Joint Effect of Attribution and Proxemics on
Interpersonal Attraction
Study on Personal Space Invasion by a Telepresence Robot
Master’s Thesis in Artificial Intelligence
Radboud University Nijmegen
Faculty of Social Sciences
J. Snippe
s4423534
Supervision:
dr. W.F.G. Haselager (Radboud University)
dr. G. Englebienne (University of Twente)
J.H. Vroon, Msc (University of Twente)
October 27, 2016
Acknowledgements
I am very grateful that I have been given the opportunity to do this amazing project as my final master’s project. I would like to express my gratitude to my three supervisors. Firstly, and most importantly, I would like to thank my daily supervisor, Jered Vroon, for his contagious enthousiasm, valuable guidance and his devotion to this project. Secondly, I would like to thank Gwenn Englebienne for his useful ideas and critical view. Thirdly, I would like to thank Pim Haselager for his inspiring suggestions and remarks. I would like to acknowledge here that we submitted a paper of this study for the Human-Robot Interaction Conference 2017 in Vienna.
Furthermore, there are a lot of people who willingly helped me with the collection of data: the participants that took the time to come to the lab, respondents of the online questionnaire and video annotators, thank you all. I would like to acknowledge that seven male students helped me by recording the videos I eventually used in the experiment, without these videos, I could not have performed the experiment.
Finally, I would like to thank Almar Snippe for his statistical advice and Gerben van Houwelingen for reviewing spelling and grammar of this thesis and for his mental support during the whole project.
Josca Snippe,
Abstract
When users share autonomy with a telepresence robot questions arise as to how the behavior of the robot is interpreted by local users. We investigated how a robot’s violations of social norms under shared autonomy influence the local user’s evaluation of the robot’s pilot users. Specifically, we examined how attribution of such violations to either robot or pilot user influences the social perception of the pilot user. Using personal space invasion as a salient social norm violation, we conducted an experiment (n=40) using a Wizard of Oz technique to investigate these questions. Participants saw several people introducing themselves through a telepresence robot, while personal space invasion and attribution were manipulated. Due to technical issues, we were forced to use two different robots with behavioural differences. We found a significant (p=0.007) joint effect of the manipulations on interpersonal attraction with one of the robots. No such effect was found for the other robot; data from video annotation suggested that less accurate robot trajectories may have modified attribution. Our results offer insights into the mechanisms of attribution in interactions with a telepresence robot as a mediator.
Contents
1 Introduction 5
2 Theoretical background 7
2.1 Mobile Robotic Telepresence Systems . . . 7
2.1.1 Autonomous behaviour of an MRP . . . 7
2.1.2 MRP and social norms . . . 8
2.2 Proxemics . . . 8
2.2.1 Personality and relations . . . 8
2.2.2 Proxemics and Impression Formation . . . 8
2.2.3 Proxemics for Artificial Agents . . . 9
2.2.4 Robot Approach . . . 9 3 Methods 10 3.1 Conditions . . . 10 3.1.1 Robot behaviour . . . 10 3.1.2 Attribution . . . 10 3.2 Task . . . 11 3.2.1 Stimuli . . . 11 3.2.2 Deceptions . . . 12 3.3 Data Collection . . . 12 3.3.1 Questionnaires . . . 12 3.3.2 Video data . . . 13 3.4 Materials . . . 13 3.4.1 Robots . . . 13 3.4.2 Room . . . 14 3.4.3 Software . . . 14 3.5 Procedure . . . 14 3.6 Participants . . . 15 4 Results 16 4.1 Descriptive Analysis . . . 16 4.1.1 Manipulation Check . . . 16 4.1.2 Favourite candidate . . . 17 4.2 Interpersonal Attraction . . . 17
4.2.1 Between subjects effects . . . 18
4.3 Personality traits . . . 18
4.4 Regression Analysis . . . 19
4.4.1 Interpersonal Attraction . . . 19
4.4.2 Personality Traits . . . 19
5 Qualitative Analysis of
Robot Differences 21
5.1 Questions and Hypotheses . . . 21
5.2 Qualitative Analysis . . . 22
5.3 Qualitative Results . . . 22
5.3.1 Compensatory Distancing Behaviour . . . 23
5.3.2 Robot Navigation Properties . . . 23
5.4 Summary of qualitative results . . . 24
6 Discussion 25 6.1 Limitations . . . 25
6.2 Contributions and Conclusions . . . 26
7 Appendix 30 7.1 Questionnaires . . . 30
7.1.1 Between-questionnaire . . . 30
7.1.2 Post-questionnaire . . . 31
7.1.3 Demographics about Participant . . . 31
7.1.4 Post-pilot-questionnaire . . . 32
7.2 IPAS per condition . . . 33
7.2.1 New Robot . . . 33
7.2.2 Old Robot . . . 33
7.3 Video Annotation Guide . . . 34
7.3.1 Introduction . . . 34
7.3.2 Properties of Robot Approach . . . 34
7.3.3 Participant’s Reaction to Robot Approach . . . 34
7.3.4 Back Channel Information . . . 34
7.3.5 Subjective Annotation . . . 34
7.3.6 Comment Box . . . 35
7.4 Video Annotation Agreements Tables . . . 35
7.4.1 Properties of Robot Approach . . . 35
7.4.2 Participant’s reaction to Robot Approach . . . 36
7.4.3 Back channel information . . . 38
1
|
Introduction
Imagine you want to be somewhere, but cannot physically go there – e.g. visiting school when you are in a hospital, a meeting on another continent, or joining activities in an eldercare facility if you’re too tired to leave your room. Mobile Robotic Telepresence systems (MRPs) might be the answer. They consist of video conferencing equipment mounted on a mobile robotic base; enabling remote users (pilot users) to be represented by a robot on a remote location and interact with the people there (local users), instead of being physically present [20]. The end goal of MRPs is to let the users forget that the interaction is mediated by a robot.
Controlling MRPs can put a high load on the pilot user, distracting them from the conversation; this can be remedied by introducing autonomous behaviours. Several studies suggest that semi-autonomous navigation is preferable for the pilot user [20, 28].
What effect do such autonomous behaviours have on local users and their impression of the pilot user? The behaviour of an MRP can be crucial for social interaction, since it could intuitively affect the impression one makes in the remote location. To our knowledge there exists no prior work investigating such effects.
Several social norms can be violated by autonomous behaviour of an MRP, such as (intimate) personal space invasion. For example, if an MRP (autonomously) stands too close to the local user, this might affect how the pilot user is perceived by the local user.
Intuitively, it also plays an important role who the local user thinks to be accountable for the violation of the social norm. It seems more appropriate to attribute the violation to the pilot user if the pilot user is manually controlling the robot, than if the robot is navigating autonomously. The local user needs to interpret whether the behaviour of an MRP was the intention of the pilot user, whether it could be attributed to –should be blamed on– a technical issue or a human error, or in the case of autonomous behaviour, the intention of the MRP. It thus appears that this can be a key modifier; we will here refer to it as attribution.
