Open Collections

UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Negotiating with robots : meshing plans and resolving conflicts in human-robot collaboration Moon, AJung 2017

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata


24-ubc_2017_september_moon_ajung.pdf [ 8.29MB ]
JSON: 24-1.0348225.json
JSON-LD: 24-1.0348225-ld.json
RDF/XML (Pretty): 24-1.0348225-rdf.xml
RDF/JSON: 24-1.0348225-rdf.json
Turtle: 24-1.0348225-turtle.txt
N-Triples: 24-1.0348225-rdf-ntriples.txt
Original Record: 24-1.0348225-source.json
Full Text

Full Text

Negotiating with RobotsMeshing Plans and Resolving Conflicts in Human-Robot CollaborationbyAJung MoonB.A.Sc., Mechatronics Engineering, The University of Waterloo, 2009M.A.Sc., Mechanical Engineering, The University of British Columbia, 2012A THESIS SUBMITTED IN PARTIAL FULFILLMENTOF THE REQUIREMENTS FOR THE DEGREE OFDoctor of PhilosophyinTHE FACULTY OF GRADUATE AND POSTDOCTORAL STUDIES(Mechanical Engineering)The University of British Columbia(Vancouver)June 2017c© AJung Moon, 2017AbstractFor both humans and robots, one of the key elements of collaboration is the collaborating agents’ability to communicate and mesh their individual plans with each other. Even after the collaborators havedecided on a joint task, the details of the task such as the how, when, and where are often determined asthe collaborative activity unfolds. The primary objective of this thesis is to enable fluent communicationbetween human and robotic agents such that they can interactively figure out unspoken details andresolve unforeseen conflicts that arise during a human-robot collaboration.The author first explores whether robot nonverbal cues inspired by human behaviours can elicitdesirable responses from a human user to interweave unspoken – yet essential – spatial-temporal detailsof an interaction. Results from a series of experiments demonstrate that a robot cue, like gaze, can havea significant influence on when a human recipient reaches out to receive an object from a robot.Subsequently, the author focuses on hesitation gestures – a type of gesture humans naturally useto express uncertainty – to explore whether members of a human-robot dyad can negotiate a desiredoutcome of an interaction through a nonverbal dialog. The author presents a reactive, real-time tra-jectory generator, the Negotiative Hesitation Generator (NHG), which has been devised to enable suchnonverbal negotiation to take place between a human and a robot. The NHG was implemented on arobot for human-robot interaction experiments where, by design, spontaneous resource conflicts oftenarose between the two agents. Results from these studies suggest that use of the NHG can enable a typeof nonverbal negotiation of resource conflicts to take place. They also demonstrate how such real-timenegotiations between a human-robot dyad can lead to a faster resolution of conflicts and a significantlyimproved outcome of the collaborative task, without jeopardizing the safety of the user.This thesis advances our understanding of the influence that nonverbal robot behaviours can haveon human users. It also demonstrates the feasibility and efficacy of nonverbal negotiations as a mode ofinteraction for human-robot collaboration.iiLay SummaryCommunication is essential to successful collaboration between people. The same is true for human-robot collaboration. The objective of this thesis is to investigate how robots can use nonverbal be-haviours to better communicate and collaborate with people.In this work, a robot was programmed to use human-inspired gaze cues as a nonverbal communi-cation mode while handing over an object to a person. Experimental results suggest that people reachfor the object earlier when a robot uses human-inspired gaze behaviours than when it does not. In afollowing experiment, a robot was programmed to appear hesitant when it and a person reached for thesame object simultaneously during a collaborative task. Results suggest that such behaviour can allowa robot to quickly and safely resolve the conflict with a person.Based on this work, future human-robot collaborations can be designed to be more effective andsafer for both parties.iiiPrefaceThis thesis is comprised of a number of collaboratively written publications. Chapter 3 containsedited versions of the following:• Moon, A., Troniak, D. M., Gleeson, B., Pan, M. K. X. J., Zheng, M., Blumer, B. A., MacLean,K., Croft, E. A. (2014). Meet me where I’m gazing: How shared attention gaze affects human-robot handover timing. In International Conference on Human-Robot Interaction (pp. 334-341).Bielefeld, Germany: ACM/IEEE. [105]1• Zheng, M., Moon, A., Gleeson, B., Troniak, D. M., Pan, M. K. X. J., Blumer, B. A., Meng, M. Q.H., Croft, E. A. (2014). Human behavioural responses to robot head gaze during robot-to-humanhandovers. In 2014 International Conference on Robotics and Biomimetics (pp. 362-367). Bali,Indonesia: IEEE. [167]The author, AJung Moon, contributed to [105] in the design of the experiments in collaboration withBrian Gleeson and Minhua Zheng. The author led the analysis of the results with the assistance ofMinhua Zheng. Daniel Troniak spearheaded the technical implementation of the system used in theexperiment along with Benjamin Blumer, and Matthew Pan. Minhua Zheng, in collaboration with theauthor, conducted a qualitative analysis of the same experiment published in [167].A follow-up study (Study 6) of the material presented in Chapter 3 has been included in Section A.1as a supplementary research material. The follow-up study has been published in:• Zheng, M., Moon, A., Croft, E. A., Meng, M. Q.-H. (2015). Impacts of Robot Head Gaze onRobot-to-Human Handovers. International Journal of Social Robotics, 7(5). [168]Minhua Zheng and the author collaboratively designed the in situ experiment for [168]. The exper-iment and the majority of the analysis were conducted by Minhua Zheng with the supervision of theauthor and Elizabeth A. Croft.The studies presented in Chapter 4 are being prepared for submission as a journal publication:• Moon, A., Billard, A., Van der Loos, H. F. M., Croft, E. A., (n.d.). Development of NegotiativeInteraction for Nonverbal Resolution of Human-Robot Conflicts. (in preparation)1This paper has won the Best Paper award at IEEE/ACM Conference on Human Robot Interaction 2014ivThe author led the design and analysis of the experiments with the supervision of Elizabeth Croft andH. F. Machiel Van der Loos. Aude Billard supervised the data exploration of the first of the two human-subjects experiments (Studies 3 and 4) as well as the development of the Negotiative Hesitation Gener-ator (NHG) controller inspired by the results of the experiments.The final study (Study 5) discussed in Chapter 5 is also being prepared for submission as a journalpublication:• Moon, A., Croft, E. A., Van der Loos, H. F. M., Billard, A. (n.d.). Bidirectional Interweavingof Subplans using Negotiative Interaction in Human-Robot Collaborative Assembly. (in prepara-tion)The author designed the experiment and conducted the analysis of the results. Aude Billard supervisedthis process along with Elizabeth Croft and H. F. Machiel Van der Loos.The University of British Columbia (UBC) Behavioural Research Ethics Board approved all userstudies conducted as part of this thesis (H10-00503, “HRI Cues”). Study 5 presented in Chapter 5has been conducted at Ecole Polytechnique Fe´de´rale de Lausanne (EPFL), Lausanne, Switzerland. It hasbeen approved by EPFL Human Research Ethics Committee (HREC: 001-2016 / 12.012016) in additionto the approval from UBC.Financial support for this research was provided by the Natural Sciences and Engineering ResearchCouncil of Canada, the Canada Foundation for Innovation, the UBC Institute for Computing, Informationand Cognitive Systems (ICICS), and the European Union project, AlterEgo, under the grant agreement#600010.vTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiLay Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ixList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xGlossary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiiAcknowledgements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xiv1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 Examples: Physical Environment and Human Behaviour . . . . . . . . . . . . . . . . 31.2 Research Questions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51.3 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62 Background and Motivating Literature . . . . . . . . . . . . . . . . . . . . . . . . . . . 72.1 Human-Robot Collaboration and Communication . . . . . . . . . . . . . . . . . . . . 72.2 Robots as a Mechanism of Influence . . . . . . . . . . . . . . . . . . . . . . . . . . . 92.2.1 Unidirectional Influence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102.2.2 Bidirectional Influence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133 Unidirectional Interweaving of Timing and Space in Robot to Human Handovers . . . . 153.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173.2.1 Human-Robot Handover . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173.2.2 Gaze in Human-Robot Interaction . . . . . . . . . . . . . . . . . . . . . . . . 18vi3.3 Study 1: Observing Gaze Patterns in Human-to-Human Handovers . . . . . . . . . . . 193.3.1 Experimental Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193.3.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193.3.3 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213.4 Study 2: Impact of Human-Inspired Gaze Cues on First-Time Robot-to-Human Handovers 223.4.1 Physical Handover Cues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223.4.2 Experimental Gaze Cues . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223.4.3 Experimental Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243.4.4 Technical Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 253.4.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 263.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 313.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 334 Development of Negotiative Interaction for Nonverbal Resolution of Human-Robot Con-flicts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354.2 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 374.2.1 Persistency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 404.2.2 Social Signal Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 414.3 Study 3: Observing Hesitations in Human-Human Dyads . . . . . . . . . . . . . . . . 414.3.1 Experimental Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 424.3.2 Results and Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 444.4 Exploring Human Hesitation Trajectories . . . . . . . . . . . . . . . . . . . . . . . . 444.4.1 Pre-Processing and Segmentation . . . . . . . . . . . . . . . . . . . . . . . . 454.4.2 Sample Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 464.4.3 Data Exploration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 474.4.4 Feature Differences in Reach and Hesitation Motion Samples . . . . . . . . . 494.4.5 Understanding Hesitation Loops . . . . . . . . . . . . . . . . . . . . . . . . . 514.4.6 The Four Cases of Hesitation Loops . . . . . . . . . . . . . . . . . . . . . . . 534.5 Design of the Negotiative Hesitation Generator . . . . . . . . . . . . . . . . . . . . . 574.6 Study 4: Validating the Negotiative Hesitation Generator . . . . . . . . . . . . . . . . 584.6.1 Experimental Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 604.6.2 Technical Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 634.6.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 644.6.4 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 714.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 725 Study 5: Bidirectional Interweaving of Subplans using Negotiative Interaction in Human-Robot Collaborative Assembly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 745.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74vii5.1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 765.2 Experimental Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 785.2.1 Experimental Task . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 785.2.2 Questionnaire and Interview . . . . . . . . . . . . . . . . . . . . . . . . . . . 795.2.3 Pilot Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 805.2.4 Experimental Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 825.3 Technical Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 825.3.1 Trajectory Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 825.3.2 Safety Features and Motion Tracking . . . . . . . . . . . . . . . . . . . . . . 845.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 855.4.1 Subjective Experience . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 855.4.2 Collaborative Task Performance . . . . . . . . . . . . . . . . . . . . . . . . . 885.4.3 Trajectory and Behaviours . . . . . . . . . . . . . . . . . . . . . . . . . . . . 915.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 945.6 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 986 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 996.1 Unidirectional Interweaving of Subplans . . . . . . . . . . . . . . . . . . . . . . . . . 1006.1.1 Limitations and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . 1016.2 Bidirectional Interweaving (Negotiation) of Subplans . . . . . . . . . . . . . . . . . . 1016.2.1 Limitations and Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . 1036.3 Efficient Human-Robot Interaction (HRI) vs. Preferred HRI . . . . . . . . . . . . . . . 1046.4 Broader Implications and Future Directions: Inclusivity and Bidirectionality . . . . . . 105Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107A Supplementary Investigations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 121A.1 Study 6: Impact of Gaze on Non-naı¨ve Handover Behaviour . . . . . . . . . . . . . . 121A.1.1 Experimental Procedure and Hypotheses . . . . . . . . . . . . . . . . . . . . 122A.1.2 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125A.2 Supplementary Investigations on Human Hesitations . . . . . . . . . . . . . . . . . . 125A.2.1 Correlation between Features . . . . . . . . . . . . . . . . . . . . . . . . . . 127viiiList of TablesTable 3.1 Study 2: Ranking of questionnaire results on participant preference of robot gaze cues 28Table 4.1 Study 3: Internal reliabilities of the self-reported measures . . . . . . . . . . . . . . 62Table 4.2 Study 4: Results summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66Table 5.1 Study 5: Internal reliabilities of the self-reported measures . . . . . . . . . . . . . . 86Table 5.2 Repeated-measures ANOVA results on self-reported human perception measures . . 87Table 5.3 Repeated-measures ANOVA results on HR team’s number of lentils processed . . . . 91Table 5.4 Repeated-measures ANOVA results on Human-Robot (HR) team throughput . . . . . 92Table A.1 Optimum λ values obtained from Shooting Algorithm . . . . . . . . . . . . . . . . 126Table A.2 List of features with non-zero weights from the shooting algorithm employed onhuman hesitation trajectory data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 128Table A.3 Number of significant results found for each feature across the four runs of t- andWelch tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129Table A.4 Ratio of the main participant’s Euclidean distance to target at zero velocity crossings 130ixList of FiguresFigure 1.1 Human-human and human-robot hesitation demonstration . . . . . . . . . . . . . 4Figure 3.1 Human-human handover demonstration . . . . . . . . . . . . . . . . . . . . . . . 20Figure 3.2 Distribution of participants’ reaching and gaze behaviour . . . . . . . . . . . . . . 20Figure 3.3 Demonstration of the experimental set-up and the three conditions at the handoverlocation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23Figure 3.4 Depiction of handover gaze cues . . . . . . . . . . . . . . . . . . . . . . . . . . . 24Figure 3.5 Experiment system flow diagram . . . . . . . . . . . . . . . . . . . . . . . . . . . 26Figure 3.6 Observed HR handover timeline . . . . . . . . . . . . . . . . . . . . . . . . . . . 27Figure 3.7 Distribution of participants’ reaching and gaze behaviour . . . . . . . . . . . . . . 30Figure 4.1 Overview of the process taken to design and test the NHG . . . . . . . . . . . . . . 36Figure 4.2 Demonstration of the Acceleration-based Hesitation Profile (AHP) . . . . . . . . . 39Figure 4.3 Study 3 experimental setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43Figure 4.4 Acceleration profiles of negotiative hesitations . . . . . . . . . . . . . . . . . . . 45Figure 4.5 Representative participant motions presented as Euclidean distance trajectories, d1(t),with respect to the two target locations, Tm1 and Tm2 . . . . . . . . . . . . . . . . . 47Figure 4.6 Illustration of an Support Vector Machine (SVM) model used to classify hesitationmotions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51Figure 4.7 Overlay of hesitation samples demonstrating the presence of hesitation loops . . . 52Figure 4.8 Overlay of reach samples in δ (t) vs. δ˙ (t) state space. . . . . . . . . . . . . . . . . 53Figure 4.9 Distribution of Kickback Distance (KD) collected from the 134 samples of hesita-tions in (Nhes≥4 = 192) that encircle δ˙ (t)=0 . . . . . . . . . . . . . . . . . . . . . 54Figure 4.10 The four cases of hesitation loops demonstrated in the δ˙ (t) vs. δ (t) state space . . 55Figure 4.11 Simulation of the NHG implementation with quintic splines . . . . . . . . . . . . . 59Figure 4.12 A screen capture of a video shown to participants in Study 4 . . . . . . . . . . . . 61Figure 4.13 Outline of the conditions tested for Study 4 . . . . . . . . . . . . . . . . . . . . . 62Figure 4.14 Perceived Hesitancy across different levels of KD and Re-attempts (RA) . . . . . . 67Figure 4.15 Perceived Persistency by Kickback Distance (KD) values . . . . . . . . . . . . . . 68Figure 4.16 Perceived Animacy by Kickback Distance (KD) values . . . . . . . . . . . . . . . 69Figure 4.17 Perceived Anthropomorphism by Kickback Distance (KD) values . . . . . . . . . . 70xFigure 4.18 Perceived Dominance by Kickback Distance (KD) values . . . . . . . . . . . . . . 71Figure 5.1 Experiment set-up of Study 5 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79Figure 5.2 Experiment procedure of Study 5 . . . . . . . . . . . . . . . . . . . . . . . . . . 80Figure 5.3 Robot motions without encountering any conflict in the Stop and Negotiate conditions 84Figure 5.4 Task completion time for each condition . . . . . . . . . . . . . . . . . . . . . . . 90Figure 5.5 Number of re-attempts observed in a trial . . . . . . . . . . . . . . . . . . . . . . 93Figure 5.6 Participant and robot motions during a Stop condition trial with a conflict of resource 94Figure 5.7 Participant and robot motions during a Negotiate condition trial with a conflict ofresource . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95Figure A.1 Timeline of the robot’s gripper motion and head gaze for the conditions in Study 6 123Figure A.2 Study 6 experiment set-up . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123Figure A.3 Regularization path from Shooting Algorithm applied to the four sets of motionsamples tested . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 126Figure A.4 ROC curve of one of SVM models using three features . . . . . . . . . . . . . . . 127Figure A.5 Correlation between Hesitation and Persistency scores obtained from the Mechani-cal Turk survey (Study 3). . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130xiGlossaryAHP Acceleration-based Hesitation Profile, a characteristic trajectory profile commonly observed in aparticular type of hesitation gesture as elaborated in [104].ANOVA Analysis of Variance, a set of statistical techniques to identify sources of variability betweengroupsDS Dynamical SystemHH Human-Human conditionHR Human-Robot conditionHHI Human-Human InteractionHRI Human-Robot InteractionHCI Human-Computer InteractionH2T Human to TargetICC Intra-class Correlation CoefficientKD Kickback DistanceLDS Linear Dynamical SystemMLM Multi-level ModellingNHG Negotiative Hesitation GeneratorRA Re-attemptsREML Restricted Maximum LikelihoodROS Robot Operating SystemR2T Robot to TargetSSP Social Signal ProcessingxiiSVM Support Vector MachineTS Trigger StateZVC Zero Velocity CrossingxiiiAcknowledgementsThis thesis would not have come to fruition without the help of many writers and academics inmy life. These mentors, friends, and family members fought alongside me as I learned to silence myinner sceptics. They include: My mother, Mi-Ja Park, who continues to fight her own battles as sheexpresses the life of expatriates through fiction; Franc¸ois Garillot, whose love, commitment to providefreshly brewed cups of coffee, and gifts of backup systems proved to be invaluable as I shipped the manychapters in this thesis; Shalaleh Rismani, who not only shared a daily thesis writing tracker to keep bothof us on track, but also motivated me with daydreams about what is awaiting us at the end of the tunnel;John Havens, Executive Director of the IEEE Global Initiative for Ethical Considerations in ArtificialIntelligence and Autonomous Systems, who has been there to support my role as a chair of one of theinitiative’s committees as I intermittently disappeared to focus on my thesis.I would also like to thank: My father, Young-Gu Moon, who sparked my interests in technologyand has never second guessed or doubted my technical abilities as an engineer even in the face of myfailures; My sister, Hyo-Jung Moon, who was there for me the day I moved across the country to startmy research career in Vancouver as well as the eleventh time I moved between continents, cities andapartments as part of this journey.I am grateful to the following individuals who provided technical support and assistance throughoutthe studies outlined in Chapters 3 and 4: Minhua Zheng, Daniel Troniak, Brian Gleeson, Matt Pan,Benjamin Bloomer, Karon MacLean, Maneezhay Hashmi, Bikram Adhikari, Fernando Ramirez, LarsBollmann, Katelyn Currie, Jake Rose, Justin Hart, Junaed Sattar, Eric Fu, and John Petkau.I would also like to thank Aude Billard and my labmates at Learning Algorithms and SystemsLaboratory (LASA), EPFL, who generously welcomed me into the lab and provided me with criticaland insightful discussions that led to the work outlined in Chapters 4 and 5. I would especially like tothank Sina Mirrazavi, Nadia Figueroa, Mahdi Khoramshahi, Nicolas Sommer, Lucia Pais Ureche andGuillaum De Chambrier at LASA who were always there to help diagnose numerous technical issues.Lastly, I would like to thank my advisors Elizabeth A. Croft and Mike Van der Loos, whose men-torship and support extended much beyond the contents of this thesis and are the very foundation of theresearcher I have become today.xivChapter 1IntroductionIn 1994, at the American Association for Artificial Intelligence’s National Conference on ArtificialIntelligence, Barbara J. Grosz delivered her presidential address titled “Collaborative Systems.”[56]Taking inspiration from Lochbaum [90], who stated that “a buddy is better than a slave,” that is, asystem that works with a user to problem solve is better than one that only takes and follows orders,she encouraged the research community to take interest in human-machine collaboration and shared hervision of what such a system would look like [56]. Since then, the field of robotics has been motivatedto develop interactive collaborative robotic systems in part as a means to address the global, impendingshortage of labour that is expected as a result of rapidly ageing demographics in the industrializedworld.1Today, industrial robotic manufacturers such as KUKA (Augsburg, Germany), ABB (Zu¨rich, Switzer-land) and Franka Emika (Munich, Germany) are spearheading the vision of robots working side byside with human workers in manufacturing facilities without barricades. They envision such Human-Robot (HR) collaboration to yield higher throughput per human worker. Rethink Robotics’ (Boston,MA) collaborative assembly robotic systems,2 for example, have already been changing manufacturingenvironments and processes in ways that computers and single-purpose automation systems could not[77].In step with the rise of collaborative robotic systems for manufacturing has been an increase in thenumber of robotic products promising to enter our homes and offices as social companions and assis-tants. For instance, Pepper is a 20 degree-of-freedom (DOF) humanoid robot from SoftBank Robotics(Tokyo, Japan) that has been marketed in 2016 as a “genuine day-to-day companion whose number onequality is his ability to perceive emotions” [133]. Jibo is a 3-DOF robot advertised as “The World’s FirstSocial Robot for the Home,” and its founders raised over $3.7M USD through a 2014 crowd-funding1 According to a July 2015 report from Statistics Canada [143], more Canadians are over the age of 65 (16%) than underthe age of 15 (16%). Canada and the United States have the lowest proportion of the population over the age of 65 comparedto other countries in the G7. The population ageing phenomenon is found globally, with Japan having 26% of the populationbeing 65 and older as of 2015. The field of robotics has been especially motivated to provide a technological solution to theprojected societal consequences [92].2 Rethink Robotics is one of the few companies to develop and deploy interactive manufacturing robots. Called Baxter andSawyer, they can be operated safely with humans in their workspace.1campaign [21]. Both robots are targeted at small businesses and the consumer market.The rise in social and collaborative robotic systems is in part enabled by the development of newinterfaces. Compared to the 1990s, when the keyboard, mouse, and physical buttons dominated theway people interacted with everyday technologies, two decades of research and development in roboticsand computer science have introduced new user interfaces and ubiquitous technologies that allow formore natural and seamless interaction with today’s electronic devices. An iconic example is the waypeople interact with smart phones. With the advancement of natural language processing, people canprovide verbal commands to their smart phones and other electronic devices to accomplish simple taskssuch as scheduling reminders, navigating the streets, or checking the weather forecast without the needto type commands or press numerous buttons (e.g., Google Now [55], Apple’s Siri [7], and voice-command-enabled cars allow for a hands-free experience). However, while our interaction with small orso-called ‘smart’ electronic devices has been deeply integrated into the daily lives of those in developedeconomies, interactive robots have yet to achieve the same penetration into the consumer market. Thehuman-machine collaboration Grosz envisioned has yet to be realized in robotics.What does collaboration comprise and what challenges remain? In Bratman’s model, supported andre-iterated by Grosz and Kraus [58], there are four key elements to a collaboration:3 mutual respon-siveness, commitment to the joint activity, commitment to mutual support, and meshing of subplans[20]. All of these elements require collaborating agents to communicate with each other, and implythe dynamic (i.e., constantly changing or evolving) nature of a collaborative process. For example, anagent’s state of commitment to the joint activity and attitude toward mutual support can change overtime, thereby breaking a collaborative process into separate, individual processes. Moreover, whileindividual agents may have their own subplans that are framed to serve a joint goal, details (e.g., spa-tiotemporal information) of the subplans are often not explicitly communicated to the other parties atthe onset. These details need to be communicated and interwoven with those of the other collaboratingpartners who themselves have subplans to help achieve the joint goal. This interweaving process in-cludes negotiating the details of a subplan when an agent’s subplan conflicts with those of the others orwhen an unforeseen problem arises [56]. Given the foundational role communication plays in establish-ing the four elements of collaboration, the author posits that the development of natural communicationmechanisms for meshing and negotiating details of subplans is key to realizing fluid and effective HRcollaboration.Unlike interaction with other electronic devices, robots are embodied in hardware, occupying phys-ical space and manipulating objects in that space. The physical embodiment of robots not only providesnew opportunities for what technology can deliver for human society – such as physically assisting anolder person rise from a chair or lifting heavy objects to a more ergonomic position for human work-ers in manufacturing facilities – but also a unique modality for interaction and communication throughphysical presence and kino-dynamic behaviour that can facilitate an HR collaboration.The main objective of this thesis is to investigate the role human-inspired robot nonverbal cues can3 Bratman used an alternate term, shared cooperative activity, to describe the type of collaborative activity pertinent to thisdiscussion. While researchers have used other terminologies to describe similar multi-agent activities, this thesis uses the termcollaboration for consistency.2have in the interweaving and negotiating of subplans between interacting and collaborating human androbotic agents. In particular, this work is focused on the last of the four elements of collaboration(meshing of subplans) where the intention to collaborate, the task at hand, and the roles assigned to eachof the agents are not left to question, yet details of the interaction such as the how, when, and whereremain to be determined.1.1 Examples: Physical Environment and Human BehaviourTo take advantage of the physical embodiment of robots, roboticists are tasked with the problem ofhow a robot should best use, occupy, and share spaces and objects with human users in the environ-ment. Communication in Human-Human (HH) collaboration includes both verbal and nonverbal means.HRI researchers who focus on nonverbal communication, in particular, examine Human-Human Interac-tion (HHI) in careful detail, extracting and characterizing interactions that are usually taken for grantedin our everyday interactions with other humans. This process helps Human-Robot Interaction (HRI)researchers to discover curious ways in which we humans get along and perform activities with eachother that dynamically help shape the outcome of the interactivity, be it psychological, behavioural,performance-oriented and so on.Take, for example, the case of an able-bodied adult, Jane, walking down a street and another adult,John, walking on the same narrow street from the opposite direction. Jane may notice John and decideto alter her path closer to the right side of the street in order to avoid a head-on collision with John,a stranger. In turn, John may notice this and change his path slightly to his right (i.e., to the oppositeside of the street), thereby giving both of them enough room to pass by each other without a collision,or a near-collision. In this scenario, there is a clear resolution on what behaviours result in fluent, safe,and collision-free interaction between the two agents. Jane acts as the main agent that leads what Johnshould do to make the sharing of the space (the street) work between the two.The effect of how humans influence one another is even more pronounced when considering the caseof Jane and John noticing each other in close proximity at the same time. In order to avoid an imminentand unpleasant head-on collision with a stranger, the two pedestrians may unintentionally choose thesame edge of the street to yield to the other. Observing this, they further decide to yield again andquickly shuffle to the other edge of the street, thereby finding themselves in the same and unresolvedsituation of imminent collision. This livelock can continue until their yielding behaviour falls out ofsynch or one of them decides to explicitly yield the right of way to the other.These series of events, finding ourselves in an unexpected conflict of a resource (the resource beingspace, in the above-mentioned cases) and having uncertainties about how it should be resolved,4 causesus to respond to the situation in interactive and communicative ways. The quick shuffling behaviour inthe second scenario is an example of hesitation behaviours that one often observes in human interac-tions. In HHI, they often result in resolution of such conflicts using only nonverbal cues to negotiate the4 For example, a four-way stop intersection in the traffic law of the province of British Columbia specifies that the firstcar to arrive at the intersection has the right of way. In the case of multiple cars arriving at the same time, it is expected thatthe right-most car in the intersection has the right of way first. However, much of our everyday interactions do not includeprescribed rules about who should yield the right of way.3Figure 1.1: Hesitation behaviours discussed in this thesis are set in the context of two agents reach-ing for the same resource at the same time. The photo on the left depicts such an instancestudied in Chapter 4 where two people are reaching for the same deck of cards at the sametime. The photo on the right illustrates an analogous resource conflict scenario in an HRI inwhich a human and a robot reach for the same object at the same time. The HR scenario isexplored in Chapter 5.outcome.Ongoing work in the field of HRI related to robot physicality and presence are focused on the designfeatures of a robot that best elicit specific responses from human users [3, 19, 22]. Findings from thesestudies provide points of reflection on how we affect each other’s behaviours and how, increasingly,robots are becoming a source of that influence. As outlined further in Section 2.2, previous studies in HRIhave explored ways in which robot nonverbal cues can help establish joint attention, share joint intention,and express internal states. Building on these studies, this thesis focuses on design and implementationof nonverbal communication cues within the context of HR collaboration as a means to interweave andnegotiate about subplans of an activity.Herein, the term interweaving is used as a shorthand to refer to the process in which multiple collab-orating agents mesh their independently formed subplans with those of the others in order to accomplishthe shared goal. It is used as an umbrella term encompassing processes that are unidirectional – suchas the first scenario of Jane and John where Jane elicits a response from John to navigate the street in amutually agreeable manner – and bidirectional – where Jane and John both influence each other througha nonverbal negotiation5 of how they want to share the narrow street. While the source of influence inHHI may not be explicitly clear in many of our daily interactions with others, this is made explicit inHRI through the way in which we design interactive robotic systems. For the purpose of this thesis, hes-itation is also an interactive behaviour that demonstrates ways in which two agents can negotiate for theresolution of a conflict within a shared, physical environment. Figure 1.1 provides a visual illustration ofthe context in which hesitation behaviours are studied in this thesis. The author asserts that investigatingthe feasibility of analogous behaviour to enable HR negotiations helps address the question: “In whatways should a physically embodied robot exert influence on our behaviours and decisions?”5 The term negotiation used within the context of this thesis is discussed more in detail in Chapter 2.41.2 Research QuestionsThis thesis explores both unidirectional and bidirectional types of interweaving within the contextof physical HR collaboration. Robot use of gaze cues and hesitation gestures are investigated as twocoordination mechanisms that can help facilitate the interweaving process. In the unidirectional modelof interweaving, the author investigates robot use of gaze cues to elicit desired responses from humans.In considering the bidirectional model, nonverbal negotiation between the two members of an HR pairusing hesitation gestures is studied as a mechanism that allows interactive coordination of subplans.The thesis is framed around the following research questions:1. Unidirectional Interweaving Can a robot provide humanlike nonverbal cues to influence peo-ple’s behavioural responses to an interaction while the interaction is taking place?2. Bidirectional Interweaving (Negotiation) Can a robot nonverbally negotiate with a person aboutwhat should happen in an interaction? Can an HR negotiation contribute to an improved HRcollaboration?To address the first question, the author considers the activity of robot-to-human handover of anobject as an interaction in which the when and where of the collaborative task is to be implicitly com-municated to the human recipient (Chapter 3). While previous studies suggest that robot use of nonver-bal cues can affect human behaviours in HR handovers, thereby providing a strong support for the firstquestion, exploration of the question in this thesis explicitly addresses the human behaviours duringthe interweaving process of the interaction – that is, coordinating when and where the person shouldmove his/her hand to receive the object the robot is handing over. The author examines effectivenessand efficacy of a robot’s use of human-inspired nonverbal cues to elicit a response by the collaboratingagent (human), in order to successfully complete the handover task. Like many other HRI studies thathave preceded this work, the robot in this interaction does not leave room for the person to initiate andinfluence the robot and its behaviours.6 The order of the task to be performed and division of roles inthe task are both made clear, and the communication cues used to interweave the when and where partof the collaboration are unidirectionally provided by the robot to the human recipient.To address the second question, the author explores an HR collaborative assembly context where,by chance, both human and robot reach for the same resource at the same time and are left to theirown devices to resolve the conflict at hand. Hesitation behaviours similar to the second scenario ofJane and John above are investigated as a means for an HR pair to negotiate with each other. In thisinteraction, the agents must effectively figure out the priority order to access a shared resource thatoften, spontaneously, comes into conflict for the collaborating HR pair. In order to have a nonverbaldialogue leading to a negotiated conflict resolution the two agents must bidirectionally communicateand respond to each other in real-time.6 This statement is not meant as a generalization of the work accomplished in the field of HRI, but only that there arenumerous work that share the open-loop interaction model employed in this study. The field has seen some progress inhuman-to-robot handover interactions, where the user takes the role of initiating the interaction (e.g., [35, 147]).51.3 Thesis OutlineThe remainder of this thesis is organized into five chapters. Chapter 2 provides a general and over-arching literature review on HRI with a focus on the influence that interacting agents exert onto eachother.Chapter 3 is dedicated to the first of the two research questions listed above. It describes two studiesthat have been conducted to understand some of the cues humans use in HH handovers and evaluatewhether human-inspired robot behaviours in robot-to-human handovers can elicit desired behaviouralresponses from human participants. In particular, the studies help address whether a robot’s unidirec-tional, nonverbal communication cue can convey the details of when and where the transfer of an objectfrom robot to human should take place.Chapter 4 and Chapter 5 focus on the second of the two research questions. Analogous to theprocess in Chapter 3, two studies are presented in Chapter 4. The first study serves to observe howhumans negotiate and resolve resource conflicts with one another using hesitation gestures. With abetter understanding of characteristic features of human hesitation gestures, a novel trajectory generator,the Negotiative Hesitation Generator (NHG), was then designed and implemented to control a robotarm. The second study outlined in the chapter evaluates the efficacy of the designed controller and itsparameter values. Subsequently, an in-person HRI study is presented in Chapter 5: participants wereinvited to collaborate with the robot, which exhibited human-inspired negotiative hesitation behavioursbased on the NHG. Results from this study demonstrate that human participants do yield to and negotiatewith robots in a collaborative activity.This thesis concludes with Chapter 6 highlighting the main contributions of the thesis, namely:advancing the field’s knowledge of nonverbal HRI, especially in robot-to-human handovers and hesita-tions; a novel hesitation controller that allows an HR pair to negotiate about and interactively and safelyresolve resource conflicts in an HR collaborative assembly scenario; and demonstrating the efficacy ofnonverbal negotiations as an efficient and fluent mode of interaction.6Chapter 2Background and Motivating LiteratureThis chapter presents relevant, motivating literature from fields of study that frame the researchquestions investigated in this thesis. It first situates the contributions of this thesis on HR communicationand collaboration (Section 2.1). Previous studies in HRI are discussed in terms of unidirectional andbidirectional modes of influence previously documented in the field (Section 2.2).As outlined in Chapter 1, the work presented in this thesis involves investigations of two differenttypes of nonverbal communication cues – gaze cues used in robot-to-human handovers (unidirectional)and hesitations for HR collaboration (bidirectional). Each of them independently contributes to a betterunderstanding of the particular type of communication cues used in HRI. Therefore, previous work thatframes the contributions specific to the type of cues investigated in the following chapters are presentedin the relevant chapters. This chapter, on the other hand, paints a broader picture of the literature leadingto the main contributions of this thesis.2.1 Human-Robot Collaboration and CommunicationRobots have been performing dull, dirty, demanding, and dangerous tasks for decades, particularlyin manufacturing facilities.1 While the industry practice of delegating tasks to robotic systems is notnew, most of the traditional robots have been limited to industrial applications and confined to operatewithin work cells that physically separate the robot’s workspace from humans. In comparison, a growingnumber of robots today are being designed and marketed to work both inside and outside of manufactur-ing environments and without physical separation from humans. An iconic example that demonstratesthe wide-ranging application of today’s robotics technology can be found at Yotel [166], a hotel on thedensely populated Manhattan Island, New York, NY. At the lobby of the hotel is an IRB 6640 (ABBGroup, Zurich, Switzerland), an industrial robot capable of handling 200 kilograms of weight, that hasbeen modified to perform the tasks of storing and retrieving luggage for guests [36] – tasks typicallyperformed by human porters in most hotels. This delegation of task frees human employees to focus onother tasks, and releases them a potentially injurious lifting task. Paralleled by the widening range of1 Robots have been deployed to perform tasks in place of humans, such as handling nuclear materials, since 1940s [135].For more historical coverage of use of robots see [135].7application domains is an increase in the type of tasks that require robots to share spaces and objectswith people in a safe and effective manner. For example, medical drug delivery robots, such as HOSPI(Panasonic Asia Pacific, Singapore), perform the traditionally human task of fetching and deliveringdrugs from one part of a hospital to another [2]. HOSPI is built to autonomously navigate through hos-pital corridors that are shared by human staff and patients. It also interacts with doctors and nurses tohandle the drugs that have been delivered.These new robotic systems2 – systems that operate outside the confines of work cells – are creatingnew technological relationships in our society. While robots can perform certain tasks that are difficultor undesirable for humans to perform in a tireless manner, there are numerous tasks that humans canaccomplish with ease that pose complex technical challenges for a robot. For instance, loading a dish-washer may be a trivial task for most able-bodied human adults. However, getting a robot to load thedish washer is a technical challenge that requires the implementation and integration of object recogni-tion, grasping, and manipulation capabilities on a robot. In fact, the dish-loading activity was used asone of the tasks for a robotics competition in 2010 [18]. Therefore, enabling robotic systems to success-fully interact and collaborate with users can help complement the human and robotic agent’s strengthsand weaknesses and create a synergy that maximizes the benefits robotics has to offer.A number of fields of study ranging from Psychology and Cognitive Science to Entomology doc-ument the efficacy of collaboration in human, animal, and insect societies [20, 37, 56–58, 154, 162].Likewise, findings from studies in HRI suggest that effective collaboration between a human and a robotcan reduce task completion time and improve accuracy, quality, enjoyability, and safety of a task incomparison to solo task performance [19, 128]. Previous work also suggests that introducing robots thatwork alongside humans may be preferred to those that replace humans [150].3 For example, Wong et al.