Individual differences in attitude toward robots predict behavior in human-robot interaction

1 Introduction
Robotic agents are already present in many aspects of our everyday-life, whether they 
greet us at the airport [1], assist in elderly care [2] or work side-by-side with human 
employers in manufacturing [3]; and wider applications are probably going to emerge. 
Therefore, humans face the need to act more and more often in the presence of robots
or to interact with them. In situations with social or work-related responsibility, it appears important to know how people behave in the presence of robots and how this 
eventually diverges from individual situations.
Although evidence showed that the presence of social agents in the environment may 
improve performance, it may also create difficulties [4, 5]. For instance, in human-human interaction, a well-known phenomenon is social facilitation, i.e. the fact that the 
presence of another human can enhance performance [4]. However, this effect is dependent on the nature of the task [6]. For complex and difficult tasks, for example, the 
presence of a co-agent can lead to social inhibition, i.e. deteriorated performance [5]. 
Similar effects have been demonstrated with artificial agents [7, 8]. As has been demonstrated before in other aspects of social cognition [9], embodiment seems to affect the 
social presence effect, as presence induced by images of social agents did not affect 
performance [10]. These studies, however, have used fairly simple tasks, such as arithmetic operations, that do not accurately resemble the complexity that characterizes
practical applications of robots. Indeed, most of the applications in which a human has 
to interact with a robot rather deal with uncertainty and risky contexts.
Another crucial aspect to examine in HRI is how individual differences in robot perception translate to differences in behavior [11, 12]. Indeed, individual differences, 
such as personality traits or attitudes, have been found to affect the perception and acceptance of robots in social situations [13–16]. For example, higher levels of neuroticism seem to predict the preference of more machine-like robots [13]. Similarly, differences in assigning human-like characteristics to non-human agents lead to differences in the trustworthiness assigned to a robot [14]. A full range of individual differences, like anxiety, perfectionism or religious fundamentalism, have been proposed to
explain feelings of eeriness towards the robot [15]. Differences in trustworthiness perception of a robot have been demonstrated to predict behavior in a human-robot team 
[16-17].

1.2 Aim of the study
Our study aimed at investigating whether human behavior is influenced by the presence 
of a robotic agent in a task necessitating risk-taking, and how this influence is moderated by individual differences in the attitude towards robots. Therefore, we asked participants to perform a game in which they had to stop a balloon from exploding alone 
or playing with the Cozmo robot (Anki Inc., San Francisco).We used a non-anthropomorphic robot in order to avoid that differences in assigning human-likeness might affect trustworthiness toward the robot [14]. We expected Cozmo to improve the performance of participants in the task due to social facilitation effects as a function of individual differences.
2 Methods
2.1 Participants
Thirty-two healthy adults participated in this study. Data of one participant were not 
analyzed because the robot crashed during the experimental session. The remaining 
sample consisted of thirty-one participants (12 male, 1 left-handed, age range: 19 -44 
years, M = 23.85, SD = 4.81). All had normal or corrected- to-normal vision. Partici-pants provided informed written consent before participation, received financial reimbursement and were debriefed after the experiment. The study was approved by the 
local ethical committee (Comitato Etico Regione Liguria).
2.2 Apparatus and Materials
Participants were seated in front of a desk on which lay a computer screen (22 inches 
diagonal, 1366x768 pixels resolution, 59 Hz refresh rate). The Cozmo robot was placed 
directly in front of them (see Fig. 1). For both the participant and the Cozmo robot, the 
response device was one of the Cozmo cubes on top of which an in-house-built onekey-keyboard was mounted.

The Cozmo robot
The Cozmo robot (Anki Inc., San Francisco) is a commercial robot designed for educational purposes. It consists of a horizontally moveable head with an LED screen on 
which eyes are displayed, four wheels, three LEDs on the back and a horizontally moveable lift. Cozmo is controlled by an application compatible with iOS and Android. A 
Python 3.6-based Software development kit (SDK) can be used to program Cozmo. 
During the experiment a mobile Android device with the Cozmo application in 
“SDK mode” was used to control Cozmo, connected to a laptop through the Android 
Debug Bridge (adb) as described in [18]. Commands were sent to the Cozmo application by using OpenSesame Version 3.1.9 [19], running on Python 3.6. 