We here report on an experiment investigating how the interaction between the users is influenced by behaviour of a mediating MRP, since advancing this interaction is the main goal of the design of MRPs. We investigated whether the autonomous personal space invasion by an MRP can harm the human-human interaction. Will the local user hold the pilot user accountable for autonomous personal space invasion by an MRP? Particularly, will the local user form a more negative impression of the pilot user when the telepresence system autonomously invades the local user’s personal space?
Studies on the equilibrium model [3, 5] predict that when one’s personal space is invaded, one will show compensatory distancing behaviour, such as averting gaze or stepping away. For a robot to invade personal space, it has to be perceived as a social actor or sufficiently human-like [30]. To test whether the MRP –as a representative of the pilot user– is perceived as a social actor that can invade personal space, we formulate the first hypothesis;
Hypothesis 1: When an MRP invades the personal space of the local user, the local user will show compensatory distancing behaviour.
When personal space is invaded by someone, this person is considered to be an intruder [10]. If the personal space invasion of the robot is indeed perceived as such, and assuming that the effects of such a personal space invasion carry over to the pilot user, we would further expect that;
Hypothesis 2a: when an MRP invades the personal space of the local user, the local user will form a more negative impression of the pilot user.
We will later operationalise this ‘impression’ by using the Interpersonal Attraction scale as a measure [24].
More specifically, the effects of the personal space invasion on the impression of the pilot user intuitively seems to also be dependent on the attribution of those behaviours to either the robot or the pilot user. The more you attribute negative behaviour to a person, the more influence this behaviour has on your judgment of this person. We capture this intuition in our third hypothesis;
Hypothesis 2b: When the local user attributes the personal space invasion by an MRP to the pilot user, the local user will form a more negative impression of the pilot user.
We will again operationalise ‘impression’ by using the Interpersonal Attraction scale as a measure [24]. Telepresence robots might become a common medium to contact relatives or employees on the other side of the world in the next decades, therefore it is important to design these robots and their behaviours well. Additionally, autonomous behaviour or assisted navigation are ways to make the use of MRPs more user friendly, this is why the possible implications for these behaviours should be investigated. This study provides answers to general questions concerning MRPs: it will provide insight into the mechanism of attribution in an Human-Human interaction with an MRP as a mediator.
We conducted an experiment with an MRP using a Wizard of Oz technique. Both interaction distance (Close or Far) and Attribution (Autonomous or Manual) were manipulated, resulting in a 2x2 within subjects design. The collected quantitative data consists of the Interpersonal Attraction Scale and scores on perceived Personality Traits (Extroversion, Aggressiveness, Friendliness and Dominance). Additionally, qualitative data in the form of video recordings of the experiment were collected. After analysing the qualitative data, hypothesis 1 can be tested and after analyzing the quantitative data hypothesis 2a and 2b can be tested.
In this thesis, we first manifest a theoretical background for the present study (Chapter 2). We then introduce our experiment, wherein we test the (joint) effects of personal space invasion and attribution on interpersonal attraction in interactions mediated by an MRP (Chapter 3). Our results support some of our hypotheses (Chapter 4). Technical issues forced us to use two different robots with minor differences, of which we found an effect, which indicated the importance of ‘trajectory accuracy’ for attribution (Chapter 5). Our work contributes various insights into the mechanism of attribution in an interaction with an MRP as a mediator (Chapter 6).
2
|
Theoretical background
In this section, we manifest a theoretical background for the present study. First, some studies about MRPs, their autonomous behaviours and possible social norms are discussed. Additionally, since the experiment will entail Personal Space Invasion, literature about Proxemics and the relation to personality and impression formation are deliberated. At last, some literature about proxemics in the context of robot approach is discussed.
2.1
Mobile Robotic Telepresence Systems
An MRP, already introduced in Chapter 1, enables a pilot user who is remotely controlling the MRP, to interact with a local user in a local environment [20]. The pilot user is thus represented by the MRP in the local environment and the pilot user can feel telepresent in the local environment, which is the sense of facing the remote local user in the same room. Figure 2.1 illustrates the interaction setup enabled by an MRP.
The interaction enabled by an MRP is complex since there are numerous different components: Human-Robot Interaction between the pilot user and the MRP, Human-Computer Interaction between the pilot user and the control system interface of the MRP, Human-Robot Interaction between the local user and the MRP and Human-Human Interaction between the pilot user and the local user. This complexity makes it challenging to study the effects of MRP behaviour on its users.
The main aim of MRPs is to advance social interaction between its users. Ideally, use of an MRP can lead to forgetting that the pilot user is not personally present in the same room [20]. Being socially represented and therefore forgetting that one is not physically present, should therefore be the goal of telepresence systems.
2.1.1
Autonomous behaviour of an MRP
For novice users it can be quite intensive to control an MRP and therefore the goal of advancing social interaction might not be reached successfully. When the task of driving the MRP has a high work load, the pilot user might not be able to focus entirely on the social interaction. Therefore, many projects, like the TERESA project [34], focus on the development of automatic social behaviour, such as social navigation and social conversation. It has been shown [28], that a system with semi-autonomous navigation control
Figure 2.1: Interaction setup enabled by an MRP. The pilot user is represented by the MRP in the local environment.
and semi-autonomous people tracker, yielded significantly higher user satisfaction compared to a system without assisted control.
Unfortunately, autonomous behaviour of MRPs has a downside as well. User-controlled movements of MRPs produce a stronger feeling of social telepresence for the pilot user than automatically generated movement [26].
While the effects of autonomous behaviour on the pilot user have thus been studied, we are not aware of any previous work investigating the effects on local users. This project looks into the effects on the local user. What happens if the autonomous behaviour violates social norms in the local environment?
2.1.2
MRP and social norms
While conversing through an MRP, social norms and standards should still be taken into account. If social norms are violated by an MRP or when the MRP is not working correctly, this can have consequences for the social interaction, like the following studies show.
If a technology, or in this case an MRP, violates social norm, the violation is viewed as social incompetence and it can be offensive [7]. It has been shown [23], that when volume of an MRP was too loud, local users described that the pilot user was disturbing workplace. Despite the fact that the pilot user was not aware of the fact that the volume was too loud, the pilot user was held responsible for this act trough the MRP.
Not only MRP behaviour can lead to different perceptions of a pilot user, failure of video conferencing itself can have this effect as well. In [7], it is shown that a delay of audio and video causes users to evaluate a speaker as less interesting, less pleasant, less influential, more agitated and less successful in their delivery, even if they did not notice the asynchrony itself.
2.2
Proxemics
Proxemics is the study of personal space and interpersonal distance. Hall [10] defined the term personal space as a psychological and physical buffer zone towards other people. The distance people take towards another person is affected by several factors, such as gender, environmental factors, age and personality traits [1].
Several models exist to explain and predict proxemic behaviour. One example is the equilibrium model that predicts that if someone invades your personal space, you will show compensatory distancing behaviour [3].
Since proxemic behaviour can be influenced by many factors and is very dependent on personal differences, it is challenging to study the effects of autonomous proxemic behaviour by an MRP on its local users.