[163] conducted a study in which a small humanoid robot, Nao (SoftBank Robotics, Tokyo, Japan),recited stories to teenagers either by itself or with a human storyteller. When the robot told the story byitself, it used gestures and gaze to perform the scenes of the story. When the robot was collaboratingwith the human storyteller, the robot narrated the story while the person performed the scenes. Althoughthe content of the story was the same in both conditions, the participants preferred the second conditioninvolving human-robot collaboration. The researchers attribute this result to the fact the human and therobot complement each other’s strengths and weaknesses in interacting with the participants.As discussed in Chapter 1, communication is an essential component of collaboration. However,the impact robot communication cues have on an HR collaboration can be complex and is the subject ofon-going research. For instance, recent studies report that performance improvements of an HR collabo-ration can be salient in complex collaborative tasks than in simple tasks in which the improvements may2 The term New Robotics has indeed been used to refer to this shift from traditional industrial robotic systems to thegrowing number of robotic systems meant to be deployed and operated outside the confines of work cells [137].3 To say that the public’s acceptance of robots that work alongside people is higher than those that work in place of peopleis a crude generalization. A survey-based study conducted by Takayama et al. [150] provides empirical evidence that peopleprefer robots that work with people than robots that replace people. The findings from the study were later contested by asubsequent study conducted by the same researchers [72]. The researchers note in discussing the contrasting results from thetwo studies that much more work is required to understand how this preference of robots’ roles in our society varies acrossdifferent demographics. The author’s previous work also demonstrates that public acceptance of robotics technology and itsrole in our society can vary according to the application domain in question [98, 102, 118–120].8not be present at all. Admoni et al. [3] investigated the impact a robot’s nonverbal gestures (pointingand gazing) can have on performing tasks of different difficulty levels. Employing a task that involvesmemorizing and following a set of instructions given by the robot, the authors varied difficulty levelsof the task by adding more steps to the instruction or by introducing interruptions. The results of thisstudy demonstrate that the improvement of the participant’s task performance with the robot’s use ofnonverbal gestures was much more pronounced when the task was difficult than when it was easy. Thepositive effect nonverbal HR communication seems to have on team performance of complex rather thansimple tasks is also echoed in two HRI studies conducted by Gleeson et al. [52]. They investigatedwhether the nonverbal communication of tapping and pushing to control a robot in a bolt insertion taskoutperforms the traditional button-based interface to achieve the same objective. The results of theirstudies demonstrate that for a simple insertion task the physical, tap-and-push interaction with the robotdid not improve the team’s task performance, nor did participants prefer the traditional button-basedinteraction. However, it did significantly outperform the button-based interface when the task was lessscripted and more complex to execute.Results from these studies suggest that a robot’s behaviours can be designed to enable humans toachieve more when a task is performed with robots. They also suggest that much work remains to bedone in understanding the nature of the impact communicative cues can have on an HR collaboration.2.2 Robots as a Mechanism of InfluenceAs briefly mentioned in Chapter 1, the physical embodiment of robots makes them a powerfulsource of influence for human behaviours compared to artificial intelligence software alone. This sectionprovides a short review of the role robot behaviours can play in unidirectionally and bidirectionallyinfluencing human behaviours and decisions.The distinction the author makes between unidirectional and bidirectional influence is based onhow the interactive system is designed. Unlike HHI, in which it is often hard to measure how much oneperson is affecting the other, the directionality of influence in HRI can be discussed much more explicitlyby considering the design of the robotic system. How much or in what ways the system incorporatesimplicit or explicit input signals from the user is a direct consequence of the designers’ decisions. Onthe other hand, the amount and type influence a user is subjected to during an HRI are not factors that areexplicitly controlled by the users. A HRI system can be designed to be reactive to human commands andbehaviours, such that human users unidirectionally influence the system. A system can also be designedto proactively elicit desired responses from humans without itself reciprocally being influenced by theperson.This thesis is focused on the influence robots can have on humans in the process of interweaving de-tails of a collaborative task. Therefore, in Section 2.2.1, only the literature in which the robot is designedto unidirectionally influence humans is discussed. In discussing bidirectional influence (Section 2.2.2),the author refers to systems that are designed to influence and be influenced by humans through implicitnonverbal cues.92.2.1 Unidirectional InfluenceThe field of HRI has uncovered a variety of ways that robot behaviours can unidirectionally affecthumans. For instance, Ju and Sirkin [70] demonstrated that when a physical gesture by a robot iscompared against an analogous gesture displayed on a screen, the effect the gesture has on humanbehaviour differs drastically. They conducted a study where a robotic kiosk placed at a bookstore anda building lobby either physically gestured using its hands or projected the gesture on its screen. Theauthors found that the robot engaged up to twice as many passersby when it physically gestured versuswhen it used a screen-based gesture. In fact, robots can significantly affect human behaviours just by itspresence. In a recent study, Hoffman et al. [62] found that humans cheat just as much in the presenceof a robot as they would with a person. The robot used in the study was built for the specific purposeof conveying social presence rather than demonstrating any monitoring behaviour. Yet, the participantscheated less when a person or the robot was present than when the participants were left alone in a room.This finding echoes the media equation put forth by Reeves and Nass [130], which suggests that humanstreat machines as though they are social beings. Moreover, studies have found that a robot’s motion cantrigger different affective as well as physiological responses to a human observer even when the startand end points of the robot’s motions remain the same [83].A robot’s ability to affect humans is a necessary condition for the robot to interweave details of atask with a person in an HR collaboration. Studies in HRI suggest that implementing communicationbehaviours on a robot, often inspired from human or animal communication behaviours, can positivelyimpact the HR team’s performance and the user’s perception of the robot [22, 24, 52, 134]. The positiveimpacts of robot nonverbal cues in HR collaboration are often attributed to the fact that implicit andexplicit nonverbal gestures help establish joint attention [65], share joint intention [45, 99, 155], andexpress the robot’s internal states [22, 45, 134] to the users in an intuitive manner. Romat et al. [134]conducted a study where a human participant and a robot were tasked to collaboratively build a towerusing Duplo blocks. Due to the kinematic constraints of the robot, the participant needed to help therobot reach the desired blocks placed outside the robot’s range of motion by understanding the robot’sneeds and moving the blocks closer to the robot. The authors found that when the robot providednonverbal cues to indicate its need to grab blocks outside of its reach the participants were significantlyquicker in providing the desired assistance than when the robot did not exhibit the cues. Likewise,Breazeal et al. [22] demonstrated that a robot’s use of implicit nonverbal gestures4 (e.g., shrug and gaze)used in conjunction with explicit nonverbal gestures (e.g., pointing gestures) significantly improves theuser’s understanding of the robot’s internal states and the HR team’s overall task performance.The type of influence that is especially relevant to this thesis is how nonverbal modes of commu-nication affect an HR dyad as the agents interweave spatiotemporal details of a collaborative task. Arobot’s use of gaze cues, for example, has been demonstrated to provide supplementary informationto human users in in-person HRI, thereby facilitating the interweaving process. In an HRI experiment4 Here, Breazeal et al. [22] describe implicit nonverbal gestures as nonverbal behaviours that convey information inherentin the agent’s behaviour rather being purposefully communicated. This is distinguished from explicit nonverbal gestures inwhich the agent intends to convey specific information to the recipient. This is analogous to the way implicit interaction isunderstood in Ju and Takayama [71] and other HRI literature.10involving a table-top manipulation task, Boucher et al. [19] explored whether a robot’s use of gaze cuescan decrease the reaction time of human collaborators whose role was to manipulate items that were ver-bally announced by the robot. The results of their study suggest that when the robot used head and eyegaze, subtly conveying which item was being attended to, the participants reacted significantly faster inreaching for the desired items than when the gaze cues were occluded from the participants. The authorsattributed this result to the fact that robot gaze cues can serve the function of establishing joint attentionand help human observers predict what the next desired item may be. This echoes the finding from astudy conducted by Mutlu et al. [113]. In this study, a humanoid robot, either a Geminoid or a Robovie[4], played a guessing game where the participant had to identify which of the items on a table the robotchose for them to guess. The authors found that the participants performed significantly better whenthe robots provided quick gaze cues as though giving a hint. Both of these studies demonstrate thatthe robot’s use of gaze cues is effective in nonverbally communicating the desired target object to theparticipant. Based on the results of these studies, gaze cues implemented on a robot can likely conveyspatiotemporal information to a human collaborator.In exploring the first research question (Can a robot provide humanlike nonverbal cues to influencepeople’s behavioural responses to an interaction while the interaction is taking place?) in Chapter 3,human-inspired gaze cues are implemented on a robot in a robot-to-human handover context. Thestudies presented in the chapter help test the hypothesis that nonverbal supplementary cues, such a gaze,that are exhibited during a robot-to-human handover interaction can influence when and where humanrecipients will reach out to accept objects handed over to them.Many of the robot nonverbal behaviours that positively impact task performance are designed basedon observed human and other animal behaviours [23, 61, 64, 71, 151]. This is not surprising given thatthe human ability to read someone else’s internal states and intentions and engage in collaborative tasksare developed in infancy and through mimicry of other humans.5 Anthropomorphic communicationcues in a robotic system, therefore, have a better chance of being understood by human users. In thisthesis, the author also looks to human behaviours to inspire the design of robot nonverbal behaviours foran improved HR collaboration. In particular, this thesis focuses on two different types of nonverbal cuesthat can be superimposed on the functional motions required to complete a task: in the studies presentedin Chapter 3, various gaze behaviours are used to supplement the functional motions of a robot handingover an object to a person; in Chapter 4 and Chapter 5, hesitation gestures are superimposed on a robot’sreaching motion as it moves toward a button to be pressed. Results from these studies also support thathuman-inspired nonverbal cues help improve the performance of the HR collaboration.In addition, there is evidence that the impact robots have on human behaviours may be attributedto how we neurologically respond to physical motions performed by robots. In neuroscience, it hasbeen established that motion properties of one person can activate the mirror neurons of an observer,and influence the properties of motions exhibited by the observing person. Recently, the field of HRIis gathering evidence that this powerful neurological phenomenon can be present in the context of HRI5 Numerous examples in developmental psychology provide evidence that such skills are established in infants as youngas 12 months old, who mainly use nonverbal, protosocial behaviours to communicate with adults and other children [31, 154].11– that is, motions of a robot can activate mirror neurons of the humans who observe them and, in turn,influence the properties of motions performed by the observers. In a study conducted by Bisio et al.[17], the human participants were asked to observe either a humanoid robot or a person demonstratea series of reach motions. The agents performed the reach motion either to a specified location or totransfer an object from one point to another.6 After the observation, the participants were asked toperform a reach, with which the experimenters measured the similarity of the participants’ trajectoryprofile (velocity) with that of the motions demonstrated to them. The results of this study suggest thatregardless of whether a motion was demonstrated by a person or a robot, the participants mimickedthe quality of the motion of what was demonstrated to them, except when the demonstrated motioncontained a non-biological velocity profile. This demonstrates that the phenomenon of motor contagion,where the motions of one agent are mirrored in the motions of another, is also present when the motionis demonstrated by a robot. While the motion contagion can be bidirectional in HHI, in HRI the directionof contagion is unidirectional (a robot one-sidedly affects the motions of the interacting person) if themotions of the robot are not designed to adapt to that of the person interacting with it.Therefore, in designing robots that will socially and collaboratively interact with us, it is crucialthat we continue to explore the nature of human responses to robot behaviours and how they are similaror distinguished from the way people respond to other technological devices. This new understandingof human responses with respect to robots will not only allow us to better contemplate the societalimplications of the technologies, including new social norms that can be formed with the presence of arobot, but also allow HRI practitioners to make informed design choices.2.2.2 Bidirectional InfluenceOne of the studies that motivated this thesis is an online survey conducted by the author and hercolleagues [30, 106]. In this survey, the authors were interested in finding out what a mail-deliveryrobot should do when it encounters conflicts with a person in using an elevator. The participants ofthe survey were given variations of a scenario where a large humanoid robot is carrying mail to bedelivered to an important person on another floor. While the robot necessarily needs to use the elevatorto travel between floors, it cannot share the elevator with a person due to space and safety reasons. Theparticipants were asked what the robot should do when it is delivering mail and a person is alreadyusing or waiting for the elevator. In addition to the location of the person with respect to the elevator,a combination of two other factors (urgency of the mail delivery task, and the status of the person – ina wheelchair, carrying heavy items, or neither) was used to generate multiple versions of the scenario.The participants were given the following choices: do nothing, yield to the person, ask the person toyield, or engage in a dialogue with the person. The sample size of the study was too small to make ageneral claim. However, when presented with a highly contentious situation (e.g., an urgent mail is tobe delivered and a person in a wheelchair is already using the elevator) the participants showed a clearpreference that the robot should engage in a dialogue with the person to figure out whether to yield.Answers to dilemmas vary from one person and context to another. However, as presented in the6 The demonstration took place in the same room rather than using a video recording.12example of Jane and John in Chapter 1, people engage in verbal or nonverbal dialogues with the othersinvolved in an uncertain situation to negotiate a way forward. The results from the abovementionedstudy suggest that an effective way to handle conflicts and contentious situations arising in an HRI maybe to enable the collaborating robot and human to negotiate a solution.In order to engage a robot in a nonverbal HR negotiation, the robotic system must not only beable to express its internal states but also have a means to react to human responses in real-time. Mutualresponsiveness is, in fact, one of the key elements of collaboration identified by Bratman [20]. Nikolaidiset al. [115] tested the impact of the bilateral relationship in an HR collaboration context. They developedand tested a mutual adaptation method that allows the HR pair to adapt to each other without an explicitchange of turns. In their experiment, the participants were asked to collaboratively lift and transfer alarge table with a robot while navigating through a narrow doorway. The authors tested three differentconditions to vary the directionality of influence between the human and the robot: fixed, in which therobot used a fixed strategy in maneuvering the table; mutual-adaptation, in which the robot guided theparticipant toward a strategy when the participant is adaptive, and follows the guidance of the participantwhen s/he is not adaptive; and a cross-training condition, in which the robot uses one strategy in onephase of the task, and follows the user’s strategy in another. Results from this study provide empiricalevidence that mutually adaptive systems – in which the system asserts one strategy while allowing theuser to negotiate for another – yields significantly improved team performance without sacrificing theuser’s subjective perception of the robot.This study provides positive evidence that a robot that influences and is influenced by its userscan improve the outcome of the collaborative task. This thesis posits that a robot that nonverballynegotiates with the user in the process of arriving at a joint goal can also improve the outcome of the HRcollaboration.The research presented in Chapter 5 extends this much under-explored line of research by testingartificially generated hesitation trajectories (developed in Chapter 4) as a mechanism for a robot tononverbally negotiate for a solution to a resource conflict with a human user.2.3 SummaryThis chapter presented a broad overview of the motivating literature framing this thesis. Acknowl-edging the rising trend for and potential usefulness of collaborative robotic systems that share spacesand objects with humans, the challenge remains for the field of HRI to develop communication mecha-nisms that can help realize that potential. Previous studies demonstrate that nonverbal cues implementedon a robot can help communicate internal states of a robot (e.g., the robot’s need for assistance froma person) and positively affect the task outcome of an HR collaboration. While the use of subtle non-verbal cues has previously been employed to convey details of a task, such as which item is relevantfor the activity at hand, the impact such cues can have on conveying spatiotemporal details of an in-teraction – thereby unidirectionally eliciting the person to move to the desired location at the desiredtime – remains to be investigated. The author hypothesizes that a robot’s use of humanlike nonverbalcues can help interweave spatiotemporal details of a task, and tests this in a robot-to-human handover13scenario in Chapter 3. There are only a limited number of examples in the HRI literature that demon-strate the usefulness of bidirectional influence in an HR collaboration. However, results from the fewstudies suggest that bidirectional influence in an HR team can improve the HR collaboration in terms oftask outcome. These findings inspire the hypothesis that HR negotiation, which necessitates the agentsto influence each other using nonverbal gestures, is a possible and desirable mode of interaction in HRcollaboration. Chapter 4 and Chapter 5 of this thesis explore the usefulness of hesitation gestures as atype of communication cue that enables both agents in an HR dyad to influence each other in a resourceconflict scenario for an improved HR collaboration.14Chapter 3Unidirectional Interweaving of Timingand Space in Robot to Human Handovers3.1 IntroductionThis chapter1 addresses the interaction of a robot handing over an object to a person (robot-to-humanhandover). This interaction is chosen as a means to explore first of the two research questions explored inthis thesis: “Can a robot provide humanlike nonverbal cues to influence people’s behavioural responsesto an interaction while the interaction is taking place?” More specifically, the studies presented in thischapter focus on the extent to which a robot’s use of nonverbal cues helps elicit desirable responses froma human user as a means to unidirectionally interweave the spatiotemporal details of the interaction.Object-handover is an important interaction to study in HRI. Enabling successful and fluent hand-overs between an HR pair is crucial in order for robots to take on more assistive roles for humans athomes and workplaces. Many application scenarios, including manufacturing and home environments,can involve situations where it is useful for a robot to fetch and hand over an object to a person. Im-plementing an effective HR handover interaction, however, is a challenge. In HH handovers, a greatvariety of subtle signals mediates the handover event. Body position, hand and arm pose, gaze, and gripforce are used to communicate not only the intent to engage in a handover, but also when and where thehandover is to occur. These subtle signals help create a fluent and fast interaction while ensuring thatthe object is not dropped (e.g., [15, 34, 86, 108, 145, 146]). When a robot does not provide appropri-ate cues, HR handovers can fail in a variety of ways: people do not recognize that the robot is givingthem an object [27], objects can be dropped [34], or people can feel uncomfortable or unsafe during the1 c©ACM/Springer. The majority of this chapter has been modified/reproduced, with permission, from the publishedarticle Moon et al. [105]. Supplementary research on the same topic has been conducted by the author in collaboration withMs Minhua Zheng, who took lead authorship. Zheng et al. [167] and Zheng et al. [168] have been published based on thesupplementary research. The majority of this chapter’s content is on the first publication, Moon et al. [105], for which AJungMoon led the project and authorship.Findings from Zheng et al. [167], a qualitative analysis of the study conducted in Moon et al. [105], is presented as part ofthe results in this chapter. A follow-up study conducted afterwards, Zheng et al. [168], is briefly mentioned in this chapter aspart of the discussion presented in Section 3.5. A more detailed summary of Zheng et al. [168] is presented in Appendix A.1.15interaction [42, 82].Robot-to-human handover is an interaction that requires both agents to coordinate their motionthrough space and time to accomplish their shared objective. While the desired outcome of a handover(an object offered by a giver is successfully handed over to a receiver without dropping the object) isoften well understood by a giver-receiver pair at or before the onset of the interaction, precise spatiotem-poral details about when and where the person should reach out to grasp the offered object are usuallynot explicitly communicated by the interacting agents. In everyday HHI, a giver-receiver pair seems tonaturally reach an agreement about these details as the interaction takes place and often without a needfor verbal dialogue. In HRI, such natural process needs to be understood and appropriate behavioursimplemented onto a robot for the interacting agents to interweave the unspoken details with each other.Studies presented in this chapter address whether gaze can be used to augment an HR handover event,subtly communicating handover location, handover timing, and providing acceptable social interactionsignals to modulate the human recipient’s behavioural decision on when and to where s/he should reachfor the offered object. Gaze cues, in either HHI or HRI, have proven to be efficient for communicatingattention [81, 110]. During a handover, givers use verbal or nonverbal cues to direct the receiver’sattention to an object. Successful handovers typically take place when the two parties achieve sharedattention on the same object. Previous studies [86, 145, 146] indicate that gaze can be used by robotsto signal handover intent to users before the handover event. However, these studies did not explore theeffect of robot gaze during the handover on the timing of the handover event.It is hypothesized that the use of human-inspired gaze cues during HR handover can influence aperson’s behavioural decision on handover timing and the subjective experience of the handover byimplicitly increasing communication transparency and perception of naturalness for the interaction. Thischapter presents two handover studies conducted to investigate the effect of gaze cues on handovertiming during an HR handover event.The first study, Study 1 (Section 3.3), was conducted to observe the type of gaze patterns humans usewhen one hands over an object to another. Results from Study 1 informed and inspired the gaze patternsimplemented for HR studies 2 and 6. In Study 2, a PR2 humanoid robot used two different human-inspired gaze patterns observed from Study 1 to address whether the use of gaze cues affects people’sfirst-time behavioural and self-reported responses to HR handover. Results from Study 2 suggests thatthe subtle and supplementary gesture of gaze does significantly affect an untrained human recipient’sdecision on when to reach for the offered object, thereby eliciting a faster, more fluent HR handoverinteraction. In addition to understanding participants’ first-time handover response, it is important tounderstand whether robot gaze cues can have lasting effect on robot-to-human handovers even after aseries of handovers. Following Study 2, the HRI experiment in Study 6 presented in Appendix A.1reaffirmed the effect of gaze in non first-time responses to handover events and investigated whether theeffect of gaze persists in repeated HR handovers. Positive results from these studies serve as a motivationfor the studies in the following chapter, Chapter 4, where the outcome of the interaction is uncertain andneeds to be resolved by the agents. All of the studies mentioned in this chapter have been approved bythe University of British Columbia Behavioural Research Ethics Board.163.2 BackgroundThis section outlines a body of work specific to HR handovers and the use of gaze in HRI. Inparticular, to home in on the discussions relevant to the three studies presented, the focus is on studieswhere a robot is the giver and a human the receiver.3.2.1 Human-Robot HandoverPrevious research in HR handovers can be broadly categorized by the aspect of the handover underinvestigation: approach for handover, handover trajectory and pose, and the handover event itself.Studies of approach for handovers consider situations where a mobile robot navigates towards ahuman to initiate a handover. These studies generally focus on human preference for approach directionsand on creating robot behaviours that clearly communicate the intent to initiate a handover. Basili etal. studied how humans hold objects as they approach for a handover [15]. Koay and Sisbot [82]studied human preferences for coordinated arm-base movement in handover approach. Mainprice et al.[91] designed an approach planner that considers the mobility of the receiver. While the studies in thischapter do not involve a robot’s handover approach (i.e., the participants approached our robot in ourexperiment), findings from the above studies guided the experimenters’ decision on placement of therobot, as discussed in Section 3.4.1.Other researchers have investigated handover trajectory and pose, reporting guidelines for how arobot arm should be positioned for handover and how that position should be achieved. In a seriesof studies, Cakmak et al. [28] and Strabala et al. [145] studied how handover trajectories and finalhandover poses can best signal the intent to initiate a handover. They found that the final handover poseshould feature a nearly fully extended arm in a natural (human achievable) pose with the elbow, wrist,and distal point on the object positioned, respectively, from closest to furthest away from the body in allthree dimensions. The object should be held in its canonical orientation (right side up) and positionedto allow easy grasping by the human. A related study emphasized the importance of the physical cuesin HR handovers, showing that poorly designed handover poses and trajectories were often unsuccessfulin communicating intent and ultimately resulted in handover failure [27]. They found that intent is bestcommunicated by having high contrast between the pose used for holding the object and the pose usedfor handing over the object. In the studies presented in this chapter, the above guidelines were followedin the design of handover pose and trajectory, as described in Section 3.4.1.Other studies have investigated the velocity profile of handover motions and have found that trajec-tories that minimize end-effector jerk make people feel safer in handover interactions [42, 66]. Otherstudies of handover trajectory include a human-based potential field planner for handover trajectories[74].Chan et al. [34] studied the actual handover event, measuring grip and load forces in HH handoversand using these data to design a robust robot handover grip controller that imitates human handoverbehaviour. This controller has been adapted for use in Studies 2 and Gaze in Human-Robot InteractionGaze is an important and useful cue in HHI. People repeatedly look each other in the eye during so-cial interaction and people do not feel that they are fully engaged in communication without eye contact[41]. Studies in psychology have shown various functions of gaze in social interaction, such as seekingand providing information, regulating interaction, expressing intimacy, exercising social control, etc.[41, 81, 124]. Gaze can be named differently in different social situations [110]; for example, mutualgaze or eye contact is defined as two people looking into each other’s face or eye region [161], while de-ictic gaze or shared visual attention is defined as one person following the other’s direction of attentionto look at a fixed point in space [26].Previous work has shown the importance of gaze in HRI. For example, Staudte and Crocker [144]demonstrated that humans react to robot gaze in a manner typical of HHI. Since gaze behaviour is closelylinked with speech [8], much work has been done on the conversational functions of gaze in HRI [84,89, 94, 111, 112, 158]. Gaze is particularly effective in regulating turn-taking during HR conversation.Kuno et al. [84] developed gaze cues for a museum guide robot to coordinate conversational turn-taking.Matsusaka et al. [94] used gaze cues to mediate turn-taking between participants in a group conversation.Another large body of literature focus on using gaze to direct people’s attention in HRI [16, 59,67, 131, 140]. Gaze was combined with pointing gestures in [16, 59, 67] to direct people’s attention,which the authors believed would make the interaction more human-like [16] while minimizing misun-derstanding [59]. In Rich et al. [131] four types of “connection events” were identified from HHI videos,namely directed gaze, mutual facial gaze, adjacency pairs and backchannels. Implementing them in anHRI game showed a high success rate in forming HR connection or joint attention. In Sidner et al. [140],people directed their attention to the robot more often in interactions where gaze was present, and peoplefound interactions more appropriate when gaze was present.Introducing gaze cues can also benefit HRI in other ways. In Mutlu et al. [111] and Skantze et al.[141], gaze increased human performance in certain HR tasks. In Kuno et al. [84] and Sidner et al.[140], gaze heightened HR engagement and in Liu et al. [89], gaze cues contributed to the perceivednaturalness of a communicating robot.In the study of HR handovers, other researchers have shown that gaze can be useful in communi-cating the intent to initiate a handover. Lee et al. [86] studied human motion and gaze cues as peopleapproached each other for handovers. They found that people looked at the object or the receiver as theyapproached the receiver. Strabala et al. [146] examined the signals that humans use to communicatehandover intent before a handover takes place. They initially acknowledged gaze as one of the impor-tant features that mark the difference between different phases in handover, but they did not find gazeto be an effective predictor of handover intent. In contrast, Kirchner et al. [80] demonstrated how robotgaze can be effective in targeting an individual recipient out of a group of people for a robot initiatedhandover. Atienza and Zelinsky [9] augmented handover interactions with gaze cues, demonstrating asystem that allowed a human to request an object for handover by looking at it.While the above studies address gaze in pre-handover cuing and communication of intent to hand-over, the studies outlined in this chapter examine the use of gaze during the handover event. Although18the effectiveness of gaze in regulating handover intent remains an open question, gaze may have a posi-tive effect when used during the handover event. Gaze can help establish shared attention to a locationin space and could elicit human receivers to move their hands to the desired location. Hence, gaze maybe useful in improving the handover itself by establishing shared attention and influencing timing atwhich the human recipient reaches out to receive the object being handed over.3.3 Study 1: Observing Gaze Patterns in Human-to-Human HandoversAn HH study, Study 1, was conducted to help understand what kind of gaze people use in HHhandovers. In this study, a pair of subjects handed over a water bottle to each other multiple times. Theexperimenters observed the gaze behaviour of the giver during the handover by collecting and analyzingvideo recordings of HH handovers (see Figure 3.1). While other researchers have observed gaze in HHhandovers (e.g., Strabala et al. [146]) before the handover event, the study presented in this sectionaugmented these previous results by focusing on gaze during the handover event.3.3.1 Experimental ProcedureTwelve volunteers (10 male, 2 female) participated in this study. The giver was asked to hand overten bottles from a side table to the receiver one at a time. The receiver was asked to bring the bottlesto a collection box about two meters behind them one at a time, requiring him/her to walk away fromthe giver between handovers. This process repeated until all ten handovers were completed. Eachparticipant performed the role of the giver, then was paired with another participant and performed therole of the receiver, resulting in twelve giver-receiver pairs (120 handover recordings in total).In order to collect human gaze patterns that can inform the design of nonverbal HR handovers, thegiver and receiver were instructed not to talk during this process. The giver was also instructed to pick upthe bottles from the side table only after the receiver returned from the collection box and had put his/herhands on the table. By requiring the receiver to turn and walk away, the common attention between thegiver and the receiver was interrupted after each handover and participants needed to re-connect for thenext handover.3.3.2 ResultsAnnotation of a frame-by-frame video analysis of the givers’ gaze patterns indicates that the giver’sgaze during a handover can shift between three positions: the object being transferred, the expectedhandover position, or the receiver’s face. Figure 3.2 shows a typical timeline of the five gaze patterns(gaze direction and timing) and corresponding frequencies observed from this study.The following describes the five gaze patterns:Shared Attention Gaze (Attn) The most frequent gaze pattern (68% of all handovers observed) con-sists of the giver gazing at a projected handover location as s/he reaches out to execute the hand-over. After picking up the bottle, the giver turns to face the receiver, looks down at a midpointbetween the giver and the receiver, and keeps the gaze there until the receiver takes control of the19Figure 3.1: Demonstration of two frequently observed gaze behaviours from the HH handoverstudy, Study 1. a) shared attention gaze: the giver looks at the location where the hand-over will occur, and b) face gaze: the giver looks up at the receiver’s face. The five gazepatterns observed in Study 1 consist of different combinations of these two gaze behaviours.( c©2014 IEEE/ACM)Figure 3.2: Giver’s gaze patterns observed from HH handovers. Attn: continual shared attentiongaze; Face: continual face gaze; Turn-Taking: long shared attention gaze followed by ashort face gaze; ShortFace-Attn: short face gaze followed by a long shared attention gaze;LongFace-Attn: long face gaze followed by a short shared attention gaze. ( c©2015 Springer)20bottle. This midpoint is approximately where the handover takes place. There is no eye contactbetween the giver and the receiver throughout this handover gaze pattern (see Figure 3.1a). Thegaze towards handover location is labeled shared attention gaze, to indicate the use of gaze todraw the subject’s attention to the handover location.Face Gaze (Face) In some other (10%) handovers, the giver gazes at the receiver’s face, perhaps toestablish an eye contact, throughout the handover. This gaze behaviour towards the receiver’sface is labeled face gaze.Turn-Taking Gaze (Turn) Some (9%) of the handovers observed showed a slight variation of theshared attention gaze. In addition to gazing at a projected handover location while reachingout, the giver also looked up to make eye contact with the receiver near the end of the handovermotion (face gaze), at approximately the time that the receiver made contact with the bottle (seeFigure 3.1b).ShortFace-Attn In 8% of the handovers, the giver looked at the receiver’s face and quickly glanced atthe bottle when the receiver is about to touch the bottle.LongFace-Attn In 5% of the cases, the giver glanced at the receiver before but not during handover,and shifted the gaze to the handover location when the receiver is about to touch the bottle.3.3.3 DiscussionThe results of this study suggest that humans use a variety of gaze patterns while handing over anobject to another person. In general, the giver tends to shift his/her gaze from the object being handedover to the receiver’s face (face gaze), the projected location at which the handover should take place(shared attention gaze), or a combination of the two. The shared attention gaze in the Attn, Turn-Taking,ShortFace-Attn and LongFace-Attn patterns can be interpreted as serving the function of communicatingwhere the physical transfer of the object should happen. The long face gaze in the Face and LongFace-Attn patterns serves a similar function as the face gaze in verbal conversation of regulating a turn [80];in verbal conversations, the speaker typically ends his/her utterance with a sustained gaze at the listener,signaling willingness to hand over the speaker role, while in this case, to hand over the object. The shortface gaze in the Turn-Taking and ShortFace-Attn patterns serves a monitoring function [80], appearingto observe whether the receiver is paying attention to, or is ready for, the transfer of the object.These results inspire the hypothesis that an implementation of analogous gaze cues for an HR hand-over could serve similar functions and help the HR dyad interweave the spatiotemporal details of theinteraction during the handover. Hence, the gaze patterns observed in this study informed the design ofexperimental conditions in Study 2. Section 3.5 provides more discussions of the results from the twostudies contrasting HHI and HRI.213.4 Study 2: Impact of Human-Inspired Gaze Cues on First-TimeRobot-to-Human HandoversTo examine the impact robot gaze has on first-time human receiver behaviour, the experiments inStudy 2 and 6 employed a PR2 humanoid mobile robotic platform (Willow Garage Inc., Menlo Park,CA) with a pan-tilt head and two 7-DOF arms, each with a two-fingered, 1-DOF gripper. The followingsection (Section 3.4.1) outlines the physical handover cues the PR2 used; Section 3.4.2 describes thethree experimental gaze conditions inspired from Study 1 and selected for HR handovers in Study 2; andSection 3.4.3 and 3.4.4 outline the experiment design and technical implementation.3.4.1 Physical Handover CuesBased on findings from Basili et al. [15] and Koay and Sisbot [82], the robot in Study 2 was posi-tioned such that it was facing the participant approximately 1 meter away.The robot executed the handover with its right gripper, as recommended in Koay and Sisbot [82].At the beginning of each handover, the robot starts its motion at the grasp position with its end-effectorprepared to grasp a bottle sitting on a table at the robot’s right side. When the subject is ready, theend-effector grabs the bottle (marking a start time, t = 0 of the interaction), then moves the bottlehorizontally to a position in front of the robot’s centreline (ready position). Then the robot moves fromthe ready position forward to the handover location. Joint-angle goals of the grasp position, readyposition, and handover location are predefined such that when the robot’s end-effector is extended, thearm is positioned in a natural pose: the elbow located below the shoulder, and the gripper located belowthe distal point on the bottle, as shown in Figure 3.3. The handover location is designed in accordwith the recommendations of previous work [80, 86]. The three locations are constant for all three gazeconditions. While other researchers have proposed handover controllers that adapt to the position of thehuman’s hand, for example Erden et al. [47], a constant handover location and gaze cues that vary onlyduring handover events is used in this study.When the robot’s arm reaches the handover location, the robot waits for a participant to grasp andpull up on the object. The force the gripper exerts on the bottle is a linear function of the downwardgravitational force exerted by the bottle as described by Chan et al. [34]. Thus, as the receiver takesthe weight of the bottle, the robot releases its grip (marked as the release time). The PR2’s fingertippressure sensor arrays were used to realize Chan et al.’s handover controller. Finally, after releasing theobject, the robot returns to the grasp position, ready to grasp and lift the next object.3.4.2 Experimental Gaze CuesIn this study, the PR2 robot expressed gaze through head orientation. Imai et al. [68] showed thatrobot head orientation, called head gaze, can be an effective substitute for human-like gaze and that headorientation is interpreted as gaze direction. A single object was used for the handovers to minimize anypossible confusion regarding the robot’s gaze direction.This study involved testing of three different gaze patterns in HR handovers, as shown in Figure 3.4.22Figure 3.3: Demonstration of the experimental set-up and the three conditions at the handover lo-cation: a) No Gaze; b) Shared Attention; and c) Turn-Taking. An array of infrared sensorswas located at the edge of the table. The red dotted lines represent the location where sub-ject’s reach motion is detected. Subjects stood at a specified location marked on the floor.( c©2014 IEEE/ACM)In all conditions, the robot’s gaze tracks its end-effector from the grasp position to the ready positionas though the robot is attending to the acquisition of the bottle. When the end-effector arrives at theready position, the robot’s head is tilted downwards towards the end-effector. Only when the robot armtransitions between the ready position to the handover location does the robot transfer its gaze accordingto the following gaze patterns:The No Gaze (None) condition is our baseline. The robot head remains looking down towards theground while the end-effector extends forward for the handover.The Shared Attention (Attn) gaze condition models the most frequently observed gaze pattern fromStudy 1. When the robot starts to move from the ready position to the handover location, itsmoothly transitions its gaze from the bottle to the location in space where the handover willoccur, as an implicit cue intended to direct the human’s attention towards the projected handoverlocation. This condition is necessary to test the hypothesis that shared attention can be establishedthrough gaze during handovers, and that doing so benefits the handover interaction. Establishingshared gaze at an object or location can serve to direct shared attention (e.g., Imai et al. [67]) andcan aid in the successful execution of HR cooperative tasks (e.g., Skantze et al. [141]).The Turn-Taking (Turn) gaze condition is also derived from the HH handovers in Study 1. When thehandover trajectory begins, the robot smoothly transfers its gaze to the handover location, as in theShared Attention condition. Afterwards, as though it is giving a turn to the participant, it shifts itsgaze up to the human’s face in a quick motion, reaching the final gaze position at approximatelythe same time that the handover motion completes. Here two hypotheses are tested: that (a)this gaze shift can cue handover timing, and (b) looking at the face can improve the subjectiveexperience of the handover. This type of gaze shift has been shown to be a meaningful HR turn-taking cue [29] and mutual gaze can increase the sense of engagement and naturalness in HRI[89, 140].23Figure 3.4: Depiction of gaze cues by the head of the PR2 robot: No Gaze (None), Shared Atten-tion Gaze (Attn), and Turn-Taking Gaze (Turn). In the Turn condition, the robot shifts itsgaze from the handover location to the human’s face midway through the handover motion.