 25–36 (2016) 22. Johnson, W.L., Rickel, J.W., Lester, J.C.: Animated pedagogical agents: Faceto-face interaction in interactive learning environments. International Journal of Artificial intelligence in education 11(1), 47–78 (2000) 23. Kim, Y., Baylor, A.L.: A social-cognitive framework for pedagogical agents as learning companions. Educational Technology Research and Development 54(6), 569–596 (2006) 24. Koedinger, K.R., Corbett, A., et al.: Cognitive tutors: Technology bringing learning sciences to the classroom (2006) 25. Lakoff, G., N´ un ˜ez, R.E.: Where mathematics comes from: How the embodied mind brings mathematics into being. Basic books (2000) 26. Lamon, S.J.: Rational numbers and proportional reasoning: Toward a theoretical framework for research. Second handbook of research on mathematics teaching and learning 1, 629–667 (2007) 27. Lester, J.C., Stone, B.A., Stelling, G.D.: Lifelike pedagogical agents for mixedinitiative problem solving in constructivist learning environments. User modeling and user-adapted interaction 9(1), 1–44 (1999) 28. Mott, B.W., Lester, J.C.: U-director: a decision-theoretic narrative planning architecture for storytelling environments. In: Proceedings of the fifth international joint conference on Autonomous agents and multiagent systems. pp. 977–984. ACM (2006) 29. Murray, R.C., VanLehn, K.: Dt tutor: A decision-theoreticdynamic approach for optimal selection of tutorial actions. In: International Conference on Intelligent Tutoring Systems. pp. 153–162. Springer (2000) 30. Pearl, J.: Probabilistic reasoning in intelligent systems (1988) 31. Piaget, J., Inhelder, B.: The psychology of the child, vol. 5001. Basic books (1969) 32. Prendinger, H., Ishizuka, M.: The empathic companion: A character-based interface that addresses users’affective states. Applied Artificial Intelligence 19(3-4), 267–285 (2005) 33. Roscoe, R.D., McNamara, D.S.: Writing pal: Feasibility of an intelligent writing strategy tutor in the high school classroom. Journal of Educational Psychology 105(4), 1010 (2013) 34. Shaw, E., Ganeshan, R., Johnson, W.L., Millar, D.: Building a case for agentassisted learning as a catalyst for curriculum reform in medical education. In: Proceedings of the International Conference on Artificial Intelligence in Education. pp. 509–516 (1999) 35. Tartaro, A., Cassell, J.: Authorable virtual peers for autism spectrum disorders. In: Proceedings of the Combined workshop on Language-Enabled Educational Technology and Development and Evaluation for Robust Spoken Dialogue Systems at the 17th European Conference on Artificial Intellegence (2006) 36. Vanlehn, K., Lynch, C., Schulze, K., Shapiro, J., Shelby, R., Taylor, Treacy, Weinstein, A., Wintersgill, M.: The andes physics tutoring system: Lessons learned. International Journal of Artificial Intelligence in Education 15(3), 147–204 (2005)

WalkNet: A Neural-Network-Based Interactive Walking Controller Omid Alemi

and Philippe Pasquier

School of Interactive Arts + Technology Simon Fraser University, Surrey, BC, Canada {oalemi, pasquier}@sfu.ca

Abstract. We present WalkNet, an interactive agent walking movement controller based on neural networks. WalkNet supports controlling the agents walking movements with high-level factors that are semantically meaningful, providing an interface between the agent and its movements in such a way that the characteristics of the movements can be directly determined by the internal state of the agent. The controlling factors are defined across the dimensions of planning, affect expression, and personal movement signature. WalkNet employs Factored, Conditional Restricted Boltzmann Machines to learn and generate movements. We train the model on a corpus of motion capture data that contains movements from multiple human subjects, multiple affect expressions, and multiple walking trajectories. The generation process is real-time and is not memory intensive. WalkNet can be used both in interactive scenarios in which it is controlled by a human user and in scenarios in which it is driven by another AI component. Keywords: agent movement · machine learning · movement animation · affective agents