We will here give a brief overview of how proxemics has been found to influence impression formation, and how it has been used in interactions with artificial agents.
2.2.1
Personality and relations
Personality traits that play a role in proxemics are extroversion [44], agreeableness and neuroticism [37]. Extrovert personalities allow people to approach closer to them than introvert personalities do [44]. Positive affect, friendship and attraction are associated with smaller interpersonal distance [36]. High vertical dimension (dominance, power and status) is associated with smaller interpersonal distance [11].
2.2.2
Proxemics and Impression Formation
It has been shown [27], that interpersonal distance has an effect on impression formation of personality during an interview setting. With larger personal distances, aggressiveness, friendliness, extroversion and dominance of the confederate were rated lower [27]. The tested interpersonal distances were 2, 4, 6, and 8 feet (60cm, 1m 20cm, 1m 80cm and 2m 40cm).
Not all studies show effects of interpersonal distance on impression formation [40]. Tesch tested interview settings in social and personal space zones and measured the impression formation of the confederate with FIRO-B [33] and the adjective checklist and found no significant effects between the two proxemic conditions.
In a virtual reality study [6], users judged the personality and interpersonal attitudes of virtual agents in first encounters, i.e. the first 12,5 seconds of an interaction. Smile, gaze and proxemics of the virtual agent were manipulated and perceived extroversion, friendliness and likability were tested. Results show that proxemic approach has a main effect on the extroversion judgment and smile has a main effect on the judgment of friendliness.
2.2.3
Proxemics for Artificial Agents
There has been a wide range of work investigating proxemics for robotic and virtual agents, of which we will here only discuss a few.
It has been shown [46], that the personal space zone in Human-Robot Interaction has the same circular shape and the same size as the personal space zone in Human-Human Interaction. Experience with robots decreases the distance to maintain towards robots [37].
Social norm, like personal space invasion, can only be violated by entities with a certain human likeness, or social actors, therefore the human likeness of robots and other technology is crucial [30]. In [15], participants reacted stronger to personal space invasion by robots in negative and positive emotional responses than to personal space invasion by humans. Therefore, it is concluded that social norms in Human-Robot Interaction are different from social norms in Human-Human Interaction.
Non-humanlike characters in virtual reality were approached closer than humanlike characters [5]. It has been shown [7], that personal space can even be invaded by pictures of human faces. Pictures of faces that appeared close, a close-up picture or a small physical distance to the picture, were evaluated more intense, both positively and negatively. Additionally, attention and memory of the pictures of faces that appeared more close, were enhanced.
In [19], proxemic and gaze behaviours of characters in virtual reality were found to have an effect on intimicy, in line with what the aforementioned equilibrium theory would predict.
Together, these works show that proxemics play an important role in the perception of such artificial agents, similar but not always equal to the role proxemics play in human-human interaction.
2.2.4
Robot Approach
Similar findings have been reported when proxemics are used in the context of approach. It has been shown [42], that approach distances for humans approaching a robot and a robot approaching a human are comparable, although the same study shows that there are indications that humans tend to approach robots more closely than they allow robots to approach them. Another study [22] shows that the freely chosen minimal frontal interaction distance with a mechanical robot is greater than 0.45m and therefore outside of the intimate zone.
In another study [17], the interaction distance, power distance and task distance were manipulated to investigate how social distance shapes Human-Robot Interaction. How the interaction distance was perceived depended on the power distance of the robot: user experience was high when the supervisor robot was physically close and when the subordinate robot was physically distant. Additionally, how the interaction distance was perceived also depended on task distance, user experience was high when the cooperating robot was physically distant and when the competing robot was physically close.
In conclusion, for various artificial agents, people did not choose or feel comfortable with distances of less than 45cm, which suggest that using such distances would violate social norms. In line with the social zones of Hall [5], a distance of 70-80cm seems to be socially normative for artificial agents.
3
|
Methods
The experiment had a 2x2 within subjects design. The factors that were manipulated were proxemics (the robot approached the participant at a close or far distance) and the attribution (the participant was either told that the robot navigated autonomously or by the candidate). So there were four conditions: ManualClose, ManualFar, AutonomousClose and AutonomousFar. All participants were tested in all four conditions in counterbalanced order.
In advance of the experiment, a pilot study was conducted to test the experimental setup and to incorporate input given by the participants into the final experimental choices. Four participants (2 female) were tested in the pilot study and their questionnaire data and video data were collected. The questionnaires that were used for the pilot study included the questionnaires that were used for the final experiment and an additional post-pilot-questionnaire to provide input on experimental choices, see appendix. The pilot study was successful and the participants showed compensatory distancing behaviour. Some issues came to attention and were changed for the design of the final experiment and will be discussed throughout this section.
3.1
Conditions
3.1.1
Robot behaviour
In all four conditions, the robot navigated towards the participant. It either stopped at 70-80 cm (far condition) from were the participant stood, as shown in figure 3.1, or at 30-40 cm (close condition), as shown in figure 3.2. A grid of markers was placed on the entire floor of the experimental room at every 30cm to control the proxemics. The driver confederate could both keep track of those proxemics indicators via the camera of the robot itself, as on the video recorders placed in the room.
Since the manipulated robot behaviour is the approach proxemics, the robot did not move during the interaction, even when the participant showed compensatory behaviour which increased the proxemics. The strategy that was used by the driver confederate, throughout the experiment, was to choose the final position before moving towards the participant, since participants often stepped away, the desired proxemics was often not even reached.
The navigation of the robot used a Wizard of Oz technique. In the first two pilot study experiments, the experimenter navigated the robot which turned out to be completely clear to the participants. For the other two pilot study experiments, a driver confederate that was invisible to the participants navigated the robot, this turned out to be a convincing Wizard of Oz design and was used for the rest of the experiment.
The pilot study also pointed out that the robot occasionally had trouble with refreshing the video of the navigation software. Due to this the robot could not be navigated correctly, it had to stand still and wait for the video to be refreshed, or go on without knowing were the robot exactly was. For this reason, it was decided to connect one of the external video cameras in the experimental room to a computer screen in the driver confederate room. If the video of the robot would freeze, the driver confederate could use the video provided by the camera to navigate the robot.
3.1.2
Attribution
The participant was told at the beginning of the experiment that the robot navigated autonomously or by the candidate (depending on the experimental condition). Halfway through the experiment, between the second and third trial, the attribution switched.
Figure 3.1: Far interaction distance Figure 3.2: Close interaction distance
3.2
Task
The experimental setting was introduced to the participants as an event to choose a new roommate. The participants were told that four male candidates were going to introduce themselves to the participant and the participant would fill in a questionnaire after each introduction round and afterwards they could choose which of the candidates was their favourite.
3.2.1
Stimuli
Four videos of male students with an age between 21 and 23 were used for the experiment. The video duration varied between 1m11s and 1m21s. All videos started with a few seconds of silence in which the robot could navigate and ended with a few seconds of silence as well. During the videos, the candidates introduced themselves and told something about their hobbies and explained what kind of roommate they would be. The order in which the stimuli occurred during an experiment were counterbalanced over participants using a balanced Latin square design.