( c©2014 IEEE/ACM)The following hypotheses are tested using these conditions:Hypothesis 3.1 Robot use of gaze during robot-to-human handovers can establish shared attention andimprove handover timing.Hypothesis 3.2 Robot shift of gaze from a projected handover location to the recipient’s face will cuehandover timing.Hypothesis 3.3 Robot gaze towards human recipient’s face during a handover can improve the subjec-tive experience of the handover.3.4.3 Experimental ProcedureThis study involved a paired-comparison HR handover experiment in a controlled room. The studytook place on the day of a university orientation event such that many and diverse participants couldbe rapidly recruited during the public event. The experiment was structured as a balanced incompleteblock design (v = 3, b = 96, r = 64, k = 2, λ = 32)2 to both support rapid trials (maximum 5 minutes)2 These variables indicate the structure of a balanced incomplete block design and are necessary for statistical analysis:v =number of treatments (i.e., three conditions – No Gaze, Shared Attention, and Turn-Taking conditions – were tested);24and include only first-time reactions: each participant evaluated one of the three condition pairings.Condition order was randomized and presentation order counterbalanced among trials.Participants provided informed consent then entered the room where verbal instructions were given(Figure 3.3). They were told to stand at a marked area facing the robot and informed they wouldparticipate in a handover interaction. Participants were also told that the robot would pick up the waterbottle placed beside it and hand it to them. They were asked to take the bottle from the robot wheneverthey felt it was the right time to do so. To avoid unintended cueing, during handovers the experimenterssat out of the field of view of participants.After receiving the first bottle, participants placed the bottle in a box approximately 3 meters behindhim/her. This served as a washout between handovers, breaking the participant’s focus on the robotand the handover, as was done previously by Cakmak et al. [27]. Participants then returned to the samemarked area in front of the robot and participated in a second handover. Participants were permitted tokeep the last bottle given to them by the robot.During each handover, the following events were timestamped: start of robot motion (start time),end of robot motion (end of motion time), start of release of the robot’s gripper (release time), and theparticipant’s first reach for the object (reach time) as measured by the motion sensor array described inSection 3.4.4.After the second handover, participants left the room and completed a short questionnaire comparingthe two handovers on three subjective metrics: overall preference, naturalness, and timing communi-cation. For each of the following three questions, participants were asked to select either the first orsecond handover:1. Which handover did you like better, overall?2. Which handover seemed more natural?3. We think that timing is important in human-robot handovers. Which handover made it easier totell when, exactly, the robot wanted you to take the object?Participants could also provide additional comments.3.4.4 Technical ImplementationThe control of the PR2 consisted of the Robot Operating System (ROS) [1] with a series of softwaremodules coordinated via the Blackboard architectural design pattern [60] (Figure 3.5). One modulecontrolled the robot’s arm and another, its head. The head-control module provided object trackingfunctionality for bringing the water bottle to the ready position, and a smooth, fast gaze transition(average 90 degrees/second) functionality to enable the Shared Attention gaze and Turn-Taking gazeconditions during the handover motion.b =number of blocks (i.e., observations from a total of 96 participants was analyzed); r =number of replicates (i.e., eachcondition was tested on a total of 64 participants); k =block size (i.e., each participant saw two conditions); and λ = r(k−1)/(v−1).25Figure 3.5: Experiment system flow diagram. ( c©2014 IEEE/ACM)An independent module logged quantitative measurements of robot’s start time, end of motion time,and release time.An array of three passive infrared motion sensors (SEN-08630, SparkFun Electronics, Boulder, CO)configured as a light curtain was placed at the edge of the table (Figure 3.3), and was used to detect thestart of the participant’s reach (reach time) triggered by the participant’s hand crossing the table edge.An Arduino microprocessor relayed the sensor reading to the PC controlling the robot. Sensor readingswere logged and time-synchronized with the robot.3.4.5 ResultsA total of 102 volunteers participated in our experiment. Six records were rejected due to the sub-jects’ failure to follow instructions. Therefore, the following analyses include data from 96 participants(63 male, 33 female; age M = 23,SD = 5.59). Due to a technical error, reach time was not logged inthe second handover for five of the participants. This did not affect the analysis of handover timing,since the focus of this study on first-time responses requires reach time measures from only the firsthandovers. No other technical failures occurred, and all handovers were successful (no bottles weredropped).26Figure 3.6: HR handover timing results. All times are measured with respect to the robot’s startof motion at t = 0. The dashed line at 2 seconds indicates the end of robot motion at thehandover location. Reach time indicates the participant’s reach toward the proffered objectcrossing the infrared sensors. Note that in the case of the Attn condition, participants startto reach before the robot has reached handover location. The mean reach time for the Attncondition is significantly earlier than that of the None condition. The error bars indicate 95%confidence intervals. ( c©2014 IEEE/ACM)Handover TimingFigure 3.6 shows the distribution of three key times: the robot’s end of motion time, participant’sreach time, and robot’s gripper release time. All times are measured relative to start time.The following results are from a one-way Analysis of Variance (ANOVA) conducted on participants’reach time across the three conditions. Only the reach time collected during the first of the two handoversperformed by each participant is used in this analysis. This is due to a significant learning effect observedbetween the first and second handover trials (t(90) = 4.21, p < .001, d = 0.43), where reach time isearlier in the second handovers. Focusing on only the first handovers also allowed the analysis of trulyfirst-time handover behaviours, which is the focus of this study. The entire robot motion from the graspposition to the handover location consistently took 2.02 seconds (SD = 0.01).Participants’ reach time varied across the three gaze conditions (F(2,93) = 6.49, p < .005) as plot-ted in Figure 3.6; post-hoc analyses used a Bonferroni correction. Participants reached for the objectsignificantly earlier with Shared Attention (M = 1.91, SD= 0.52) than with No Gaze (M = 2.54, SD=0.76) (p < .005). Note that the mean reach time for Shared Attention occurs before the robot hasstopped moving at the handover location (reach time < end of motion time). These results supportHypothesis 3.1. No significant differences were found between Shared Attention and Turn-Taking(M = 2.26, SD = 0.79), or between Turn-Taking and No Gaze.Subjective ExperienceContrasting overall preference, perceived naturalness, and timing communication across the threegaze patterns during handovers involved Durbin’s test [46] – analogous to a Friedman test for rank data,but adapted to balanced incomplete block designs – on the aforementioned questionnaire data.Mann-Whitney U tests revealed no significant gender effects (overall preference: U = 935.0, p =.23, r = .12; naturalness: U = 918.5, p = .18, r = .14; timing communication: U = 935.5, p = .22,r = .12). One-sample Wilcoxon signed rank tests allowed observation of potential bias in selecting27Table 3.1: Ranking of questionnaire results. Each cell represents the number of people who chosethe row condition over the column condition. * indicate pairwise comparisons that are signif-icant to p < .10 (none were significant to p < .05). Note that participants’ bias to select thesecond handover experience was observed regardless of experiment condition.Overall PreferenceTurn Attn NoneTurn-Taking 0 21* 19*Shared Attention 11 0 17No Gaze 13 15 0NaturalnessTurn Attn NoneTurn-Taking 0 20* 19Shared Attention 12 0 19No Gaze 13 13 0Timing CommunicationTurn Attn NoneTurn-Taking 0 21* 18Shared Attention 11 0 19No Gaze 14 13 0the first or second handover experience in the questionnaire. Results show a significant bias towardsselecting the second handover on the timing communication metric (Z = 2.22, p< .05) and a weak trendto select the second handover on both overall preference and naturalness metrics (Z = 1.62, p= .11 andZ = 1.41, p= .16, respectively). The rank data collected using the questionnaire is insufficient to correctfor this bias statistically.Given this general bias to select the second handover, finding statistical significance to α = .10 inquestionnaire results is also noteworthy. Hence, observation of trends (results having p < .10) is alsoreported. See Table 3.1 for a summary of the results.Overall Preference: The results did not show a significant difference in user preference across the threegaze conditions (T 2 = 2.04, p = .14). However, one-tailed pairwise comparisons demonstrate a trendfor preference toward Turn-Taking over No Gaze (p < .10) and Shared Attention (p < .10) conditions.Naturalness: While the results show no significant difference in perceived naturalness of the handoversacross the three gaze conditions (T 2 = 1.82, p = .17), participants tended to choose Turn-Taking asmore natural than Shared Attention (p < .10) but not over the No Gaze condition.Timing Communication: No significant differences were found in the perceived communication oftiming across the gaze conditions (T 2 = 1.65, p = .20). However, participants tended to choose Turn-Taking over Shared Attention (p < .10), but not over No Gaze, as easiest to communicate handover28timing.In total, 59% of all participants provided additional comments (optional) on the questionnaire.Twelve subjects who experienced the Turn-Taking condition explicitly used words such as “head mo-tion”, “eye contact” or “looking at me” and expressed the condition in a positive light (e.g., P90 com-pared No Gaze with Turn-Taking: “During second handover [Turn-Taking], robot made eye contact,which made it easier to tell when the bottle should be taken.”; P10 compared Shared Attention andTurn-Taking: “I liked it when robot looked at me. That confirms it’s good to take.”). However, anothertwelve subjects expressed that they did not notice any difference between the conditions.Video Analysis of Participants’ Reaching BehaviourIt was unexpected to find that the Turn-Taking condition did not yield a significantly earlier reachtime than the No Gaze condition, while the Shared Attention condition did. This led to the hypothesisthat there are subtle yet important differences in participants’ behavioural responses to different robotgaze cues during HR handovers. In order to understand this result, a frame-by-frame video analysisof the recordings of the 97 participants3 (Turn-Taking: 33, Shared Attention: 32, No Gaze: 32) wascarried out using ELAN [95] with a special attention to the participant’s reaching (hand) behaviourand gaze behaviour. Two coders annotated the videos with partial overlay. The inter-coder reliabilitywas evaluated through Intra-class Correlation Coefficient (ICC), and showed substantial agreement (allICCs > .80, p < .01).The coders identified a participant’s start motion (the time when the participant’s hand starts tomove to the bottle, as a secondary measure of the reach time) and touch bottle times (the time whenthe participant’s hand touches the bottle). Figure 3.7 illustrates the distribution of these measures withrespect to the robot’s behaviours. A one-way ANOVA indicates that participant start motion varies acrossthe conditions (F(2,94) = 4.94, p < .01); post-hoc analysis using Bonferroni correction indicates thatparticipant start motion is significantly earlier with Attn (M = 3.29, SD = 0.50) than with Turn-Taking(M = 3.70, SD = 0.62) and No Gaze (M = 3.68, SD = 0.65), but no significant difference was foundbetween Turn-Taking and No Gaze. This is consistent with the findings reported in Section 3.4.5 thatrelied on reach time measured with IR sensors.Video Analysis of Participants’ Gaze BehaviourThe video analysis suggests that 92% of the participants looked at the robot’s head at least onceduring the handovers. These behaviours are always observed as quick glances, rather than long stares.The distribution of the glances with respect to the robot’s behaviours is shown in Figure 3.7.For all conditions, the glances are clustered around the time when the robot gripper is at the readyposition. In the Turn-Taking condition, a cluster of glances occurs after the robot starts shifting its gazeto the participant’s face (after t = 3.6 s). While this cluster of glances could be due to the robot’s gaze,it is also possible that the trigger for these glances is internal (e.g., the participants wanted to make3 One of the six participants rejected in the above analysis was still valid for the purposes of the video analysis. Hence, thevideo analysis includes 97, instead of 96 participants’ data.29Figure 3.7: Distribution of participants’ reaching behaviour (participant start motion and partic-ipant touch bottle) and gaze behaviour (glance at robot head) with respect to the robot’sgripper behaviour (marked by vertical solid lines) and gaze behaviour (marked by verticaldashed lines). ( c©2014 IEEE)sure it was an appropriate time to touch the bottle) since the glances are also close to the participanttouch bottle time. Result from a Chi-squared test on the number of glances between t = 3.6 s andparticipant touch bottle time indicates that participants look at the robot head more in the Turn-Takingcondition (12 out of 33) than in the Shared Attention (3 out of 32) and No Gaze (7 out of 32) conditions(X2(2) = 6.77, p < .05). This result suggests that the robot’s face gaze in the Turn-Taking conditioncatches some participants’ attention and elicits them to make eye contact with the robot.Could this be why the participants reach for the object later in Turn-Taking than in Shared Attentioncondition? Figure 3.7 shows that many participants have already started to reach for the bottle beforethe robot starts shifting its gaze to their face (before t = 3.6 s). In the Turn-Taking condition, 58%of the participants (19/33) started to reach before t = 3.6 s, so their participant start motion time canonly be affected by the robot gripper motion and shared attention gaze; 15% of the participants (5/33)started to reach during the transition period (during t = 3.6 to 4.0 s), while 27% participants (9/33)started to reach after t = 4.0 s, which could also be triggered by the robot’s face gaze. Six of the nineparticipants who started to reach after the face gaze (after t = 4.0 s) made eye contact with the robotbefore starting to reach for the bottle. Hence, participant start motion time from these individuals, toa considerable extent, was delayed. Therefore, it is possible that the late appearance of the robot’sface gaze contributes to the delayed average participant start motion time and average reach time in30the Turn-Taking condition, compared with the Shared Attention condition. One can test this hypothesisby changing the timing of the robot’s gaze toward the participants’ face, more specifically, introducingthe robot’s gaze to the receiver’s face earlier in the handover interaction, and measuring whether thismodification elicits an earlier reach time from participants.Since a cluster of glances from participants was observed around the time when the robot was at theready position regardless of conditions, it can be suspected that a face gaze immediately after the readyposition would be more likely to be noticed by the participants. Interestingly, when referring to resultsfrom Study 1 (Section 3.3), this proposed gaze pattern is effectively the Face pattern.3.5 DiscussionBuilding on previous work that studied communication of intent to handover using gaze, the studiespresented in this chapter delved into the use of gaze during a handover (i.e., after the intent to handoveris already communicated and while the handover is taking place). A handover interaction typicallyinvolves well-defined role assignments (a giver and a receiver) and a clear sequence of actions that musttake place (the giver grabs and transports the object, and then the receiver receives the object from thegiver). However, even in such a well-defined interaction between two agents, the details of when andwhere must be interwoven dynamically in order for the interaction to be successful. As an effort tounderstand what nonverbal cues may help the interweaving process, Study 1 was conducted to identifygaze patterns human givers use when handing over an object to another person. Results of this study notonly provide five different gaze patterns human givers use in an HHI handover, but it also suggests thatthe observed patterns more often than not involve gazing at the projected handover location. FollowingStudy 1, Study 2 explored the impact of robot gaze on HR handover timing and user perception of thehandover experience. Results show that participants reached for the proffered object significantly earlierwhen the robot performed a shared attention gaze at the projected handover location. In fact, participantsreached for the object even before the robot had arrived and stopped at the handover location (a meanof 0.11 seconds before the end of motion time). This is in contrast to the No Gaze condition where themean reach time is 0.52 seconds after the robot’s end of motion.In Study 2, participants were explicitly told that the robot would be handing over objects to themand that they were to take the object from the robot. In addition to this foreknowledge, the robot usedhighly contrasting poses between the ready position and the handover location which, according toCakmak et al. [27], makes the robot’s intent to hand over the bottle very clear. Hence, it is unlikely thatthe observed difference in timing between the gaze conditions is due to uncertainties in understandingthe robot’s handover intent. Rather, the results suggest that the robot’s gaze at the projected handoverlocation supplements the communicated intent with implicit information on where the handover shouldtake place. This may be helping to establish shared attention on the handover location even before therobot arrives there, naturally allowing participants to respond and meet the robot at the location earlierthan when such a cue is absent. Thus, the result best supports an increase of fluidity in the execution ofthe handover as it takes place.However, the role of mutual gaze used in the Turn-Taking condition required further investigation.31At the beginning of the robot’s handover motion, the robot expresses the same locational, shared at-tention gaze in both the Shared Attention and Turn-Taking conditions. Hence, it was surprising to findthat the reach time of the Turn-Taking condition is not significantly earlier than that of the No Gazecondition. This finding was supported even when the similar measure, participant start time manuallycollected from the video analysis, was used to verify this result.The following two hypotheses were tested that involve the Turn-Taking condition: that the Turn-Taking gaze would cue handover timing, and that looking at the participant’s face would improve thesubjective experience of handover. While there is a trend suggesting that the robot’s gaze directed atthe face improves the subjective experience of the handover, it remained uncertain whether the sharedattention gaze – instead of the face gaze – in the Turn-Taking condition is the main influencer of thehandover timing. It may be that the face gaze employed in this study served the function of acknowl-edgment rather than the intended function of giving a turn to the participant. There are also behaviouraldifferences in participants’ reaction to Shared Attention and Turn-Taking conditions. According to thevideo analysis of participants’ gaze behaviour, a number of participants waited for the robot to gaze atthe subjects’ face before starting to reach across the table.This raised questions about how human reach time is affected by the timing of the robot’s gaze.How much would varying the robot’s gaze timing affect human reach time? Is the timing of the robot’sgaze a more dominant cue than the location the robot is gazing at? That is, would a robot that shifts itsgaze from the object directly to the person’s face during handovers have the same effect as the SharedAttention condition?Also, it is important to note that results from Study 2 alone are representative of first-time participantresponses only, where novelty effects may have motivated the participants to observe the robot morecarefully than they would if they were more familiar with the robot. Unsurprisingly, there is a significanttraining effect in the reach time data, as well as a bias toward describing the second handover experiencemore favourably in the questionnaire regardless of the condition experienced. Some of the participants’comments suggest that in certain cases, people did not pay attention to the head of the robot at all.Indeed, it is suspected that in many HH handover scenarios, especially those that are repetitive or trained(e.g., passing a baton in a sprint relay race), people do not use gaze cues at all and still succeed in objecthandover. Thus, it was hypothesized that robot gaze cues might not have the same effect on trained orfamiliarized users.This is effectively the Face condition in Study 6 (Appendix A.1), a follow-up study to Study 2 thatcontrasts non-naı¨ve participant responses to handover events between the Attn, Face, and LongFace-Attngaze patterns (Figure 3.2). Results from Study 6 reaffirm findings from Study 2 that gaze significantlyaffects the timing at which participants reach for the object offered by the robot. It demonstrates thatthis effect holds true even after the training effect of HR handovers can no longer be observed throughrepeated handover trials. This implies that the even a subtle gesture of gaze used by a robot during ahandover event helps communicate the non-trivial information about the details of the interaction (inparticular, the information about when the handover can take place) such that the resulting HRI is morefluent and efficient. However, as presented in Appendix A.1, the LongFace-Attn condition, which is32designed to be a variant of the Face condition, did not perform as well as the Face condition itself.This result is similar to the findings on reach time in Study 2. Given that LongFace-Attn is a variant ofFace, just like how Turn-Taking is a variant of Attn, it was expected that the two conditions would elicitsimilar reach times. This indicates that, although gaze does affect the timing involved in a handover,the nature of its effect is not a simple function of the type and timing of gaze implemented on the robot.Further investigation is necessary to understand this relationship fully.Would there be changes in participants’ reach direction if the robot gazed at a different location?Without conducting an additional study, it is difficult to tell, with accuracy, if and when shared attentionis established with the participant. A separate experiment with a gaze tracking device would help answerthese questions, but this is beyond the scope of this study.Although the earlier reach time of participants in a handover may seem more similar to natural, un-scripted HH handovers, this may not necessarily be desired in some HR handover situations. Dependingon the handover controller implemented on a robot, handover timing may need to be controlled suchthat people naturally grab the object only when it is safe to do so. Many of the handover controllersthat modulate the release time of the object are built for cases where the robot’s gripper is already atthe handover location before people grab the object. A situation where the object is grabbed before therobot is ready to release the object could lead people to pull hard on the object, possibly damaging ordropping the object, or resulting in a negative perception of the robot.3.6 ConclusionMany previous studies provide evidence that a robot’s use of nonverbal cues can affect human be-haviours. The two-part investigation outlined in this chapter extends our understanding of the phenom-ena to cases where the cues are given during an atomic interaction. In this chapter, robot-to-humanhandovers are used as a means to investigate whether subtle nonverbal cues used by a robot can effec-tively influence details of human action for a successful, fluent HRI. Within the handover context, thedetail in question is the timing of the person’s reach to receive the offered object.There were three main contributions to the understanding of HR handovers from the work outlinedin this chapter. The results from Study 1 contribute to the understanding of the type of gaze that humanstend to use during HH handovers. Results from Study 1 identified five different types of gaze used duringHH handovers. The most frequently observed gaze pattern (Attn) from the giver was from the objectto be transferred to the projected location where the handover should take place. Results of Study 2demonstrate how a robot’s use of human-inspired gaze expressions during HR handovers can affect thetiming of the handover event in first-time participant responses. The study provides empirical evidencethat a human-inspired gaze pattern (Attn) implemented on a robot can elicit a human receiver to reachfor and retrieve the proffered object earlier than when no gaze (None) cues were provided. Study 6,a follow-up study presented in Appendix A.1, further extends this finding and demonstrates that suchan effect persists even after repetitions of HR handovers has taken place. These studies contribute to abetter understanding of the effect of gaze on HR handovers.In a broader context, findings from these studies support the hypothesis that a robot’s use of nonver-33bal cues during an interaction can be used to interweave details of the interaction. That is, the studiesprovide evidence that nonverbal robot cues can not only be used to communicate its “intent” prior totaking an action or to communicate details about its action independent of the user, but also elicit desiredbehaviours from the user while the agents are already engaged in the interaction.However, the robot’s gaze cues used in this study were designed as an open loop interactive be-haviour. The robot’s use of gaze was not designed into a feedback loop to respond to the person’s gaze.Further investigation is required to explore the full extent to which a robot’s use of nonverbal cues canhelp negotiate details of an interaction while the interaction is taking place. Investigations outlinedin the next chapter delve into this domain by focusing on a proactive and responsive robot nonverbalbehaviour designed to interactively negotiate for a solution to unforeseen resource conflicts.34Chapter 4Development of Negotiative Interactionfor Nonverbal Resolution ofHuman-Robot Conflicts4.1 IntroductionIn the previous chapter, a robot’s use of gaze cues during robot-to-human handovers was discussedas a means of affecting a human’s behavioural response with regards to when to reach to receive theobject. In the experimental handover context, there is no contention or ambiguity regarding the outcomeof the interaction (e.g., to successfully transfer the object to the person).However, in many everyday activities, it is common for interacting individuals to experience con-flicts regarding priority to access shared resources. When the question of who should get access to theresource first is uncertain, people often nonverbally communicate with each other to negotiate a solution.For example, when two people reach for the same object at the same time, they often manage to resolvethe conflict by one yielding the object to the other or by claiming the object before the other obtainsit. This chapter considers: What should a robot be programmed to do when such conflicts occur in anHRI? Answers to this question can vary depending on the context of the situation at hand. Investigationspresented in this and the next chapter focus on a robot’s use of interactive negotiation behaviours as apractical solution to the problem.Whereas the previous chapter investigated the efficacy of nonverbal cues for interweaving subplansin HRI with predefined roles and outcomes, studies presented in this chapter focus on an interactivescenario when the outcome itself – more specifically, the question of who should access the sharedresource first – is not predefined and requires negotiation. As Chapter 1 introduced, negotiation refers tothe bidirectional interweaving of subplans. As is clear from the dictionary definition, “Discussion aimedat reaching an agreement” [121], negotiation is a process that requires bidirectional communicationbetween the parties involved. Within the context of this thesis, only nonverbal communication anddialogue is relevant.35Figure 4.1: Overview of the process taken to devise the Negotiative Hesitation Generator (NHG)that can produce human-inspired, artificial hesitation motions for robotic manipulators. Thearrows represent the direction of information from one process that is used to inform the next.To investigate nonverbal negotiations between a human and a robot, this and the following chapterfocus on the nonverbal gesture of hesitation. As Section 4.2 presents in more detail, hesitation be-haviours in HHI are used to communicate uncertainties between people in their everyday life. They havebeen studied as a type of behavioural indicator in psychology and linguistics, and increasingly are beingused within the HRI community as a behavioural measure to examine user confusion and uncertainties.Presented in this chapter are two studies the author conducted (Studies 3 and 4) to investigate thepossibility of using hesitations as a means to dynamically negotiate conflicts in HR collaboration. Fig-ure 4.1 provides an overview of the process described in this chapter.In Study 3 (Section 4.3), the author conducted a two-part experiment with the aim of observingnegotiative hesitation gestures humans naturally exhibit in HHI. This study helped confirm and betterunderstand human use of hesitations as an immediate and nonverbal means to negotiate imminent re-source conflicts. In the first part of Study 3, an HH dyad experiment was designed to elicit numerousinstances of hesitation gestures in a task that involves natural sharing of resources. In the second partof the study, video recordings from the HHI experiment was used to conduct an online survey to mea-sure the level of hesitancy and persistency third party human observers can perceive from the collectedhuman gestures.With the results and collected data from Study 3, the author subsequently explored the collected andlabelled time series data from the study to identify key features that are characteristic of naturally elicitedhuman hesitations (Section 4.4). This data exploration process led to the discovery that the subset of36hesitation gestures of interest in this thesis share a common trajectory pattern. This pattern, referred toas hesitation loops (Section 4.4.6), can be reproduced with an implementation of three trajectory-relatedelements. These elements shaped the design of an artificial negotiative hesitation trajectory generatorfor a robotic system called the Negotiative Hesitation Generator (NHG) (Section 4.5).An online, video-based perception study, Study 4, was then conducted to validate the NHG. It wasimplemented on a 7-DOF articulated robotic platform as a means to generate robotic resource conflictresponses that are perceived as humanlike and expressive of hesitation (Section 4.6). Results from thisstudy provide an empirical support that the NHG-generated trajectories are capable of producing robotbehaviours that are perceived to be significantly more hesitant, persistent, animate, and anthropomorphicthan a conflict response consisting of smoothly stopping and pausing.Taken together, the work presented in this chapter contributes to a better understanding of kinematicfeatures of hesitation behaviours. It also provides an empirically validated means of generating robotconflict responses that are perceived as hesitations by human observers. Subsequently, human partic-ipants were invited to interact with the devised robotic system for an in-person experiment (Study 5)presented in Chapter 5. This helped investigate whether an interacting HR pair can negotiate and resolveconflicts pertaining to shared resources using purely nonverbal means, thereby dynamically determiningthe outcome of the conflict through interaction.4.2 BackgroundHesitations are social signals often observed as a form of disfluencies and uncertainties in human andanimal behaviours. In Psychology, Doob [44] defined hesitation in the temporal domain, as the time thatelapses between a stimulus and a corresponding response to the stimulus. Researchers in linguistics referto filled pauses or disfluencies in speech, such as uhs and ums, as hesitations [39, 93, 97, 160]. Rober, aclinical psychologist and family therapist, describes a variety of verbal and nonverbal behaviours – suchas prolonged silence, a glance, a sigh, and a pause in the flow of the conversation – as hesitations. Hesuggests that such verbal hesitations are related to the communicator’s wavering willingness to speakabout a matter due to an internal conflict that has not been verbalized [132]. In Human-ComputerInteraction (HCI), hesitation has been used as a measure of a computer user’s experienced difficultywith designed interfaces. Reeder and Maxion [129], for example, developed a simple algorithm thatdetects abnormally long pauses between inputs from keyboard and mouse to detect user hesitationsduring use of a computer interface. This automated detection of hesitation in user behaviour reduced upto 96% of an experimenter’s time required for computer interface analysis.Given the usefulness of understanding and detecting hesitation behaviours, linguists have used ma-chine learning approaches to detect hesitations in speech such as disfluency and filled pauses [139, 156,157, 164]. Being able to identify and recognize hesitations is increasingly being recognized as valuablein the study of HRI. Analogous to the HCI studies mentioned above, such natural human behavioursare immediate indicators of interrupted flow and increased uncertainty in an interaction [10, 32, 125].Researchers have employed the notion of hesitation behaviours as the duration of pause between actionsin HRI [12, 13, 76, 96, 142, 142].37For instance, Bartneck et al. [12] conducted a study in which the participants were asked to turn offa robot while the robot begged them not to. The researchers measured the amount of time it took theparticipant to shut down the robot as an expression of the participant’s hesitation in following throughwith the task. In another HRI study, human participants and a humanoid robot engaged in a verbalinteraction as part of an experiment on verbal turn-taking [142]. The robot used either a passive oractive mode of interrupting a conversation with the participant in an attempt to give or take speakingturns between the agents. In the active condition, the robot actively interrupted the participants to takeover the turn to speak, and it ignored the participants’ analogous interruptions to do the same. In thepassive condition, the robot hesitated (paused its actions) as a means to yield the turn to the participantwhen s/he interrupted the robot. In its yielding of the turn to speak, the robot was programmed tohesitate by pausing its actions for a given period until the participant either seized the turn from therobot or yielded to the robot. Results from this study demonstrate that people speak more when therobot is passive and yielding its turns than when the robot is active and interrupting the user.In addition to the hesitations measured and expressed as pauses, literature in sports, zoology, andHRI provides evidence that hesitation also includes certain kinematic behaviours in humans and animals[43, 104, 114, 116, 165]. However, only a few scholars have studied kinematic hesitation gestureswithin and outside the context of HRI [43, 103, 104, 114, 116, 148, 165]. This is in part due to the addedcomplexity of processing kinematic hesitation signals. Kinematic hesitations have high dimensionalityand provide a noisier signal than auditory pauses or the durations measured between keyboard/mouseinputs used in HCI. One of the few studies that focused on kinematic hesitation gestures in HRI is theauthor’s previous work [103, 104].Within the framework of HR collaboration involving reaching motions, the author asserts that hes-itations can be exhibited as a response to interruptions of the motions. The authors of Moon et al.[104] analyzed acceleration trajectories of the wrist in a subset of human hesitation gestures, called R-type hesitations, and formulated characteristic features from the trajectories as a series of cubic splinesin the acceleration domain called the Acceleration-based Hesitation Profile (AHP). R-type hesitationsinvolve the hesitating agent to immediately yield to the other agent upon encountering a resource con-flict. Trajectories of these hesitations are characterized as having three key points that represent thetemporal location and amplitude of acceleration at launch, braking, and yielding phases of a hesitationmotion. Figure 4.2 illustrates sample R-type trajectories. Using the AHP, the authors generated human-inspired R-type hesitation trajectories for a robot. They used an analytic approach to extract characteris-tic features of the most basic type of human hesitation trajectories that, when used to generate artificialhesitations for HR collaboration, are perceived as hesitations and clearly distinctive from other similarrobot motions. Findings from Moon et al. [104] demonstrate that human observers correctly recognizethe artificially generated robot-hesitation having the AHP as hesitations and can distinguish them fromother similar robot motions. However, the type of hesitation behaviour (R-type) investigated in Moonet al. [104] was limited to the simplest form of hesitation behaviour, and one that is not meant to elicitnonverbal dialogue among the interacting HR pair.11 By definition, R-type hesitations are comprised of reacting to the conflict with a jerky halting motion followed by an38Figure 4.2: Demonstration of the AHP that characterizes R-type hesitations. Human R-type hesi-tation motions shown here are captured at the wrist and characterized in acceleration space.R-type hesitations involve the hesitating agent to immediately yield to the other agent uponencountering a resource conflict. The reference S-type motion (a Successful reach-retractmotion) is characterized by a relatively symmetrical accelerate-decelerate, and then accel-erate back to the starting point, motion. R-type hesitation trajectories (samples 1-4 shown)were characterized as four splines that connect the three key points – t1 and a1, t2 and a2,and t3 and a3 – and the start and end points of the motion. The three key points representthe temporal location and amplitude of acceleration at launch, braking, and yielding phasesof a hesitation motion. The ratios between these variables in acceleration space have beendetermined from observed R-type human hesitation trajectories.In developing a robotic system that can use hesitation gestures to resolve resource conflicts withhumans interactively, it is important to equip the robot with a vocabulary of hesitation gestures thatproactively engages the user in a nonverbal dialogue. One method of enriching the vocabulary of hes-itation gestures for a robot is to understand and mimic how humans negotiate a solution to resourceconflicts using hesitations (negotiative hesitations). Once a set of trajectory features is identified fromnegotiative human hesitations, these features can be used to generate analogous robot trajectories for HRnegotiation of resource conflicts. The studies presented in this chapter contribute to this end by collect-ing negotiative hesitation motion samples from a human-subjects experiment and examining trajectoryfeatures that can be implemented in an HR collaboration context.immediate yielding of the resource before any dialogue can take place.394.2.1 PersistencyOne of the elements of negotiative hesitations that distinguish them from R-type hesitations is theexpressed level of persistent interest and prolonged need for a resolution of the conflict at hand. Bypersistency, the author refers to the quality that is expressed in the persistent state of an agent or the actof persisting in something (i.e., how persistent is an agent?). Here, the author refers to the dictionarydefinition of persistence as “the fact of continuing in an opinion or course of action in spite of difficultyor opposition,” and “the continued or prolonged existence of something” [122]. While persistenceis synonymous to nouns such as perseverance, tenacity, endurance, and tirelessness, the concept ofpersistency discussed here pertains to the spectrum of persistence conveyed in a communicative signal– in particular, in hesitations.Persistency expressed as part of human behaviour has been studied in a variety of contexts. In childdevelopment literature, in particular, persistence is accepted as a means to determine, measure, and an-alyze intentional communication, especially in infant preverbal communication signals [51, 53, 85]. Instudies involving human infants, persistency is measured as repeating of or augmenting a version of acommunicative gesture when the infant’s initial attempt to communicate with the recipient fails [51, 53].Persistence is also one of two criteria for an infant’s gesture to be considered an intentional communi-cation [85].2 Scholars in child development suggest that intentional communication – evidenced bypersistence, among other behaviours – is a means for the communicator to manipulate the interlocutorrather than a mere attempt to engage in a conversation [54]. This suggests that negotiative hesitationsby a robot that – in contrast to R-type hesitations – express a higher level of persistency are likely tobe perceived by humans as intentional communication, possibly leading to the emergence of nonverbaldialogues.Persistence in communicative gestures is also closely related to negotiation behaviours. Golinkoff[53] states that persistence is seen as a proactive behaviour that is a necessary component of ongoingnegotiation between mother and infant. Golinkoff [53] also suggests that even at infancy, humans ne-gotiate with their mothers. An infant’s persistent communication with the mother elicits this type ofnegotiation, which allows the infant to use communicative cues to manipulate its interlocutor/mother.In this context of infant-adult negotiation, what is being communicated by an infant is the infant’s de-sires and intents. The negotiation between them takes place for the purpose of the infant influencingthe decision-making of the interlocutor (e.g., a child requesting for an adult to bring a toy that is out ofreach).In this work, the author investigates persistency as it pertains to negotiative hesitations. A level ofpersistency is assumed to be expressed in negotiative hesitations since negotiative hesitations requirethe hesitating agent to have a persistent interest in resolving a conflict. Therefore, the studies presentedin this chapter explore persistency as a concept related to, but not trivially correlated with, hesitancyexpressed in negotiative hesitations.2 Elaboration of the original communication signal upon failure to communicate is the other criteria.404.2.2 Social Signal ProcessingScholars who attempt to generate readily recognizable social robot behaviours often employ em-pirical methods that involve human subjects. For example, Lim et al. selected four parameters (speed,intensity, regularity, and extent), and varied them across three modes of expressing emotion (gesture,verbal expression, and music) [88]. Their aim was to determine whether these four parameters could beused to generate artificial expression of emotion across the three modes. The studies presented in thischapter are similar in terms of studying the qualitative content of communicative, time-series signals. Tovalidate the usefulness of parameters in emotion expression, Lim et al. [88] measured the percentage ofhuman subjects who recognized the designed, artificial emotional expression correctly. However, suchpercentage measures for validating the efficacy of parameters studied are not sufficiently informative toallow one to create new signals of the same quality.Studies in Social Signal Processing (SSP), or socially aware computing, share the similar goal ofanalyzing time-series signals from human subjects using algorithms. In SSP, sensed signals from anindividual, such as audio or visually observable behaviours, are used to infer otherwise hard-to-measureinternal states of the individual computationally [136, 160]. Some of the popular work in SSP includeinferring individuals’ states and actions during group discussions (e.g., dominance or interest level) withdynamic Bayesian networks [109] or layered Hidden Markov