Questionnaires
To assess the participant’s attitude towards robots, before the experiment we administered three different questionnaires. All of the questionnaires were presented on a computer screen, using OpenSesame Version 3.1.9 [19] and responded to with a standard 
computer mouse. The questionnaires were: 

• the Frankenstein Syndrome Questionnaire (FSQ; [20]) is composed of thirty items, 
  measuring the “Frankenstein Syndrome”, the fear of creating new entities. Four different subscales are evident in the questionnaire: “General Anxiety toward humanoid 
  robots”, “Apprehension toward social risks of humanoid robots”, “Trustworthiness 
  for developers of humanoid robots” and “Expectation for humanoid robots in daily 
  life”. The subscales have medium to good internal consistency, the questionnaire in 
  general, however, demonstrated good reliability [21]. Higher scores stand for a more 
  severe Frankenstein Syndrome, i.e. more negative attitude towards robots.
• the Negative Attitude Towards Robots Scale (NARS; [13]) consisting of fourteen 
  items, organized in three different subscales: “Negative Attitudes toward Situations 
  and Interactions with Robots”, “Negative Attitudes toward Social Influence of Robots” and “Negative Attitudes toward Emotions in Interaction with Robots”. The questionnaire demonstrated high internal consistency and validity. Higher scores 
  stand for more negative attitude towards robots.
• the Robotic Social Attributes Scale (RoSAS; [22]) consisting of eighteen items in 
  which participants are presented with an adjective (e.g. social) and have to express on 
  a 9-point Likert scale how much they believe that the adjective can be used to describe 
  a robot. Items are classified into three psychometrically validated subscales: 
  “Warmth”, “Competence” and “Discomfort”. While higher scores on the subscales 
  “Warmth” and “Competence” are related to a more positive attitude towards robots, 
  higher scores on the “Discomfort” subscale signify a more negative attitude.

2.3 Task and Trial Procedure
The task was based on the Balloon Analogous Risk Task [23]. Participants played a 
game, in which they had to stop an inflating balloon before it exploded when reaching 
a pin on the top of the display. In every trial the participants would lose points from an 
initial amount of 4000 points, with the amount of lost points depending on when the 
balloon was stopped. The later the balloon was stopped, the less points were lost. The 
exact number of lost points per trial was randomly chosen from a range of points, depending on four different clusters of sizes at which the balloon was stopped (see Table 
1). The maximal amount of points was lost when the balloon exploded. The goal of the 
game was to save the maximal amount of points; therefore the best strategy would have 
been to wait as long as possible before stopping the balloon. In 50% of trials, participants played alone (individual trials), while in the remaining trials also Cozmo was in 
charge to stop the balloon inflation (joint trials). Cozmo was programmed to act only 
in the 60% of joint trials. When Cozmo stopped the balloon, it would lose the respective 
amount of points, whereas the participant would lose no points. In case of an explosion 
both agents would lose the points. Cozmo always acted when approximately 90% of 
the inflation time were reached. In joint trials, the best strategy for the participant would 
have been to wait for an action of Cozmo and only react in trials, in which it did not, to 
prevent an explosion of the balloon. However, participants could not predict a priori 
when Cozmo would act.
At the beginning of each trial participants were told whether they were playing alone 
(individual trial) or with Cozmo (joint trial) with a text presented on the screen for 1000ms (see Fig. 2). During the first trial of each block Cozmo either went to sleep (transition from joint to individual trial) or woke up and approached the cube (transition from 
individual to joint trial). After the initial instruction about the type of trial, a sketchpad 
displaying “The trial is starting” was presented on the screen, so participants could prepare for the beginning of each trial. Then, a fixation point was presented for a random 
duration of 800-1000 ms (note that the fixation point duration was randomly set at the 
beginning of each trial and kept constant for all fixation points presented in the respective trials). After that, the images of a pin and of the balloon at its starting size were 
presented for 500 ms. Following, the balloon started inflating. The inflation speed was 
variable across and within trials, in order to make the explosion time not predictable. 
In each block ten different inflation speeds were used. Additionally, during the inflation 
sequence, two grey circles were presented on the bottom right and top left corner. When 
a response was given one of the circles turned blue, indicating whose reaction was
counted (with top left corner representing Cozmo and bottom right corner representing 
the participant). After a response was executed, the balloon was displayed in its final 
size for 1000 ms. If the balloon exploded, an image representing the bursting event was 
presented. Then a fixation point was followed by a sketchpad (2000 ms) showing the 
amount of lost points. 
Participants were explicitly instructed that their goal was to save as many points as 
possible to defeat previous participants and they would not be able to defeat Cozmo 
since it was playing in only half of trials. This was done to try to avoid that they would 
perceive the task as a competition between themselves and Cozmo.
The task consisted of 180 trials presented in 18 blocks of 10 trials each. The type of 
trials (individual or joint) was manipulated across blocks. The order of the blocks was 
randomly selected. A practice session of 6 trials was administered before the experiment. During the practice participants experienced one trial in which Cozmo was not 
reacting.