The videos were recorded with the webcam of a Samsung series 9 laptop with the software Cyberlink Youcam. We made sure that, while capturing the videos, the distance from face to the camera was similar for all videos, so all faces would appear the same size on a screen. Additionally, the distance towards the camera was as small as possible, so the faces would appear to be approximately real life size on the screen of the robot.
All four participants of the pilot study reported that they recognized that the stimuli were pre captured videos instead of life people talking. We decided not to change the stimuli because of feasibility and consistency issues.
Additionally, all four participants of the pilot study chose the same favourite candidate. This might indicate that there was one video that was much more likable than the others. Therefore an online questionnaire was set up to determine average scores for all videos that were used and the videos that were not selected for the pilot study.
Online questionnaire
The questionnaire was presented via surveymonkey and the videos were put online in a private channel on youtube, via a link in surveymonkey the respondents could watch the videos. In total 7 videos were recorded of which originally 4 were used in the pilot study. One of the candidates did not give permission to put his video on youtube, so this video was left out of the rest of the experiment. The six videos were not played in random order.
Because of privacy issues and possible participant issues, the questionnaire was only shared with people that already indicated that they wanted to fill in the questionnaire. Eventually, 8 respondents (6 female) filled in the online questionnaire. The age of the respondents varied between 19 and 60 with a mean of 36 years.
The respondents were asked to watch the video and then they had to indicate whether they already knew the person in the video, if that was the case, they were not allowed to fill in the rest of the questionnaire about that particular video. Next, the respondent was asked to fill in the Interpersonal
Attraction scale and rate the candidate on the four personality traits that were used in the pilot study as well. The mean scores of IPAS and the Personality Traits are presented in Table 3.1.
Table 3.1
Candidate IPAS Extroversion Aggressiveness Friendliness Dominance Number of respondents → 1 34.26 2.33 1.38 5.50 1.75 8 2 40.7 4.33 1.67 5.00 4.00 3 → 3 38 4.13 2.13 5.38 2.75 8 → 4 34.34 2.83 1.83 5.00 2.83 6 → 5 37.85 4.8 2.14 5.14 2.42 7 6 49.27 5.86 1.71 6.14 3.42 7 Based on these results, the four most similar videos were selected (indicated by →). Although the group of respondents of this online questionnaire does not resemble the group of participants of the experiment, the outcomes of the online questionnaire are in line with the outcomes of the pilot study and the presumed rankings in response to the pilot study. Candidate 6 is left out, because it is a very positive outlier in both the online questionnaire and the pilot study, since this was the video that was chosen as a favourite by all four participants of the pilot study. Candidate 2 is also left out, because there are only 3 respondents for this video, which was because the other respondents already knew this person. The video that was not included in the online questionnaire, because no permission was provided, was left out. So we decided to use the videos of 1, 3, 4 and 5 in the final experiment.
3.2.2
Deceptions
There were several deceptions in this experiment. First of all, we tried to make the participant believe that the candidates made a live connection with the robot, instead of playing a prerecorded video, the participant was told that the wifi connection was very bad. As a result of this the candidates would not be able to hear the participant, so the participant could not ask questions or have a conversation with the candidate. There was thus a deception concerning the videos and the pretext concerning the wifi connection.
Another deception is the Wizard of Oz technique in general, there was a driver confederate responsible for the robot navigation. The participant was told at the beginning of the experiment that the robot navigated autonomously or by the candidate (depending on the experimental condition). Halfway through the experiment, between the second and the third trial, the participant was told that due to the bad wifi connection, there was an error so the robot from that point on would be navigating autonomously or by the candidate (depending on the experimental condition). We made sure that there was only one switch of attribution during an experiment, since otherwise it could be confusing or unbelievable. So, there was a deception about how the robot was actually navigated and another pretext about the wifi connection.
3.3
Data Collection
Quantitative data has been collected in the form of questionnaire data and qualitative data has been collected in the form of video data.
3.3.1
Questionnaires
All the questionnaires used for this study are translated into Dutch with the help of a native speaker of English and one of Dutch. The Dutch version of the questionnaires can be found in the appendix. The questionnaires were presented via surveymonkey.
Between-questionnaire
After each introduction round, the participant was asked to fill in the between-questionnaire, which consisted of three parts. The first part of the between-questionnaire, consisted of the Interpersonal Attraction scale [24]. We chose this scale since it gives a measure of how much the local user is attracted to the pilot user, which seems to be relevant for the setting of choosing a roommate as well. It is expected,
see hypothesis 2 and 3, that proxemics and attribution affect interpersonal attraction. The scale consists of three parts: Social Attraction, Physical Attraction and Task Attraction. For this study, only Social Attraction and Task Attraction are used, since physical attraction does not seem to be relevant. Each of the parts consists of 10 questions, which were reduced to the 5 questions with the highest loading (more than .49).
The second part of the between-questionnaire is a set of personality traits of which the participant had to answer whether the candidate appeared to have the personality trait, similar to [27]. In [27], it is already indicated that proxemics affect these personality traits.
In the last part, the participants had to rate how suitable they thought the candidate would be as a roommate.
Post-questionnaire
After all four questionnaires about the candidates were filled out, the participant had to answer which one of the candidates they would choose as a new roommate and why. This question was used to provide insight in the distribution of favourite candidates.
After that, the participants were asked how they felt during the interaction with the robot and whether they would want another interaction with the robot. These neutral question were used to make the participant think about the robot and the interaction without triggering the personal space invasion directly. After answering these questions, the question appeared whether the robot came too close and when. At last, the participant was asked for each candidates individually, whether the candidate was responsible for the navigation of the robot. These last two questions were used as manipulation checks. Demographics about participant
The last questionnaire entails general information about the participant (age, gender, nationality and education). A reduced version of the Big Five-questionnaire is used, similar to [6], to get a score on Agreeableness, Neuroticism and Extroversion. Additionally, the NARS is reduced as well, only the items with the highest loading per sub scale are used.
3.3.2
Video data
Video recordings of the experiment were made to be able to check particular features later, if this would turn out to be necessary. Two camera’s were placed in the experimental room, one facing the floor to capture the proxemics and one camera facing the participant to capture body posture and facial expressions. Both camera’s were pointed at the participant, meaning that the first part of the robot approach was not visible on the recordings.
3.4
Materials
3.4.1
Robots
A Giraff 4.0T telepresence robot (from 2014) was used to run the experiment, this robot is later referred to as the New robot. This robot was situated in the docking station between the trials. As in the pilot study, the software interface of the navigation software tended to freeze during the navigation, that is why the floor-camera was connected to a computer screen via an hdmi-cable so the driver confederate could use that camera to navigate the robot when the software froze. This robot was connected to the control software via a local network.
After 20 participants, the battery of the robot died and we had to switch to an older version of the Giraff 3.3 (from 2012), which is later referred to as the Old robot. This robot had a wheel that did not function correctly, which affected the speed, precision and fluency of the navigation. Instead of using the docking station, this robot was plugged into the charger by the experimenter by hand, also between trials. This robot could also be navigated via the servers of Giraff, so we used the campus network, which resulted in a somewhat unstable connection.