Models [117] on the individuals’ speakingenergy and body language.While the primary goal of SSP is to perceive measurable social signals to infer internal states fromhumans, the design of interactive, social robots has the added challenge of needing to respond to theinferred human states with behaviours that express appropriate social signals easily understood by hu-mans. Hence, a number of well-performing approaches in SSP are not directly transferable to HRI.3 It isnot always possible for a robot to exhibit the mix of signals needed to produce a social signal understoodby human observers (e.g., the robot may not have a speaker to verbalize a message, or a finger to pointat an object) [160]. Therefore, in contrast to employing algorithms that may require the use of numerousfeatures from a social signal, it is advantageous to identify the minimum number of features that a robotcan generate to convey desired social signals to human observers. In exploring the human motion datacollected from Study 3 (Section 4.4.3), the author attempted to find such a minimum set of features thatcan be used to generate artificial, negotiative hesitation gestures for a robot.4.3 Study 3: Observing Hesitations in Human-Human DyadsThis section describes a human-subjects experiment conducted to collect a large set of naturallyexhibited, negotiative human hesitations. Section 4.3.1 describes the experiment involving a motioncapture system (VICON, Vicon Motion Systems Ltd., Oxford, UK, [159]) and the experimental taskthat was designed to trigger multiple instances of conflicts about shared resources between two human3 For example, it is often suggested that combining different classifiers, each of which classifies different aspect of the sameproblem, performs well and is recommended in SSP. While this recommendation is useful in classifying even complex socialsignals, generating the mix of signals for the robot to exhibit in order to send the same social signal to the human observersremains a challenge in social robotics.41participants. As presented in Section 4.3.1, the collected set of human behaviours was then labelledthrough a video-based online study using the Amazon Mechanical Turk platform [5]. This online studywas designed to collect the varying degrees of perceived persistency and hesitancy people perceive fromthe recorded instances of hesitations. These samples of hesitation gestures were then used to explore thecommon features that are present in the trajectories. The exploration of the trajectory data is presentedin the following section (Section 4.4).4.3.1 Experimental ProcedureThe author conducted an HH paired experiment in a controlled environment. Paired participantswere brought into a laboratory and introduced to their partners. Eight pairs (N=16, 3 male-male, 2female-female, 3 female-male pairs) of volunteers participated in this study.The experiment employed nine VICON cameras (six T40, two T160, and a synchronized digitalcamera) to capture participants’ motion at 100 Hz (50 Hz for the digital camera). Both participantsin each pair wore seven reflective markers on the joints of their dominant arm and hand as shown inFigure 4.3. This allowed the collection of Cartesian coordinates for each marker with respect to apredetermined reference point in space. Knowing the significant role wrist trajectories played in theprevious work, Moon et al. [104], participants wore two markers on either side of the wrist for moreaccurate measurement of this joint.After being instrumented with the VICON markers, participants stood facing each other across atable and played a modified version of Solitaire, a card sorting game. On the table was a deck ofrandomly ordered cards organized into two piles. In front of each participant were two aces of the samecolour (either red or black). The participants’ task was to order two full sets of cards hierarchically (e.g.,ace, 2, 3, and so on) starting from the two aces given to them. They had to order the cards according toan alternating colour pattern, following the rules of the traditional Solitaire. The participants were toldto finish the task as fast as they could. They were also instructed to access the two piles of cards in analternating (left - right or right - left) order and to grab or return only one card at a time. This requirementfor alternating access to the shared resource (two piles of cards) added cognitive load on the participantsand prevented them from accurately keeping track of which pile of cards the other person would reachfor next. By the design of the game, both participants needed to frequently access the shared piles ofcards to finish the task, resulting in multiple natural occurrences of hesitations along with numerousreaching motions exhibited by each. Each pair of participants played the game twice, swapping thecolour of the starting aces in the second round.Labelling Human Hesitation GesturesThe experimenter watched all subject motions from the video recordings of the experiment and man-ually segmented and labelled participant motions that were deemed hesitations. In total, this qualitativeprocess yielded 302 trajectories of hesitation gesture samples (171 from subjects on the left and 131from the right side of the digital camera).Subsequently, a video-based online survey was conducted using the segments of videos labelled42Figure 4.3: A side view of the HHI experiment captured from VICON’s synchronized digital cam-era. Each pair of participants stood facing each other with a table between them and wereasked to play a modified game of Solitaire. The numbers indicate the seven markers, usedto track the participants’ motion, placed on the following locations: 1 – shoulder (scapularacromion), 2 – elbow (humeral lateral epicondyle), 3 – right wrist (ulnar styloid), 4 – leftwrist (scaphoid), 5 – knuckle (metacarpophalangeal joint of the second finger), 6 – upperfinger (interphalangeal joint, between first and second phalanges, of the second finger), and7 – lower finger (interphalangeal joint, between second and third phalanges, of the secondfinger).as hesitations in order to identify which samples of motion were perceived to express lower or higherdegrees of persistency and hesitancy. Participants for this online survey were recruited using the Ama-zon Mechanical Turk platform [5]. They were required to understand basic English to follow writteninstructions, and have good vision to be able to watch video recordings of human gestures.Each video contained a segment identified as a hesitation plus a second before and after the motionsegment used to provide the context of the motion. The survey participant watched and scored a randomselection of twenty hesitation videos on the following two seven-point scales:• Persistency “How persistent do you think the person on the LEFT is?” (1- Not persistent at all, 7- Extremely persistent)• Hesitancy “How hesitant do you think the person on the LEFT is?” (1 - Not hesitant at all, 7 -Extremely hesitant)We changed the word LEFT to RIGHT, as appropriate, to highlight the specific individual of interest inthe video.434.3.2 Results and DiscussionBoth the VICON study and the online survey were approved by the University of British Columbia(UBC) Behavioural Research Ethics Board (H10-00503).A total of 300 participants completed the online survey. For each segment of motion (video), ap-proximately 20 individuals (M = 19.8, SD= 3.25) provided both persistency and hesitancy scores. Theaverage persistency and hesitancy scores were computed for each, and the score averages range from1.68 to 6.26 (M = 4.32) for hesitancy and 2.10 to 6.58 (M = 3.94) for persistency. The average standarddeviation of the hesitancy and persistency scores range from 0.83 to 2.39 (M = 1.58) and 0.67 to 2.37(M = 1.52), respectively. These are a reasonable range of scores and standard deviation considering the7-point Likert scale data. The data collected from this study is used to analyze human motion data asdiscussed at length in the following sections.4.4 Exploring Human Hesitation TrajectoriesTo generate artificial hesitation behaviours for a robot, it is necessary to understand what trajec-tory features distinguish hesitation gestures from uninterrupted reach motions. The author’s previousresearch [103, 104] demonstrated that there is a characteristic acceleration profile in R-type hesitationsthat is different from uninterrupted reach motions. R-type hesitations are comprised of an individual’sreach trajectory toward a target, and interruption during the reach that leads to an immediate halting ofthe reach. This is followed by a retraction (yielding) motion that returns the agent back to the startingpoint of the reach. Based on this characteristic, human-recognizable robot hesitations can be success-fully generated even for a 7-DOF robot4 with a non-anthropomorphic morphology [104].However, the same research showed that the same trajectory features are absent in negotiative hesi-tations (formerly introduced as P-type in Moon et al. [103]), and that the acceleration profiles of nego-tiative hesitations collected in the study are similar to that of uninterrupted reaches. Figure 4.4 demon-strates this finding. Although the negotiative hesitations may seem indistinguishable from reach motionsin the acceleration domain, the results of Study 3 suggest that humans can perceive and recognize thesubtle differences between the two types of motions. This result suggests that negotiative hesitationtrajectories must be characterized in a different domain or using a different set of trajectory features.This section presents the author’s two-part exploration of the negotiative hesitation trajectory datasetas an attempt to identify characteristic features from the trajectories. The first process employs infer-ential statistics to understand the nature of negotiative human hesitations better. The second processtries to determine the metric of the collected human trajectories that can be used to inspire the designof hesitation behaviours for a robot. As is common in exploratory work such this, much of the analysesconducted on the collected human trajectories did not produce positive results. Some of these failedinvestigations led to the ultimate findings outlined in this section. Appendix A.2 documents these failedanalyses. Appendix A.2 also presents detailed results of the final analyses described in this section soas not to distract the reader from the main findings.4 WAMTM, Barrett Technologies, Cambridge, MA, USA.44Figure 4.4: Acceleration profiles of sample negotiative hesitations, formerly referred to as P-typehesitations in [103, 104]. Two samples of the hesitations are shown here along with a sampleof an uninterrupted reach motion (S-type). These samples demonstrate the negotiative hesi-tation gestures’ similarity to an uninterrupted reach motion in this domain. ( c©2014 IEEE)In the following section, Section 4.4.1, the author provides a description of the segmentation andpre-processing conducted on the stream of trajectory data collected in Study 3. The processed samples ofthe negotiative hesitations and uninterrupted reach motions were selected and divided into four samplesets. This division of sample sets allowed the author to conduct analyses that require a balanced numberof samples as well as those with robustness to unbalanced number of samples (Section 4.4.2). Subse-quently, 75 trajectory features were computed and analyzed from the motion samples. Section 4.4.3describes this process.Results from this exploration point to two metrics with which the hesitation trajectories can bedistinguished from reach motions. This result, discussed in Section 4.4.4, informs the subsequent inves-tigations on negotiative hesitation motions (Section 4.4.5) and ultimately, the design of the NHG.4.4.1 Pre-Processing and SegmentationBefore analyzing the collected trajectories from Study 3, the raw recordings of human motionswere processed and prepared as follows. First, the recorded Cartesian coordinates of the joints weretransformed with respect to the coordinates of the shoulder. The average of the two markers placed oneither side of each participant’s wrist was computed for a more consistent treatment of the wrist jointmeasurement. The trajectories were filtered using a Butterworth low-pass filter at the cutoff frequencyof 10 Hz, the highest frequency at which voluntary human motion can occur. After a number of failedanalyses using trajectories of all joints, the author narrowed the focus of the analyses to the wrist joint.This decision was based on the findings from the author and colleagues’ earlier work, Moon et al. [103],demonstrating that robot motions generated by tracing the wrist trajectories of human hesitations are45also perceived as expressive of hesitation.Herein, M(t) and P(t) refer to the Cartesian coordinates of the main and partner participants’ tra-jectories, respectively. The orientation of the wrist was ignored to simplify the analysis. The mainparticipant is the one whose motion segment is labelled as a reach or a hesitation. The partner par-ticipant’s trajectory refers to whatever motion the partner was exhibiting during the main participant’sreach or hesitation. Also, dimensionality-reduced expressions of the relative wrist motions from the3D Cartesian coordinates (α1(t) for the main and α2(t) for the partner participant) were produced byselecting the dominant component from the Principal Components Analysis.One of the key challenges when working with human motion trajectories in an unconstrained exper-imental task, such as the one described in the previous section, is the need to divide a stream of motiondata into segments that can be labelled and studied as a supervised learning problem. Since the partic-ipants’ motions during the experiment were not constrained spatially (they could move about the taskspace in whatever direction they chose), it was imperative to segment the vast amount of the collectednatural reach motions in a systematic way such that they could be compared to segments of hesitationmotions.Different methods were employed to tackle this problem for each motion type. To segment reachmotions, coordinates of practical target locations for each participant were computed. Clusters of eachparticipant’s trajectory data within 150 mm (approximately half of the distance participants travelledin a reach) from the known centre locations of the two card decks (targets) were identified. Theserepresent the locations (Tm1 and Tm2 for the main, and Tp1 and Tp2 for the partner participant) wherethe participants’ wrists were located when they grabbed a card from or placed a card onto the deck.Compared to the known centre location of the target objects, T1 and T2, these locations correspond tothe Cartesian points in space the person tried to reach for, based on his/her preferred way of taking cardsfrom the deck. These locations were typically near one of the four corners of the card closest to theparticipant’s right hand. Figure 4.5 illustrates an example output of the segmentation algorithm used.Since it was not clear which patterns exist in marking the start and end of a hesitation, these sampleswere manually segmented by watching the video recordings.4.4.2 Sample SelectionThe author excluded outliers from the 302 hesitation samples collected in Study 3. These outliershad≥ 3 SD in velocity maximum and range. With the above-mentioned procedure, a total of 1706 reachand 298 hesitation motion segments (N(hes≥0) = 298) were collected and selected for analysis. That is,only 15% of the total samples, Nub1 = 2004, are hesitation samples. To account for the large unequalsample size, the author employed a random number generator to randomly select the same number ofreach samples from a participant as hesitation samples from the same participant (balanced total samplesize, Nb1 = 596).Since one individual manually segmented the hesitation motions, there is a possibility of bias fromthe individual. To account for this possibility, the author used the hesitation scores collected fromMechanical Turk for each of the segments to filter out samples with hesitancy scores below the median46Figure 4.5: Segmentation of a stream of a representative participant’s motion shown in Euclideandistance trajectories, d1(t), with respect to the two target locations, Tm1 (solid line) and Tm2(dashed line). Unlike the known centre location of the two card decks, T1 and T2, Tm1 andTm2 represents the participant’s distance to the Cartesian points in space the person tried toreach for (e.g., bottom right corner of the left deck). Points indicated by o and * are thezero velocity crossings of the Euclidean trajectory that coincide with the start and end ofa participant’s reach motion. The shaded portion indicates an area manually labelled as ahesitation.(hesitancy < 4). This yielded a total of 192 hesitation samples (N(hes≥4) = 192), resulting in Nb2 = 384and Nub2 = 1898.In summary, this process created four sample sets (Nb1, Nub1, Nb2, Nub2) with which one can investi-gate salient features that set hesitation samples apart from reach motions. The following section outlinesthis data exploration process.4.4.3 Data ExplorationTo explore trajectory features that are characteristic of hesitation gestures, the author obtained andinvestigated a set of 75 features from the segmented sample trajectories of both hesitation and reachmotions. This initial set includes the five types of features – maximum (max), minimum (min), mean(µ), amplitude (A), and number of zero crossings (ρ) – computed for the following fifteen metrics:• d1(t)= ‖M(t)−Tm‖, the main participant’s Euclidean distance to target, and its first (d˙1(t)) andsecond derivative (d¨1(t)),• α1(t), the main participant’s principal component of the segment along the direction of travel, andits first (α˙1(t)) and second derivative (α¨1(t)),• d2(t)= ‖P(t)−Tp‖, the partner participant’s Euclidean distance to the same target, and its first(d˙2(t)) and second derivative (d¨2(t)),• α2(t), the partner participant’s principal component of the motion segment along the direction of47travel, and its first (α˙2(t)) and second derivative (α¨2(t)),• δ (t)= ‖M(t)−Tm‖−‖P(t)−Tp‖, the difference between the main and partner participants’ Eu-clidean distances to their respective target locations, and its first (δ˙ (t)) and second derivative(δ¨ (t)).Here, Tm and Tp refer to the static target location relevant for the segment of motion computed foreach participant. It is implied that the variables Tm and Tp refer to the appropriate one of the two customtarget locations5 relevant to the particular motion segment. The purpose of examining d1(t), d2(t),δ (t), and their derivatives is to find out which of the metrics deserve further analysis as an attempt tocharacterize and produce hesitation motions.With the 75 features considered, the Shooting Algorithm was used as a regularization method [50]for building a logistic regression model of the two types of motions. This process allowed the authorto eliminate a large number of the 75 features that do not distinguish one motion type from the other.The optimum λ value, a tuning parameter for the shooting algorithm, used was 16 for all sample setsconsidered. The regularization path used to obtain the optimum value is presented in Appendix A.2. Asis a standard practice, the features used were normalized to have a zero-mean with unit variance acrossall samples considered. A total of twenty-one features were found to have a non-zero weight (> 0.0001).See Table A.2 in Appendix A.2 for the full list of the features and their weights computed for the foursample sets.Here, it is important to note that, due to the definition of δ (t), some of the 75 features considered arelinearly correlated with each other. Hence, one can expect that some of the results from the inferentialstatistics and shooting algorithm will be redundant. That is, for a feature computed across metrics d1(t),d2(t), and δ (t), the shooting algorithm would return non-zero weights in two out of three metrics. How-ever, rather than eliminating one of the metrics entirely from the analysis, the author chose to exploreall of these features. This is because these features represent distances, velocities, and accelerationsin physical space. Therefore, investigating which two of the three metrics return with a larger weightfrom the shooting algorithm analysis can be useful, because it helps identify the features that are mostnumerically sensitive to change and likely most meaningful for implementation in real-life.The features from the shooting algorithm that yielded non-zero weights at the optimum λ wereselected as features worthy of further analysis. Using combinations of these features, the author em-ployed a Support Vector Machine (SVM) having a linear kernel to build a logistic regression model. Thisprocess helped examine the significance of the features in classifying one motion type from the other.In addition, the author conducted t-tests on the balanced sets of samples (Nb1 and Nb2) and Welchtests6 on unbalanced set of samples (Nub1 and Nub2). The inferential statistical analyses were conductedon computed max, min, µ , A, and ρ of d1(t), d2(t), and δ (t) metrics. This analysis helped better testsome of the assumptions about the nature of negotiative hesitation behaviours. Although only minor5 One target location belonging to the left card deck, and the other belonging to the right card deck, as explained in theSection Welch test serves the same function as that of t-test, while computationally accounting for the differences in sample sizein the distributions being compared.48differences were expected between the results of the t- and Welch tests, both were conducted to gaugepossible biases in the random sampling of the reach motions in generating the balanced sample set. Thefull results of the t-tests and the list of features with significant results are presented in Table A. Feature Differences in Reach and Hesitation Motion SamplesThis section presents a summary of key findings from the t- and Welch tests followed by SVMclassification models that were built to confirm saliency of selected features further. The two metrics ofinterest for further analysis were selected based on the results of the SVM and informed the discussionof trajectory characteristics of hesitation samples in the next section.Inferential Statistics for Understanding Significant Trajectory DifferencesThis section presents a brief summary of the inferential statistics conducted on the four sample sets.While all 75 features were tested, only 47 features showed significant differences7 between hesitationand uninterrupted reach motion samples, and not all of the significant results are non-trivial findingsworthy of discussion. Hence, the full results of the statistical tests are presented in Table A.3, and onlythe discussion of the inferential statistics on the metrics related to d1(t), d2(t), and δ (t) are presentedhere.d1(t)– Main participant’s Euclidean distance to target: Feature differences observed from the mainparticipant’s reach and hesitation trajectories provide empirical confirmation of some of the assumptionsabout hesitations. First, hesitation motions have a significantly larger minimum distance to the target,min(d1(t)), and smaller range of motion travelled, Ad1(t), on average than uninterrupted reach motions.This indicates that, as expected, the participants travelled a smaller distance and did not get as close tothe target when hesitating as compared to when they successfully reached the target. Observations ofzero crossings, ρd1(t), suggest that there are multiple zero velocity crossings for a majority of hesitationmotion segments in the sample sets. A total of 179 hesitation motion segments have two or more zerovelocity crossings, of which 112 have more than three, 56 more than four, and so on. Table A.4 presentsthis distribution.d2(t)– Partner participant’s Euclidean distance to target: There is a significant difference in the dis-tribution of the partner participant’s minimum and average distance to the target (min(d2(t)) and µd2(t),respectively) in hesitation and reach motions. In the segments of motion where the main participant ishesitating, the partner participants tend to be closer to the target than in segments where the main partic-ipant is fully reaching. This finding supports the hypothesis that the spatial state of an interacting agentsharing the same space influences behavioural responses – in this case, hesitations – of another. For thepurposes of this thesis, this relationship indicates that the partner participant’s location with respect tothe target is a key influencer of the main agent’s hesitation motion. Moreover, a significant portion of7 All inferential statistics reported here were conducted with α < 0.05.49Ad2(t), the range of motion covered by the partner’s motion, for the main participant’s reach segmentsare found to be zero or near-zero. This indicates that the partner participants were often stationary ornot moving along the axis toward or away from the target in these samples. This is in contrast to thelarger Ad2(t) for cases where the main participant is hesitating.δ (t)– Difference between d1(t) and d2(t): δ (t) is a measure of the difference between the main andthe partner participant’s distance to their respective target locations. Based on the experimental setup ofStudy 3, it is also a measure of who is closer to the target and by how much. Results of the statisticaltests suggest that the hesitation motion segments have a significantly larger and positive max(δ (t)), themaximum difference between the main and the partner participant’s distance to the target, on averagethan reach motion segments. This indicates that the main participant is farther away from the targetthan the partner when the participant is hesitating. For reach, the average of max(δ (t)) hovers aroundzero, indicating an unbiased mix of cases where one participant is farther away from or closer to thetarget than the other. Complementing this spatial dynamics is the finding that min(δ (t)) has a largernegative value in reach motions than in hesitations. This indicates that, for reach segments, the partneris typically farther away from the target than the main participant. This is consistent with the result ofAd2(t) described above in which the participants were interpreted to be stationary or not moving towardor away from the target when the main participants fully reached the target.From the abovementioned results, it is clear that the main participant’s hesitation motions are af-fected by the trajectories and spatial locations of the partner participant with respect to the shared re-source. This suggests that in characterizing negotiative hesitations – unlike the R-type hesitations thatcan be characterized from a single agent’s wrist trajectories as a standalone motion, irrespective of thestate or presence of another agent – one must take into account the spatial context of the workspaceshared by the dyad.Verification of Salient Features with SVMA number of different combinations of features were used to classify one motion type from the otherusing the SVM approach. For each logistic regression model built with an SVM, a 4-fold cross validationwith a 50% train/test ratio was conducted. To obtain a fair measure of the performance of the models,only the two balanced sets of samples, Nb1 and Nb2, were used in this process. The ν parameter8 wastuned to have approximately 15% of the total samples designated as support vectors. Appendix A.2presents the full results of the SVM models.In this process, two features, max(d˙1(t)) and µδ˙ (t), stood out from the rest as strong contributorsfor accurately classifying hesitations from reach motions – max(d˙1(t)) represents the maximum speedthe main participant was moving toward/away from the target, and µδ˙ (t) represents the average of howthe main participant’s speed toward/away from the target compared against the speed of the partnerwith respect to the target. In particular, using only these two features for the Nb2 sample set yields a8 The parameter ν ∈ (0,1] in the Support Vector method controls for the number of samples used as support vectors in theregression. See Chalimourda et al. [33] for more detail.50Figure 4.6: Illustration of an SVM model with Nb2 = 384, ν = 0.3, and SV = 66. Featuresmax(d˙1(t)) and µδ˙ (t) are used. The highlighted samples (thick circles) represent supportvectors. Class 1 refers to reach and Class 2 to hesitation samples. White circles represent thehesitation samples that have been properly classified using the algorithm.test accuracy of 86% with a 50% test/train ratio and 82% with 75% test/train ratio (CI : 69.4,87.0 andCI : 68.8,92.3 for the 50 and 75% ratios respectively).9 Figure 4.6 presents this SVM model.This finding alone does not suffice for producing artificial hesitation behaviours for a robot. How-ever, the saliency of these features suggests that the trajectory metrics d˙1(t) and δ˙ (t) could providepromising results toward this end upon further analysis. These results motivate the follow-up inves-tigation outlined in the next section that ultimately leads to the design of a human-inspired hesitationtrajectory generator.4.4.5 Understanding Hesitation LoopsVisual investigation of hesitation and reach samples in d˙1(t) and δ˙ (t) presents trajectory patternsthat are common to the majority of hesitation samples but are absent from reach gestures. That is, whenhesitation samples are plotted in the state space of δ˙ (t) and δ (t), the trajectories tend to have the shapeof a loop (see Figure 4.7). This is in contrast to reach trajectories visualized in the same state space asshown in Figure 4.8.To verify this feature as characteristic of hesitation motions in contrast to reach motions, the authorcounted the number of samples that have a looped shaped in the state space. Results show that a total of9 SVM parameters ν = 0.3 and SV = 67(17%) were used for this model.51Figure 4.7: Overlay of hesitation samples demonstrating the nature of the δ˙ (t) vs. δ (t) plots tohave circular loops around the zero δ˙ (t) axis. Zero Velocity Crossing (ZVC) of the mainparticipant is marked in red circles. Blue crosses represent the start of the motion segment.As can be observed in sample A0L2, not all but most motions labelled as hesitation looparound δ˙ (t)=0. A0L2 in particular consists of a hesitation that has larger sideways thanback-and-forth motion with respect to the target.134 hesitation samples out of Nhes≥4 = 192 share this looped trajectory that is absent in reach gestures.Further analysis of the location and direction of the loops suggests that all of the loops in hesitationsamples encircle δ˙ (t) = 0. Due to the linear relationship between δ˙ (t) and d˙1(t), these loops can alsobe observed in d˙1(t) vs. δ (t) space.With the sample hesitations that have these loops, the author computed the Euclidean distance trav-elled by the main participant between the two zero crossings of the encirclement. This distance repre-sents how much the main participant retracted, if any, while the negotiative hesitation took place. Basedon the computation, the distribution of this distance, Kickback Distance (KD), has a mean of 19.6 mm(SD = 11.8). However, as shown in Figure 4.9, KD of human hesitations are not normally distributed.This raises the question whether implementing a generator that exhibits a KD of 19.6 mm would producecommunicative negotiative hesitation gestures for a robot.It is important to note that, since behaviours humans perceive as hesitations can take on manydifferent forms, including lack of motion, the characteristic loops referred to as hesitation loops inthis section are not proposed as a feature that is present in all hesitations. Instead, they are presentedas characteristic features of a subset of hesitation behaviours studied in this work, and put forth ascomponents that can be implemented onto a robot to generate human recognizable artificial hesitationbehaviours.52Figure 4.8: Overlay of reach samples in δ (t) vs. δ˙ (t) state space. As shown in the figure, reachsamples demonstrate a fewer number of circular loops around δ˙ (t)=0 than negotiative hesi-tation samples.The ultimate goal of having such dynamic and reactive behaviours in a robot is to be able to generatenegotiative behaviours in HRI that result in an interactive resolution of conflicts. To this end, hesitationloops are discussed in more detail in the following section and are used to frame the design of the NHG.4.4.6 The Four Cases of Hesitation LoopsGiven the discovery of hesitation loops, one can formulate an intuitive understanding of the contextembedded in the trajectory in state space.First, trajectories that move in the positive direction of state space, or crossing δ (t)= 0 from theleft to the right quadrants, can indicate that the main participant was moving toward the target or thepartner was moving away from the target. In the former case, the trajectory must travel in the upperquadrants (δ˙ (t)> 0);10 in the latter case, the trajectory must travel in the lower quadrants (δ˙ (t)< 0).11Since hesitation motions of the main participant are of focus, only the former case is interesting for ourdiscussion.Trajectories that move in the negative direction in the state space represent either that the partnerparticipants were getting closer to the target than the main participant, or that the main participant wasmoving away from the target while the partner remained stationary. Likewise, only the trajectories that10 d˙1(t)> d˙2(t)11 0− d˙2(t)< 053Figure 4.9: Distribution of KD collected from the 134 samples of hesitations in (Nhes≥4 = 192) thatencircle δ˙ (t)=0. The distribution is non-normal with most of the values occurring between 5to 30 mm.move from right to left while the partner is moving faster than the main participant (δ˙ (t)< 0)12 arerelevant to the discussion of hesitations.There are four cases of observed hesitations that describe the spatial contexts in which hesitationsoccur. Figure 4.10 provides a visual demonstration of the cases. This section outlines the HH dyad’sinteraction dynamics that are represented in these cases and what the entering and exiting directions ofa hesitation loop mean in these contexts.Case 1 This case represents hesitation samples starting with a state of (δ (t)> 0, δ˙ (t)< 0). In this case,δ (t)> 0 indicates that the main participant is farther away from the target than the partner at the onsetof a hesitation (d1(t) > d2(t)). δ˙ (t)< 0 suggests that the main participant is moving slower toward thetarget than the partner. Indeed, it is also possible that the same starting state in state space can representthe main participant moving faster away from the target than the partner who may be moving toward oraway from the target him/herself. However, these non-contentious scenarios are not considered here.1312 d˙1(t)< d˙2(t)13 All of the hesitation samples considered in this investigation have a resource conflict to resolve. This quality is ensuredfrom the manual segmentation method employed for the hesitation samples, which involved qualitatively observing recordingsof human motions during occurences of resource conflicts.54Figure 4.10: The four cases of hesitation loops demonstrated in the δ˙ (t) vs. δ (t) state space.Each of the cases represent a spatial context between the two interacting agents’ motions asoutlined in Section 4.4.6. All hesitations loop around δ˙ (t)= 0.From this starting location, the main participant remains farther away from the target than the partnerδ (t)< 0, but closes the distance gap as the main participant gains speed and catches up to the speed of thepartner (crosses δ˙ (t)= 0 in the upward direction). The main participant is briefly faster than the partner,but quickly slows down (crosses δ˙ (t)= 0 in the downward direction). In some of the cases, the mainparticipant immediately speeds up again with respect to the partner, suggestive of the main participant’ssecond attempt at accessing the target (see Case 1 (a) in Figure 4.10). In other cases, the main participantslows down to d˙1(t)= 0, with the hesitation motion segment ending with δ˙ (t)= 0−d˙2(t). This is shownas Case 1 (b) in Figure 4.10.55Case 2 Hesitation samples in Case 2 have a starting state of (δ (t)< 0, δ˙ (t)≤ 0). Often the value ofδ˙ (t) is near zero, indicating that the partner participant may be engaged in some other activity (e.g.,shuffling cards sideways on his/her side of the table) or remaining stationary when the main participantstarts to reach for the target (i.e., d˙1(t)∼ 0). Depicted in Figure 4.10 is the main participant locatedalmost equidistant to the target as the partner at the onset of the motion. Then, the main participantmoves faster and gets closer to the target than the partner when the partner starts his/her motion ormoves faster towards the shared target. Either by a brief speeding up of the partner’s motion or slowingdown of the main participant’s motion (hesitation), the partner moves faster toward the target beforethe main participant re-attempts to continue his/her reach. As illustrated in Figure 4.10 Case 2(a), theparticipants may go through multiple cycles in which one of the participants moves slower than the other,crossing δ˙ (t)= 0 downward. Alternatively, the partner may yield to the main participant as depicted inFigure 4.10 Case 2(b). Going through multiple cycles of the trajectory pattern is indicative of multiplere-attempts by the main participant to gain access to the target resource. The number of times thisre-attempt takes place is heretofore referred to as Re-attempts (RA).Case 3 Cases 3 is a variation of Cases 1 described above. In Case 3, the main participant reaches forthe target, but the speed toward, and subsequently, the distance to the target is overtaken by the partnerwho moves faster – perhaps despite having started his/her reach later than the main participant. Thisdynamics brings the pair’s trajectory to cross the δ˙ (t)= 0 state downwards once before crossing it againin the manner described in Case 1.Case 4 Likewise, Case 4 is a variation of Case 2, where the dyad crosses δ˙ (t)= 0 upward beforegoing through the same downward hesitation loop in Case 2. Case 4 illustrated in Figure 4.10 starts ata state where the main participant is closer to the target than the partner. However, the partner’s reachtoward the target overtakes the dynamics, until the main participant’s motion toward the target overtakesthe speed of the partner. Their speed and respective distances to the target lead them to eventually gothrough the same hesitation loop as described in Case 2.Understandably, because δ (t) and δ˙ (t) combine the effects of both the main and the partner partici-pants, the zero crossings of the system depend on both agents’ actions. Hence, generation of hesitationloops in HRI requires a translation of this trajectory pattern into a system in which only one of the agents’(robot’s) motions can be precisely controlled – needless to say, motions of the other agent (human’s)remain outside of our control.For the purpose of developing a trajectory generator that can recreate such hesitation dynamics inHRI, one can detect when such encirclement of δ˙ (t)= 0, such as detection of the zero crossing, occursand manipulate the robot’s reaching motion to produce the kickback motion (move backwards by a KD)that completes the loop. Herein, a δ˙ (t)= 0 occurrence is referred to as the Trigger State (TS) to designatethe state at which a hesitation behaviour of a robot is to be triggered.564.5 Design of the Negotiative Hesitation GeneratorTo investigate the efficacy of using negotiative hesitations in HRI as a mode of resolving resourceconflicts, interactive behaviours of a robot must be able to generate the elements of human hesitationbehaviours mentioned in the previous section. This artificial hesitation behaviour must also be imple-mented on a robot in such a way that the system can respond to human hesitations in real-time to allownegotiative nonverbal dialogues to emerge.Among other alternatives, the author chose to use the Linear Dynamical System (LDS) approach togenerate reference trajectories of artificial hesitations. A LDS is defined as the following:x˙ = Ax+b (4.1)In this differential equation, A is a matrix and b is a vector describing a vector field in the designatedstate space. These parameters represent the linear dynamics that exist between the state variable, x∈Rd ,and the rate of change expressed in x˙. Hence, iterative computation of the differential equation results inthe determination of the trajectory of the linear system through the state space. In robotics, x typicallyrepresents unambiguous states of a robot, such as joint angles. Therefore, computed states and statevelocities can be used as reference trajectories with which to control a robot.One of the reasons for choosing to implement artificial hesitation behaviours with an LDS is that theDynamical System (DS) approach offers reference trajectory generators that are reactive, real-time, androbust to disturbances or changes to the environment. In contrast, the author and colleagues’ previouswork used trajectories generated using splines [104]. The splines function as an interpolation betweentwo boundary states defined early in the motion trajectory. Hence, unexpected disturbances that push therobot out of its planned trajectory, for example, typically result in an undesirable, high-jerk movementof the robotic platform. This happens as the system tries to catch up to larger than planned differencebetween the state that has been planned for its next time step and the current, disturbed state of therobot. In comparison, trajectories that are generated using a DS iteratively uses the robot’s current statefor the computation of the desired next state. With a DS that converges to a target location, any suddendisturbances to the system only changes the value of the input state used to compute the desired nextstate of the system. Since the reference state for the next time step is computed based on the vectorfield, sudden changes to its input states are seamlessly handled. This highly reactive nature of DS-basedsystems has also been showcased in complex real-time obstacle avoidance tasks, as well as tasks thatinvolve catching flying target objects [79, 101].With the development of algorithms that can guarantee the stability of a DS during the training phaseof the system, such as Stable Estimator of Dynamical Systems [78], the DS has also been used to encodea variety of tasks for a robot as a teach-by-demonstration method [48, 63, 123]. Given its potential todevelop a large vocabulary of activities for robots to perform, implementing hesitation behaviours on aDS opens up the possibility for artificial hesitation behaviours to be triggered during tasks not envisionedor tested in this thesis. With this in mind, the author designed a DS-based trajectory generator that takesinto consideration point-to-point motions involved in reaching for a shared resource. Such motions are57not only foundational to many tasks, but they are also the main type of motion used in the type ofresource conflicts discussed in this thesis.Coincidentally, implementation of the TS – the state at which the controller should trigger a hesita-tion behaviour – and KD – the distance to retract as part of the hesitation behaviour – in a LDS yieldsa simple control architecture. As discussed in the previous section, for hesitation loops to be generatedin an HRI, one must remember that only one of the agents, the robot, is within the scope of our control.Moreover, δ (t) and δ˙ (t) represent ambiguous states that can be translated into an infinite combinationof states in the physical world.14 Therefore, rather than devising a control signal in δ (t) and δ˙ (t) space,one can monitor the state space to detect the TS in which an encirclement of the δ˙ (t)= 0 is desired.To do this, the author employed a simple LDS in d1(t) vs. d˙1(t) space (d˙1(t)= Ad1(t)) that can beused to reach for a target, and implemented an upper layer module that monitors for the TS in δ (t) vs.