1

Introduction

Data-driven movement animation manipulation and generation techniques use recorded motion capture data to preserve the realism of their output while providing some level of control and manipulation. This makes them more suitable for generating affect-expressive movements, compared to the physics-based approaches to modelling and generating movement animation. Data-driven techniques bring the possibility of augmenting a corpus of motion capture data so that human animators have more assets at their disposal. Furthermore, one can use movement generation models in interactive scenarios, where a human user or an algorithm controls the behaviour of the animated agent in real-time. With the increasing demand for content for nonlinear media such as video games, a movement controller that supports generating movements in real-time based on the given descriptions has applications in AI-based agent animation, interactive agent control, as well as crowd simulation. © Springer International Publishing AG 2017 J. Beskow et al. (Eds.): IVA 2017, LNAI 10498, pp. 15 -24 , 2017. DOI 10.1007/978-3-319-67401-8_2

15



O. Alemi and P. Pasquier

Data-driven methods allow for manipulation of the motion capture data, either by concatenating, blending, or learning and then generating data. Concatenation methods repeat and reuse the movements in a motion capture corpus by rearranging them, making longer streams of movements from shorter segments. In blending, the representations of two or more motion capture segments are combined to create a new segment that exhibits characteristics from the blended segments. Compared to other techniques, machine learning models are better at generalizing over the variations in the data and generating movements that do not exist in their training corpus. Some of the machine learning techniques also provide mechanisms for controlling and manipulating what is being generated, making them suitable for controlling virtual agents. The body of the research on machine-learning-based movement generation has some challenges. Controlling the movements of an agent requires a description of the movement to be generated, and a machine learning model that is capable of mapping those descriptions to movement, in real-time. In this regard, the majority of the works suffer from one or more of the following: (1) they do not support controlling the generated movements (e.g., [3]), (2) they only support controlling a single factor (e.g., [1]), (3) the controlling factor is often not clearly defined with respect to an agent’s internal state(e.g., [13]), or (4) the generation process is computationally and/or memory intensive (e.g., [14]). In order to overcome the above limitations, we present WalkNet, a walking movement controller for animated virtual agents. At its core, WalkNet uses a neural network to learn and generate its movements based on a set of given controlling factors. The factors are chosen to work directly with the internal state of the agent, corresponding to the planning, expression, and personal movement signature dimensions of movement. In future, we intent to extend the model to support controlling the functional dimension as well. The agent can plan its walking movements based on any given trajectory. The affective state is modelled by the valence and arousal dimensions of affect. Furthermore, the movement generation model is capable of exhibiting distinctive personal movement signatures (styles). This allows for using the same model for a group of agents that each portray a different character. The main contributions of our approach are summarized below: – Walknet provides control over multiple dimensions of movement in a single model. – Learned over a limited sample of affective states (i.e., only high, neutral, and low points) and only two human subjects, WalkNet learns a generalized space of affect and movement signature. – The generation process is real-time. Unlike graph and tree based structures, there is no need for search or optimization to generate desired movements.

2

Background and Related Work

Controlling Movement Generation In data-driven movement-generation approaches, different techniques, and the combinations of them are used to control