Figure 3.3: Schematic overview of experimental setup (not on scale). The circles from the participant represent the Far en Close stop distances. The figure illustrates a straight trajectory and a curved trajectory (as often occurred when using the Old Robot)
Figure 3.4: Robot in experimental room
3.4.2
Room
The room that was used for the experiments, is the demo lab for human media interaction at the University of Twente. Figure 3.3 shows the overview of the experimental room and figure 3.4 shows the robot during the experiment, turned around towards the participant and with a video playing.
3.4.3
Software
For controlling the Giraff robot, software provided by the Giraff company was used. The new Giraff robot was navigated using Pilot local software via a local network and the old Giraff robot was navigated using Giraff pilot software on the University’s wireless connection. Both software packages navigate the same way, only the Internet connection and Giraff servers differ.
Manycam 5.2.0 was used as a virtual webcam to play videos on the Giraff.
3.5
Procedure
The participant was welcomed into the room by the experimenter. The experimenter then gave the introduction: first some information about the robot and the way it was navigated (depending on the attribution condition). After that, the experimenter explained the experimental setup and task, then the experimenter explained that the interaction would be filmed but that there would be no right or wrong behaviour. When the participant had no questions, the participant was asked to fill in the informed consent form. The experimenter turned on the camera’s.
When the participant was ready and stood on the floor, the experimenter initiated the first interaction video (new robot: pushing red button, old robot: using touch screen) and walked away. Then the robot turned around and approached the participant while playing the first interaction video. After the introduction, the robot made a u-turn and drove back to the starting location. During the retreat of the robot, the experimenter appeared to ask the participant to fill in the first questionnaire.
This interaction was repeated four times for each participant, with four different candidate videos. Between the second and the third video, the attribution switched, see section 3.2.2.
After the fourth trial, the participant was asked to fill in the last questionnaire. The experimenter turned off the camera’s. When the participant was finished with the post-questionnaire, the experimenter explained the goal of the experiment in a debriefing.
3.6
Participants
40 participants were tested for this experiment of which 17 were female. The mean age was 22.46 with a standard deviation of 2.87. The youngest participant was 19 years old and the oldest participant was 31. One of these participants only performed three of the four trials, since the New robot shut down after the third trial and could not be restarted. This participant was asked to fill in the post-questionnaire anyway, which data was used in the final analysis. Another participant only performed two of the four trials, due to technical complications with the New robot. Three additional participants could not start the experiment and were excluded completely.
In total 20 participants were tested with the New robot (of which one only completed two trials and one only completed three trials) and 20 participants were tested with the Old robot.
4
|
Results
In this chapter we present the results of the quantitative data analysis. First, we will discuss some descriptive analysis: the manipulation checks of the Proxemic condition and the Attribution condition, according to the questionnaire data, and the distribution of the chosen favourite candidate.
Secondly and most importantly, the results of the analysis of the Interpersonal Attraction Scale will be discussed. According to the Friedmann’s test, the non-parametric equivalent of the repeated measures ANOVA, the experimental condition had no significant effect on the Interpersonal Attraction Scale (IPAS). However, since we used two different robots for the experiment (section 3.4.1), we also compared the scores between the robots and this turned out to have a significant effect on the IPAS. Additionally, according to a repeated measures ANOVA, the experimental condition, Attribution and Proxemic, had a significant joint effect on the IPAS when tested with the New robot, while there was no such an effect for the Old robot.
Although there was no clear hypothesis formulated, the scores on the Personality traits are analysed as well. Additionally, some multiple regression analysis for both IPAS and the Personality traits are performed.
4.1
Descriptive Analysis
4.1.1
Manipulation Check
In the post questionnaire (section 3.3.1) the participants were asked whether they thought the robot came too close and if this was the case they had to answer an open question about when the robot came too close. Additionally, the participants were asked in a closed question for each candidate individually, whether this candidate was responsible for the robot navigation. These two questions were used for the manipulation checks for respectively the proxemic condition and the attribution condition. However, since the question for the proxemic condition check was open, not all participants gave a clear answer that could be translated into a manipulation check.
Proxemic condition
38 of the 40 participants indicated that the robot came too close, so only 2 participants reported that the interaction distances were suitable. In total, 8 participants reported that the robot came too close in all conditions. To the question when the robot came too close 9 participants did not give a clear answer, so these answers are left out of the decision table. The answers of the participant that did not complete the experiment correctly were left out. All 4 trials of each participant, of which 2 were intended to be close and 2 were intended to be far, are displayed in the decision table 4.1. The total number of responses is 120.
There is a significant (p<0.001) correlation between the manipulated and perceived proxemics (Pearson Correlation Coefficient 0.568).
Attribution condition
All 4 trials of each participant, of which 2 were intended to be manual and 2 were intended to be autonomous, are displayed in the decision table 4.2. Two participants did not complete all four trials, but instead completed 2 and 3 trials (section 3.6). The total number of responses is 157.
Table 4.1: Manipulation check for proxemic conditions Manipulated
CLOSE FAR Perceived CLOSE 45 11
FAR 15 49
Table 4.2: Manipulation check for attribution conditions Manipulated
Manual Autonomous Perceived Manual 65 10
Autonomous 12 70
There is a significant (p<0.001) correlation between the manipulated and perceived attribution (Pearson Correlation Coefficient 0.720).
4.1.2
Favourite candidate
The distribution of the chosen favourite candidate, the answer to the question which candidate the participant would choose as a roommate, is shown in figure 4.1 Candidate. It appears that the order in which the video’s occurred also had a big effect on the answer of this question, as is shown in figure 4.1 Trial. So, it seems that the first and the last video had the biggest chance to be chosen as the favourite candidate. Although, this question did not contain free recall, since there were pictures of all candidates presented along with the question, the effect that the first and last videos were chosen more often than the second and the third might be related to the primacy and recency effect in free recall [4].
Figure 4.1: Histograms of the number of occurrences a certain candidate or trial was chosen as a favourite candidate.
4.2
Interpersonal Attraction
In this section the analysis of the data obtained from the Interpersonal Attraction Scale (IPAS) will be discussed.
To check for internal consistency, we first calculated the Cronbach’s alpha for the IPAS. The score was 0.869, which is considered to be good. In advance, the IPAS scores were calculated for each trial per participant. The IPAS score turned out to be strongly correlated (p<0.001) with the score on the question whether the candidate was suitable as a roommate (Pearson Correlation Coefficient .753).
To perform a repeated measures ANOVA, the data needs to be normally distributed. A Shapiro Wilk test was performed on the seperate data of all four conditions. The data of one condition was significantly different (p=0.013) from a normal distribution, so a repeated measures ANOVA could not be performed. The non-parametric alternative, the Friedman’s test, was performed instead. There was no statistically significant difference in IPAS on Proxemics or Attribution, X2(2) = 0.804, p = 0.849.
Figure 4.2: Mean IPAS per condition for New Robot and Old Robot with 95% confidence intervals. The joint effect of Proxemics and Attribution is significant (p=0.007) for the New Robot.