δ˙ (t) state space. Knowing the location of the shared target resource, T , the values of d1(t) and d˙1(t) canbe translated to Cartesian positions and velocities along a desired path of motion. This transforms theambiguous one-dimensional states, d1(t) and d˙1(t), to the unambiguous position and velocity M(t) and∇M(t).Once a TS is detected, the upper layer algorithm moves the equilibrium point, or the point of con-vergence, of the LDS by a KD. The devised system that produces reference trajectories using this controlregime is the Negotiative Hesitation Generator (NHG). To demonstrate that such a layered approachproduces the desired encirclement of δ˙ (t)= 0, this regime was simulated in MATLAB (The MathWorksInc., Natick, MA, USA) with quintic splines that represent two agents reaching for the same target.As shown in Figure 4.11, the generated reference trajectory from the simulation encircles the δ˙ (t)= 0observed from HH negotiative hesitations. The upper layer module was simulated using splines, sincethis was practical due to the off-line nature of the simulation. A spline-based implementation of the TSand KD for real-time, in-person HRI will require added complexities in the control architecture to ensureresponsiveness of the system without generating undesirable, residual motion. The system architectureto implement the NHG using a LDS is described in detail in Section 4.6.2.In the next section, Section 4.6, an online study is used to demonstrate that this particular reactivesystem can communicate hesitation to third party observers. In the following chapter, Chapter 5, theauthor determines the system’s efficacy in triggering negotiative responses from humans in in-personHRI.4.6 Study 4: Validating the Negotiative Hesitation GeneratorAlthough the design of the NHG is based on the observations of human motions from Study 3, theefficacy of the trajectories generated by the NHG has yet to be tested. Does the NHG produce humanlikehesitation responses? Does the LDS implementation of the elements of hesitation convey a level ofhesitation a human can observe? These are some of the grounding questions that need to be addressedto justify the use of the NHG for bidirectional, negotiative interaction between an HR dyad.14 δ (t) and δ˙ (t) are functions of relative Euclidean distances to the target. Hence, δ (t)=2 cm can mean one agent beinganywhere in a 2 cm radius from the target and the other being 4 cm away.58Figure 4.11: Simulation output demonstrating the NHG implementation using quintic splines. Twoquintic splines are generated to represent reach motions of two agents reaching for theshared target a fraction of a second (0.05 s) apart from each other. The main agent’s quinticspline is interrupted when a TS is detected. This triggers the main agent to move back byKD before reaching for the target again. This results in a circular motion in the δ (t) statespace (representing difference between the main and the partner participants distance totheir respective target locations) around (0, 0) of δ˙ (t)= 0.As outlined in Section 4.5, a premise of the NHG is in the presence of hesitation loops found inthe δ˙ (t) vs. δ (t) state space. When the state space is interpreted with the context of an in-personHRI, it allows one to understand the dynamic, spatial interaction between the two agents. In this statespace, the NHG calls for the implementation of the following three elements: TS, a state at which humanagents tend to hesitate in response to an imminent resource conflict; KD, the amount of distance thehesitating agent retracts as part of the hesitation behaviour before re-attempting to access the resourcein conflict; and re-attempts (RA), the maximum number of times the robot should re-attempt to accessthe conflicted resource before yielding. While TS refers to a state at which hesitation behaviour shouldbe triggered, implementation of an NHG involves assigning values of KD and RA, both of which areadjustable parameters. In the HH hesitation samples collected in Study 3 (Section 4.3), a range of valuesof both KD and RA were observed. Hence, within the range of observed parameter values, it is usefulto investigate what pairs of values for these parameters are acceptable for the NHG to generate artificialrobot motions that convey a state of hesitation.59In Study 4, presented in this section, the efficacy of a set of KD and RA parameter values are evaluatedagainst each other, in addition to a smooth stopping motion typically implemented in robotic collisionavoidance behaviours. The following hypotheses are tested:Hypothesis 4.1 Robot hesitation responses generated by the NHG are perceived to be more hesitantand persistent than a smooth stopping behaviour.Hypothesis 4.2 Robot hesitation responses generated by the NHG are perceived to be more animateand anthropomorphic than a smooth stopping behaviour.Hypothesis 4.3 Robot hesitation responses generated by the NHG are perceived to be more dominantand useful than a smooth stopping behaviour.Hypothesis 4.4 Robot hesitation responses generated with a human-inspired kickback distance (0 ≤KD < 19 mm) are perceived to be more hesitant than those with KD outside this range.Hypothesis 4.5 Robot hesitation responses showing a larger number of RA are perceived to be morepersistent, dominant and useful than those with a larger RA.Obtaining empirical support for these hypotheses would validate the use of the trajectory gener-ator as a mechanism that is adequate for creating artificial robot hesitation responses for in-personinteractions with humans. In addition, it would provide a measure of the dominance, animacy, and an-thropomorphic qualities perceived by human observers in the generated trajectories. These perceptionmeasures provide valuable information to consider in implementing robot behaviours for an improvedHRI experience. The remainder of this section outlines the experimental procedure used to conduct theonline survey (Section 4.6.1), the technical system used to implement the NHG (Section 4.6.2), and em-pirical and qualitative findings from the analysis (Section 4.6.3). Section 4.6.4 discusses implications ofthe findings in detail.4.6.1 Experimental ProcedureThis online survey-based study was structured to be similar to Study 3. The participants wererecruited via Amazon’s Mechanical Turk system to watch and report on their impression of a seriesof videos. All survey participants gave consent online by explicitly selecting the option “I consentto participate.” Upon giving consent, participants provided their demographical information (age andgender). Afterward, participants were asked to watch a short introductory video before proceedingwith the survey. This introductory video was made to provide the context of the interaction that theparticipants need to know to understand the videos that followed. The introductory video showed theexperimenter and the robot, facing each other and collaborating on an assembly task. In front of therobot were two liquid pumps, and between the two agents was a dispenser operated with a button ontop. Near the experimenter was a bin containing uncooked lentil beans of two different colours. Thefollowing text appeared at the bottom of the screen:60Figure 4.12: A screen capture of the introductory video shown to participants in Study 4. All othernon-introductory videos shown to the participants were captured from the same cameraangle, showing only the experimenter’s arm, the robot, and the experimental setup.You will be watching a person and a robot working on a collaborative task. Inthis task, the person sorts and places items onto the dispenser in the centre.Meanwhile, the robot helps by handling the liquid pumps, and flushing the dis-penser using the black button. Occasionally, the person and the robot reach forthe dispenser at the same time.Near the end of the video, the video showed the robot and the experimenter reaching for the dis-penser at the same time. See Figure 4.12 for a screen capture of this segment of the introductory video.The video faded to black before revealing who got the right of way.In the following pages of the survey, participants watched a total of fifteen videos. All videos startedwith the motion of the robot returning from the dispenser after successfully having reached and pressedit. Afterward, each of the videos showed the robot pressing on one of the liquid pumps, then reaching forthe dispenser again. But this time, the experimenter reached for the dispenser as well, creating a conflictof resource. The videos showed the robot responding to the conflict with either a smooth stoppingbehaviour, or a motion generated by the NHG with one of the fourteen parameter pairs (see Figure 4.13for the full list of conditions). For the purpose of testing Hypothesis 4.4, values 0, 10, 19, and 40 mmare used to represent the range of KD observed from Study 4. The KD value of 19 mm represents anapproximate average of the distribution of KD observed in human hesitation behaviours. The KD values0 and 40 represent the two extremes of the distribution.Ideally, the parameter pair would include five recordings for each value of KD, spanning RA valuesfrom 0 to 4. However, videos for the parameter set (KD= 10, RA= 4) and (KD= 40, RA= 0) were lostfrom our recording due to a technical failure. They could not be re-recorded due to logistical constraintsof the international collaborative research.15 In addition, the KD value of 0 inherently does not provide15 All of the videos were recorded at Ecole Polytechnique Fe´de´rale de Lausanne (EPFL) at a location and with an experi-mental setup that are no longer accessible to the author.61Figure 4.13: Outline of the conditions tested for Study 4. The structure of this study is a non-standard two-factor factorial design that has a control condition outside of the two factorsof interest. The conditions tests are marked with an ‘X.’Table 4.1: Internal reliabilities of the self-reported measures used for Study 3 are presented here.Hesitancy and Persistency measures only used a single item and their Cronbach’s alpha valuecannot be computed. All except the Usefulness measures used in this study have a Cronbach’salpha ≥ .70.Measures Cronbach’s ItemsalphaHesitancy N/A HesitantPersistency N/A PersistentDominance .92 Dominant, AggressiveUsefulness .66 Useful, EfficientAnimacy .78 Artificial, MechanicalAnthropomorphism .88 Artificial, Machinelikeany back and forth motion, rendering the value of RA meaningless. Given the lack of the re-attemptbehaviour, the robot to quickly retreat to its home position as soon as an imminent conflict is detected.The parameter set (KD = 0, RA = 0) serves as a reference point to study the main effects of KD and RAas is explained in more detail in the results section (Section 4.6.3). The smooth stop response is used asa control condition with which to contrast all NHG-based trajectories.Presented below each video were a set of questions that echo the questionnaire used in Study 3.The first nine questions were designed to collect six self-report measures: three questions for measuringAnthropomorphism and Animacy, derived from the Godspeed Questionnaire [14]; four questions formeasuring Dominance and Usefulness, two perceived team measures from Moon and Nass [107]; andHesitancy and Persistency measures.16 Table 4.1 presents the questionnaire items used for this study.Findings of Study 3 suggest that responses to Persistency and Hesitancy may not be enough todescribe the quality of motion the participants pay attention to. Therefore, a textbox was included tocollect optional, open-ended qualitative responses to the question “What adjectives would you use to16 The Hesitancy and Persistency measures were used in Study 3 with word pairs, (Hesitant – Not hesitant), and (Persistent– Not persistent). The same word pairs were used in this study.62describe the motion of the robot observed in the video? (Optional)”.It was also made explicit that participants could replay the video as many times as they liked tocomplete the survey. Each video was 16 to 22 seconds in length. All participants watched all fifteenvideos in random order. Based on a pilot study, the total time to complete the survey was estimated tobe less than 30 minutes. Participants were offered a financial reward of $1.60 USD for completing thesurvey.4.6.2 Technical ImplementationThe experimental system consisted of a 7-DOF KUKA LWR 4+ (KUKA Robot Group, Augsburg,Germany) manipulator controlled using the Fast Research Interface (FRI) at a 1 kHz frequency, and anOptiTrack (NaturalPoint Inc. DBA OptiTrack, Corvallis, OR) motion capture system operating at a 120Hz sampling rate. The motion capture system was used to track markers on the experimenter’s hand andforearm.In all of the conditions, the robot’s task involved reaching and retraction motions to and from thetwo liquid pumps and the dispenser button. Depending on the experimental condition, the robot usedeither a Hermite quintic spline or the NHG to generate the reaching motion to and from the dispenser asfurther described below. For all conditions, once the robot reached the dispenser button or one of thepumps, it was programmed to press on the button/pump using a Hermite quintic spline. This ensuredthat all aspects of the generated robot motions remain equal for all conditions except for the reachingmotion and the robot’s response to conflicts.In the fourteen conditions that involve values of KD and RA, the LDS implementation of the NHGwas used to generate reference Cartesian position trajectories of reaching motions. The robot was pro-grammed to re-attempt (i.e., reach for the resource again immediately after returning to the startingposition) a maximum of four times before giving up its access to the resource and resuming its task oftending to the liquid bottles.In the Smooth Stop (control) condition, the robot’s reaching motion was generated using a Hermitequintic spline. This technique takes the start and end states of the robot and the desired travel completiontime to interpolate between the two points in a minimum-jerk manner. The resulting trajectory allowsfor the robot to trace smooth reference trajectories considered humanlike [49]. When the robot detectsa conflict near the onset of its reach motion, it exhibited the R-type hesitation using the AHP. Whenthe conflict was detected after the time window t1 (see Figure 4.2), it was programmed to remain at itscurrent location, leading it to come to a full stop and hold its position for 1 second before retreating.Since this condition does not employ the notion of the TS used in the NHG to determine the state at whichthe robot should trigger a conflict response behaviour, the system was programmed to use proximity anddirection of motion between the robot and the experimenter to detect possible HR resource conflict states.When the robot and the experimenter moved toward each other and were closer than 5 cm from eachother, the robot was programmed to remain at the current location such that it comes to a full stop. After1 second of pause in the stopped state, the robot retracted back to its home position.63System ArchitectureThe system architecture employed for this study was implemented in Robot Operating System (ROS)[126]. This section provides a high-level overview of the architecture. Three main components werecustom developed for this experiment:1. a communication node dedicated to receiving, transforming, and publishing sensed informationfrom OptiTrack,2. an action server that registers and maps requested actions from action clients onto a set of prede-fined action policies that ultimately generate desired reference trajectories, and3. a client action node that acts as the master script to trigger different robot actions at appropriatetimes.First, the data stream from OptiTrack was broadcast through a local area network and converted intoa data format appropriate for further use of the data in ROS. The ROS action client-server protocol wasemployed to interface the trajectory generator algorithms with the FRI interface controlling the robot.The experiment’s main action client acted, in part, as a server for smaller action clients. The main clientis responsible for the continuous monitoring of human motion captured via OptiTrack. When initialized,it launches a dedicated node to receive and convert the stream of OptiTrack data into a local coordinatesystem as well as Euclidean distance to the shared resource (dispenser). The final ROSTOPIC containinghuman motion information is ultimately communicated to the action server to trigger different trajectoryresponses for the robot. The smaller action clients trigger various behaviours, such as reach or retract,to the main client, which in turn requests actions to be serviced by the action server. The action server,upon receiving a request, generated and communicated the reference trajectories for the FRI controlinterface to actuate the robot.4.6.3 ResultsA total of 50 people participated in the study (16 females, 33 males, 1 preferred not to disclose).On average, it took approximately 18 minutes for the participants to complete the survey (M =18:27,SD =9:03). The age of the participants ranged from 21 to 52 (M = 32, SD = 7.0).As mentioned in Section 4.6.1, this experiment was structured as a non-standard, two-factor, facto-rial design where variables KD and RA serve as the two crossed factors. The five levels (0, 1, 2, 3, 4)of RA are factored with the four levels of KD (0, 10, 19, 40). Two elements of this experiment structuremake this factorial design non-standard. First, none of the values of KD and RA apply to the controlcondition, requiring the analysis to include a control group outside the KD-RA factorial. Second, thereare some combinations of KD-RA that are missing from the experiment due to the reasons describedearlier. While this is not ideal, subjecting the participants to a smaller number of conditions shortenedthe length of the within-subjects study, thereby avoiding potential effects of pencil-whipping.1717 Pencil-whipping within the context of survey-based studies refers to the tendency of participants to become carelessor provide answers to questions without giving much thought, often due to boredom, and produces undesirable noise in thedataset.64To test the aforementioned hypotheses, the responses to the six measures from the survey are an-alyzed using a regression analysis with a linear mixed effects model. A Multi-level Modelling (MLM)using the lme4 package in R with Restricted Maximum Likelihood (REML) was used for model fitting.Given that the structure of this study is a non-standard two-factor factorial design (see Figure 4.13 forthe structure of the conditions), a more familiar statistical modeling such as ANOVA cannot be used.On the other hand, MLM allows for data from such non-standard study designs to be captured such thatinferential statistics can still be conducted. Keeping in mind the repeated measures aspect of this study(all participants saw all 15 videos), Participant was used as a random factor and KD and RA as fixedfactors. In this analysis, the mean responses to the (KD=0, RA=0) pair, µ0, is used as the reference pointwith which the responses to the remaining parameter pairs are modelled:(KD = 0,RA = 0)∼ µ0 (4.2)Smooth Stop∼ µ0+ τSmooth Stop (4.3)(KD = α,RA = 0)∼ µ0+KDα (4.4)(KD = α,RA = β )∼ µ0+KDα +KDα : RAβ (4.5)Here, τSmooth Stop refers to the difference in survey response between the Smooth Stop and the (KD=0,RA=0) condition; KDα refers to the main effect of KD = α ∈ [10, 19, 40] compared to KD=0 where αis one of the non-zero KD levels tested; RAβ refers to the main effect of RA= β ∈ [1, 2, 3, 4] comparedto RA=0, where β is one of the non-zero RA levels tested; KDα : RAβ is the interaction effect betweenKDα and RAβ . This approach allows us to identify the effects of the NHG conditions against the controlcondition while taking into account the main effects of KD and RA.The internal reliability of the measures was computed for all of the standardized questionnairesused. All except the Usefulness measure have a Chronbach’s α > .70 (see Table 4.1). Therefore, theUsefulness measure is excluded from the analysis.All of the statistical analyses presented below use a significance level of α = .05. The Tukey methodwas used to conduct post-hoc analyses. The presentation of the results below is organized according tothe questionnaire measures. Table 4.2 provides a summary of the findings.HesitancyResults from the MLM analysis on Hesitancy demonstrate that the NHG-generated robot motions areperceived to be significantly more hesitant than that of the control condition (F(1, 692) = 30.4, p <.001), providing support for Hypothesis 4.1. It also shows that there are significant differences in Hes-itancy across the different values of KD (F(3, 692) = 3.98, p < .01) and RA (F(4, 692) = 7.80, p <.001). These main effects of KD and RA can be interpreted independently since the interaction effectbetween the two is negligible (X2(6) = 10.5, p = .11).1818 This Chi-square result reflects the difference between the two regression models built to predict the same responsevariable – in this case, Hesitancy. For all of the response variables discussed in this section, a regression model that includesan interaction effect between KD and RA (Equation 4.5) is compared against a model that does not. A significant finding in theChi-square result from this analysis indicates that the interaction effect of KD and RA plays a significant role in predicting the65Table 4.2: Findings from Study 4 are summarized here in the order of the hypotheses tested.Measure Result ConclusionHypothesis 4.1 Robot hesitation responses generated by the NHG are perceivedto be more hesitant and persistent than a smooth stopping behaviour.Hesitancy (F(1,692) = 30.1, p < .001) SupportedPersistency (F(1,686) = 16.3, p < .001) SupportedHypothesis 4.2 Robot hesitation responses generated by the NHG are perceivedto be more animate and anthropomorphic than a smooth stopping behaviour.Animacy (F(1,692) = 11.9, p < .001) SupportedAnthropomorphism (F(1,692) = 10.4, p < .01) SupportedHypothesis 4.3 Robot hesitation responses generated by the NHG are perceivedto be more dominant and useful than a smooth stopping behaviour.Dominance Significant effect in the opposite direction Not supportedUsefulness Internal reliability not sufficient InconclusiveHypothesis 4.4 Robot hesitation responses generated with a human-inspiredkickback distance (0≤ KD < 19 mm) are perceived to be more hesitantthan those with KD outside this range.Hesitancy Not supportedHypothesis 4.5 Robot hesitation responses showing a larger number of RA are perceived to bemore persistent, dominant and useful than those with a larger RA.Persistency Significant KD & RA interaction effect InconclusiveDominance Significant KD & RA interaction effect InconclusiveUsefulness Internal reliability not sufficient InconclusivePost-hoc analyses across the levels of KD indicate that robot motions generated with KD=0 areperceived to convey a significantly higher level of Hesitancy than those with KD=10 and KD=19 (seeFigure 4.14). This is interesting given that the robot does not perform any re-attempts in the (KD=0,RA=0) condition – the only condition with a value of KD= 0 – and, therefore, is the least negotiativeof all NHG conditions. This result does not support the Hypothesis 4.4 that robot hesitation responsesgenerated with a human-inspired kickback distance (0≤KD< 19 mm) are perceived to be more hesitantthan those with higher values of KD. Rather, this result demonstrates that there are significant differenceson Hesitancy among the human-inspired values of KD. In addition, the results suggest that Gendersignificantly affects the perception of Hesitancy (F(1, 692) = 6.50, p < .01). Across the conditionstested, female participants rated robots to be more hesitant than males (Z =−2.89, p < .01).response variable.66Kickback Distance40mm19mm10mm0mmSmooth StopMean Hesitancy210-1-2Error Bars: 95% CI4321None# of ReattemptsFigure 4.14: The perceived Hesitancy scores (a 5-point Likert scale) collected for the differentlevels of KD and RA. The scores are higher for the case of KD= 0 than most of the otherNHG-generated motions. Analyses of the scores across all NHG-generated motions suggestthat there are significant differences in Hesitancy among the human-inspired values of KD.PersistencyThe analysis of Persistency scores suggests that there is a significant interaction effect betweenKD and RA (X2(6) = 12.65, p < .05). This indicates that the effects of factors KD and RA cannot beinterpreted independently from each other. Hence, the results do not offer a straight-forward support forHypothesis 4.5 that robot hesitation responses showing a higher number of RA are perceived to be morepersistent than those with lower RA values.Nonetheless, the analysis provides an empirical support that the NHG-generated robot motions areperceived to be significantly more persistent than smooth stop motions (F(1, 686) = 16.3, p < .001).As shown in Figure 4.15, the Persistency score is higher for KD= [10, 19] for most non-zero RA valuesthan that of the control condition or the reference (KD=0, RA=0) condition, although this is not true forall values of RA.Gender is not found to be a significant factor of Persistency (X2(2) = 4.87, p = .09).Animacy and AnthropomorphismResults of the analysis on Animacy and Anthropomorphism suggest that the NHG-generated motionsare perceived to be more animate and anthropomorphic than the smooth stopping motions in the control67Kickback Distance40mm19mm10mm0mmSmooth StopMean Persistency210-1-2Error Bars: 95% CI4321None# of ReattemptsFigure 4.15: Perceived Persistency of different KD by RA values. There is a significant interac-tion effect between KD and RA. Nonetheless, NHG-generated motions are perceived to besignificantly more persistent than the motion implemented in the Smooth Stop condition.condition (Animacy: F(1, 692)= 11.9, p< .001; Anthropomorphism: F(1, 692)= 10.4, p< .01). Thissupports Hypothesis 4.2. No significant interaction effect between KD and RA is found for either of themeasures (Animacy: X2(6)= 6.24, p< .40; Anthropomorphism: X2(6)= 5.21, p= .52). No significantGender effect is found in either of the measures (Animacy: X2(2) = 2.54, p = .28; Anthropomorphism:X2(2) = 4.28, p = .12).The results of the analysis on Animacy and Anthropomorphism, along with Hesitancy mentionedabove, provide strong support for Hypotheses 4.1 and 4.2 that robot conflict responses generated by theNHG are perceived to be more humanlike hesitations than the smooth stopping behaviour typically usedin industries. Figure 4.16 presents perceived Animacy, and Figure 4.17 presents perceived Anthropo-morphism scores.DominanceResults of the regression analysis on Dominance provide evidence rejecting Hypothesis 4.3. Con-trary to the hypothesis, the results suggest that the NHG-generated motions are perceived to be signifi-cantly less dominant than the smooth stopping motions of the control condition (F(1, 686) = 35.2, p <.001). The results also demonstrate that the male participants rated the robot’s motions to be more dom-inant than the females did (X2(2) = 6.55, p < .05). In exploring the effects of KD and RA on perceived68Kickback Distance40mm19mm10mm0mmSmooth StopMean Animacy2.01.00.0-1.0-2.0Error Bars: 95% CI4321None# of ReattemptsFigure 4.16: Perceived Animacy of different KD by RA values. The NHG-generated behaviours are,in general, perceived to be more animate than that smooth stopping behaviour demonstratedin the Smooth Stop condition.dominance of NHG-generated motions, the results find that there is a significant interaction effect be-tween the two variables (X2(6) = 31.6, p < .001). Hence, the effects of factors KD and RA cannot beinterpreted independently from each other, and support for Hypothesis 4.5 remains inconclusive. SeeFigure 4.18 for the distribution of the Dominance measure.Qualitative FeedbackThe adjectives the survey participants optionally submitted were collected for qualitative analysis.Keeping in mind the principles of grounded theory, two experimenters coded the adjective entries andextracted emerging themes with respect to the above-mentioned hypotheses. The responses capturedusing an open text field were mostly adjectives, as requested in the questionnaire. However, some par-ticipants included full sentences. The author included these responses in the analysis. Since these entrieswere not a mandatory part of the survey, only 50% of the participants provided at least one adjectivethroughout the survey. Most of these participants submitted responses to four or fewer conditions. Giventhe small number of responses collected, this qualitative analysis serves the purpose of complementingthe quantitative findings discussed above.The types of adjectives that participants used fell into approximately three categories: adjectives de-scribing the perceived expression or conveyed personality of the robot (e.g., hesitant, confident, scared),69Kickback Distance40mm19mm10mm0mmSmooth StopMean Anthropomorphism2.01.00.0-1.0-2.0Error Bars: 95% CI4321None# of ReattemptsFigure 4.17: Perceived Anthropomorphism of different KD by RA values. Echoing the resultsof Animacy, NHG-generated motions are perceived to be more anthropomorphic than thesmooth stopping behaviours of the robot in the Smooth Stop condition.motion quality of the robot (e.g., smooth, jerky), perceived anthropomorphism (e.g., humanlike, calcu-lated, artificial), and perceived performance of the robot (e.g., efficient, smart). In all of the conditions,one of the words ‘hesitant’ and ‘persistent’ was mentioned by at least one participant in all but condition(KD=40, RA=4).When the participant responses to Smooth Stop and (KD=0, RA=0) are compared, both of the robotmotions are described as hesitant. However, whereas the Smooth Stop condition is described with wordshinting at low anthropomorphism (e.g., awkward, cautious, mechanical and robotic), motion quality of(KD=0, RA=0) is described with positive adjectives such as safe and smart. The Smooth Stop conditionis described as expressing a sense of dominance (e.g., aggressive and assertive), whereas the (KD=0,RA=0) condition is described to be apprehensive and wary, despite being responsive (e.g., pushover,scared, submissive). Based on the quantitative results mentioned above, the (KD=0, RA=0) conditionreceived the lowest mean Persistency and Dominance scores, while scoring highly on the Hesitancymeasure. The qualitative description of the robot’s motion in this condition supports and complementsthese results. The Smooth Stop condition also received the lowest Animacy and Anthropomorphismscores, which is corroborated in the qualitative finding.In comparison, the participants perceived the NHG-generated motions with KD=10 and 19 to beefficient, calculated, and cautious. The robot motions with KD=19, in particular, were described as more70Kickback Distance40mm19mm10mm0mmSmooth StopMean Dominance2.01.00.0-1.0-2.0Error Bars: 95% CI4321None# of ReattemptsFigure 4.18: Perceived Dominance of different KD by RA values. Contrary to Hypothesis 4.3,NHG-generated motions were perceived to be less dominant than the smooth stopping alter-native. There is an interaction effect between KD and RA as demonstrated by the inconsistentdistribution of Dominance across the two and aware across the RA values, except for RA=4, which was described to be submissive andunsure. The NHG-generated motions with KD=40 were described as timid and shy. This affirms the lowDominance scores collected for these conditions. In addition, KD=40 motions with lower RA valueswere perceived as compromising and slow, while those with higher RA values were seen as scared.4.6.4 DiscussionThe purpose of this study was to validate the efficacy of the NHG as a mechanism to generatehuman recognizable robot hesitation behaviours, and to establish the effects of KD and RA on humanperception of the generated motions. The main finding of this study is that the NHG-generated motionsare indeed distinguished from smooth stopping behaviours of a robot. The motions produced by theNHG were perceived to be more hesitant and persistent. They were also observed to be more animateand anthropomorphic. These results provide empirical support for using the NHG for in-person HRIwhere human behavioural responses to the artificial hesitations can be observed.However, there were also surprising findings. First, counterintuitively, the (KD=0, RA=0) conditionreceived the highest Hesitancy score. The qualitative understanding of the participants’ perception ofthe motion offers some explanation for this finding (i.e., the robot motion with this parameter setting71was considered to constitute a highly apprehensive and hesitant behaviour). Moreover, this conditionreceived the lowest Persistency score. This is consistent with the concept of persistency established inSection 4.2, in which a higher value of RA would be considered to exhibit a higher Persistency. Thisinterpretation does not generalize across all levels of KD, as demonstrated by the significant interactioneffect of KD and RA observed in the Persistency scores. Given that this condition did not include anyre-attempt behaviours and, therefore, was non-negotiative in nature, it does not make sense to use thisset of parameters to investigate in-person HR non-verbal negotiations. On the other hand, the human-inspired KD values, 10 and 19, received high Persistency scores across the various RA values. Moreover,taking into account variabilities by RA values, KD=19 received a relatively high level of Hesitancy incomparison to other values of KD in general. This also coincides with the average KD value discovered inHH hesitations (see Section 4.4.5), and confirms the efficacy of using this value for generating artificialhesitations with the NHG.Second, it was unexpected to find that, contrary to Hypothesis 4.3, the motions produced by theNHG are perceived to be less dominant than that of the smooth stopping motions. This is surprisingbecause the smooth stopping motion employed in this study only had a short pausing behaviour beforethe robot retracted back. This is in contrast to motions generated by the NHG with non-zero RA valuesthat exhibited multiple back-and-forth motions that could be perceived as more dominant than a pause.In observing the distribution of Dominance scores, in general, motions with a high RA value are seen tobe less dominant than those with a lower RA – although this trend is not significant. This result suggeststhe possibility that the resulting negotiative hesitation behaviours designed to exhibit multiple RA areperceived to be timid and apprehensive.4.7 ConclusionThe work presented in this chapter establishes a framework with which one can investigate the rolerobot negotiative hesitations can play in an interactive HR resolution of resource conflicts.To this end, naturally occurring negotiative hesitations in humans were collected from an HHI ex-periment (Study 3). A video-based online study involving 300 participants helped rate each sample ofhesitations on its expressed level of Hesitancy and Persistency. Exploration of the numerous featuresof the human hesitation trajectories revealed that, unlike R-type hesitation behaviours investigated inthe author’s previous work, negotiative hesitations take into consideration motions of both agents. Thisled to the discovery of hesitation loops, which are trajectory patterns found in the majority of the ne-gotiative hesitation samples collected in Study 3. The design of the NHG was inspired by the hesitationloop trajectory patterns. It was designed such that the change in the relative distance between the twointeracting agents to the shared target is used to determine when the robot should trigger its hesitationbehaviour.In implementing the NHG for in-person HRI, it is important that the trajectory output of the NHGbe established as communicative of hesitation. Study 4 helped validate the efficacy of the NHG-basedhesitation behaviours. It provides empirical evidence that robot motions generated by the NHG areperceived to be more hesitant, animate, and anthropomorphic than those of a smooth stopping behaviour72of the same robot. In addition, it helped better understand the relationship various parameter values ofthe NHG have on human perception of the robot motions generated. For example, the results suggest thata KD value of 19 mm, an approximate average of the KD values found in human negotiative hesitations,is acceptable for producing communicative hesitation gestures.Taken together, this chapter contributes to a better understanding of negotiative hesitations naturallyfound in HHI. It also provides a validated trajectory generator with which one can explore the dynamicsthat could emerge as HR nonverbal negotiation of resource conflicts. In the following chapter, the authortakes the first step in investigating such in-person HRI using NHG-generated robot hesitations.73Chapter 5Study 5: Bidirectional Interweaving ofSubplans using Negotiative Interaction inHuman-Robot Collaborative Assembly5.1 IntroductionIn Chapter 3 the author addressed the unidirectional perspective of how robots exert influence onour behaviours in a robot-to-human handover context. Chapter 4 presented a novel mechanism, theNHG, that permits exploration of the dynamics of bidirectional interweaving (negotiation) of subplansbetween an HR collaborating dyad. The results of a video-based online survey, Study 4, suggested thatthe artificial robot hesitation behaviours generated using the NHG are observed as expressing a state ofhesitation. This provides crucial building blocks for investigating the second research question of thisthesis: “Can a robot nonverbally negotiate with a person about what should happen in an interaction?”Study 5 presented in this chapter builds on the previous chapter’s investigation of hesitation to addressthis question. The main focus of this in-person study is to investigate whether the generated hesitationbehaviour of a robot can be used in an HR collaboration to interactively negotiate for access to a sharedresource with a human user in real-time.Given the numerous studies on nonverbal HRI – including the results of Study 2 (Chapter 3) –that suggest that robot motions can influence human behaviours, there are reasons to believe that theanswer to the research question is trivial and affirmative. However, previous work in HRI also providesevidence of interactions where human users dominate and jeopardize the efficacy of the interaction, oreven obstruct the robot from performing its task.1 For example, in a study involving the use of robotsin a mall environment in Japan, Brscic´ et al. [25] found that the robots were subject to both verbal andphysical harassment by children. In order to prevent physical damage to the robots, the researchersdevised a novel path-planning algorithm that preemptively prevents robots being abused by children.1 Brscic´ et al. [25] defined the term robot abuse to refer to the increasingly documented “persistent offensive action, eitherverbal or nonverbal, or physical violence that violates the robot’s role or its human-like (or animal-like) nature.”74A previous in-person experiment conducted by the author and her colleagues [105] also providesa relevant example. In the experiment, an HR pair engaged in a collaborative assembly task in whichthey often coincidentally reached for the same resource at the same time. The robot was programmedto respond to the resource conflict in one of three ways: ignore and risk the possibility of colliding withthe user; exhibit an R-type hesitation behaviour and retract; or trigger an emergency-like stop responsebefore retracting. Results from the study showed that when the robot immediately yielded to the user(as was the case in the conditions involving hesitation or emergency-like stopping behaviours), the HRpair took significantly longer to complete the assembly task than when the robot ignored and sometimeseven collided with the user to accomplish the task.2 These studies suggest that timid behaviours in arobot may not be useful or desirable in HRI, and that such interactions with a robot, over time, couldcause the robot to be ignored by users.There are also reports about robots deployed outside of laboratory environments that echo this trend.For example, Dietmar Exler, the current Chief Executive Officer of Mercedes-Benz USA, noted thatbullying of self-driving vehicles by humans is one of the hindrances when deploying such vehicles ontothe roads [100]. He remarked that human drivers may take advantage of autonomous, safety-prioritizingdriving patterns. This concern is shared by others in the automotive industry. Volvo, for example, statedthat the self-driving vehicles to be pilot tested on the streets of London in 2018 would not be flagged asself-driving because of this concern [38]. This decision by Volvo came after the release of a report bythe London School of Economics and City University of London on public perception of, and attitudestoward, self-driving vehicles [152]. The report indicates that drivers who have a combative style ofdriving also feel that they could take advantage of self-driving vehicles. This potential for humans tosuboptimally share the road with self-driving vehicles results from the human expectation that thesevehicles will strictly adhere to safe driving practices at all times, and will always yield to drivers whocut them off.Hence, if we wish to see a future in which robots share our physical world in friendly and efficientways, it is important to investigate novel behaviours that robots can use to proactively address conflictsabout spaces and resources shared with humans. One possibility is to design communicative behavioursthat robots can use to negotiate for their right of way or access to a shared resource.As demonstrated in Chapter 4, humans use hesitation behaviours to interactively negotiate for so-lutions to spontaneously occurring resource conflicts. The description of the NHG captures aspects ofthese human hesitation characteristics. The results of Study 4 helped validate the NHG’s effectiveness ingenerating artificial hesitation motions for a robotic arm. Building on the findings from Chapter 4, thischapter presents a within-subjects study, Study 5, that contributes to a better understanding of hesitation-based HR negotiations. Study 5 consists of an in-person HRI experiment in which human participantsinteracted with a robot in an HR collaborative assembly task. Each participant performed the same taskfour times, twice with a robot responding to the resource conflicts with artificial hesitations (the Nego-tiate condition), and twice with the robot smoothly stopping to avoid collision with the participant (theStop condition).2 Safety measures were taken for this experiment such that collisions with the robot were harmless.75The following hypotheses were tested:Hypothesis 5.1 Robot conflict responses generated by the NHG are perceived more favorably than thesmooth stopping behaviour often used in industries.Hypothesis 5.1a Smooth stopping behaviours are perceived as less animate, anthropomorphic, andlikeable than hesitation responses generated by the NHG.Hypothesis 5.1b Responses generated by the NHG are perceived to be less dominant, more emotionallysatisfactory, and more useful than a smooth stopping behaviour.Hypothesis 5.2 Hesitation responses generated by the NHG enable faster completion of the task than asmooth stopping behaviour.Hypothesis 5.3 Hesitation responses generated by the NHG do not threaten perceived safety, nor jeop-ardize the actual safety, of the interaction in comparison to a smooth stopping behaviour.Hypothesis 5.4 Hesitation responses generated using the NHG will lead to a faster resolution of re-source conflicts with humans than can be achieved with a smooth stopping behaviour.Continuing the discussion of the literature on hesitation, Section 5.1.1 provides a literature reviewon hesitations observed as behavioural responses exhibited by humans and animals. Subsequently,Section 5.2.1 outlines the experimental task in more detail, including a pilot study that was conductedto inform the experiment design (Section 5.2.3). Following the experimental procedure outlined inSection 5.2, technical details of the robotic system are presented in Section 5.3. Findings from thisstudy are discussed in Section 5.4. These findings not only support the hypothesis that nonverbal HRnegotiation is possible using this paradigm, but also provide evidence that such negotiative interactioncan improve the performance of the HR team without jeopardizing the perceived or actual safety ofthe human. The implications of the results are discussed in Section 5.5, followed by a conclusion inSection BackgroundThe discussion of hesitations in Section 4.2 establishes that verbal and nonverbal expression ofhesitation behaviours have been studied in psychology as behaviours closely linked with uncertainty orindecision. It also suggests that observing human hesitations and expressing analogous robot hesitationsare increasingly becoming useful in the study of HRI. This section extends the discussion on hesitationby focusing on the nature of nonverbal hesitation behaviours and their relationship to the stimuli thattrigger them.Hesitations in the form of a kinetic dialogue are found in primitive animals as well as humans. Forexample, Levin [87] studied the approach-withdrawal behaviour of fish to investigate how fish decide toflee, pursue, or ignore something when the difference between prey, predator, and the natural environ-ment is difficult to perceive. While a fish should pursue its prey, it must flee from its predators in order76to survive. Levin [87] reports that fish stop swimming in uncertain cases. Analogous hesitations in hu-mans have also been studied as a kinetic response to situations that require either action or cessation ofaction. Netick and Klapp [114] observed hesitation behaviours in human participants engaged in a con-tinuous tracking task. In investigating what behavioural processes are involved in hesitation behaviours(which they define as “a reaction in which an ongoing action is halted”), Netick and Klapp posit thathesitations are a response to a stimulus, rather than a part of the strategic processing of the stimulus.They theorize that rather than a strategic and controlled response to a stimulus, kinetic hesitations arereflexive behaviours that occur when attention is diverted from an on-going task, thereby interruptingcontrol over the task, which results in the freezing of an action.3This theory of the reflexive nature of human hesitation is crucial to the understanding of the rolethat hesitation-based kinetic negotiations can play in HRI. First, it presents hesitations as a universalbehaviour that can be used to kinetically resolve conflicts between an HR pair over a shared resource.Second, having a hesitation-based kinetic dialogue may be intuitive even for a naı¨ve user, given thathesitation is a universal response that can be observed and understood by humans even from the motions(or lack thereof) of primitive animals such as fish.Moreover, the reflexive nature of hesitation responses makes hesitation a promising technique toenable a robot to kinetically engage in a dialogue with humans without its actions being undesirablydominated – or bullied as described in the introduction. For example, in Austin, Texas, a cyclist and anautonomous car unintentionally engaged in a nonverbal dialogue as they negotiated each other’s rightof way at a four-way stop [40]. The car had stopped slightly earlier than the cyclist and had the right ofway. The cyclist recounts:“The car remained motionless for several seconds and I continued to maintain my balancewithout moving. As the car finally inched forward, I was forced to rock the handlebars tohold my position. Apparently, this motion was detected by one of the sensors and the carstopped abruptly. I held the bike in balance and waited for another several seconds and thecycle repeated itself ... the car inched forward, I shifted my weight and re-positioned the barsand the car stopped. We did this little dance three separate times and the car was not evenhalfway through the intersection.”In this incident, the system was confused about the state of the cyclist who balanced his bicycle byrocking the handlebars rather than putting his foot down on the ground. It is not an unfamiliar experiencefor most drivers to have nonverbal dialogues such as this one at an intersection that somehow result in aresolution of the conflict using purely nonverbal means. This example suggests that equipping a robotwith an ability to negotiate its