WalkNet: A Neural-Network-Based Interactive Walking Controller

17

and manipulate the characteristics and qualities of motion capture data. These include organizing the data using specialized data structures, such as motion graphs [8,5], as well as blending and interpolating multiple segments. Regarding the machine learning models, there are multiple ways that they support controlling the generation: 1) Train a separate model for each point in the factor space. Each model is trained only on the data that correspond to that particular point, thus only imitating the same factor value. To control the generation, one has to switch between the models. 2) Using a parametric probability distribution, in which the parameters of the distribution are a function of the controlling factors [6], allows for controlling the statistical characteristics of the generated data. 3) By designing the machine learning model in a way that provides a mechanism for a factor variable to control the characteristics of the generated movements. In particular, Factored Conditional Restricted Boltzmann Machine (FCRBM) uses a context variable (Figure 3.b) that controls the behaviour of the network through gated connection between the observations and the hidden variables [12]. Machine Learning Methods for Movement Generation Machine learning models that are used for learning and generating motion capture data range from dimensionality reduction (DR) techniques (e.g., [10]), to the Gaussian Process Latent Variable Models (GPLVMs) (e.g., [14]), Hidden Markov Models (HMMs) (e.g., [2]), temporal variations of the Restricted Boltzmann Machines (e.g., [12,1]), Recurrent Neural Networks (e.g, [3]), and Convolutional Autoencoders combined with Feed-Forward Networks (e.g., [7]). DR techniques do not handle the temporality of the motion capture data. Furthermore, the dimensionality-reduction-based techniques rely on preprocessing steps such as sequence alignments and fixed-length representation of the data. The main limitation of the GPLVMs is that they demand heavy computational and memory resources, which makes them unsuitable for real-time generation. HMMs overcome the limitations of the two aforementioned families of models but provide a limited expressive power regarding capturing the variations in the data. Neural networks provide a better expressive power than HMMs. Convolutional Autoencoders have shown promising results in generating motion capture data and offline controlling [7]. Factored Conditional RBM (FCRBM), with its special architecture that is designed to support controlling the properties of the generated data, has shown to be able to generate movements in real-time, and learn a generalized space of the movement variations [12,1]. Affect-Expressive Movement Generation Taubert et al. [11] combine a Gaussian process latent variable model (GP-LVM) with a standard HMM that learns the dynamics of the handshakes, encoded by the emotion information. Samadani et al. [10] use functional principal component analysis (FPCA) to generate hand movements. Alemi et al. [1] train an FCRBM to control the valence and arousal dimensions of walking movements. Our Approach We build WalkNet on top of the previous work by the same authors [1], extending the affect-expression control with the walking planning



O. Alemi and P. Pasquier

 


!
"
!
!(


!&"
"#"


 


(
"#!""


'#
 


 


!&"#!#


#
 


  


"!


 





(





"



 


 


 !""
!


 


! (


!


##
$"
'



 (


Fig. 1. The affect model described by valence and arousal dimensions with the 9 zones recorded in the training data. The mapping to the categorical emotion labels are based on Plutchik and Conte [9]. H : high, N : neutral, L: low, V : valence, and A: arousal.

and personal movement signature. Our work differs with graph-like structures as it does not require to build an explicit and fixed data-structure, does not require search and optimization for generating movement, and does not require storing the movement data for generation. It also differs from the work of Crnkovic-Friis and Crnkovic-Friis [3] as it provides a mechanism to control the generated data. It allows for real-time and iterative generation compared to the work of Holden et al. [7].

3

Training Data

For training the model, we use a set of motion capture data that provides movements with variations in walking direction (planning), the valence and arousal levels (expression), and the personal movement signature. As we could not find a publicly available motion capture database that provides movements with such variations, we recorded our own set of training data. The complete data set is publicly accessible in the MoDa database1 . The training data includes the movements of two professional actors and dancers (one female, one male). Each subject walks following a curved figure-8shaped path. The turning variations in this pattern allow the machine learning model to learn a generalized space of turning directions. To capture a space of affect-expression, each subject performs each movement with nine different expressions along the valence and arousal dimensions [9], shown in Figure 1. 1

http://moda.movingstories.ca/projects/29-affective-motion-graph

19

WalkNet: A Neural-Network-Based Interactive Walking Controller



 
 

 

 

 





 

   


  

  




  









Fig. 2. The WalkNet controller, embedded in an agent model.

Using the dimensional representation of affect over the categorical systems allows for interpolation and extrapolation of the affect states, as well as transitions. Each valence and arousal combination is repeated four times to capture enough motor variabilities. The original motion capture data consists of a skeleton with 30 joints, resulting in 93 dimensions including the root position, with their rotations represented in Euler angles. The data is captured at 120 frames-per-second. We use exponential maps [4] to represent joint angles to avoid loss of degrees-of-freedom and discontinuities. We replace the skeleton root orientation and translation by the delta values of the translational velocity of the root along the floor plane, as well as its rotational velocity along the axis perpendicular to the floor plane. We remove the dimensions of the data that are constant or zero and downsample the data to 30 frames-per-seconds. The final data set used for the training consist of 18 motion capture segments (2 subjects × 9 affective states), containing 37,562 frames in total, with 52-dimensional frames.