Since there were two different robot used for the experiment (section 3.4.1), the scores between the robots are compared as well and since the robot causes a significant difference in IPAS, the scores for both robots are considered separately.
Comparison between the two robots
A Mann Whitney U test (MWU test) was performed on the IPAS to compare the scores of the two robots. The combined scores of the two groups were tested with a MWU test, and differ significantly (p=0.042). Since the Shapiro Wilk test indicated that the two distributions can be considered as normally distributed, an independent t-test was performed as well. According to the Shapiro Wilk test, however, the standard deviations are quite different (21.6 and 16.6). An independent t-test is performed and both when equal variances are assumed and when this assumption is not met, the difference is not statistically significant (p=0.071 and p=0.075).
Since the data of the two robots are significantly different, which robot is used can be seen as a (unintended) between subjects factor. Therefore, two repeated measures ANOVA were performed on the dataset, one on the data of the New robot and one on the data of the Old robot. First the Shapiro Wilk test was performed to indicate whether the data was normally distributed and whether an ANOVA could be performed, all four conditions of the two groups could be considered as normally distributed.
In the data of the New Robot, there was a statistically significant joint effect of Proxemics and Attribution (p=0.007) on IPAS. There was no statistically significant effect of Proxemics (p=0.665) and Attribution (p=0.514) separately on IPAS. In the data of the Old Robot, there was no significant effect of Proxemics (p=0.971), Attribution (p=0.737) or Proxemics and Attribution together (p=0.471). Figure 4.2 shows the mean IPAS per condition of both robots.
4.2.1
Between subjects effects
Some statistical tests were run to investigate whether there were between subjects effects for the IPAS. No such effects were found.
4.3
Personality traits
Since the data collected for the personality traits is non-parametric, the non-parametric equivalent of the repeated measures ANOVA, the Friedman’s test, was used to investigate within subjects effects between the conditions, however, no significant effects of experimental condition were found. The test was performed on Extroversion (p=0.301), Agressiveness (p=0.208), Friendliness (p=0.919) and Dominance (p=0.293).
Since there was a significant difference of IPAS between the two robots, we compared the scores on personality traits of the two robots.
Comparison between the two robots
Similar to the comparison of IPAS between the two robots (section 4.2), the scores for the personality traits were also combined per participant to compare the scores of the two robots. The MWU test was performed to compare the scores between the two robots of Extroversion (p=0.965), Friendliness (p=0.156), Dominance (p=0.385) and Agressiveness (p=0.188). The personality traits scores are not significantly different between the New robot and the Old robot.
To investigate whether there were within subjects effects within the two groups, the Friedman’s test was performed for all four personality traits tested with the New robot and tested with the Old robot. The p-scores are respectively for the New robot and the Old robot: Extroversion(p=0.146 / p=0.514), Agressiveness (p=0.151 / p=0.132), Friendliness (p=0.612 / p=0.852), Dominance (p=0.901 / p=0.135). There is no significant within subjects effect found on personality traits.
4.4
Regression Analysis
We performed a regression analysis on both Interpersonal Attraction and Personality Traits. The data that was used as covariates consisted of data from the participant (self-rated personality traits, NARS, gender and Experience with robots) and data from the experimental conditions (Candidate number, Trial number, which Robot is used and Condition).
The scores of the personality traits of the participants were each based on 4 questions with likert scales. Chronbach’s alpha scores were calculated for self-rated Extroversion (0.865), self-rated Agreeableness (0.313) and self-rated Neuroticism (0.358). The Chronbach’s alpha for the 3 questions from the NARS was (0.453). Despite the fact that these Chronbach’s alpha scores are low or even unacceptable, the scores of each participant were used in the regression model.
The data collected on education was left out the analysis, since the question turned out to be ambiguous.
4.4.1
Interpersonal Attraction
We performed a multiple regression analysis to find which of the covariates are significant predictors for IPAS. To predict the Interpersonal Attraction, the following covariates turned out to be significant predictors:
• Candidate is a significant predictor of IPAS, despite the pre-check in the online questionnaire to select the most similar videos. This seems to be in line with the findings in section 4.1.2.
• Robot is a significant predictor of the Interpersonal Attraction score, this is in line with the findings in section 4.2.
• Age is a significant predictor for the Interpersonal Attraction score. Since the questionnaire consists of questions like ‘This person could be a friend of mine’, and ‘This person would not fit into my group of friends’ it is not surprising that age plays a role in assigning these scores. When the participant is much older than the candidate, it would generally be less likely that they would be friends.
4.4.2
Personality Traits
A multivariate linear regression analysis was performed to indicate which covariates are significant predicors for the personality trait scores of Extroversion, Friendliness, Dominance and Aggressiveness. The significant predictors can be divided into a few groups:
• Personality of participant. The regression analysis pointed out that the personality of the participant is a significant predictor for the personality traits the participants attribute to the candidates. However, note that the Chronbach’s alpha’s were unacceptably low for Agreeableness and Neuroticism (Section 4.4).
Extroversion of participant is a significant negative predictor for Friendliness, a significant positive predictor for Dominance and a significant negative predictor for Aggressiveness.
Agreeableness of participant is a significant negative predictor for Friendliness and a positive predictor for Dominance.
Neuroticism of participant (a high rating means that a participant is more relaxed and a low rating means that a participant is more neurotic) is a positive predictor for Friendliness. • Trial is a significant positive predictor for Extroversion and Friendliness. So, the candidates that
were presented in later trials, were rated more Extrovert and Friendly.
• Condition is a predictor for Friendliness. How to interpret this is not visible in the B-value only, since the different conditions are represented with a discrete number (1,2,3,4). The mean scores for Friendliness per condition are CloseAutonomous=5.575, CloseManual=5.325, FarAutonomous=5.625, FarManual=5.175.
• Robot is a positive predictor for Dominance. The scores for Dominance are significantly higher for the old robot, compared to the new robot.
• Experience is a positive significant predictor for Aggressiveness.
4.5
Summary of quantitative results
The IPAS from the experimental conditions were not significantly different according to the Friedman’s test. However, when considering the scores for the two different robot separately, according to the MWU test the IPAS scores for the different robots differ significantly (p=0.042), there is an significant joint effect of proxemics and attribution together (p=0.007) for the New robot.
There were no effects of the experimental condition found on the scores for Personality Traits. To investigate what caused the difference in IPAS between the two robots, the recordings of the experiments were annotated by human annotators, which will be discussed in the next chapter.
5
|
Qualitative Analysis of
Robot Differences
In the quantitative analysis, a significant within subjects effect was found for the IPAS obtained with the New robot, while no such effect was found for the Old robot. There are several possible explanations for this difference between the two robots. Some physical differences between the robots that could have been responsible for behavioural differences are discussed here.
Firstly, the navigation of the Old robot appeared to be worse than the navigation of the New robot. Where the New robot approached the participant in a nearly (not perfectly) straight line, the Old robot often made curved trajectories (see figure 3.3).