right of way using hesitation behaviours could lead to the development ofa natural and interactive mode of conflict resolution for HRI.3 To support this claim, Netick and Klapp [114] provide evidence that electromyographic recordings of human hesitationsfrom their study resemble the patterns of electromyographic signals that are expected in active cessation of movement.775.2 Experimental ProcedureThis section outlines the procedures of a within-subjects HRI experiment conducted in a controlledlaboratory environment to examine the impact that negotiative robot behaviours have on real-time HRcollaboration. This project was conducted as a collaboration between the Learning Algorithms andSystems Laboratory (LASA) at EPFL and the Collaborative Advanced Robotics and Intelligent Systems(CARIS) laboratory at UBC. The experiment itself was conducted at LASA, EPFL, Lausanne, Switzerland.The devised experimental context consisted of a collaborative task requiring both the participantand a robotic manipulator to frequently access a shared resource. The author conducted a pilot study,consisting of the Yield and Negotiate conditions, to fine-tune the HR collaborative task devised for thisstudy. The pilot study also served to inform the design of a control condition, the Stop condition, suchthat the negotiative robot behaviour, the Negotiate condition described below, is compared against anon-trivial alternative. This selection process also helped minimize the number of tested conditions totwo in the full study, thereby limiting the duration of the experiment to an hour and minimizing possibleeffects of fatigue.Participants were recruited from within and outside of the EPFL campus and were offered 20 CHFcompensation for their participation. At the beginning of the experiment, participants provided informedconsent and were asked to fill out a short demographics questionnaire on age, gender, dominant hand,and their familiarity in working with robotic arms. They were then introduced to the 7-DOF roboticarm, KUKA LWR 4+ (KUKA Robot Group, Augsburg, Germany), and the experimental task to beperformed in collaboration with the robot. Afterwards, participants sat facing the robot arm with a smalltable between them. Figure 5.1 illustrates the HR workspace setup. The experimenter placed clusters ofreflective OptiTrack markers on the participants’ dominant forearms and hands to track the participants’motions in real-time. The entire experiment was video recorded. It was also time-synchronized withhuman and robot trajectory recordings via integration of the OptiTrack system and the robot’s jointtrajectory readings.5.2.1 Experimental TaskThe participants were told that the goal of the experimental task was for the participant and therobot to collaborate with each other to create a small concoction. The concoction was to consist oftwo different liquids4 and orange-coloured dry lentil beans. The participant’s job was to sort orange-coloured lentils from a mix of orange and green lentil beans in a bin and place them onto a dispenserin front of them. The robot’s job was to pump the two liquids in a specific order into a transparent cup,and also to press the knob in the centre of the dispenser so as to flush the orange lentils into the cup.The participants were told that the robot had a fixed number of pumps it needed to perform before theconcoction would be considered complete, and that the team’s performance would be measured basedon how many orange lentils were in the concoction at the end of the trial, as well as on how quickly theteam finished the task. They were also told that incorrect-coloured lentils in the concoction would be4 Coffee and water were used, although not specified to the participant.78Figure 5.1: Experiment set-up of Study 5. The robot and a participant sat facing each other witha dispenser mechanism located on the table between them. The shared use of the dispenserallowed spontaneous HR resource conflicts to occur throughout the experiment.counted as a penalty for the team. The dispenser was intentionally made of transparent material so thatparticipants could monitor the progress of the task based on the amount of liquids and solids depositedinto the cup. This also served to provide a visual confirmation that the lentil sorting by the participant,and the liquid pumped by the robot, came together as one final product.Each trial started with the robot reaching for the dispenser and pressing on the dispenser knobto demonstrate the robot’s reaching motion. The participants were asked to start immediately afterthis demonstration. Meanwhile, the experimenter sat diagonally behind the participants, out of theirimmediate field of view to ensure that the experimenter would not cause unintentional visual bias. Thisarrangement allowed the experimenter to observe the workspace continuously, so as to be able to triggeran emergency stop if necessary.The two experimental conditions, described in detail in Section 5.2.4, were randomized in such away that each participant encountered the two conditions in the first two trials of the experiment inrandom order. They also encountered the two conditions again in another random order in the lasttwo trials of the experiment. In total, each participant completed four trials of the experimental task.Figure 5.2 provides a visual overview of this procedure.5.2.2 Questionnaire and InterviewAt the end of each trial, the participants were accompanied away from the robot to fill out a ques-tionnaire. In order to avoid participant bias toward the robot in their reports (as reported as possible inReeves and Nass [130]), the room was configured in such a way that the participants faced away fromthe robot while filling out the questionnaire. This also served as a washout period during which theparticipant was distracted while the experimenter prepared the robot and the experimental workspace79Figure 5.2: Experiment procedure overview of Study 5. The two conditions (Stop and Negotiate)were randomly assigned to the trials for each session, such that the participants encounteredboth conditions at least once in Session 1 and once again in Session 2. The questionnairesbetween the trials served as a wash-out period. The transition between the two sessions werenot announced to the participants.for subsequent trials.The questions in the questionnaire were selected from two widely used questionnaires: 1) God-speed questionnaire [14] used to measure Animacy, Anthropomorphism, Perceived Safety, Likeability,and Perceived Intelligence and 2) Moon (no relation to the author) and Nass questionnaire [107] usedto evaluate human-machine teamwork including Dominance, Usefulness, and Emotional Satisfaction.Internal reliability values from a similar previous in-person experiment, Moon et al. [104], informed theselection of the most promising set of questions for the eight measures. These standardized user per-ception and teamwork measures helped collate and contrast the participants’ impressions of the robotacross the two experimental conditions.At the end of the fourth and last trial, the experimenter conducted a semi-structured interview tocollect qualitative feedback on the participants’ perception of the robot, and preference for and inter-pretation of different robot behaviours in the two experimental conditions. The interview was videorecorded with the camera angled in such a way that it captured the participants’ hand gestures withoutrecording their faces.5.2.3 Pilot StudyA pilot study was conducted in order to design conflict response behaviours for the robot that wouldbe comparable to the devised hesitation behaviours. Two conditions were tested in the pilot study: Yieldand Negotiate.Yield The Yield condition is comprised of the same components of the control (Smooth Stop) conditionimplemented and tested in Study 4. In this condition, the robot was equipped to respond usinga human-inspired, R-type hesitation controller called the Acceleration-based Hesitation Profile(AHP) controller [104]. By design, the controller requires the robot to automatically yield to theperson upon detection of an imminent conflict with no continued nonverbal dialogue between therobot and the person. In addition, this controller is limited in terms of the relative time-framewithin which the human-inspired hesitation response can be triggered. If the robot detects that80a conflict persists past the time window immediately after the robot has peaked in its forwardacceleration toward the person (practically immediately after the robot has started its reachingmotion from its starting position), the robot was programmed to come to a stop (a smooth stop)and hold its position for one second before retreating. The robot was programmed to re-attempt(i.e., reach for the resource again immediately after returning to the starting position) a maximumof three times before giving up on access to the resource and resuming its task of tending to theliquid bottles.Negotiate The Negotiate condition was designed using the negotiative hesitation controller described inthe previous chapters and further discussed in Section 5.2.4. In contrast to the Yield condition, therobot does not yield to the person immediately upon detection of a conflict, and actively reattemptsto access the shared resource until a resolution of the conflict is reached. Similarly to the Yieldcondition, the robot was programmed to re-attempt access to the resource three times before givingup and returning to the starting position and resuming its other tasks.Using the procedure described above, the experimenter conducted the pilot study with seven biasedparticipants (two females, five males).5 The average age of the participants was 31.1 years old, witha relatively high familiarity-with-robots score of 3.14, indicative of the fact that the participants hadrobotics background. Due to the small number of samples used for this pilot study, no inferentialstatistical analyses were conducted. Also, the robot was programmed to complete a total of fourteendispensing motions instead of the forty used in the full experiment. Performance measures suggestedthat there could be significant differences between the Yield and Negotiate conditions in how fast the HRteam completed the task. The average duration spent per trial for the Negotiate condition is much shorter(M = 149.18, SD= 8.39) than that of the Yield condition (M =184.41, SD= 21.12). A qualitative reviewof video-recorded interviews with the pilot study participants suggests that the Negotiate condition ispreferred over the Yield condition.Participants also reported that the first two trials of the experiment were exciting because of thenovelty effect of working with the robot, and perceived the last two trials to be more representativeexperiences of working with the robot in the two conditions. This supported the experiment designdecision to conduct a within-subjects, repeated-measures experiment in which participant responses tothe same experimental conditions are collected twice.Although the results from the qualitative study provide optimistic results for the Negotiate condi-tion, the participants’ qualitative feedback hinted that the immediate retraction behaviour of the Yieldcondition is perceived as inefficient. This retraction behaviour also unfairly disadvantages the respectivetrials’ measured completion time for the Yield condition, because the extra distance the robot must travelfor the retraction is absent from the Negotiate condition. Hence, the control condition was redesignedas the Stop condition outlined in Section 5.2.4. The Stop condition, used in the main experiment, doesnot have the extra travel time required by the robot to return to the starting position.5 These participants were colleagues and acquaintances of the experimenter, thereby having a chance of providing biasedresponses to the experiment.815.2.4 Experimental ConditionsInformed by the pilot study, the following near-collision responses were implemented to form thetwo experimental conditions: Stop and Negotiate.Stop The Stop condition was designed as a control against which to compare the more dynamic Ne-gotiate condition. Instead of the NHG, which was built using a linear dynamical system (refer toSection 4.5), this condition was a modified version of an already tested hesitation controller fromMoon et al. [104]. In the original controller, the author had employed a series of quintic splineswith parameters derived from human hesitations. However, the original controller only generatedhesitations that led to automatic yielding behaviour on the part of the robot. Also, due to the pa-rameter constraints that are necessary to generate this yielding hesitation, the robot must resort toa different type of collision avoidance behaviour after it has travelled a short distance. Hence, inthe Stop condition, the original controller was modified such that the robot triggers the yield-typehesitation if it can do so. Otherwise, the robot stopped as soon as a near collision situation wasdetected, and paused at that position for 1.0 second before reaching again. This generated a be-haviour where the distance travelled by the robot in each condition is the same, except that thegenerated robot motion while negotiating for access to the shared resource was different.Negotiate The Negotiate condition uses an implementation of the NHG controller developed in Chap-ter 4. This controller allows the robot to exhibit human-inspired hesitation behaviour in which therobot’s persistent interest in the shared resource is expressed. In this condition, the robot reachedfor the dispenser, and when a near collision situation was detected the robot moved away, in adirection opposite to its original direction, by a kickback-distance empirically determined fromChapter 4, before attempting to reach for the shared resource again. To avoid a possible livelock6by perpetual hesitation on the part of the robot and the participant, for each instance of hesitation,the robot repeated this kickback motion for a maximum of 15 times before retracting to its startingposition and yielding to the participant.5.3 Technical ImplementationThe same robotic system described in Section 4.6 was used in this study. Instead of the Smooth Stopcondition, equivalent to the Yield condition tested in the pilot study, the Stop condition was implemented.Details of the trajectories generated for each condition are outlined in Section 5.3.1. The safety featuresof the system are presented in Section Trajectory GenerationThe Stop condition employs the spline-based trajectory generator used in Moon et al. [104]. Thecontroller uses quintic Hermite splines derived from a cubic acceleration profile. When hesitation is6 The author uses the definition of livelock provided by Zo¨bel [169] who defines it as “all forms of imprevisible delay,often called starvation, permanent blocking and indefinite delay.”82triggered very early in the robot’s reach, four different quintic splines are stitched together with con-tinuous boundary conditions. This generates a smooth quintic-based trajectory that a robot can exhibitto immediately and anthropomorphically yield the resource to the participant. However, this particularsystem requires resource conflict to be detected at the onset of the robot’s reach, limiting the robot’s ca-pability to respond to conflicts that are detected later on in the interaction. Hence, for conflicts detectedafterwards, the spline-based trajectory generator was designed to smoothly stop and pause the robot for1.0 second, wherever the robot might be at the time. Unlike the pilot study, in which the robot returnedto its starting position upon detection of a conflict, the robot in the Stop condition was designed to con-tinue to reach for the target (the dispenser) after the pause, thereby generating a smooth and pause-basedre-attempt behaviour.In order to generate a negotiative conflict response for the Negotiate condition, the NHG, a LDS (de-fined by x˙ = mx+b) implementation of the characteristics observed in human hesitations in Section 4.4was employed. The LDS was tuned such that one round of reach, press, and retract motion took 3.2 sec-onds, matching the time it takes the robot to complete the same action in the Stop condition. Figure 5.3demonstrates an NHG-generated and a spline-generated trajectories without encountering any conflict.The linear slope, shown in the Robot to Target (R2T) plot of the figure, demonstrates the use of theLDS in reaching toward the target object. The lower right-hand corner also demonstrates a sharp changein velocity. This is because the robot moves faster the farther it is from the target due to the nature ofLDS. Hence the robot quickly reached its desired velocity at the beginning of its reach motions. Thecurved trajectory shown on the upper left-hand side of the plot demonstrates the robot’s pressing ofthe dispenser button. This particular motion used the same quintic trajectory generator as in the Stopcondition. This was to ensure that the only difference between the two conditions was in the robot’sconflict response, rather than other extraneous task-related trajectories such as pressing of the liquidpumps.Upon detection of a conflict, an upper layer module modulating the LDS moved the target locationof the DS to 19 mm (i.e., Kickback Distance (KD)= 19 mm) behind the current location of the robot.This choice of the NHG kickback parameter, KD, was based as the average KD distance observed inhuman hesitations. This effectively moved the robot backwards by KD before the upper layer modulemoved the target location of the DS again to the dispenser button for a re-attempt to access the resource.Figure 5.6 shows conflict response trajectories generated for the Stop and Negotiate conditions. In orderto prevent the robot from reaching a livelock condition in which the robot is in motion but is neverable to successfully reach the resource – presumably because of a participant refusing to yield – theexperimenter counted the number of re-attempts (RA) made by the robot and triggered the robot toyield/retract after 15 re-attempts. This value of RA, a value over and above what is observed in HHI, waschosen as a large value that will help contrast the maximum observed RAs in HRI against that of HHI.It was assumed that an HR pair would not reach a situation where the robot would use all 15 RAs. SeeFigure 5.7 for a sample of a trajectory with multiple RAs generated during the experiment.83Figure 5.3: A spline-generated and NHG-generated robot motions (d1(t), d˙1(t)) without encoun-tering any conflict in the Stop and Negotiate conditions. As defined in Chapter 4, d1(t)=‖M(t)−T‖, in which M(t) is the main agent’s (robot’s) Euclidean distance to the static targetT , and d˙1(t) represents its first derivative. Shown in both R2T plots are the robot’s motiontoward the target from 0.15 m away, reaching the target, and pressing down on the button(shown as the curved upside-down J shape) before returning to the original position. a) illus-trates a spline-generated motion used for the Stop condition, and b) shows an NHG-generatedmotion used for the Negotiate condition. The plots on the right-hand column represent therespective Euclidean distances over time (d1(t), d2(t), and δ (t)). d2(t) represents the Eu-clidean distance of the participant to the target, and δ (t) represents that difference betweend1(t) and d2(t).5.3.2 Safety Features and Motion TrackingThe system also included technical features to ensure the safety of the participants interacting withthe robots. For example, the control panel of the KUKA LWR robot is equipped with an emergencybutton. The button must be pressed in order for the robot to operate. Hence, a release of the buttonbrings the robot to a total stop immediately. To use this safety constraint, the experimenter sat near theworkspace and continuously held the button depressed, ready at all times to release it.In addition, the motion capture system utilized in Section 4.6, OptiTrack, was also used for thisstudy. The motion capturing system provided a stream of the participant’s hand and arm locations at asampling frequency of 120 Hz with a known latency of 4.2 ms to sense the participant movement. TheOptiTrack system yields easily identifiable values (series of zeroes) when the system fails to detect thedesignated markers accurately. The experimental system was designed to detect this anomaly such thatthe robot would come to a full and immediate stop upon losing connection with the sensor, or beingunable to detect the participant’s hand/arm location.84Previous studies report that the average maximum speeds at which humans feel safe while workingin close proximity to a robot’s working envelope is 41 to 64 cm/s depending on the size of the robot [75].Since it is not desirable to threaten the participants, the robot was set with a velocity limit of 41 cm/s,such that the robot would stop its motion immediately if it is commanded to move at a higher speed.Note that this maximum speed is higher than what international standards – ANSI/RIA R15.06-2012 [6]and ISO 10218-1:2011 [69] – consider as a safe enough speed for a person to withdraw from hazardousmotion (25 cm/s).Since the purpose of this study is to test the collision avoidance mechanism, the aforementionedsafety features were implemented with the aim of preventing any collision between human and robot.However, in the case where the robot comes to a stop and the person’s hand continues to travel towardthe fully stopped robot, the end effector of the robot was padded with a soft cushion-like material,such that the participant’s contact with the robot was soft and harmless. This approach and velocitylevel have been used in a previous study Moon et al. [104] where, in certain experimental conditions, apadded moving robot collided with the participants without causing any physical harm.In addition, the range of participant’s hand motion during the manipulation task had a maximumof 6 cm overlap with the robot’s end-effector at their fully extended positions. This overlapping regionof the workspaces is also the end of a robot’s reaching motion toward the shared object. Hence, bydesign, the robot always moved below its peak velocity within the workspace region it shared with theparticipants.5.4 ResultsOut of the 39 participants recruited, data from only 33 participants (14 females and 19 males) wereanalyzed. Data from the remaining six participants were rejected due to technical failures. The partici-pants were 20 to 40 years of age (M = 26, SD = 5.6). In contrast to the biased participants recruited forthe pilot study, most of the participants indicated that they were not at all familiar working with robots(M = 0.73, SD = 1.2).This section is structured based on the type of collected measures: Analyses of participants’ self-reported subjective experiences from questionnaire responses and interviews are presented in Sec-tion 5.4.1; quantitative measures of the HR teams’ performance, such as task completion time andnumber of conflicts triggered, are given in Section 5.4.2; differences observed in human behavioursacross the two conditions are in Section 5.4.3. References to the hypotheses stated in Section 5.1 aremade throughout these sections as appropriate. Possible type I errors for multiple pairwise comparisonswere corrected using the Bonferroni method.5.4.1 Subjective ExperienceIn this section, Hypothesis 5.1 is tested using the self-reported subjective experience of the partici-pants, which was analyzed using the questionnaire and interview responses.85Self-Reports on Perception of a Robot as a PartnerAs presented in Table 5.1, all eight of the self-report measures used in this study have internalreliability greater than 0.7, and are all deemed reliable for analysis. A repeated-measures ANOVA withCondition and Encounter as within-subjects fixed effects was employed to analyze the measures and issummarized in Table 5.2. Overall, the participants viewed the Stop condition to be more favourable.Regarding Hypothesis 5.1b, participants perceived the Negotiate condition responses to be sig-nificantly more dominant than the Stop condition. However, there are significant interaction effectsbetween Condition and Order on Usefulness (F(1, 32) = 9.66, p < .01) and Emotional Satisfaction(F(1, 32) = 6.32, p < .05) measures. This indicates that the perceived Usefulness and Emotional Sat-isfaction of the conditions change from the participants’ first encounter of the condition to the second.While this significant interaction effect neither fully supports nor rejects Hypothesis 5.1b on the Use-fulness and Emotional Satisfaction measures, the results do suggest that both the Negotiate and theStop conditions received an above-median Emotional Satisfaction score (3.52 and 3.86 respectively ona 5-point scale).In contrast to Hypothesis 5.1a, the participants perceived the Stop condition to be more animate,anthropomorphic, and likeable than the Negotiate condition. The robot was also perceived to be moreintelligent in the Stop than in the Negotiate condition. However, there was no significant difference asto how safe the participants perceived the robot to be in the two conditions.There are significant order effects on some but not all of the eight measures. The participantsperceived the robot to be more useful, animate, and anthropomorphic when the conditions were en-countered for the second time (Session 2). The robot was also perceived to be significantly safer in thesecond encounters, irrespective of the condition.Table 5.1: Internal reliabilities of the eight self-reported measures are presented here. All measuresused in this study have Cronbach’s alpha ≥ .70. These values are comparable to the valuesreported by the original authors of the standardized questionnaires Moon and Nass [107] andBartneck et al. [14].Measures Cronbach’s ItemsalphaDominance .76 Aggressive, Dominant, ForcefulUsefulness .79 Efficient, Helpful, UsefulEmotional Satisfaction .83How much did you like this robot?,How much did you like workingwith this robot?, EnjoyablePerceived Safety .83 Anxious, AgitatedLikeability .86 Like, Kind, Pleasant, FriendlyAnimacy .71 Apathetic, Artificial, MechanicalAnthropomorphism .87 Artificial, Fake, MachinelikePerceived Intelligence .70Foolish, Intelligent (reverse scale),Incompetent, Ignorant86Table 5.2: Repeated-measures ANOVA results are presented for all eight perception measures withCondition and Encounter as fixed factors. Since both factors only have two levels, Mauchly’sTest of sphericity does not apply. +Likeability score is a reverse measure, in which a lowervalue indicates a more favourable score.ConditionMeasure ANOVA Negotiate, Mean (SE) Stop, Mean (SE)Dominance∗ F(1, 32) = 6.06, p < .05 2.52 (0.16) 2.11 (0.13)Usefulness F(1, 32) = 0.89, p = .35 3.68 (0.13) 3.76 (0.10)Emotional Satisfaction∗ F(1, 32) = 6.63, p < .05 3.52 (0.13) 3.86 (0.10)Animacy∗ F(1, 32) = 4.35, p < .05 2.70 (0.12) 2.93 (0.12)Anthropomorphism∗ F(1, 32) = 4.87, p < .05 2.31 (0.14) 2.60 (0.14)Perceived Safety F(1, 32) = 2.59, p = .12 3.96 (0.14) 4.21 (0.14)Perceived Intelligence∗∗ F(1, 32) = 8.16, p < .01 3.33 (0.10) 3.56 (0.11)Likeability∗+ F(1, 32) = 6.23, p < .05 2.87 (0.12) 2.52 (0.12)EncounterMeasure ANOVA 1st, Mean (SE) 2nd, Mean (SE)Dominance F(1, 32) = 0.00, p = 1.00 2.32 (0.14) 2.32 (0.12)Usefulness∗ F(1, 32) = 5.91, p < .05 3.62 (0.11) 3.82 (0.12)Emotional Satisfaction F(1, 32) = 0.02, p = .89 3.70 (0.11) 3.69 (0.11)Animacy∗ F(1, 32) = 6.91, p < .05 2.66 (0.10) 2.98 (0.14)Anthropomorphism∗ F(1, 32) = 5.42, p < .05 2.32 (0.12) 2.60 (0.15)Perceived Safety∗ F(1, 32) = 4.96, p < .05 3.94 (0.15) 4.22 (0.12)Perceived Intelligence F(1, 32) = 0.02, p = .90 3.44 (0.10) 3.45 (0.11)Likeability+ F(1, 32) = 0.64, p = .43 2.74 (0.10) 2.65 (0.11)Qualitative Analysis of User InterviewsTo supplement some aspects of the subjective experience that may not be captured in, and to validateparticipants’ responses to, the questionnaire, the experimenter conducted a semi-structured interviewframed around the following questions:1. Out of the four trials, which trial was your favourite?2. Did you feel that you and the robot were collaborating together on a task, or did you feel that youwere both working on a task independently from each other?3. Was there any point in time that you felt the robot was being too aggressive?Only 25% of the participants chose trials with the Negotiate condition as their favourite trial, com-pared to 75% who chose a Stop trial. This supports the quantitative finding on Likeability, rejectingHypothesis 5.1a. The participants were also biased in choosing later trials as a favourite, with 78% ofthe participants choosing one of the last two trials over one of the first two.In addressing the issue of the robot being too aggressive, 60% of the participants said that there wasno point at which they felt the robot to be too aggressive. The remaining 40% of the participants eithermentioned the way the robot pressed the dispenser, or the high jerks observed at the start of reaching87motions in the Negotiate condition trials. These high jerks are inevitable due to the nature of LDS, wherethe highest forward velocity is commanded from a resting state at the onset of the robot’s reach towardthe target.On the question of perceived collaboration, 57% of the participants saw the experimental task to becollaborative, whereas the remaining participants said that they worked independently from the robot,despite the fact that the final product required the contribution of both the robot and the participant.5.4.2 Collaborative Task PerformanceTo better understand the full effect of the two conditions in an HR collaborative task, it is importantto observe the HR pairs’ overall collaborative task performance. In this section, the number of con-flicts triggered within a trial, the task completion time and the teams’ throughput are discussed as teamperformance measures.Number of Conflicts TriggeredWhile the number of dispenser presses by the robot was fixed at 40 per trial, the number of conflictsthe robot encountered varied due to the spontaneous nature of the conflicts triggered. Chi-square testswere conducted on this measure across Gender, Condition, and Encounter as factors.Across the two conflict resolution responses by the robot (Condition), there is no significant differ-ence in the number of conflicts triggered. In both conditions, approximately 37% of all robot motions ina trial encountered a conflict with a participant (X2(1) = 0.796, p= .82). However, surprisingly, genderplayed a role in how many conflicts the robot had with the participants (X2(1) = 5.00, p < .05). Femaleparticipants triggered more than the expected number of conflicts, whereas male participants triggeredless than the expected number. Looking at a more detailed analysis of whether there were significantdifferences in the number of conflicts triggered across conditions within the same gender, there is nosignificant effect (p = .21 and p = .14 for female and male participant groups, respectively).However, when looking at gender differences in the number of conflicts triggered, holding condi-tions constant, there are significant differences in conflicts triggered in the Stop condition only. In theNegotiate condition, 38% and 37% of all robot motions encountered conflicts with female and maleparticipants, respectively (X2(1) = 0.054, p = .82). In the Stop condition, female participants triggereda significantly larger number of conflicts (40%) than male participants (35%) (X2(1) = 8.59, p < .01).The robot encountered more conflicts in the second encounters than in the first (X2(1) = 15.8, p <.001). While in the first two trials, the robot came across conflicts 35% of the time, it saw conflicts40% of the time in the second two trials. A Pearson’s chi-square analysis on each set of encounterswith respect to the response conditions reveals that this training effect is common in both conditions(X2(1) = 0.043, p = .84 for the first two trials, and X2(1) = 0.06, p = .81 for the second two trialsacross the two conditions).88Task Completion TimeA repeated-measures ANOVA was conducted on task completion times, a measure representing howlong it took for the HR pair to complete a trial. The start and end of a trial are defined as the robot’sstart of the first and end of the last reach (the 40th). Factors including Condition, Gender, and Encounterwere considered.The ANOVA results indicate that there is a significant difference in the task completion time be-tween the two conditions (F(1, 31) = 100.3, p < .0001) and across the two encounters (F(1, 31) =7.62, p < .01). The HR team completed each trial in the Negotiate condition significantly faster(M = 298.3, SE = 0.92) than the trials in the Stop condition (M = 320.0, SE = 2.43), supportingHypothesis 5.1b. This supports Hypothesis 5.2, that is that hesitation responses generated using theNHG enable faster completion of the task than smooth stopping behaviour. Perhaps surprisingly, the HRteams took longer in the last two trials (the second encounter, M = 311.0, SE = 1.69) than in the firsttwo (first encounter, M = 307.3, SE = 1.58), demonstrating a reverse of the direction of the order effectone would typically expect. However, as seen in Figure 5.4, this order effect is much more pronouncedin the Stop condition in comparison to the Negotiate condition. There is no significant effect of Gender(F(1, 31) = 0.663, p = .422), nor interaction effect between Condition, Gender, and Encounter.Team ThroughputIn this experiment, measure of throughput for an HR collaboration consists of the following mea-sures: the number of correctly processed and incorrectly processed lentils, and two customized through-put scores as a function of the lentil count.As per previous analyses, a repeated-measures ANOVA was conducted considering Condition, Gen-der, and Encounter. Table 5.3 provides a summary of the results. In both correctly and incorrectlyprocessed products, there are no interaction effects across Conditions, Encounters, and Gender. Resultsindicate that female participants (M = 89.4, SE = 4.26) have a significantly higher number of correctlyprocessed lentils (p < .01) than male participants (M = 72.9, SE = 3.65). Despite this significant dif-ference in the number of lentils correctly processed, there were no significant gender differences in thenumber of incorrectly processed lentils (p = .92). This suggests that female participants were able todo the sorting task more quickly, without impacting the quality of the sorting process.The robot’s response to conflict (the Condition factor) also significantly affected the number oflentils processed. While the number of correctly processed lentils was higher (F(1, 31) = 21.17, p <.001) for the Stop condition (M= 85.1, SE = 3.25) compared to the Negotiate condition (M= 77.2, SE =2.58), the number of mistakes made by the participants was also higher (F(1, 31) = 6.60, p < .05) inthe Stop condition (M = 0.33, SE = 0.08) than in the Negotiate condition (M = 0.64, SE = 0.13).Since it is highly likely that a larger number of processed lentils is correlated to a higher probabilityof making mistakes, a Penalized Output Score, computed as the number of correctly processed minusincorrectly processed lentils, was employed in the analysis. In addition, the Rate of Output (PenalizedOutput Score / Total Duration of Trial), the rate at which the participant processed the output correctly,was computed.89Figure 5.4: Task completion time for the Stop vs. Negotiate conditions. The HR team completedeach trial in the Negotiate condition significantly faster than in the Stop condition. The HRteams took longer in the second encounter than the first encounter. However, this order effectis much more pronounced in the Stop condition in comparison to the Negotiate condition.The repeated-measures ANOVA results on these customized throughput scores are presented in Ta-ble 5.4. Analysis of the penalized score indicates that there is a significant difference between Negotiate(M = 76.9, SE = 2.58) and Stop (M = 84.8, SE = 3.25). This indicates that even though the number ofincorrectly processed lentils is higher for the Stop condition, the number of correctly processed lentilsoverall remains significantly higher for the Stop over the Negotiate condition. On the other hand, thereis no significant difference in the Rate of Output of Negotiate (M = 0.258, SE = 0.009) versus Stop(M = 0.262, SE = 0.009). The fact that the Rate of Output of Stop, with a significantly larger numberof mistakes observed, is not noticeably higher than Negotiate suggests that the mistakes observed in theStop condition cannot be attributed to the larger number of lentils processed. Rather, it is highly likelythat the two conditions contributed to this result.One should note that there are no significant differences in the Rate of Output, while there is asignificant difference in the task completion times between the two conditions. This demonstrates thata significant amount of time is saved in HR resolution of conflicts when the robot exhibits negotiativehesitation behaviours instead of the stopping behaviours used in the Stop condition.90Table 5.3: Repeated-measures ANOVA results with Condition and Encounter as fixed, within-subjects factors, and Gender as a between-subjects factor are presented for the correctly andincorrectly (mistake) processed products (lentils). Results from Levene’s test suggest that thecollected data-set does not violate the homogeneity of variance assumption.Measure ANOVACorrectCondition∗∗∗ F(1, 31) = 21.2, p < .001Order∗∗∗ F(1, 31) = 25.2, p < .001Gender∗∗ F(1, 31) = 8.64, p < .01Condition*Gender F(1, 31) = 0.343, p = .56Order*Gender F(1, 31) = 0.090, p = .77Condition*Order F(1, 31) = 0.408, p = .53Condition*Order*Gender F(1, 31) = 0.138, p = .71MistakesCondition∗ F(1, 31) = 6.60, p < .05Order F(1, 31) = 1.42, p = .24Gender F(1, 31) = 0.010, p = .92Condition*Gender F(1, 31) = 0.685, p = .41Order*Gender F(1, 31) = 0.129, p = .72Condition*Order F(1, 31) = 0.523, p = .48Condition*Order*Gender F(1, 31) = 0.055, p = .82As expected, the ANOVA results show a significant order effect on the counts of lentils correctlyprocessed. Fewer correctly processed lentils were found in the first encounters (M = 76.9, SE = 2.65)compared to the second encounters (M = 85.4, SE = 3.18). Fewer incorrectly processed lentils wereobserved in the second encounters (M = 0.41, SE = 0.08) than in the first (M = 0.56, SE = 0.13);however, this is not significant (p = .24).5.4.3 Trajectory and BehavioursIn addition to the subjective experience of the participants and the performance of the HR teamconsidered above, the author also analyzed the participants’ trajectory characteristics as indicators oftheir behaviour patterns. This section outlines trajectory-related findings from segments of motion thattriggered conflict responses in the robot. Since resource conflicts were randomly triggered during theHR interaction, the number of conflicts that occurred in a trial for each subject consequently varied.To account for the unequal number of conflicts within and between subjects, the author conducteda linear mixed-model analysis using REML (Multi-level Modelling (MLM), lme4 package, using thestatistics program, R (R, The R Foundation) with Condition, Encounter, and Gender as fixed factors,and Participants as a random factor. Possible interaction effects between Condition and Encounter wereaccounted for in the model, resulting in the following mixed model:measure∼Condition+Encounter+Gender+Condition∗Encounter+(1|Sub ject). (5.1)91Table 5.4: Repeated-measures ANOVA results with Condition and Encounter as fixed, within-subjects factors, and Gender as a between-subjects factor are presented for the Penalized Out-put Score and the Rate of Output. Levene’s test on these newly created measures suggests thatthey do not violate the homogeneity of variance assumption.Measure ANOVAPenalised Output ScoreCondition∗∗∗ F(1, 31) = 18.8, p < .001Order∗∗∗ F(1, 31) = 26.2, p < .001Gender∗∗ F(1, 31) = 8.66, p < .01Condition*Gender F(1, 31) = 0.268, p = .61Order*Gender F(1, 31) = 0.105, p = .75Condition*Order F(1, 31) = 0.470, p = .50Condition*Order*Gender F(1, 31) = 0.125, p = .73Rate of OutputCondition F(1, 31) = 0.995, p = .33Order∗∗∗ F(1, 31) = 20.4, p < .001Gender∗∗ F(1, 31) = 9.25, p < .01Condition*Gender F(1, 31) = 0.201, p = .66Order*Gender F(1, 31) = 0.006, p = .94Condition*Order F(1, 31) = 0.001, p = .98Condition*Order*Gender F(1, 31) = 0.262, p = .61Discussed below are the factors that were found to be significant predictors of the measurement ofinterest.Does it take longer to resolve conflicts when a robot uses negotiative hesitations?One of the hypotheses (Hypothesis 5.4) of this study is that the amount of time it takes for a resourceconflict to be resolved between an HR pair would be smaller for the Negotiate condition in comparisonto the Stop condition. As stated before, the amount of time it takes for the robot to reach for and returnfrom the dispenser is the same in both conditions. Hence, any residual time it takes for the robot toperform a reach-and-return motion when a conflict is detected is a measure of how long the HR pair tookto resolve the conflict in the segment of motion.The MLM analysis of this duration measure indicates that Condition is indeed a significant factorin determining how long a robot’s motion segment in conflict took place (X2(2) = 78.2, p < .001).7Motion segments in the Stop condition are predicted to take an average of 1.38 seconds (SE = 0.068)longer than that of an equivalent motion segment in the Negotiate condition, and supports Hypothesis5.4. This finding is despite the fact that the robot re-attempted a greater number of times during aconflict in the Negotiate condition than in the Stop condition. As shown in Figure 5.5, only in theNegotiate condition did the robot trigger its maximum programmed RA value of 15 before yielding to7 This means that the regression model, Equation 5.1, provides a significantly better prediction of duration than a regressionmodel without Condition as a factor.92Figure 5.5: The number of trials that triggered RA values greater than zero are presented here in acumulative manner. For example, Participant 301 in Trial 1 interacted with the robot in theNegotiate condition and had a conflict situation that resulted in triggering of all 15 RAs. Thisincident is reflected as a trial in all fifteen bar charts for the Negotiate condition. Since therobot’s maximum RA value was set to be 15, the robot yielded to the participant afterwards. Ahigher number of RAs were triggered under the Negotiate condition than the Stop condition,although the conflicts were resolved much more quickly in the Negotiate condition than theStop condition.the participant. These extreme instances reflect cases where curious participants intentionally kept theirhand out near the dispenser to see what the robot would do next.8 Other than these extreme cases, all ofthe HR conflicts were resolved before the robot fully yielded to the participant.There was no significant interaction effect between Condition and Encounter, and no significantmain effects of Gender and Encounter in the MLM analysis.Do participants stay farther away from the shared resource in negotiative HRI?One possible indicator of how safe participants perceived the robot to be is the location of their handswith respect to the workspace. If the participants kept their hands farther away from the shared resource,it would provide a behavioural indication of their perceived safety or comfort level with respect to theHRI.Results from the MLM analysis indicate that Condition is a significant predictor of mean Human to8 In fact, in two out of the three trials represented in Figure 5.5, the same subject triggered the maximum RA value in twoseparate trials.93Figure 5.6: One representative participant’s behavioural response to a resource conflict during aStop condition trial (Subject: 1, Trial: 2, Motion: 38). As illustrated in the Euclidean Dis-tances plot, the participant (d2(t)) hesitated upon the robot’s start of pause and yielded theright of way. The peak shown at (0.10, 0.0) of the R2T plot demonstrates the point at whichthe robot stopped to pause before proceeding to the target location.Target (H2T) distance (X2(2) = 53.2, p < .001). However, Encounter and Gender are not significantpredictors of average H2T distance. The participants’ mean Euclidean distance (measured from the wristof the dominant hand to target:H2T) during conflict segments is smaller in the Negotiate condition thanin the Stop condition by 1.3 cm (t(1870) = 6.22, p < .001). Participants, on average, had their handscloser to the dispenser in Negotiate than in Stop.A more telling indicator of a person’s comfort level with a robot is the minimum Euclidean distanceto the target (minimum H2T). However, analysis of this measure shows that there is a significant inter-action effect between Encounter and Condition (t(1870) =−2.23, p < .05). The use of both Condition(X2(2) = 6.50, p < .05) and Encounter (X2(2) = 6.13, p < .05) significantly improves prediction ofminimum H2T; however, these factors cannot be interpreted independently from each other. Hence,this analysis provides only a partial support for Hypothesis 5.3, that robot hesitation responses gener-ated using the NHG do not threaten perceived safety of the HRI in comparison to a smooth stoppingbehaviour.5.5 DiscussionWith the in-person HRI study presented in this chapter, the author attempted to address the second oftwo research questions stated at the beginning of this thesis: “Can a robot nonverbally negotiate with aperson about what should happen in an interaction? Can an HR negotiation contribute to an improved HRcollaboration?” The previous chapter described how the NHG was devised and validated as a mechanism94Figure 5.7: One representative participant’s behavioural response to a resource conflict during aNegotiate condition trial (Subject: 1, Trial: 1, Motion: 21). As illustrated by the EuclideanDistances plot, the participant (d2(t)) reached for the target and retracted right away as therobot triggered its negotiative hesitation behaviour before proceeding to access the target.The loop shown in the R2T plot demonstrates one re-attempt (RA) by the robot in this segmentof interaction. This is also shown as the cluster of points encircling δ˙ (t)=0.0 in the δ (t) vs.