4

The Walking Controller

System Overview As shown in Figure 2, at the core of the WalkNet, the movement generator, a Factored, Conditional Restricted Boltzmann Machine (FCRBM), generates a continuous stream of movement. The movement stream is modulated by a set of controlling factors, determined from the internal state of the agent or through external commands. From an agency perspective, we organize these factors into different dimensions, mainly the function, the planning, the expression, and factors that together make the personal movement signature of the agent. WalkNet does not make any assumptions on how the agent



O. Alemi and P. Pasquier

movement descriptor is set. Thus, making it flexible to be integrated into various agent models for different applications. Agent Movement Descriptor We use an agent movement descriptor AMD to formalize the contributing factors to the agent’s movements at time t along the dimensions of function (F ), planning (P ), expression (E), and personal movement signature (S): AMDt = Ft , Pt , Et , St  In WalkNet, the F is always set to walking. The planning dimension of walking is defined by Pt = Dt  where Dt represents the direction that the agent intends to walk towards, relative to its current orientation. The expression dimension is defined by Et = Vt , At  where Vt and At stand for valance and arousal levels at time t respectively. Currently, we use the actor/performer’s identity as a proxy to model the personal movement signature, through a weighted combination of a K-dimensional vector, representing K subjects: St = {It1 , It2 , .., ItK |

K 

Itk = 1}

k

We recognize that this is a simple way of capturing movement signature. In the future, we plan on learning a representation that captures the personal movement signature. Training Data Annotation Here we describe how we annotate our data to capture different states of the factors in the agent movement descriptor. As we have two human subjects in our training data, we use a 2-dimensional label with a one-hot encoding scheme for the movement signature. We use the valence and arousal representation of affect to annotate the expression of affect. Each movement segment in the training data is labeled with low, neutral, and high for both their valence and arousal levels. After experimenting with different ranges, we use the values of 1, 2, and 3 to represent low, neutral, and high levels in the annotations. Although the training labels are discrete, the valence and arousal values are continuous in nature, and for the generation, one can specify any real value within the range of [0, 4], as the FCRBM is able to interpolate or extrapolate between those discrete states. For annotating the heading direction, we determine the labels using a method that is inspired from Kover et al. [8]. For the label at frame t, considering the projection of the traveled path of the skeleton root on the ground floor, we select two points on the path, one at a very close distance to the current location, and another one at a slightly further location from the current location (Figure 3.a). We calculate the angle between the two lines that result from connecting the two chosen points and use this angle as a measure of the heading direction. After scaling the angle to have a value between -1 and +1, directions towards the right of the subject are associated with positive numbers, and directions towards the left of the subject are associated with negative numbers.

21

WalkNet: A Neural-Network-Based Interactive Walking Controller 





 

ht

zt




vt

v 0

If the coefficient ga is 0, the variable bounds are in any case constrained to amin ≤ xa ≤ amax , as defined in the basic model. With R = 0 the bounds are strict and the linear problem is harder to solve, while with R = 1 there is no more effect of the capping and the “one-by-one increment” behavior arises. Figure 3 (right) shows the behavior with R = 0.1. To verify that the capping mechanism improves the precision, we recomputed the MSEs table with the capping enabled and a relaxation set at 0.25. Although the solve rate decreased from 99.8% to 84.3%, the average error decreased from 51.9% to 31.7% (sd 5.1%). Evaluating the coercion. We evaluated the behavior of the coercion mechanism for all of the three traits in both p and R selection modes. For each trait (D, T, or A), we solved the traits-to-attributes problem with and without the capping

311

Generation of Virtual Characters from Personality Traits Table 5. Correlation gain between noCap and Cap conditions.

noCap trait mean (sd) D T A

p Cap mean (sd)

R gain

0.619 (0.345) 0.992 (0.005) 60.29% 0.776 (0.179) 0.820 (0.339) 5.68% 0.781 (0.176) 0.895 (0.186) 14.60%

noCap Fisher’s p mean