Secondly, it also took more time for the Old robot to navigate to the final position, compared to the New robot. Due to this, in almost all cases the verbal video introduction started during the approach, while in the approach with the New robot, the video started while the robot was at least nearer to the participant.
Thirdly, since the Old robot had a non-functioning wheel, it often made a turn to the right after starting the approach to the participant (see figure 3.3). As a result, the Old robot was not always exactly facing the participant when the verbal video introduction started.
These three physical differences between the robots can have had influences on interpretation. First of all, it might be possible that because of the non fluency of the Old robot’s trajectory, the participant thought all robot behaviour was due to technical issues. Secondly, since impression formation takes place in the first seconds of a conversation [6], it is possible that when the Old robot had not invaded the personal space of the participant during the impression formation, the robot behaviour had no effect on the impression formation as measured by the Interpersonal Attraction.
Two additional differences between the robots, that might have had an effect, are that the Old robot was unplugged by the experimenter between the trails and the older design of the Old robot. However, the effects of these two possible explanations cannot be annotated using the video data of the recordings of the experiment.
To investigate which of the above explanations is the most probable explanation for the significant difference in IPAS between the robots, the recordings of the experiments were annotated by human annotators.
Additionally, we still needed to test hypothesis 1, stating that participants show compensatory distancing behaviour when the MRP invades personal space. Therefore a part of the video analysis concerns compensatory distancing behaviour.
5.1
Questions and Hypotheses
The complete list of questions used for the video annotation as well as the additional information for the annotators can be found in Appendix 7.3.
The first three questions of the video annotation concern the properties of the robot approach: whether the robot drives in a (more or less) straight line, whether the robot is facing the participant when the video introduction starts and whether the robot is facing the participant in the final position. It is hypothesised that the old robot is more often not facing the participant when the video starts, as well as in the final position compared to the new robot. Additionally, it is hypothesised that the old robot’s navigation was worse compared to the new robot and that this resulted in a trajectory towards the participant that was not more or less straight.
The question about the participant’s reaction to the robot approach will provide information to test hypothesis 1. “When the robot reaches the final position, the participant: Steps Aside(2 feet)/ Steps Towards robot(2 feet)/Steps Away from Robot(2feet)/Hesitates to step(1foot)/Leans Away From Robot/Leans Towards Robot/Does not move.” Only the answer ’Steps Away from Robot’ was considered as compensatory distancing behaviour. The other categories were included to make distinctions between those behaviours that were not considered to be compensatory distancing behaviours and actually stepping away from the robot.
The question concerning back-channel information concerns the engagement of the participant in the conversation. The question is merely a check which of the nonverbal and verbal back-channel cues occur in the recordings. Giving back-channel information might indicate that the participant was not yet aware of the fact that the candidate was in fact a video.
The last three questions are the most subjective questions. To investigate whether the participant might have attributed the robot behaviour to the human navigator or to a technical issue, the video annotators were asked to rate the human navigator’s skills and whether they thought there was a technical issue causing bad robot behaviour. These questions might provide insight in the interpretation of the participants. When there was bad robot behaviour it could either be a clear human error or a clear technical issue or something in between. So the subjective interpretation of the human annotator is necessary for the analysis of this question. It is hypothesized that the behaviour of the Old robot was attributed to technical issues, while the behaviour of the New robot was attributed more to the navigator. This is based on the results that the IPAS was significantly higher for the old robot compared to the new robot.
The last subjective question asks the annotator about the interpretation of the speed of the robot, whether it is too fast or too slow. This question is added after reading comments in the questionnaire of the experiment about the fast approach of the new robot. It is possible that the difference in speed of the approach is the crucial difference between the two robots. It is also hypothesized that the speed of the old robot is relatively slow, while the speed of the new robot is relatively fast.
5.2
Qualitative Analysis
Three human annotators were asked to annotate the video’s. Annotator 1 evaluated all 149 video fragments, while Annotator 2 evaluated 35 fragments and Annotator 3 evaluated 30. All three annotators started with the same 15 fragments in slightly different order. For the analysis of the data, the data from Annotator 2 and 3 were used for the calculation of the inter-rater agreement with Annotator 1 and when this was substantial (>0.6), it was considered to be validated to analyse the data by Annotator 1. So, in the following section only the data from Annotator 1 is reported when there was a substantial Cohen’s Kappa.
Additionally, a correlation with the IPAS is calculated for the questions with a substantial inter-rater agreement.
5.3
Qualitative Results
All three annotators reported after the annotation to be unaware of the fact that two different robots were used in the experiment and thus were visible in the video data.
All agreements tables and Cohen’s Kappa’s of all annotator pairs of all questions can be found in Appendix 7.4. When there were several question categories, one category had to be chosen to calculate the Cohen’s Kappa. Some questions, in particular the subjective questions, obtained almost no agreement, therefore no kappa was calculated.
Table 5.1 shows the Cohen’s kappa’s for the questions with a substantial inter-rater agreement. Although the inter-rater agreement of Annotator 1 and 2 of the question concerning Orientation was not substantial but instead moderate, we decided to analyse the data of this question nonetheless since the inter-rater agreement between Annotator 1 and 3 is substantial.
Table 5.1: Questions with substantial inter-rater agreements.
Question Kappa A1-A2 Kappa A1-A3 When the robot reaches the final position, the participant:
Steps Away from Robot(both feet)/Shows Other Behaviour (Steps Towards robot(both feet)/Steps Aside(both feet)/Hesitates to step(one foot)/Leans Away From Robot/Leans Towards Robot/Does not move).
0.706 1
Trajectory: Robot drives in (more or less) straight line. 0.713 0.615 Orientation: Robot is facing participant when video
introduction starts.
0.440 0.615
5.3.1
Compensatory Distancing Behaviour
The inter-rater agreement for the classification of ‘Steps away from Robot’ is substantial. Wald 95 % confidence intervals are calculated for the chance of ‘Steps away from robot’. The number of times participants stepped away from the robot per condition, data obtained from the video annotation, is presented in table 5.2 as well as the lower and upper bounds of the Wald intervals.
Since the Wald confidence intervals do not overlap, participants stepped away significantly more often in the Close condition compared to the Far condition.
Table 5.2: Video annotation data concerning Compensatory Distancing Behaviour and the corresponding lower and upper bounds of the Wald intervals.
Steps away from robot Other Lower bound Upper bound Close 40 35 0.418 0.648
Far 5 69 0.010 0.126
5.3.2
Robot Navigation Properties
The Cohen’s Kappa’s for the questions about the Trajectory of the robot approach and the Orientation of the robot when the verbal video introduction starts, are substantial. The data obtained from the questions about trajectory and orientation of the robot’s navigation, is presented in table 5.3 as well as the lower and upper bounds of the Wald intervals.
For both questions the Wald confidence intervals of the two robots do not overlap, so the differences in Trajectory and Orientation between the robots are considered to be significant.
Table 5.3: Video annotation data concerning Trajectory and Orientation and the corresponding lower and upper bounds of the Wald intervals.