δ˙ (t) plot, which is absent in the H2T plot.that enables such negotiation-based conflict resolution to take place in HRI. This chapter presents howan implementation of the NHG to generate artificial, hesitation-based conflict responses from a robotwas used in in-person HR collaborative assembly.HR negotiations lead to faster conflict resolution without jeopardizing the perceived and actualsafety of the userThe results of the experiment provide empirical evidence that a robot’s resource conflict responsegenerated using the NHG contributes to an efficient resolution of a conflict and supports an affirmativeanswer to the research question. For instance, the analysis of the collaborative task performance supportsHypothesis 5.2 – that participants completed the task significantly faster in the Negotiate than in theStop trials. This is despite the fact that the number of conflicts triggered in the two conditions is notsignificantly different.It is also important to note that there is a training effect (in the reverse direction) in the number ofconflicts triggered per trial, and in the task completion time. In both conditions, more conflicts weretriggered in the second session than in the first. This suggests the possibility that repeated exposure toboth types of conflict responses may result in the participants ignoring the robot and dominating theworkspace completely, as was the case observed in Moon et al. [105]. This brings us back to the reportsof human users abusing self-driving vehicles’ systematic, safety-prioritizing responses discussed in the95Introduction [40, 100]. While hesitation-based kinetic dialogue in HRI has been proposed as a solutionto a human’s potential unwillingness to yield to a robot, thereby undesirably dominating the interaction,the results of this study suggest that repetition of such dialogue may be ineffective. With the increase inthe number of conflicts triggered in the second session (40% of robot’s reach motions were interrupted,in comparison to 35% in the first session), trials in the second session for both conditions took longer tocomplete than those in the first.Upon a closer look at this trend, there is a case to be made for employing a persistent or pushyrobot in resolution of HR resource conflicts. As seen in Figure 5.4, there is a much larger increase inthe task completion time of the Stop condition from the first to the second session in comparison to thatof the Negotiate condition. This suggests that, even if human users tend to inevitably take advantageof conservative, safety-driven robotic systems after repeated encounters, a negotiative behaviour of therobot can help it to reach resolution of a conflict in a more efficacious manner. In support of this, theresults show that the resource conflicts were resolved more quickly in the Negotiate condition than theStop condition, which supports Hypothesis 5.4. Hence, while the participants may have started to takeadvantage of the robot’s conflict response in both conditions in the second session, this did not hinderthe overall performance of the team in the Negotiate condition due to better handling of the resourceconflict using the negotiative mode of interaction.At this point, it is important to note that a robot behaviour that is perceived as more persistentand dominant does not always translate to an HRI that is threatening, harmful, or disliked. From thequestionnaire responses reporting on the perceived dominance of the robot, the Negotiate condition isconsidered more dominant than the Stop condition (supporting Hypothesis ??). The mean Dominancescore for the Negotiate condition was 2.52 on a 5-point scale – approximately the middle of the scale.However, behavioural measures on the participants’ average distance to the shared resource show thatthe participants kept their wrist closer to the resource in the Negotiate condition than they did in theStop condition. This demonstrates that the participants were, in general, comfortable interacting withthe robot in the Negotiate condition. If they felt threatened by the robot’s motions in the Negotiatecondition, one would expect the average distance to the shared resource to be larger in comparison tothe Stop condition.Qualitative analysis of the interview further supports that while the robot may be perceived as moreaggressive in the Negotiate condition, it was never seen as too aggressive or threatening. The interviewresults also suggest that the source of the perceived dominance is not necessarily the robot’s conflictresponse behaviour. For example, the participants referred to the way the robot was pressing the dis-penser button while discussing dominance of the robot, although the same pressing behaviour was usedin both conditions. This suggests that the perceived dominance of the robot may stem from anothersource, or that the jerky behaviour of the robot at the onset of reach motions in the Negotiate conditioncaused participants to project a perception of dominance to other parts of the robot’s motion, such aspressing of the dispenser by the robot. Moreover, the Perceived Safety measure from the questionnaireresponses also demonstrates that the robot behaviours in both conditions score highly (averages of 3.96and 4.21 on a 5-point scale for Negotiate and Stop conditions, respectively), and that there is no sig-96nificant difference in this measure across the two conditions (p = .12). Hence, the results provide amultitude of evidence for Hypothesis 5.3, namely that hesitation responses generated using the NHG donot threaten the perceived safety nor jeopardize the actual safety of the interaction9 in comparison to asmooth stopping behaviour.Perception of the interaction with negotiative robot behaviours is not superior to that of smoothstopping behaviours, but is still viewed in a positive lightSelf-reports of participants’ experience with the two conditions were measured using the question-naire response and the interview. Results of these measures provide some mixed findings. First, Emo-tional Satisfaction and Usefulness (two out of the three team perception measures collected) for both theStop and the Negotiate conditions were above the median (above 3.5 on 5-point scales). Both conditionsprovided a positive experience in terms of emotional satisfaction and perceived usefulness of the sys-tem. However, participants reported a significantly higher Emotional Satisfaction for the Stop conditionthan the Negotiate condition. In addition, there was no significant difference in the Usefulness score ofthe two conditions. These results fail to support Hypothesis 5.1b that the hesitation responses generatedusing the NHG are perceived to be more emotionally satisfactory than a smooth stopping behaviour.However, the fact that the Negotiate condition scores highly on both measures is a positive indicator thatinteracting with a robot that uses the NHG to negotiate for a resource conflict with them is not perceivedas threatening, but rather as emotionally satisfying and useful.Second, the author hypothesized (Hypothesis 5.1a) that the robot in the Negotiate condition will beperceived in a more favourable light than the Stop condition. However, the results indicate that the robotin the Stop condition was perceived to be more animate, anthropomorphic, likeable, and intelligent thanin the Negotiate condition. This fails to support Hypothesis 5.1a, and is in contrast to the results ofStudy 4, in which the NHG-generated robot responses were perceived to be more anthropomorphic andanimate than a non-NHG alternative. This finding suggests that a perception gap exists between thirdparty observation of and in-person interaction with a robot that uses the NHG.The higher Animacy and Anthropomorphism scores of the Stop condition may be owing to the factthat the reach trajectories of the condition were designed using quintic splines. Quintic splines generateinherently smooth and humanlike trajectories compared to the LDS-generated trajectories of the Nego-tiate condition, which provide a jerkier motion at the onset of a robot’s reach. The higher Likeabilityscore of the Stop condition in comparison to the Negotiate condition may be related to the higher Anthro-pomorphism and Animacy scores of the Stop condition. The collected Likeability score is not subjectto a significant training effect. This suggests that the participants’ preference for the Stop conditionwas retained throughout the experiment. This is also supported by the results of the post-experimentinterviews.It is also interesting to note that the participants were divided about their perception of the task aseither collaborative or independent. Results from the interview show that 57% of the participants per-9 With the description of the safety features of the system presented in this chapter, no physical collision or safety issueoccurred in this experiment for any of the participants.97ceived the task as collaborative, versus 43% perceiving the experimental tasks as independent tasks thatinvolve sharing of a resource. Reflecting on this feedback, it is likely that the team metrics (Dominance,Usefulness, Emotional Satisfaction) are inadequate to capture the full nature of the HR collaboration thatwas established. In describing the experiment, the participants’ perception of the collaborative task didnot involve terms such as Dominance and Usefulness. Rather, the interaction was described in termsthat are related more closely to the ideas of flow and fluency.5.6 ConclusionIn this chapter, an in-person HRI experiment was presented as a means to demonstrate the efficacyof HR nonverbal negotiations using hesitation gestures. While designed hesitation behaviours using theNHG are not preferred over a more traditional conflict response, the results of this study strongly supportthe effectiveness of this interaction mode on the basis of improved HR collaborative performance. Inaddition, while the NHG-based negotiative behaviour of the robot is perceived to be more dominant thanthe smooth stopping alternative, this perception does not equate to a negative experience or perceptionof the robot. Therefore, the results from this experiment successfully present the NHG as a proof ofconcept of a new mode of interaction that enables interactive, shared decision-making about an outcomeof a conflict in an HR collaboration. This mode of interaction also does not jeopardize the safety of theuser nor the performance of the collaboration.The journey to conducting this experiment also involved a few secondary contributions. The arti-ficial hesitation behaviours generated using the NHG were successfully tested in an in-person interac-tion with human participants, and validates its efficacy in in-person HRI. The devised system providesa proof-of-concept of real-time, negotiative conflict resolution between a collaborating human and arobot using tracking of the user’s wrist motion. It also demonstrates that the hesitation-based conflict re-sponses generated using the NHG deliver an interaction experience distinguished from smooth stoppingbehaviours used in the Stop condition.In the next chapter, the findings from the studies presented in this thesis – including the supplemen-tary Study 6, presented in Appendix A.1 – are summarized with respect to the two research questionsframing this thesis.98Chapter 6ConclusionAn interacting agent’s ability to interweave its plans with another agent is a key element of col-laboration [20]. These plans include details of a task such as the who, when, and where that must becommunicated with the other agents to successfully collaborate on the task together. Noting that thesedetails are often unspoken at the onset of a collaboration and organically determined as the task un-folds, the author asserted that fluid communication between the collaborating agents is imperative forthe agents to interweave these plans. To contribute to the development of robotic systems that collabo-rate with people in an efficient, safe, and friendly manner, this thesis focused on communication cues arobot can be equipped with to communicate and interweave these details with human users.In particular, the physical actuation capability of a robot is an essential feature that sets apart thepromise of robotics from that of other technologies. This thesis leveraged this capability by exploringthe ways in which kino-dynamic behaviours of a robot can be designed to facilitate the interweavingprocess in an HR collaboration in two exemplar situations, object handover and resource sharing. Theoften unspoken, yet essential details of the specific tasks explored in this thesis were: when and where ahuman recipient should reach out to receive an object from the robot in a robot-to-human handover, andwho should get access to the shared resource first when both a human and a robot are reaching for it atthe same time. A series of six human-subjects studies collectively explored the following two researchquestions:1. Unidirectional Interweaving Can a robot provide humanlike nonverbal cues to influence people’sbehavioural responses to an interaction while the interaction is taking place?2. Bidirectional Interweaving (Negotiation) Can a robot nonverbally negotiate with a person aboutwhat should happen in an interaction? Can an HR negotiation contribute to an improved HR col-laboration?Results from the six studies support that human-inspired nonverbal cues are an effective means tounidirectionally and bidirectionally enable the process of interweaving in an HR team. The followingsections present a summary of the findings, research contributions, and future work that pertain to eachof the questions. In addition, this chapter concludes with two points of discussion on the implicationsof the findings from this thesis.996.1 Unidirectional Interweaving of SubplansCan a robot’s use of nonverbal cues help interweave spatiotemporal details of a task? As discussedin Chapter 2, even simple motions of an automated door or presence of a robot in a room can affecthuman decisions and behaviours. Nonverbal cues of a robot have been used in various HRI contexts toestablish joint attention with and communicate a robot’s intent and internal states to a user. Building onthe findings from previous studies, this thesis focused on the role nonverbal robot behaviours can playin helping to interweave plans with a person.The investigations presented in Chapter 3 used the context of robot-to-human handover interaction.In this context, the order of the tasks to be performed and the division of roles was made explicit fromthe beginning. The desired outcome was also clear from the onset, while the spatiotemporal details ofthe interaction needed to be determined by the interacting agents while the interaction took place. Bydesign, only the participants were influenced by the interaction. That is, the robot led the interactionwith communication cues that unidirectionally influenced the participant, while the robot itself was notinfluenced by the communication signals from the participants.Study 1 (n = 12) was conducted to investigate the pattern of gaze cues humans use when handingover an object to each other. Qualitative analysis of the collected data resulted in the identification offive types of gaze patterns. This result was used to inform the design of human-inspired gaze patterns arobot can use when handing over an object to a person.The most frequently observed gaze patterns from Study 1 – the shared attention gaze – was im-plemented on a humanoid robot for an in-person HRI study, Study 2 (n =102). Study 2 investigatedwhether the robot’s use of human-inspired gaze behaviour can help interweave unspoken spatiotempo-ral details of a robot-to-human handover interaction. In contrast to previous HRI studies [86, 146] thatfocused on establishing the joint intent to engage in a handover before a handover event takes place,the investigations in Study 2 explicitly addressed the impact a robot’s gaze has on human behavioursduring a handover. The results of Study 2 demonstrate that the robot’s use of gaze can impact when andwhere the participants decide to move their hands to receive the object from the robot. The participantsreached out for the object significantly earlier when the robot exhibited the shared attention gaze thanwhen it did not provide a human-inspired gaze pattern. This suggests that the implementation of nonver-bal cues on a robot, such as gaze, can influence people’s behavioural responses to an interaction whilethe interaction is taking place. In addition, with the shared attention gaze, the participants reached tothe projected handover location before the robot had fully reached the location. This provides a glimpseof the power nonverbal cues can have in the interweaving process of an HR collaboration. By cueingthe participants to reach out earlier to meet the offered object before the robot has finished moving, theinteraction is not only more efficient but also more fluid. Results of a follow-up study, Study 6 (n =30)led by Minhua Zheng and presented in Appendix A.1, further demonstrate that this effect of gaze issustained even after a period of repeated HR handovers.These results supplement the previous work in HRI discussed in Section 2.2.1. In previous studies,gaze cues were demonstrated as effective in communicating a robot’s target object and internal states toan observing person. Findings from this thesis add to the previous literature and demonstrate that robot100gaze cues, even when exhibited during a handover interaction, can be used to subtly communicate spa-tiotemporal details of a handover to the human receiver and elicit a desired kinodynamic behaviour fromthe person. This work not only contributes to increasing the fluidity of robot-to-human handovers, butalso inspired the investigations on how a robot’s use of nonverbal gestures can impact the bidirectionalinterweaving process of an HR collaboration.6.1.1 Limitations and Future WorkThe role of gaze cues explored in Studies 1, 2, and 6 only focused on the context of robot-to-humanhandovers. This limits the level at which one can generalize the findings of the studies’ results withrespect to the impact human-inspired robots cues. Much more extensive studies involving various con-texts and nonverbal cues would be necessary before being able to draw general and detailed conclusionsabout the role a robot’s human-inspired nonverbal cues can have on unidirectionally eliciting desirablehuman behaviours for collaboration.In addition, although the author and her colleagues proposed and led the qualitative analyses toidentify the gaze behaviour of the participants during the handover event, the quality of information thatcan be gathered on a participants’ gaze seen from recordings of a digital camera is limited. In the future,a gaze detector would be necessary to gather an accurate insight about when the participant understandsthe robot’s communication of the desired spatiotemporal details.While limited in scope, however, the results of this work provided empirical support that even thesubtle, and supplementary nonverbal cues such as gaze can play a sigificant role in improving the fluencyof an interaction and informing the interweaving process.6.2 Bidirectional Interweaving (Negotiation) of SubplansWould people negotiate with a robot? Would an HR negotiation lead to an improved outcome of aninteraction? In Chapters 4 and 5, the author focused on negotiative hesitation behaviours as a means toinvestigate whether a human and a robot can affect each other to negotiate for a desired outcome of aninteraction. A series of experiments was conducted to study the dynamics that emerge during in-personHR negotiations of resource conflicts. The experiments considered a collaborative assembly scenariowhere, while necessarily sharing the same spaces and resources, the collaborating agents reached forthe same object at the same time, creating a conflict that they must resolve before they can proceed. Incontrast to the unidirectional interweaving discussed in Chapter 3, the outcome of the conflict (i.e., thequestion of who should access the object first) is not determined at the onset of the interaction.As presented in Chapter 4, a human-subjects study, Study 3 (n =16) was conducted to collect sam-ples of naturally occurring human hesitation behaviours from an HH collaboration context. Samples ofhesitations were observed by online participants (n =300) who rated them according to the expressedhesitancy and persistency of the motions captured. The results of this study demonstrate that hesitationsare indeed used as a mechanism to nonverbally and dynamically resolve the conflict of resources thatspontaneously arise in HHI.Upon analysis of the sample human hesitations and uninterrupted reach gestures collected from101Study 3, the author discovered that, when viewed in state space, a trajectory pattern exists in negotiativehesitations, called hesitation loops. These loops represent the changing speeds and distances of thetwo collaborating agents with respect to the shared target object. While the trajectory pattern can beobserved post hoc, generating the same trajectory pattern in HRI is non-trivial. Only one of the twoagents (the robot) that shape the hesitation loops is within our control.Subsequently, the author identified a set of parameters that can be used to convert the hesitationloops into a trajectory generation scheme, called the NHG, which can be used to produce negotiativehesitation trajectories for a robot. The NHG was implemented using the LDS approach to develop asystem that is highly responsive to the changes in the workspace. In Study 4 (n =50), trajectoriesgenerated by the NHG were used to validate the efficacy of the motions produced by the trajectorygenerator. The results of this video-based online study demonstrate that the motions produced by theNHG are not only perceived to be more hesitant than an industrial alternative, but also more animateand anthropomorphic. In addition, the feature space used to discover the hesitation loops presented adomain in which the interplay of motions in a nonverbal HR negotiation can be evaluated. Findings fromthe study also confirmed that the parameter values derived from observed human motions are adequatefor generating effective negotiative hesitation gestures for a robot.The investigations presented in Chapter 4 contribute to a better understanding of hesitation be-haviours that are increasingly becoming useful in the field of HRI. The design of the NHG also con-tributes to the field by providing a mechanism with which one can generate human-inspired negotiativehesitation behaviours. This allows one to study the negotiative dynamics of an HR dyad using reac-tive, hesitation-based conflict responses with parameter values that are empirically validated by humanobservers.In Chapter 5, the author used the NHG as a mechanism to study the negotiation dynamics that emergein an HR collaboration. A final in-person human-subjects study, Study 5 (n =39), was conducted with acollaborative assembly context in which the HR pair naturally and often reached for the same object atthe same time.Results from this study demonstrate that the NHG can be used to enable nonverbal HR negotiation ofresource conflicts to take place. It also illustrates how an HR pair can determine the outcome of a conflictinteractively and nonverbally. Not only did the participants yield access to the shared resource to therobot when it exhibited NHG-based hesitation behaviours, the HR team finished the task significantlyfaster and with fewer mistakes than a non-negotiative alternative, in which the robot paused its motionupon detection of a conflict. Moreover, the real-time negotiations between a human and a robot led toa faster resolution of conflicts without jeopardizing the actual safety of the user or the user’s perceivedsafety of the interaction.The findings from Study 5 empirically address the second research question. They support that anHR pair can dynamically figure out a desired outcome through nonverbal negotiation. The findings alsosuggest that, similar to how people typically behave toward another person in such conflicts, peopledo yield to the robot rather than dominating the shared space. This contributes to the field of HRIby demonstrating the efficacy of nonverbal negotiations as an efficient and fluent mode of interaction.102Furthermore, such negotiative interactions can shift the point of decision-making about an outcome ofan interaction from the designers, who are often not directly affected by the outcome, to the users, whoare engaged in the interaction in-person.6.2.1 Limitations and Future WorkResults of Studies 4 and 5 provide a positive outlook on the use of negotiations as a means for usersto take part in determining the outcome of an interaction. This shifts the decision making process fromthe designers and engineers who do not take part in the interaction to the users themselves. However,the investigations presented in this thesis did not evaluate the quality of the negotiated outcome otherthan the overall task performance. Given this work as a foundation, it would be imperative to developan experiment where an HR nonverbal negotiation can involve an obvious good or bad outcome. Suchexperiments would allow us to understand the role robot motions can play in helping people to makethe right, split-second decisions, or whether a person’s decisions are less likely to be affected by themotions of the robot in contexts where the person is presented with obvious good decisions.Note that, as further elaborated in the discussions below, the negotiative mode of interaction inStudy 5 was not favoured over the alternative. It was also perceived to be less animate and anthropo-morphic. This contradicts the findings from Study 4, where the NHG-based motions were perceivedmore favourably against a non-negotiative alternative. This contradicting finding may be indicative ofthe fact that an observer’s preference or perception of a robot motion can understandably be different ifthey are third party observers of the interaction versus an in-person participant in the interaction. It mayalso suggest that the implementation of the NHG using the LDS approach requires further improvementsin order to increase its perceived animacy and anthropomorphism.While the results of Study 5 illustrated that the NHG-based conflict responses can contribute to animprovement of team performance indicators, the results also illustrated a trend that the participantstriggered more conflicts as they got used to the experiment trials. If such negotiative behaviours are tobe implemented in repetitive collaborative task scenarios, the training effect on the number of conflictstriggered must be further investigated to evaluate if the performance improvements are maintained aftera prolonged duration of interaction with the robot.Going forward, it is important to note that the NHG provides a mechanism with which one can studythe dynamics of an HR nonverbal negotiation. For example, one of the next steps that utilize the NHGwould be to modify the parameters of the NHG to express higher or lower levels of dominance. Theimpact of the robot’s expressed dominance in a nonverbal negotiation could be examined in such inves-tigations along with an examination of the resulting quality of the HR negotiated outcome. Participants’personality types could also be used as a factor to better understand the type of negotiation dynamics thatcan be expected in an HR conflict resolution. Such investigations would provide an important insightin how human motions at a subconscious level can be affected by motions of the robot in a nonverbalnegotiation setting.1036.3 Efficient HRI vs. Preferred HRIThe results of Studies 2 (Chapter 3) and 5 (Chapter 5) illustrate that the two human-inspired nonver-bal cues designed and tested in the studies improve the efficiency of the HR collaboration. However, theauthor also found that the participants do not prefer these more efficient, human-inspired cues over theless efficient alternatives. In Study 2, the robot’s use of the human-inspired, Shared Attention gaze ina robot-to-human handover improved the efficiency of the HRI by encouraging the participants to reachfor and receive the object earlier. However, the participants did not show a significant preferrence forthis gaze over the Turn-Taking gaze, a less efficient human-inspired alternative, or not having a human-inspired gaze at all – the No Gaze condition where the robot simply looked downwards. Likewise, inStudy 5, the Smooth Stop condition, which was not designed based on human behaviours, was preferredover the NHG-generated behaviours. This was despite the fact that the NHG-generated motions werehuman-inspired and resulted in a superior team task performance.This follows a puzzling observation within the field of HRI: a robot behaviour that improves the taskperformance does not necessarily result in a more preferred subjective experience. For example, in astudy by Ragni et al. [127], a robot that provides imperfect answers in a memory game with a personwas perceived in a more positive light, while it negatively impacted the participant’s performance of thetask. In addition, Baraglia et al. [11] conducted a study where a robot helped a person on a collaborativemanipulation task in one of the following ways: the user controlled when the robot should help; therobot reactively helped the user when the need for help was detected; and the robot proactively providedhelp whenever it was able to. The researchers found that the HR team worked more fluently when therobot was proactive in helping the user. However, the users preferred it when they had the explicitcontrol of when the robot should help, despite the fact that such interaction resulted in a suboptimal taskperformance.In these studies, a user’s preference for an HRI does not seem to have a positive correlation with theperformance improvements the interaction can provide. Implementation of a robot’s behaviour, locusof control, and the nature of the collaborative tasks investigated in the abovementioned studies are alldifferent from one another. They are also factors that are likely to affect user preferences. Therefore, itis hard to make any definitive claims about this intriguing trend observed in these studies.In order to further investigate this trend, it is necessary to devise a standardized and comprehensivemethod for measuring the impact an interactive behaviour has on an HR collaboration. The two stan-dardised questionnaires used within the field of HRI – the Godspeed questionnaire [14] and the NegativeAttitudes towards Robots Scale [149] – both focus only on self-reported user perception of robots. Uti-lizing such questionnaires may be sufficient to evaluate the quality of interactive robot behaviours thatdo not emphasize functional efficiency of the resulting interaction. However, evaluating an interactivebehaviour for an HR collaboration cannot rely only on questionnaire results, and must also consider var-ious measures of task performance. For collaborative systems that involve repetitive tasks, studies withmuch longer duration of interaction may also be necessary before one can identify the eventual prefer-ence and efficacy of a collaborative system. In the future, a more comprehensive metric that combinesself-reports of user experiences along with task outcomes may be necessary in order to better contrast104one collaborative system against another. This will also help demystify the observed tradeoff betweenefficiency and preference of interactive robot behaviours in HR collaboration.6.4 Broader Implications and Future Directions: Inclusivity andBidirectionalityUnlike HHI, HRI is often a process in which one of the agent’s precise behaviours and motionsare determined by a third party (designers and engineers) who does not partake in the interactivityand are not directly affected by the consequence or the outcome of the interaction. Nonetheless, HRIpractitioners are the ones tasked with the design decisions that must be made about future interactionsthat users will have with the systems. As the author established in Chapter 2, affecting human behavioursand decision-making with a moving object is inevitable. Designing behaviours of an interactive robotis, therefore, an activity that is directly related to determining what type of systematic influence theusers will encounter. With the rise of interactive robotic systems that can manipulate our physicalenvironment, it is imperative that HRI practitioners closely examine this relationship we have with thesystem’s users. It may be harmless to implement gaze cues on a robot that can consistently elicit adesired kinodynamic response from a person in a handover interaction. A system that unidirectionallyinfluences users to make certain purchasing decisions, on the other hand, may be considered unethical[138]. This relates to the uneasy feelings depicted in many works of science fiction where a dystopianfuture with robots involves humans who are helplessly subjected to unidirectional influence of the so-called robot overlords. In reality, the robotic systems that are being developed today lie on a spectrum-of-influence model with unidirectional influence on one extreme and complete bidirectional influenceon the other. Where a system lies on this spectrum is determined by design decisions that map thesystem’s ability to sense its environment, including users around it, and its ability to respond to thesensed signals. As robots can be viewed as a new ‘designed’ species deployed into the world, it is up tothe HRI practitioners to be mindful of this influence directionality spectrum and its implications to theoutcome of an interaction [73].A number of researchers are concerned about this social distance between designers and users, andadvocate for the principle of transparency as a means to instill the right level of trust in a user of arobotic system [153]. One way to increase such transparency between the user and the system is toenable more fluent and intuitive modes of communication for the HRI. In fact, systems on any pointon the influence directionality spectrum can be designed to be transparent to the user as long as theyare equipped with appropriate abilities to convey their internal states to the users. However, one-sidedcommunication does not allow room for continuous changes that may be desired in HRI. The factthat communication is imperative to every element of collaboration is founded upon the notion thatcollaborative processes are typically dynamic in nature, allowing a team of agents to respond to changesin the situation, environment, and even internal states or desires. To realize the type of human-machinecollaboration Barbara Grosz envisioned, the dynamic nature of our everyday tasks must be somethingthat robots can accommodate.The value further supported by designing systems to be more bidirectional in its interaction with105users is that it allows room for the user to have a say in what should happen in the interaction, which canbe different from one context and person to another. When presented with an uncertain or a contentioussituation that involves a person and a robot, the author’s previous work suggests that people prefer therobot to engage in dialogues with the stakeholder than determine the interaction outcome by itself (seeSection 2.2.2). The results of Study 5 (Chapter 5) illustrate that, at the most basic level of interactionthat involves subtle movements of the hands, a robot is able to engage in a dialogue with a person towarda resolution of a conflict, even if it is not favoured over a robotic alternative. Regardless of the rapid ad-vancements in artificial intelligence and machine learning technologies, which could help determine anoptimal solution to a problem, a system that is opaque in its decision-making may be seen as a hindranceto the user who may feel the need to exercise control and have a sense of autonomy about a situationat hand. The negotiative behaviours explored in this thesis attempt to put forth inclusivity – which theauthor refers to as a system’s ability to include the stakeholders (including users) into the process ofmaking decisions that affect the collective – as a principle that can be fostered in interaction designs aswe deal with more systems that will come across unintended conflicts and unpredicted situations withpeople.106Bibliography[1] URL Accessed on 2013-09-08. → pages 25[2] E. Ackerman. Panasonic revives hospital delivery robot, May 2014. URL on 2015-01-03. → pages 8[3] H. Admoni, T. Weng, B. Hayes, and B. Scassellati. Robot nonverbal behavior improves taskperformance in difficult collaborations. In ACM/IEEE International Conference onHuman-Robot Interaction, volume 2016-April, pages 51–58. ACM/IEEE, 2016. ISBN9781467383707. doi:10.1109/HRI.2016.7451733. → pages 4, 9[4] A. Albu-Scha¨ffer, S. Haddadin, C. Ott, A. Stemmer, T. Wimbo¨ck, and G. Hirzinger. Robovie: aninteractive humanoid robot. Industrial Robot: An International Journal, 28(6):498 – 504, Aug2001. ISSN 0143-991X. doi:10.1108/01439910710774386. → pages 11[5] Amazon Mechanical Turk: artificial artificial intelligence, 2016. URL Accessed on 2016-12-10. → pages 42, 43[6] ANSI/RIA. ANSI/RIA R15.06-2012 American National Standard for Industrial Robots andRobot Systems - Safety Requirements (revision of ANSI/RIA R15.06-1999). American NationalStandards Institute, 2012. → pages 85[7] Apple Inc. iOS - Siri - Apple (CA), 2011. URL Accessed on2016-12-12. → pages 2[8] M. Argyle and M. Cook. Gaze and Mutual Gaze. Cambridge University Press, Cambridge, UK,1976. → pages 18[9] R. Atienza and A. Zelinsky. Intuitive Human-Robot Interaction Through Active 3D GazeTracking, pages 172–181. Springer Berlin Heidelberg, Berlin, Heidelberg, 2005. ISBN978-3-540-31508-7. doi:10.1007/11008941 19. URL 19.→ pages 18[10] W. a. Bainbridge, J. W. Hart, E. S. Kim, and B. Scassellati. The benefits of interactions withphysically present robots over video-displayed agents. International Journal of Social Robotics,3:41–52, 2011. ISSN 18754791. doi:10.1007/s12369-010-0082-7. → pages 37[11] J. Baraglia, M. Cakmak, Y. Nagai, R. Rao, and M. Asada. Initiative in robot assistance duringcollaborative task execution. ACM/IEEE International Conference on Human-Robot Interaction,2016-April:67–74, 2016. ISSN 21672148. doi:10.1109/HRI.2016.7451735. → pages 104107[12] C. Bartneck, M. van der Hoek, O. Mubin, and A. Al Mahmud. “Daisy, Daisy, give me youranswer do!”. Proceedings of the ACM/IEEE international conference on Human-robotinteraction - HRI ’07, (2007):217, 2007. doi:10.1145/1228716.1228746. → pages 37, 38[13] C. Bartneck, T. Kanda, O. Mubin, and A. Al Mahmud. Does the design of a robot influence itsanimacy and perceived intelligence? International Journal of Social Robotics, 1(2):195–204,Feb 2009. ISSN 1875-4791. doi:10.1007/s12369-009-0013-7. → pages 37[14] C. Bartneck, D. Kulic´, E. Croft, and S. Zoghbi. Measurement instruments for theanthropomorphism, animacy, likeability, perceived intelligence, and perceived safety of robots.International Journal of Social Robotics, 1(1):71–81, Nov 2009. ISSN 1875-4791.doi:10.1007/s12369-008-0001-3. → pages 62, 80, 86, 104, 124[15] P. Basili, M. Huber, and T. Brandt. Investigating human-human approach and hand-over.Human Centered Robot Systems, pages 151–160, 2009.doi:DOI=10.1007/978-3-642-10403-9 16. → pages 15, 17, 22[16] M. Bennewitz, F. Faber, D. Joho, M. Schreiber, and S. Behnke. Towards a humanoid museumguide robot that interacts with multiple persons. In 5th IEEE-RAS International Conference onHumanoid Robots, 2005., pages 418–423, Dec 2005. doi:10.1109/ICHR.2005.1573603. →pages 18[17] A. Bisio, A. Sciutti, F. Nori, G. Metta, L. Fadiga, G. Sandini, and T. Pozzo. Motor contagionduring human-human and human-robot interaction. PLoS ONE, 9(8), 2014. ISSN 19326203.doi:10.1371/journal.pone.0106172. → pages 12[18] S. Bouchard. Robot manipulation challenge: Clean up dining table, load dishwasher, Nov 2009.URL on 2016-12-12. → pages 8[19] J. Boucher, U. Pattacini, A. Lelong, G. Bailly, F. Elisei, S. Fagel, P. F. Dominey, andJ. Ventre-Dominey. I reach faster when I see you look: gaze effects in humanhuman andhumanrobot face-to-face cooperation. Frontiers in Neurorobotics, 6(3), 2012. ISSN 1662-5218.doi:10.3389/fnbot.2012.00003. → pages 4, 8, 11[20] M. E. Bratman. Shared cooperative activity. The Philosophical Review, 101(2):327–341, 1992.→ pages 2, 8, 13, 99[21] C. Breazeal. Jibo, the world’s first social robot for the home, 2014. URL Accessedon 2016-11-14. → pages 2[22] C. Breazeal, C. Kidd, A. Thomaz, G. Hoffman, and M. Berlin. Effects of nonverbalcommunication on efficiency and robustness in human-robot teamwork. In IEEE/RSJInternational Conference on Intelligent Robots and Systems, pages 383–388, Edmonton,Canada, 2005. IEEE. ISBN 0-7803-8912-3. doi:10.1109/IROS.2005.1545011. → pages 4, 10[23] C. Breazeal, A. Takanishi, and T. Kobayashi. Social robots that interact with people. SpringerHandbook of Robotics, pages 1349–1369, 2008. doi:10.1007/978-3-540-30301-5 59. → pages11108[24] K. O. Brien, J. Sutherland, C. Rich, and C. L. Sidner. Collaboration with an autonomoushumanoid robot: a little gesture goes a long way. In Proceedings of the 6th internationalconference on Human-robot interaction - HRI ’11, pages 215–216, Lausanne, Switzerland,2011. ACM. → pages 10[25] D. Brscic´, H. Kidokoro, Y. Suehiro, and T. Kanda. Escaping from children’s abuse of socialrobots. In Proceedings of the Tenth Annual ACM/IEEE International Conference onHuman-Robot Interaction, HRI ’15, pages 59–66, New York, NY, USA, 2015. ACM. ISBN978-1-4503-2883-8. doi:10.1145/2696454.2696468. → pages 74[26] G. Butterworth. The ontogeny and phylogeny of joint visual attention. In A. Whiten, editor,Natural theories of mind: Evolution, development, and simulation of everyday mindreading,pages 223–232. Basil Blackwell, 1991. → pages 18[27] M. Cakmak, S. S. Srinivasa, M. K. Lee, S. Kiesler, and J. Forlizzi. Using spatial and temporalcontrast for fluent robot-human hand-overs. In Proceedings of the 6th International Conferenceon Human-robot Interaction, HRI ’11, pages 489–496, New York, NY, USA, 2011. ACM. ISBN978-1-4503-0561-7. doi:10.1145/1957656.1957823. → pages 15, 17, 25, 31[28] M. Cakmak, S. S. Srinivasa, Min Kyung Lee, J. Forlizzi, and S. Kiesler. Human preferences forrobot-human hand-over configurations. In 2011 IEEE/RSJ International Conference onIntelligent Robots and Systems, pages 1986–1993. IEEE, Sep 2011. ISBN 978-1-61284-456-5.doi:10.1109/IROS.2011.6094735. → pages 17[29] E. Calisgan, A. Haddadi, H. F. M. Van der Loos, J. A. Alcazar, and E. A. Croft. Identifyingnonverbal cues for automated human-robot turn-taking. In 2012 IEEE RO-MAN: The 21st IEEEInternational Symposium on Robot and Human Interactive Communication, pages 418–423,Sept 2012. doi:10.1109/ROMAN.2012.6343788. → pages 23[30] E. Calisgan, A. Moon, C. Bassani, F. Ferreira, F. Operto, G. Veruggio, E. A. Croft, and H. F. M.Van der Loos. Open Roboethics pilot: accelerated policy design, implementation anddemonstration of socially acceptable behaviors. In We Robot2, Stanford, CA, 2013. → pages 12[31] M. Carpenter, K. Nagell, and M. Tomasello. Social cognition, joint attention, andcommunicative competence from 9 to 15 months of age. Monographs of the Society forResearch in Child Development, 63(4):i–vi, 1–143, 1998. ISSN 0037-976X. → pages 11[32] J. Cassell and K. R. Thorisson. The power of a nod and a glance: envelope vs. emotionalfeedback in animated conversational agents. Applied Artificial Intelligence, 13(4-5):519–538,May 1999. ISSN 0883-9514. doi:10.1080/088395199117360. → pages 37[33] A. Chalimourda, B. Scho¨lkopf, and A. J. Smola. Experimentally optimal ν in support vectorregression for different noise models and parameter settings. Neural Networks, 17(1):127–141,2004. ISSN 08936080. doi:10.1016/S0893-6080(03)00209-0. → pages 50[34] W. P. Chan, C. A. Parker, H. F. M. Van der Loos, and E. A. Croft. A human-inspired objecthandover controller. The International Journal of Robotics Research, 32(8):971–983, Jul 2013.ISSN 0278-3649. doi:10.1177/0278364913488806. → pages 15, 17, 22[35] W. P. Chan, K. Nagahama, H. Yaguchi, Y. Kakiuchi, K. Okada, and M. Inaba. Implementationof a framework for learning handover grasp configurations through observation duringhuman-robot object handovers. In 15th IEEE-RAS International Conference on Humanoid109Robots, Humanoids 2015, Seoul, South Korea, November 3-5, 2015, pages 1115–1120, 2015.doi:10.1109/HUMANOIDS.2015.7363492. → pages 5[36] A. Chang. NY midtown robots allow for conversation-free hotel service, May 2015. URL Accessed on 2016-12-12. → pages 7[37] P. R. Cohen and H. J. Levesque. Teamwork. Nouˆs, 25(4):487–512, 1991. → pages 8[38] S. Connor. First self-driving cars will be unmarked so that other drivers don’t try to bully them:Volvo fears that other road users could drive erratically in order to take advantage, 2016. URL on 2016-12-12. → pages 75[39] M. Corley and O. W. Stewart. Hesitation disfluencies in spontaneous speech: the meaning ofum. Linguistics and Language Compass, 2(4):589–602, 2008. ISSN 1749818X.doi:10.1111/j.1749-818X.2008.00068.x. → pages 37[40] S. Crowe. A cyclist’s encounter with an indecisive google self-driving car, Aug 2015. URL cyclists encounter with an indecisive google self driving car/P2. Accessed on 2016-10-03.→ pages 77, 96[41] D. C. L. David J. Knight, Daniel Langmeyer. Eye-contact, distance, and affiliation: the role ofobserver bias. Sociometry, 36(3):390–401, 1973. ISSN 00380431. URL → pages 18[42] F. Dehais, E. A. Sisbot, R. Alami, and M. Causse. Physiological and subjective evaluation of ahumanrobot object hand-over task. Applied Ergonomics, 42(6):785–791, Nov 2011. ISSN00036870. doi:10.1016/j.apergo.2010.12.005. → pages 16, 17[43] C. Dondrup, N. Bellotto, and M. Hanheide. Social distance augmented qualitative trajectorycalculus for human-robot spatial interaction. In The 23rd IEEE International Symposium onRobot and Human Interactive Communication, pages 519–524, 2014. ISBN978-1-4799-6765-0. doi:10.1109/ROMAN.2014.6926305. → pages 38[44] L. W. Doob. Hesitation: Impulsivity and Reflection. Greenwood Press, Westport, CT, 1990.ISBN 0313274460. → pages 37[45] A. D. Dragan, K. C. Lee, and S. S. Srinivasa. Legibility and predictability of robot motion. In2013 8th ACM/IEEE International Conference on Human-Robot Interaction (HRI), pages301–308. IEEE, Mar 2013. ISBN 978-1-4673-3101-2. doi:10.1109/HRI.2013.6483603. →pages 10[46] J. DURBIN. Incomplete blocks in ranking experiments. British Journal of StatisticalPsychology, 4(2):85–90, 1951. ISSN 2044-8317. doi:10.1111/j.2044-8317.1951.tb00310.x. →pages 27[47] M. S. Erden, K. Leblebiciog˘lu, and U. Halici. Multi-agent system-based fuzzy controller designwith genetic tuning for a mobile manipulator robot in the hand over task. J. Intell. RoboticsSyst., 39(3):287–306, Mar 2004. ISSN 0921-0296. doi:10.1023/B:JINT.0000021039.56110.c8.→ pages 22110[48] N. Figueroa, A. L. P. Ureche, and A. Billard. Learning complex sequential tasks fromdemonstration: A pizza dough rolling case study. In ACM/IEEE International Conference onHuman-Robot Interaction, volume 2016-April, pages 611–612, 2016. ISBN 9781467383707.doi:10.1109/HRI.2016.7451881. → pages 57[49] T. Flash and N. Hogan. The coordination of arm movements: Mathematical model. Journal ofNeuroscience, 5(7):1688–1703, 1985. → pages 63[50] W. J. Fu. Penalized regressions: the bridge versus the lasso. Journal of Computational andGraphical Statistics, 7(3):397–416, 1998. ISSN 10618600. doi:10.2307/1390712. → pages 48[51] K. Gillespie-Lynch, P. M. Greenfield, Y. Feng, S. Savage-Rumbaugh, and H. Lyn. Across-species study of gesture and its role in symbolic development: implications for thegestural theory of language evolution. Frontiers in psychology, 4(June):160, Jan. 2013. ISSN1664-1078. doi:10.3389/fpsyg.2013.00160. → pages 40[52] B. Gleeson, K. Currie, K. MacLean, and E. Croft. Tap and push: Assessing the value of directphysical control in human-robot collaborative tasks. Journal of Human-Robot Interaction, 4(1):95, 2015. ISSN 2163-0364. doi:10.5898/JHRI.4.1.Gleeson. → pages 9, 10[53] R. M. Golinkoff. i beg your pardon?: the preverbal negotiation of failed messages. Journal ofChild Language, 13(03):455–476, Sept. 1986. ISSN 0305-0009.doi:10.1017/S0305000900006826. → pages 40[54] R. M. Golinkoff. When is communication a meeting of minds? Journal of child language, 20(1):199–207, Feb. 1993. ISSN 0305-0009. → pages 40[55] Google. The Google app. URL Accessed on2016-12-12. → pages 2[56] B. J. Grosz. Collaborative systems. AI Magazine, 17(2):67–85, 1996. → pages 1, 2, 8[57] B. J. Grosz and L. Hunsberger. The dynamics of intention in collaborative activity. CognitiveSystems Research, 7(2-3):259–272, 2006. ISSN 13890417. doi:10.1016/j.cogsys.2005.11.006.→ pages[58] B. J. Grosz and S. Kraus. Collaborative plans for complex group action. Artificial Intelligence,86(2):269–357, 1996. ISSN 00043702. doi:10.1016/0004-3702(95)00103-4. → pages 2, 8[59] M. Ha¨ring, J. Eichberg, and E. Andre´. Studies on grounding with gaze and pointing gestures inhuman-robot-interaction. In Proceedings of the 4th International Conference on SocialRobotics, ICSR’12, pages 378–387, Berlin, Heidelberg, 2012. Springer-Verlag. ISBN978-3-642-34102-1. doi:10.1007/978-3-642-34103-8 38. → pages 18[60] B. Hayes-Roth. A blackboard architecture for control. Artif. Intell., 26(3):251–321, Aug 1985.ISSN 0004-3702. doi:10.1016/0004-3702(85)90063-3. → pages 25[61] G. Hoffman and W. Ju. Designing robots with movement in mind. Journal of Human-RobotInteraction, 3(1):89, 2014. ISSN 2163-0364. doi:10.5898/JHRI.3.1.Hoffman. → pages 11[62] G. Hoffman, J. Forlizzi, S. Ayal, A. Steinfeld, J. Antanitis, G. Hochman, E. Hochendoner, andJ. Finkenaur. Robot presence and human honesty. In Proceedings of the Tenth AnnualACM/IEEE International Conference on Human-Robot Interaction - HRI ’15, pages 181–188,111New York, New York, USA, 2015. ACM Press. ISBN 9781450328838.doi:10.1145/2696454.2696487. → pages 10[63] B. Huang, M. Li, R. L. De Souza, J. J. Bryson, and A. Billard. A modular approach to learningmanipulation strategies from human demonstration. Autonomous Robots, 40(5):903–927, 2016.ISSN 15737527. doi:10.1007/s10514-015-9501-9. → pages 57[64] C.-M. Huang and B. Mutlu. The repertoire of robot behavior: Designing social behaviors tosupport human-robot joint activity. Journal of Human-Robot Interaction, 2(2):80–102, 2013.ISSN 2163-0364. doi:10.5898/JHRI.2.2.Huang. → pages 11[65] C.-M. Huang, S. Andrist, A. Sauppe´, and B. Mutlu. Using gaze patterns to predict task intent incollaboration. Frontiers in psychology, 6(July):1049, 2015. ISSN 1664-1078.doi:10.3389/fpsyg.2015.01049. → pages 10[66] M. Huber, M. Rickert, A. Knoll, T. Brandt, and S. Glasauer. Human-robot interaction inhanding-over tasks. In RO-MAN 2008 - The 17th IEEE International Symposium on Robot andHuman Interactive Communication, pages 107–112, Aug 2008.doi:10.1109/ROMAN.2008.4600651. → pages 17[67] M. Imai, T. Ono, and H. Ishiguro. Physical relation and expression: joint attention forhuman-robot interaction. In Robot and Human Interactive Communication, 2001. Proceedings.10th IEEE International Workshop on, pages 512–517, 2001.doi:10.1109/ROMAN.2001.981955. → pages 18, 23[68] M. Imai, T. Kanda, T. Ono, H. Ishiguro, and K. Mase. Robot mediated round table: Analysis ofthe effect of robot’s gaze. In Robot and Human Interactive Communication, 2002. Proceedings.11th IEEE International Workshop on, pages 411–416, 2002.doi:10.1109/ROMAN.2002.1045657. → pages 22[69] International Standards Organization. ISO 10218-1:2011 – robots and robotic devices – safetyrequirements for industrial robots. Technical report, ISO, 2011. URL{ }detail?csnumber=51330. → pages 85[70] W. Ju and D. Sirkin. Animate objects : How physical motion encourages public interaction. InP. Thomas, H. Per, and O.-K. Harri, editors, Persuasive Technology: Lecture Notes in ComputerScience, volume 6137, pages 40–51. Springer, Berlin, Heidelberg, 2010.doi:10.1007/978-3-642-13226-1 6. → pages 10[71] W. Ju and L. Takayama. Approachability: How people interpret automatic door movement asgesture. International Journal, 3(2):1–10, Aug 2009. URL→ pages 10, 11[72] W. Ju and L. Takayama. Should robots or people do these jobs ? a survey of robotics experts andnon-experts about which jobs robots should do. In International Conference on IntelligentRobots and Systems, pages 2452–2459, San Francisco, CA, 2011. IEEE/RSJ. ISBN9781612844558. → pages 8[73] P. H. Kahn, A. L. Reichert, H. E. Gary, T. Kanda, H. Ishiguro, S. Shen, J. H. Ruckert, andB. Gill. The new ontological category hypothesis in human- robot interaction. In Proceedings ofthe 6th international conference on Human-robot interaction - HRI ’11, pages 159–160,Lausanne, Switzerland, 2011. ACM. → pages 105112[74] S. Kajikawa, T. Okino, K. Ohba, and H. Inooka. Motion planning for hand-over between humanand robot. In Proceedings of the International Conference on Intelligent Robots andSystems-Volume 1 - Volume 1, IROS ’95, pages 193–, Washington, DC, USA, 1995. IEEEComputer Society. ISBN 0-8186-7108-4. → pages 17[75] W. Karwowski, T. Plank, M. Parsaei, and M. Rahimi. Human perception of the maximum safespeed of robot motions. Proceedings of the Human Factors and Ergonomics Society AnnualMeeting, 31(2):186–190, Sep 1987. ISSN 1071-1813. doi:10.1177/154193128703100211. →pages 85[76] T. Kazuaki, O. Motoyuki, and O. Natsuki. The hesitation of a robot: A delay in its motionincreases learning efficiency and impresses humans as teachable. Human-Robot Interaction(HRI), 2010 5th ACM/IEEE International Conference on, 8821007:189–190, 2010.doi:10.1109/HRI.2010.5453200. → pages 37[77] T. Kellner. Rethink Robotics Is Freeing Robots From Their Cages, 2016. URL on 2016-11-14. → pages 1[78] S. M. Khansari-Zadeh and A. Billard. Learning stable non-linear dynamical systems withgaussian mixture models. Transactions on Robotics, 27(5):943—-957, 2011. ISSN 1552-3098.doi:10.1109/TRO.2011.2159412. → pages 57[79] S. Kim, A. Shukla, and A. Billard. Catching objects in flight. IEEE Transactions on Robotics,PP:1–17, 2014. ISSN 1552-3098. doi:10.1109/TRO.2014.2316022. → pages 57[80] N. Kirchner, A. Alempijevic, and G. Dissanayake. Nonverbal robot-group interaction using animitated gaze cue. In Proceedings of the 6th International Conference on Human-robotInteraction, HRI ’11, pages 497–504, New York, NY, USA, 2011. ACM. ISBN978-1-4503-0561-7. doi:10.1145/1957656.1957824. → pages 18, 21, 22[81] C. L. Kleinke. Gaze and eye contact: a research review. Psychological bulletin, 100(1):78–100,Jul 1986. ISSN 0033-2909. → pages 16, 18[82] K. L. Koay and E. A. Sisbot. Exploratory study of a robot approaching a person in the context ofhanding over an object. In Symposium on Multidisciplinary Collaboration for Socially AssistiveRobotics, pages 18–24, 2007. → pages 16, 17, 22[83] D. Kulic and E. Croft. Physiological and subjective responses to articulated robot motion.Robotica, 25(01):13, Aug 2006. ISSN 0263-5747. doi:10.1017/S0263574706002955. → pages10[84] Y. Kuno, K. Sadazuka, M. Kawashima, K. Yamazaki, A. Yamazaki, and H. Kuzuoka. Museumguide robot based on sociological interaction analysis. In Proceedings of the SIGCHIConference on Human Factors in Computing Systems, CHI ’07, pages 1191–1194, New York,NY, USA, 2007. ACM. ISBN 978-1-59593-593-9. doi:10.1145/1240624.1240804. → pages 18[85] D. Leavens, J. Russell, and W. Hopkins. Intentionality as measured in the persistence andelaboration of communication by chimpanzees (pan troglodytes). Child development, 76(1):291–306, 2005. → pages 40113[86] M. K. Lee, J. Forlizzi, S. Kiesler, M. Cakmak, and S. Srinivasa. Predictability or adaptivity?:Designing robot handoffs modeled from trained dogs and people. In Proceedings of the 6thInternational Conference on Human-robot Interaction, HRI ’11, pages 179–180, New York, NY,USA, 2011. ACM. ISBN 978-1-4503-0561-7. doi:10.1145/1957656.1957720. → pages 15, 16,18, 22, 100[87] L. E. Levin. Kinetic dialogs in predator-prey recognition. Behavioural processes, 40(2):113–20,Jul 1997. ISSN 0376-6357. → pages 76, 77[88] A. Lim, T. Ogata, and H. G. Okuno. Towards expressive musical robots: a cross-modalframework for emotional gesture, voice and music. EURASIP Journal on Audio, Speech, andMusic Processing, 3, 2012. ISSN 1687-4722. doi:10.1186/1687-4722-2012-3. → pages 41[89] C. Liu, C. T. Ishi, H. Ishiguro, and N. Hagita. Generation of nodding, head tilting and eyegazing for human-robot dialogue interaction. In Proceedings of the Seventh Annual ACM/IEEEInternational Conference on Human-Robot Interaction, HRI ’12, pages 285–292, New York,NY, USA, 2012. ACM. ISBN 978-1-4503-1063-5. doi:10.1145/2157689.2157797. → pages18, 23[90] K. E. Lochbaum. Using Collaborative Plans to Model the Intentional Structure of Discourse.PhD thesis, Harvard University, Cambridge, MA, USA, 1995. UMI Order No. GAX95-14804.→ pages 1[91] J. Mainprice, M. Gharbi, T. Simon, and R. Alami. Sharing effort in planning human-robothandover tasks. In 2012 IEEE RO-MAN: The 21st IEEE International Symposium on Robot andHuman Interactive Communication, pages 764–770, Sept 2012.doi:10.1109/ROMAN.2012.6343844. → pages 17[92] I. Marlow. The boomer shift: Japan’s bold steps, Nov 2015. URL Accessed on2016-11-03. → pages 1[93] J. G. Martin and W. Strange. The perception of hesitation in spontaneous speech. Perception &Psychophysics, 3(6):427–438, 1968. ISSN 0031-5117. doi:10.3758/BF03205750. → pages 37[94] Y. Matsusaka, S. Fujie, and T. Kobayashi. Modeling of conversational strategy for the robotparticipating in the group conversation. In INTERSPEECH, 2001. → pages 18[95] Max Planck Institute for Psycholinguistics. ELAN — The Language Archive. URL Accessed on 2016-11-03. → pages 29[96] R. Maxion and A. DeChambeau. Dependability at the user interface. In Twenty-FifthInternational Symposium on Fault-Tolerant Computing. Digest of Papers, pages 528–535. IEEEComput. Soc. Press, 1995. ISBN 0-8186-7079-7. doi:10.1109/FTCS.1995.466944. → pages 37[97] S. Merlo and P. A. Barbosa. Hesitation phenomena: a dynamical perspective. Cognitiveprocessing, 11(3):251–61, Aug 2010. ISSN 1612-4790. doi:10.1007/s10339-009-0348-x. →pages 37[98] J. Millar and A. Moon. How to engage the public on the ethics and governance of lethalautonomous weapons. In We Robot, Miami, FL, 2016. → pages 8114[99] G. Milliez, R. Lallement, M. Fiore, and R. Alami. Using human knowledge awareness to adaptcollaborative plan generation, explanation and monitoring. ACM/IEEE International Conferenceon Human-Robot Interaction, 2016-April:43–50, 2016. ISSN 21672148.doi:10.1109/HRI.2016.7451732. → pages 10[100] R. Mitchell. Human drivers will bully robot cars, says CEO of Mercedes-Benz USA, 2016.URL Accessed on 2012-12-10. → pages 75, 96[101] S. Mohammad Khansari-Zadeh and A. Billard. A dynamical system approach to realtimeobstacle avoidance. Autonomous Robots, 32(4):433–454, 2012. ISSN 09295593.doi:10.1007/s10514-012-9287-y. → pages 57[102] A. Moon, P. Danielson, and H. F. M. Van der Loos. Survey-based discussions on morallycontentious applications of interactive robotics. International Journal of Social Robotics, pages1–20, Nov 2011. ISSN 1875-4791. doi:10.1007/s12369-011-0120-0. → pages 8[103] A. Moon, C. A. C. Parker, E. A. Croft, and H. F. M. Van der Loos. Did you see it hesitate? -Empirically grounded design of hesitation trajectories for collaborative robots. In 2011IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 1994–1999, SanFrancisco, CA, Sep 2011. IEEE/RSJ. ISBN 978-1-61284-454-1.doi:10.1109/IROS.2011.6048228. → pages 38, 44, 45[104] A. Moon, C. A. C. Parker, E. A. Croft, and H. F. M. Van der Loos. Design and impact ofhesitation gestures during human-robot resource conflicts. Journal of Human-Robot Interaction,2(3):18–40, Sept. 2013. ISSN 2163-0364. doi:10.5898/JHRI.2.3.Moon. → pages xii, 38, 42, 44,45, 57, 80, 82, 85[105] A. Moon, D. M. Troniak, B. Gleeson, M. K. Pan, M. Zheng, B. A. Blumer, K. MacLean, andE. A. Croft. Meet me where I’m gazing: How shared attention gaze affects human-robothandover timing. In ACM/IEEE International Conference on Human-Robot Interaction, pages334–341, Bielefeld, Germany, 2014. ACM/IEEE. ISBN 9781450326582.doi:10.1145/2559636.2559656. → pages iv, 15, 75, 95[106] A. Moon, E. Calisgan, C. Bassani, F. Ferreira, F. Operto, G. Veruggio, E. A. Croft, and H. F. M.Van der Loos. The Open Roboethics Initiative and the elevator-riding robot. In R. Calo, A. M.Froomkin, and I. Kerr, editors, Robot Law, chapter 6, pages 131–162. Edward Elgar Publishin,Cheltenham, UK, Northampton, MA, USA, 2016. → pages 12[107] Y. Moon and C. Nass. How “real” are computer personalities?: Psychological responses topersonality types in human-computer interaction. Communication Research, 23(6):651–674,Dec 1996. ISSN 0093-6502. doi:10.1177/009365096023006002. → pages 62, 80, 86[108] A. Mo¨rtl, T. Lorenz, and S. Hirche. Rhythm patterns interaction - synchronization behavior forhuman-robot joint action. PLoS ONE, 9(4):e95195, Apr 2014. ISSN 1932-6203.doi:10.1371/journal.pone.0095195. → pages 15[109] K. P. K. Murphy. Dynamic Bayesian networks: representation, inference and learning. Ph.d.,University of California Berkeley, 2002. → pages 41115[110] B. Mutlu. Designing Gaze Behavior for Humanlike Robots. PhD thesis, Pittsburgh, PA, USA,2009. AAI3367045. → pages 16, 18[111] B. Mutlu, J. Forlizzi, and J. Hodgins. A storytelling robot: Modeling and evaluation ofhuman-like gaze behavior. In 2006 6th IEEE-RAS International Conference on HumanoidRobots, pages 518–523, Dec 2006. doi:10.1109/ICHR.2006.321322. → pages 18[112] B. Mutlu, T. Shiwa, T. Kanda, H. Ishiguro, and N. Hagita. Footing in human-robotconversations: How robots might shape participant roles using gaze cues. In Proceedings of the4th ACM/IEEE International Conference on Human Robot Interaction, HRI ’09, pages 61–68,New York, NY, USA, 2009. ACM. ISBN 978-1-60558-404-1. doi:10.1145/1514095.1514109.→ pages 18[113] B. Mutlu, F. Yamaoka, T. Kanda, H. Ishiguro, and N. Hagita. Nonverbal leakage in robots:communication of intentions through seemingly unintentional behavior. ACM/IEEEInternational Conference on Human-Robot Interaction, pages 69–76, 2009. → pages 11[114] A. Netick and S. T. Klapp. Hesitations in manual tracking: A single-channel limit in responseprogramming. Journal of Experimental Psychology: Human Perception and Performance, 20(4):766–782, 1994. ISSN 0096-1523. doi:10.1037/0096-1523.20.4.766. → pages 38, 77[115] S. Nikolaidis, A. Kuznetsov, D. Hsu, and S. Srinivasa. Formalizing human-robot mutualadaptation: A bounded memory model. ACM/IEEE International Conference on Human-RobotInteraction, 2016-April:75–82, 2016. ISSN 21672148. doi:10.1109/HRI.2016.7451736. →pages 13[116] Y. Ogai and T. Ikegami. Microslip as a simulated artificial mind. Adaptive Behavior, 16(2-3):129–147, Apr 2008. ISSN 1059-7123. doi:10.1177/1059712308089182. → pages 38[117] N. Oliver, A. Garg, and E. Horvitz. Layered representations for learning and inferring officeactivity from multiple sensory channels. Computer Vision and Image Understanding, 96(2):163–180, Nov 2004. ISSN 10773142. doi:10.1016/j.cviu.2004.02.004. → pages 41[118] Open Roboethics initiative. Results: Will autonomous cars create more jobs?, 2014. URL Accessed on2016-11-09. → pages 8[119] Open Roboethics initiative. The two factors in bath by robots: Privacy and control, 2014. URL on 2016-11-09. → pages[120] Open Roboethics initiative. The ethics and governance of lethal autonomous weapons systems :An international public opinion poll. Technical report, Open Roboethics initiative, Vancouver,BC, Canada, 2015. → pages 8[121] Oxford University Press. Definition of negotiation in english, 2013. URL Accessed on 2013-10-13. → pages 35[122] Oxford University Press. Definition of persistence in english, 2013. URL 5.Accessed on 2013-10-13. → pages 40116[123] A.-L. Pais Ureche and A. Billard. Learning bimanual coordinated tasks from humandemonstrations. In Proceedings of the Tenth Annual ACM/IEEE International Conference onHuman-Robot Interaction Extended Abstracts, pages 141–142, 2015. ISBN 9781450333184.doi:10.1145/2701973.2702007. → pages 57[124] M. L. Patterson. A sequential functional model of nonverbal exchange. Psychological Review,89(3):231–249, 1982. → pages 18[125] A. Peters. Spatial coordination - human and robotic communicative whole-body motions innarrow passages. PhD thesis, 2012. → pages 37[126] M. Quigley, K. Conley, B. P. Gerkey, J. Faust, T. Foote, J. Leibs, R. Wheeler, and A. Y. Ng.ROS: an open-source Robot Operating System. In ICRA Workshop on Open Source Software.IEEE, 2009. → pages 64[127] M. Ragni, A. Rudenko, B. Kuhnert, and K. O. Arras. Errare humanum est: Erroneous robots inhuman-robot interaction. In 2016 25th IEEE International Symposium on Robot and HumanInteractive Communication (RO-MAN), pages 501–506. IEEE, Aug 2016. ISBN978-1-5090-3929-6. doi:10.1109/ROMAN.2016.7745164. → pages 104[128] K. Reed, M. Peshkin, M. J. Hartmann, M. Grabowecky, J. Patton, and P. M. Vishton. Hapticallylinked dyads: are two motor-control systems better than one? Psychological science, 17(5):365–6, May 2006. ISSN 0956-7976. doi:10.1111/j.1467-9280.2006.01712.x. → pages 8[129] R. W. Reeder and R. A. Maxion. User interface defect detection by hesitation analysis.Proceedings of the International Conference on Dependable Systems and Networks, 2006:61–70, 2006. doi:10.1109/DSN.2006.71. → pages 37[130] B. Reeves and C. Nass. The Media Equation: How People Treat Computers, Television, and NewMedia like Real People and Places. Cambridge University Press, 1996. ISBN 157586052X.URL →pages 10, 79[131] C. Rich, B. Ponsleur, A. Holroyd, and C. L. Sidner. Recognizing engagement in human-robotinteraction. In Proceedings of the 5th ACM/IEEE International Conference on Human-robotInteraction, HRI ’10, pages 375–382, Piscataway, NJ, USA, 2010. IEEE Press. ISBN978-1-4244-4893-7. → pages 18[132] P. Rober. Some hypotheses about hesitations and their nonverbal expression in family therapypractice. Journal of Family Therapy, 24(2):187–204, 2002. ISSN 1467-6427.doi:10.1111/1467-6427.00211. → pages 37[133] S. Robotics. Softbank robotics: Who is pepper? URL Accessed on 2016-11-14. →pages 1[134] H. Romat, M. Williams, X. Wang, B. Johnston, and H. Bard. Natural human-robot interactionusing social cues. In Human-Robot Interaction, pages 503–504, 2016. ISBN 9781467383707.→ pages 10[135] B. Roth. Foreword. In B. Siciliano and O. Khatib, editors, Springer Handbook of Robotics2,pages v–ix. Springer-Verlag, New York, 2008. ISBN 9783319325507. → pages 7117[136] A. Sandy. Socially aware computation and communication. Computer, 38(3):33–40, 2005. →pages 41[137] S. Schaal. The New Robotics – towards human-centered machines. HFSP Journal, 1(2):115–126, Jul 2007. ISSN 1955-2068. doi:10.2976/1.2748612. → pages 8[138] M. Scheutz. The inherent dangers of unidirectional emotional bonds between humans and socialrobots. In P. Lin, K. Abney, and G. Bekey, editors, Robot ethics: the ethical and socialimplications of robotics, chapter 13, pages 205–221. MIT Press, Cambridge, Massachusetts,2012. → pages 105[139] E. E. Shriberg. Phonetic consequences of speech disfluency. In In Proceedings of theInternational Congress of Phonetic Sciences (ICPhS-99), pages 619–622, 1999.doi: → pages 37[140] C. L. Sidner, C. Lee, C. D. Kidd, N. Lesh, and C. Rich. Explorations in engagement for humansand robots. Artif. Intell., 166(1-2):140–164, Aug 2005. ISSN 0004-3702.doi:10.1016/j.artint.2005.03.005. → pages 18, 23[141] G. Skantze, A. Hjalmarsson, and C. Oertel. Exploring the effects of gaze and pauses in situatedhuman-robot interaction. In SIGdial Meeting on Discourse and Dialogue, pages 163–172.Association for Computational Linguistics, 2013. → pages 18, 23[142] J. S. Smith, C. Chao, and A. L. Thomaz. Real-time changes to social dynamics in human-robotturn-taking. In IEEE International Conference on Intelligent Robots and Systems, volume2015-Decem, pages 3024–3029, 2015. ISBN 9781479999941.doi:10.1109/IROS.2015.7353794. → pages 37, 38[143] Statistics Canada. Canada’s population estimates: Age and sex, july 1, 2015. Technical report,Statistics Canada, 2015. URL → pages 1[144] M. Staudte and M. Crocker. The utility of gaze in spoken human-robot interaction. In Workshopon Metrics for Human-Robot Interaction, pages 53–39, 2008. → pages 18[145] K. Strabala, M. K. Lee, A. Dragan, J. Forlizzi, S. S. Srinivasa, M. Cakmak, W. Garage, andV. Micelli. Towards seamless human-robot handovers. 2012. → pages 15, 16, 17[146] K. W. Strabala, M. K. Lee, A. D. Dragan, J. L. Forlizzi, and S. S. Srinivasa. Learning thecommunication of intent prior to physical collaboration. In International Symposium on Robotand Human Interactive Communication, pages 968–973, Paris, France, 2012. → pages 15, 16,18, 19, 100[147] K. W. Strabala, M. K. Lee, A. D. Dragan, J. L. Forlizzi, S. Srinivasa, M. Cakmak, andV. Micelli. Towards seamless human-robot handovers. Journal of Human-Robot Interaction, 2(1):112–132, Mar 2013. ISSN 21630364. doi:10.5898/JHRI.2.1.Strabala. → pages 5[148] C. Suda and J. Call. What does an intermediate success rate mean? an analysis of a piagetianliquid conservation task in the great apes. Cognition, 99(1):53–71, Feb 2006. ISSN 00100277.doi:10.1016/j.cognition.2005.01.005. → pages 38118[149] D. S. Syrdal, K. Dautenhahn, K. Koay, and M. Walters. The negative attitudes towards robotsscale and reactions to robot behaviour in a live human-robot interaction study. In 23rdConvention of the Society for the Study of Artificiel Intelligence and Simulation of Behaviour,AISB, pages 109–115, Edinburgh; United Kingdom, 2009. ISBN 1902956850. → pages 104[150] L. Takayama, W. Ju, and C. Nass. Beyond dirty, dangerous, and dull: What everyday peoplethink robots should do. In Proc. of Human-Robot Interaction (HRI), Amsterdam, NL, 2008.ACM/IEEE. doi:10.1145/1349822.1349827. → pages 8[151] L. Takayama, D. Dooley, and W. Ju. Expressing thought: improving robot readability withanimation principles. In Proceedings of the 6th international conference on Human-robotinteraction - HRI ’11, page 69, New York, New York, USA, 2011. ACM Press. ISBN9781450305617. doi:10.1145/1957656.1957674. → pages 11[152] C. Tennant, S. Howard, B. Franks, M. W. Bauer, S. Stares, P. Pansegrau, M. Stys´ko-Kunkowska,and A. Cuevas-Badallo. Autonomous vehicles - negotiating a place on the road: Executivesummary. Technical report, London School of Economics and Political Science Department ofPsychological and Behavioural Science; City University, London Department of Sociology,2016. URL →pages 75[153] The IEEE Global Initiative for Ethical Considerations and in Artificial Intelligence andAutonomous Systems. Ethically aligned design: A vision for prioritizing human wellbeing withartificial intelligence and autonomous systems. Technical report, IEEE Standards Association,2016. → pages 105[154] M. Tomasello, M. Carpenter, J. Call, T. Behne, and H. Moll. Understanding and sharingintentions: the origins of cultural cognition. The Behavioral and brain sciences, 28(5):675–735,Oct 2005. ISSN 0140-525X. doi:10.1017/S0140525X05000129. → pages 8, 11[155] P. D. Tremoulet and J. Feldman. The influence of spatial context and the role of intentionality inthe interpretation of animacy from motion. Perception & psychophysics, 68(6):1047–58, Aug2006. ISSN 0031-5117. → pages 10[156] V. Tsiaras, C. Panagiotakis, and Y. Stylianou. Video and audio based detection of filledhesitation pauses in classroom lectures. In 17th European Signal Processing Conference, pages834–838, Glsgow, Scotland, 2009. → pages 37[157] A. Veiga, S. Candeias, D. Celorico, J. Proenca, and F. Perdigao. Towards automaticclassification of speech styles. In H. Caseli, A. Villavicencio, A. Teixeira, and F. Perdigao,editors, Lecture Notes in Artificial Intelligence: Computational Processing of the PortugueseLanguage, PROPOR 2012, pages 421–426. Springer-Verlag, 2012. ISBN 978-3-642-28885-2.→ pages 37[158] R. Vertegaal and Y. Ding. Explaining effects of eye gaze on mediated group conversations::Amount or synchronization? In Proceedings of the 2002 ACM Conference on ComputerSupported Cooperative Work, CSCW ’02, pages 41–48, New York, NY, USA, 2002. ACM.ISBN 1-58113-560-2. doi:10.1145/587078.587085. → pages 18[159] Vicon Motion Systems Ltd. VICON, 2013. URL Accessed on2016-12-10. → pages 41119[160] A. Vinciarelli, M. Pantic, and H. Bourlard. Social signal processing: Survey of an emergingdomain. Image and Vision Computing, 27(12):1743–1759, Nov 2009. ISSN 02628856.doi:10.1016/j.imavis.2008.11.007. → pages 37, 41[161] M. Von Cranach and J. H. Ellgring. The perception of looking behaviour. In SocialCommununication and Movement. Academic Press, London, 1973. → pages 18[162] E. O. Wilson. The Insect Societies. Belknap Press of Harvard University Press, 1971. ISBN9780674454903. → pages 8[163] C. J. Wong, Y. L. Tay, R. Wang, and Y. Wu. Human-robot partnership: A study on collaborativestorytelling. ACM/IEEE International Conference on Human-Robot Interaction, 2016-April:535–536, 2016. ISSN 21672148. doi:10.1109/HRI.2016.7451843. → pages 8[164] E. S. A. S. M. H. Yang Liu. Comparing HMM, maximum entropy, and conditional randomfields for disfluency detection. In Proceedings of the International Congress of PhoneticSciences, pages 619–622, 1999. → pages 37[165] T. Yokoi and K. Fujisaki. Hesitation behaviour of hoverflies sphaerophoria spp. to avoid ambushby crab spiders. Die Naturwissenschaften, 96(2):195–200, Feb 2009. ISSN 0028-1042.doi:10.1007/s00114-008-0459-8. → pages 38[166] Yotel. Yotel, 2016. URL Accessed on 2016-12-21. → pages 7[167] M. Zheng, A. Moon, B. Gleeson, D. Troniak, M. Pan, B. Blumer, M. Meng, and E. Croft.Human behavioural responses to robot head gaze during robot-to-human handovers. In 2014IEEE International Conference on Robotics and Biomimetics, IEEE ROBIO 2014, 2014. ISBN9781479973965. doi:10.1109/ROBIO.2014.7090357. → pages iv, 15[168] M. Zheng, A. Moon, E. Croft, and M.-H. Meng. Impacts of robot head gaze on robot-to-humanhandovers. International Journal of Social Robotics, 7(5), 2015. ISSN 18754805 18754791.doi:10.1007/s12369-015-0305-z. → pages iv, 15[169] D. Zo¨bel. The Deadlock problem. ACM SIGOPS Operating Systems Review, 17(4):6–15, Oct1983. ISSN 01635980. doi:10.1145/850752.850753. → pages 82120Appendix ASupplementary InvestigationsAs discussed in the main body of this thesis, there are studies that supplement the contents of thethesis. Study 6 outlined in this Appendix is one such study the provides additional support for theimpact robot use of gaze cues can have on HR handovers (see Section A.1). In addition, presented inSection A.2 are investigations conducted on samples of human hesitation motions. These investigationshelped better understand the dataset, although they did not significantly contribute to the design ofartificial negotiative hesitation gestures of robots.A.1 Study 6: Impact of Gaze on Non-naı¨ve Handover BehaviourThis section presents a summary of a handover study that followed Studies 1 and 2 presented inChapter 3. While the results from Study 2 (Section 3.4) suggest that a robot’s use of gaze can influence ahuman’s behavioural decision on when and where to reach for the offered subject, the study was focusedon naı¨ve human responses to HR handovers. Study 6, presented in this section, investigates whether suchimpact of the robot’s use of gaze persists in non-naı¨ve handover responses. One of the aims of this studywas to reaffirm that the robot’s use of gaze during handovers positively helps the two agents interweavesubplans about when the handover should occur, thereby increasing fluency of the interaction. Whilethe results of Study 2 provide evidence of such impact of gaze, the scope of experimental result islimited to naı¨ve participants’ responses. Study 2 by itself is also insufficient to assert that the timing ofthe robot’s use of shared attention gaze or face gaze affects the timing of reaching behaviours elicitedin the human receiver. Results from Study 6 sheds light on this relationship. In particular, this studyinvestigates whether earlier timing of robot’s use of face gaze during a robot-to-human handover elicitsearlier reaching behaviours on the human receiver.The presented study was led by Minhua Zheng in collaboration with the author. It is only brieflyoutlined here in order to highlight the implications of the findings relevant to this thesis1.1 More detailed contents of this section has been published in the International Journal of Social Robotics, in which AJungMoon is a co-author: Zheng, M., Moon, A., Croft, E. A., Meng, M. Q. H., Impact of robot head gaze on robot-to-humanhandover, International Journal of Social Robotics. Minhua Zheng and AJung Moon collaboratively designed the in situexperiment, while the experiment and majority of the analysis was conducted by Minhua Zheng with the supervision of AJungMoon and Elizabeth A. Croft.121A.1.1 Experimental Procedure and HypothesesThree types of gaze patterns are implemented:Attn This condition is the same as the Attn condition employed in Study 2. The robot moved its headsmoothly and quickly from the ready position to the handover position. It stayed gazing at thehandover position afterwards, providing a prolonged shared attention gaze. Given that this wasthe best performing gaze pattern in Study 2, it is used as a baseline to compare the two newconditions (Face and LongFace-Attn) not tested previously.Face The robot moved its head from the ready position to the participant’s face, and remained gazingat the participant’s face to provide a prolonged face gaze.LongFace-Attn Similar to the Face condition, the robot shifted its gaze from the ready position to theparticipant’s face. However, afterwards, the robot moved its gaze from the participant’s face to thehandover position to arrive at the handover position approximately at the same time the gripperarrived at the handover position.In all of the gaze condition, the robot used an average velocity of 90 deg/s to shift its gaze from onepoint to another.In Study 6, the robotic platform as well as the robot’s arm motion during handovers remained thesame as in Study 2. Only the experimental gaze cues were changed in order to investigate the followinghypotheses:1. Since the use of face gaze seems to trigger gaze behaviours in humans, the Face and LongFace-Attn conditions, both exhibiting face gaze early in the handover process, will elicit an earlier reachtime from the participants than the Attn.2. Given the participants’ tendency to prefer the Turn (that includes a face gaze) over the Attn con-dition in Study 2, Participants will perceive the Face and LongFace-Attn conditions to be morefavorable than the Attn condition.The timeline of the robot’s gripper and head gaze behaviours for the conditions are illustrated inFigure A.1. The robot’s arm and gripper behaviours remained consistent across the conditions, but onlythe robot’s head motion varied according to the three conditions above. Just like the procedure used inStudy 2, in all experimental conditions, when the robot gripper moved from the grasp position to theready position, the robot head tracked the gripper; when the gripper moved from the ready position tothe handover location, the robot head moved according to the conditions.While gathering a large quantity of samples was important in the between-subjects comparison ofStudy 2, Study 6 focuses on within-subject comparisons with smaller number of subjects in order toinvestigate non-naı¨ve handover responses occurring after a set of repeated handovers. A balanced in-complete block design (v= 3,b= 3,k= 2,r= 2,λ = 1) [8] was used with six ordered pair of conditions.The condition orders were randomized across participants. Recruitment of participants involved emails,web advertisements and posters with a $5 monetary incentive. The experiment took approximately 15minutes per participant.122Figure A.1: Timeline of the robot’s gripper motion and head gaze for the conditions. The releasebottle time depends on the participant’s behaviour because the robot was programmed torelease the bottle when the participant took it. ( c©2014 IEEE)Figure A.2: Study 6 experiment set-up. The participant was instructed to take the bottle fromthe robot whenever s/he felt it was the right time to do so. After receiving the bottle, theparticipant took the bottle to a table approximately two meters behind him/her where threetubs were located. ( c©2015 Springer)123Participants read and signed a consent form before the experiment, and filled in a pre-questionnairededicated to collect demographic information. Then the experimenter briefly introduced the experi-mental procedure to the participants. The experiment included six sessions of handovers. Each sessionconsisted of two handovers and a questionnaire directed to compare the two handovers. The first threesessions included all three pairs of the conditions, and the last three sessions were a repeat of the firstthree sessions with the order of two handovers reversed in each pair. In total, each participant completed12 handovers (four trials for each condition) and six questionnaires.In each handover, the robot lifted a filled water bottle from a table located between the robot andthe participant, and handed it to the participant. The robot used gaze cues according to the conditionassigned to the trial. The participant was instructed to take the bottle from the robot whenever s/hefelt it was the right time to do so. After receiving the bottle, the participant took the bottle to a tableapproximately two meters behind him/her where three tubs were located. The tubs were labeled red,green and blue, respectively (Figure A.2). The water bottle was also labeled with one of the three colors.The participant was asked to empty the bottle into a tub that match the colour of the water bottle, putthe empty bottle into a bin, and return to the robot to start the next handover. The water pouring activityserved as a washout between handovers, and distracted the participant while the experimenter readiedthe robot for subsequent handovers.After completing two handovers, the participant compared the them on three subjective metrics(likeability, anthropomorphism and timing communication) by selecting either the first or the secondhandover to the following questions:1. Which handover did you like better?2. Which handover seemed more friendly?3. Which handover seemed more natural?4. Which handover seemed more humanlike?5. Which handover made it easier to tell when, exactly, the robot wanted you to take the object?Similar to Study 2, the participants could optionally provide additional comments.Questions 1 and 2, inspired by the Godspeed questionnaire [14], provided a measure of likeability(Cronbach’s α = 0.83). Questions 3 and 4, also inspired by [14], measured anthropomorphism (Cron-bach’s α = 0.91). Question 5, echoing one of the questions from Study 2, measured perceived timingcommunication.At the end of the experiment, an experimenter conducted a short structured interview with the par-ticipant. The first question was “Did you notice any difference between two handovers in each session?”If the participant answered “Yes”, the experimenter asked the participant to describe the difference(s).If the participant mentioned the difference in the robot’s gaze pattern or head motion by using wordssuch as “looking at”, “head”, or “gaze”, the experimenter asked about his/her opinion on or preferencefor the gaze patterns or head motions.124A.1.2 ResultsOf the 30 participants recruited, only one participant’s data was rejected as an outlier (reach timeoutside 3.6 SD). Hence, data from only 29 participants’ (17 male, 12 female; age M = 24.8,SD = 3.31)was used in the following analysis. Unsurprisingly, and echoing the findings from Study 2, there isa training effect in reach time in the first and second handovers (t(28) = 1.60, one-tailed p = 0.06).However, this disappears in later trials. Since the focus of this study is in non-naı¨ve handover responses,the analysis only includes measures from the last six HR handovers out of a total of twelve handoversperformed per participant. This effectively includes only the third and fourth times each gaze conditionwas presented to the participant. Results from a repeated-measures ANOVA confirm the findings fromStudy 2 that gaze condition significantly affects participant’s reach time during handovers. Post-hocanalyses with Bonferroni method suggest that participants reach for the object significantly earlier in theFace condition (M = 3.49,SD = 0.34) compared to Attn (M = 3.62,SD = 0.42) as well as LongFace-Attn (M = 3.61,SD = 0.44).Bradley method [9] was used to analyze the paired ranking data for likeability, anthropomorphism,and timing communication measures. The results indicate that the subjective reports vary significantlyacross gaze conditions (L > 5.99 for all three measures)2. The ranking of the conditions for the threemeasures are as follows:Likeability LongFace−Attn(0.45)> Face(0.38)> Attn(0.17)Anthropomorphism Face(0.41)> LongFace−Attn(0.39)> Attn(0.20)Timing communication LongFace−Attn(0.48)> Face(0.30)> Attn(0.22)A.2 Supplementary Investigations on Human HesitationsThis section provides investigations and results that supplement the work described in Chapter 4.As reported in Chapter 4, the author used Shooting Algorithm to find saliency of 75 trajectoryfeatures studied. The features with a non-zero weight from the Shooting Algorithm are presented inTable A.2. Figure A.3 demonstrates the regularization path of the Shooting Algorithm applied to thefour sets of motion samples tested (Nb2 = 384, Nb1 = 596, Nub2 = 1898, Nub1 = 2004). The optimum λvalue found through this approach 16 for all sample sets and is summarized in Table A.1.Those that exhibited significant mean differences in t-test or Welch test in addition to having a non-zero weight from the Shooting Algorithm were selected for analysis with SVM. Table A.3 presents thefull results of the t-tests. The analysis involving the SVM is outlined in Section A.2.In order to identify trajectory features that are different in reach and negotiative hesitation gestures,the author developed a number of SVMs using combinations of trajectory features. In addition, thevalues of ν was varied to tune the models. With the value of ν = 0.25 for the sample set, Nb1 = 596,a total of 82 (9.15%) of support vectors were used when three features (max(d˙1(t)), µδ˙ (t), and Aα1(t))were employed for an SVM. Figure A.4 illustrates the ROC curve of this model.2 Critical X2(2,0.05) = 5.99, with likeability (L = 11.03), anthropomorphism (L = 6.82), and timing communication(L = 6.61).125Table A.1: Optimum λ values obtained from Shooting AlgorithmSample Set λNb1 = 596 16Nub1 = 2004 16Nb2 = 384 16Nub2 = 1898 16(a) (b)(c) (d)Figure A.3: Regularization path for Shooting Algorithm applied to the four sets of motion samplestested. The top blue line indicates the bias variable in all cases.126Figure A.4: ROC curve of one of SVM models using max(d˙1(t)), µδ˙ (t), and Aα1(t). 9.15% of thesamples were used as Support Vectors.As mentioned in Section 4.4.3 two features (max(d˙1(t)) and µδ˙ (t)) stand out from the rest of thefeatures as strong contributors for accurately classifying hesitations from reach motions. An SVM clas-sifier using only these two features for the Nb2 = 384 sample set returns a test accuracy of 85.8%(CI : 69.4,87.0) with 50% test/train ratio, and 82.4% (CI : 68.8,92.3) with 75% test/train ratio (ν =0.3,SV = 67(17%)). This SVM model in feature space is shown in Figure 4.6.A.2.1 Correlation between FeaturesIn addition to identifying trajectory features that can be used to design artificial negotiative hesi-tation gestures for a robot, the author also sought to identify what trajectory features may be stronglycorrelated with human perception of hesitancy and persistency. Taking the Hesitency and Persistencyscores collected from Study 3, the author conducted a linear regression across the 75 features of allhesitation samples.Many of the features yielded a non-zero correlation coefficient with significance (p < 0.05). How-ever, r and R2 obtained were too small in all of the features to consider them as reliable predictors ofperceived Hesitancy and Persistency. Interestingly, the author found that the highest linear correlationoccurs between Persistency and Hesitancy scores, (r = −.92 [CI:-0.93, -0.90], R2=0.84). Figure A.5illustrates this relationship. However, this finding, although intriguing, is not useful for the purposes ofgenerating artificial hesitation trajectories. The next highest correlation with Hesitancy score was foundwith ρd¨2(t) (r = 0.35 [CI : 0.24,0.44], R2=0.12), ρδ¨ (t) (r = 0.34 [CI : 0.23,0.44], R2=0.11), and ρα2(t)(r = 0.32 [CI : 0.21,0.42], R2=0.10). No correlation with Persistency, other than Hesitancy Scores,yielded R2 > 0.1.127Table A.2: List of features with non-zero (> 0.0001) weights from the shooting algorithm. Here,S represents the number of sample sets out of four tested that have non-zero weights.FeatureSample SetSNb2 = 384 Nb1 = 596 Nub2 = 1898 Nub1 = 20041 µd¨1(t) 0 0 0.001408951 0.003123393 22 min(d¨1(t)) 0 0.000555272 0.00356456 0 23 Ad¨1(t) 0.007659523 0.004717826 0.013822345 0.03424624 44 min(d1(t)) 0.092830002 0.072352424 0.047765145 0.0524311 45 ρd1(t) 0.081822797 0.092379396 0.065071197 0.080628918 46 max(d˙1(t)) 0.15614317 0.146142143 0.129540803 0.145034212 47 µd˙1(t) 0.034778617 0.036699048 0 0 28 Ad˙1(t) 0 0 0.015197184 0.042392476 29 min(α¨1(t)) 0 0.002656939 0.002036993 0.005449252 310 max(α1(t)) 0 0 0.010195064 0.009812983 211 ρα1(t) 0.008706001 0.003330487 0.022997939 0.019261962 412 Aα1(t) 0 0 0.00079188 0.012803684 213 µδ1(t) 0.051925879 0.0585937 0.008110461 0.020604495 414 µ ¨delta2(t) 0 0 0.004541851 0.010461328 215 µd2(t) 0.043327179 0.046716392 0.014594146 0.026333731 416 µδ (t) 0.027825126 0 0.015221386 0.015401922 317 µα2(t) 0 0 0.003210239 0.007110355 218 min(α2(t)) 0 0 0.000635747 0.006038327 219 Aα2(t) 0 0.047078946 0.003604433 0 220 max(α˙2(t)) 0 0 0.003323284 0.000152331 221 Aα˙2(t) 0 0.021096329 0.001312776 0.009001647 3128Table A.3: Number of significant results (p < 0.05) found for each feature across the four runs oft- and Welch tests. The multiple runs of Welch tests on the unbalanced data are redundant.Feature Nb1 Nb2 Nub1 Nub2 Feature Nb1 Nb2 Nub1 Nub2max(d1(t)) 0 1 0 0 max(d2(t)) 4 3 4 4min(d1(t)) 4 4 4 4 min(d2(t)) 4 4 4 4µd1(t) 0 3 0 0 µd2(t) 4 4 4 4Ad1(t) 4 4 4 4 Ad2(t) 4 4 4 4ρd1(t) 4 4 4 4 ρd2(t) 4 4 4 4max(d˙1(t)) 4 4 4 4 max(d˙2(t)) 3 4 4 4min(d˙1(t)) 4 4 4 4 min(d˙2(t)) 4 4 4 4µd˙1(t) 4 4 4 4 µd˙2(t) 4 4 4 4Ad˙1(t) 4 4 4 4 Ad˙2(t) 3 4 4 4ρd˙1(t) 4 4 4 4 ρd˙2(t) 4 4 4 4max(d¨1(t)) 1 2 4 4 max(d¨2(t)) 3 4 0 0min(d¨1(t)) 4 4 4 4 min(d¨2(t)) 3 4 0 0µd¨1(t) 4 4 4 4 µd¨2(t) 4 3 4 4Ad¨1(t) 3 4 4 4 Ad¨2(t) 3 4 0 0ρd¨1(t) 4 4 4 4 ρd¨2(t) 4 4 4 4max(α1(t)) 4 4 4 4 max(α2(t)) 4 4 4 4min(α1(t)) 4 4 4 4 min(α2(t)) 4 4 4 4µα1(t) 0 0 0 0 µα2(t) 0 0 0 0Aα1(t) 4 4 4 4 Aα2(t) 4 4 4 4ρα1(t) 4 4 4 4 ρα2(t) 4 4 4 4max(α˙1(t)) 0 0 0 0 max(α˙2(t)) 4 4 4 4min(α˙1(t)) 4 4 4 4 min(α˙2(t)) 4 4 4 4µα˙1(t) 1 0 0 0 µα˙2(t) 0 0 0 0Aα˙1(t) 4 4 4 4 Aα˙2(t) 4 4 4 4ρα˙1(t) 4 4 4 4 ρα˙2(t) 4 4 4 4max(α¨1(t)) 2 3 4 4 max(α¨2(t)) 1 4 4 4min(α¨1(t)) 1 2 4 4 min(α¨2(t)) 3 4 4 4µα¨1(t) 1 0 0 0 µα¨2(t) 0 0 0 0Aα¨1(t) 3 4 4 4 Aα¨2(t) 3 4 4 4ρα¨1(t) 4 4 4 4 ρα¨2(t) 4 4 4 4Aδ˙ (t) 3 4 4 4 max(δ (t)) 4 4 4 4ρδ˙ (t) 4 4 4 4 min(δ (t)) 4 4 4 4max(δ¨ (t)) 0 0 0 0 µδ (t) 4 4 4 4min(δ¨ (t)) 1 1 0 0 Aδ (t) 0 0 0 0µδ¨ (t) 0 0 0 0 ρδ (t) 4 4 4 4Aδ¨ (t) 0 0 0 0 max(δ˙ (t)) 4 4 4 4ρδ¨ (t) 4 4 4 4 min(δ˙ (t)) 4 4 4 4µδ˙ (t) 4 4 4 4129Table A.4: Ratio of the main participant’s Euclidean distance to target at zero velocity crossingsRatio N Mean Std Var Range Min MaxZVC2 / ZVC1 179 1.0304 0.3841 0.1475 3.5800 0.2586 3.8385ZVC3 / ZVC2 112 0.9719 0.2611 0.0682 2.0391 0.3219 2.3609ZVC4 / ZVC3 56 1.1273 0.3148 0.0991 1.6052 0.7338 2.3390ZVC5 / ZVC4 28 1.0350 0.1352 0.0183 0.6157 0.8194 1.4351ZVC6 / ZVC5 20 1.0755 0.4036 0.1629 2.0191 0.5527 2.5719ZVC7 / ZVC6 10 1.0459 0.1052 0.0111 0.2880 0.9236 1.2116ZVC8 / ZVC7 7 0.9707 0.0524 0.0027 0.1316 0.8751 1.0067ZVC9 / ZVC8 4 0.9240 0.1528 0.0233 0.3097 0.6949 1.0046ZVC10 / ZVC9 4 1.0066 0.0248 0.0006 0.0538 0.9889 1.0427Mean - 1.02080 0.20380 0.05930 1.18250 0.68540 1.86790STD - 0.06100 0.14140 0.06340 1.20530 0.25910 0.97660VAR - 0.00370 0.02000 0.00400 1.45270 0.06710 0.95380Min - 0.92400 0.02480 0.00060 0.05380 0.25860 1.00460Max - 1.12730 0.40360 0.16290 3.58000 0.98890 3.83850Figure A.5: Correlation between Hesitation and Persistency scores obtained from the MechanicalTurk survey (Study 3).130


Citation Scheme:


Citations by CSL (citeproc-js)

Usage Statistics



Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            async >
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:


Related Items