Trajectory: Drives in (more or less) straight line Other Lower bound Upper bound New Robot 67 6 0.873 0.963
Old Robot 25 51 0.224 0.437
Orientation: Facing participant when video starts Other Lower bound Upper bound New Robot 67 6 0.909 1.005
Old Robot 32 44 0.327 0.516
Correlation
To investigate whether these properties of the robot navigation were responsible for the difference in IPAS, correlations with IPAS were calculated.
A point-biserial correlation was run to determine the relationship between IPAS and Trajectory and between IPAS and Orientation. It is possible that robot navigation properties of previous trials affect trust or confidence in the robot’s behaviour, since the participants interacted with the same robot for four trials. For this reason, a memory effect was used for the data collected from the video annotation; if the
robot did not drive in a straight line in the first trial, the scores for the following trials are considered to be the same, similar for the second or third trial. There was a statistically significant negative correlation between IPAS and Trajectory (rpb = -.182, n = 149, p = .026). There was a non significant negative correlation between IPAS and Orientation (rpb = -.154, n = 149, p = .068).
When the memory effect was not taken into account, there were no statistically significant point-biserial correlations between IPAS and Trajectory and between IPAS and Orientation.
5.4
Summary of qualitative results
The question concerning Compensatory Distancing Behaviour obtained substantial inter-rater agreement. Participants stepped away significantly more often in the Close condition compared to the Far condition, which confirms hypothesis 1.
The subjective questions and the question concerning Back-Channel information of the video annotation did not obtain substantial inter-rater agreement, so these questions were not analysed further.
The New robot was navigated significantly more often in a straight line than the Old robot and was also significantly more often facing the participant when the verbal video introduction started, in line with what we hypothesised in the beginning of this chapter. However, the inter-rater agreement on the question concerning Orientation was only moderate between Annotator 1 and Annotator 2. When memory effects were taken into account, the trajectory of the robot’s approach correlated with IPAS.
6
|
Discussion
We performed an experiment investigating whether a local user holds a pilot user accountable for the autonomous proxemic behaviour of an MRP. Participants had to rate Interpersonal Attraction and four personality traits of four video’s played on an MRP in four conditions (ManualClose, ManualFar, AutonomousClose, AutonomousFar). Since halfway during the experiment the battery of the robot died, we were forced to use another robot for the rest of the experiment, which introduced an unexpected but relevant between subjects factor.
The video annotation analysis indicated that in the Close condition participants stepped away from the robot significantly more often than in the Far condition. This means that participants showed compensatory distancing behaviour when the MRP invaded their personal space, confirming Hypothesis 1– when an MRP invades the personal space of the local user, the local user will show compensatory distancing behaviour.
When an MRP invades the personal space of a local user, we found that the local user is not automatically less likely to be attracted, according to the Interpersonal Attraction scale, to the pilot user. The Interpersonal Attraction is not significantly different in the Close conditions compared to the Far conditions, rejecting Hypothesis 2a– when an MRP invades the personal space of the local user, the local user will form a more negative impression of the pilot user.
When the local user holds the pilot user accountable for the invasion of personal space by an MRP, the local user will be less likely to be attracted, according to the Interpersonal Attraction scale, to the pilot user. The Interpersonal Attraction was significantly higher in the AutonomousClose condition than in the ManualClose condition when tested with the New robot, so the attribution of the robot navigation seems to play a crucial role. As figure 4.2 illustrates, results show a significant difference between the conditions ManualClose and AutonomousClose for the New robot. This suggests that local users form a more negative impression of the pilot users when a pilot user manually invades the local users personal space with an MRP, confirming Hypothesis 2b– when the local user attributes the personal space invasion by an MRP to the pilot user, the local user will form a more negative impression of the pilot user for the New robot. However, the impression formation is not affected by autonomous personal space invasion by an MRP. This answers the research question: a local user does not hold the pilot user accountable for autonomous proxemic behaviour.
That leaves the interesting additional finding that the effect only occurred for one of the robots, the group that was tested with the New robot and not at all for the group that was tested with the Old robot. After analysing the video recordings of the experiment it was found that there were visible differences in the robot approach, trajectory and orientation when the verbal video introduction started, between the two robots. There was no correlation between these properties and the IPAS directly (Section 5.3.2), however, when the memory factor was included there was a correlation between Trajectory and the IPAS. One possible explanation is that previous mistakes or technical errors can affect trust or confidence in the robot which in turn could lead to attributing all robot behaviour to technical errors. So, if the robot is not navigated properly or predictably, the robot behaviour might not be attributed correctly and does not seem to affect IPAS. Hence, attribution seems to be modified by trajectory accuracy.
6.1
Limitations
Since we did not plan to use two different robots as a between subjects factor, the final number of 40 participants is quite limited to compare two groups. Another limitation in this sense is that the experimental conditions were counterbalanced for 40 participants, which resulted in two not purely counterbalanced groups of 20 participants. This might have affected the data.
It is possible that participants did not show compensatory distancing behaviour in reaction to personal or intimate space invasion, but instead stepped away from the robot to prevent physical collisions. There is a conceptual difference in the psychological personal space invasion, which can be perceived as awkward, and almost physically hitting a person, which can be perceived as threatening. The close interaction distance was extremely small, so there would be no doubt that it was undesirable, but this might have resulted in the idea that the robot would not stop in time and would drive into the participant.
One main factor in the approach, that could have resulted in the perception that the robot was not going to stop in time, would be the speed of the robot. The navigation software of the Giraff does not control for speed precisely, so this is not kept completely constant over the experiment. Future research should investigate the effects of speed in these situations. Additionally, it was hypothesized for the video annotation, that the speed of the New robot was in general higher than the speed of the Old robot. However, the video annotation did not get sufficient inter-rater agreement on the topic of speed, probably due to the fact that it is hard to interpret speed in a video, so this did not give meaningful information. Since 8 participants (20 %) reported that the robot came too close in all four conditions, instead of only in the two manipulated close conditions, it is possible that the chosen interaction distances were too small in all conditions or that the difference between the close and far conditions was too small. Future research should point out what the appropriate interaction distances are for MRPs, and how individual differences can play a role in this.
Additionally, in this experiment no real interaction was held via the MRP, but instead, video’s were used. This means that the results are not directly generalisable to all interactions with MRPs. Future research should investigate if the effects still hold for real interactions.
6.2
Contributions and Conclusions
Attribution and Proxemics together can affect the impression of a pilot user on a local user during an interaction mediated by an MRP. This study shows that attribution plays a key role in shared autonomy, since personal space invasion itself did not affect the impression of the pilot user.
From these findings follows an important consideration for the design of shared autonomy in telepresence; it can be desirable to indicate to whom the behaviour should be attributed, especially when social norms can be violated. For when it is clear to people when a telepresence robot navigates autonomously, (unintentional) social norm violations may have a smaller effect on how the pilot user is perceived. Though this study focused on telepresence robots, similar design considerations can be extended to other cases of shared autonomy, such as (semi-)autonomous vehicles.
In all, a clear indication of human vs robot responsibility in MRP behaviour can be of great importance for social interaction.