2.4 Statistical Analysis
Each trial was classified as a “Human” trial if the participant stopped the balloon and 
as a “Cozmo” trial if Cozmo stopped the balloon. If no one stopped the balloon and it 
exploded, the trial was classified as “No reaction”. Performance was assessed by the 
amount of reactions in each trial type, reaction time (i.e. the amount of time in ms from 
the starting of the inflation until the balloon was stopped) and the points lost after the 
reaction. This behavioral data was analyzed using paired-sample t-tests. Ratings from 
the questionnaire subscales were summed up to a total score. Pearson correlations were 
conducted to examine the relationship between questionnaire (subscale and total) scores 
and the behavioral data. All analyses were performed using R Version 3.5.1 [24]. Plots 
were created using the ggplot2 package Version 3.0.0 [25].

3 Results
Participants successfully stopped the balloon in 85.4% (SD = 7.7%; see Error! 
Reference source not found.) of individual trials and 47.3% (SD = 9.5%) of joint trials. 
In 42.1% (SD = 6.7%) of joint trials they let Cozmo react. The balloon exploded in 
14.6% (SD = 7.7%) of individual trials, but only 10.6% (SD = 4.4%) of joint trials. The 
number of explosions differed significantly between the two types of trials (t29 = -3.55, 
p < .001).
Given that Cozmo’s actions were influencing the information contained in outcome 
and balloon size, only trials in which the participant successfully stopped the balloon 
(Human trials) were further analyzed. No difference was found between the points lost 
in each trial type (Joint: M = 9.05, SD = 0.85; Individual: M = 8.79, SD = 0.74; t29 = 
1.32, p = .19; see Fig. left). However, reaction times were different between the trial 
types (t29 = -3.42, p < .001), showing faster performance in joint (M = 4078 ms, SD = 
93 ms) compared to individual trials (M = 4130 ms, SD = 66 ms; see Error! Reference 
source not found. left).
The analysis of correlations between questionnaires and performance (i.e. lost points, 
balloon sizes and reaction times) showed that the FSQ total score correlated negatively 
with the amount of lost points in the joint trials only (Joint: r = .45, p < .01; 
Individual: r = .09, p = .62; see Fig. right). No significant correlations were found between the amount of lost points and each FSQ-subscale score (all ps > .XX). Finally, 
the score on the discomfort scale of the RoSAS was positively correlated to reaction 
times in joint trials only (Joint: r = .33, p = .07; Individual: r = .06, p = .76; see Error! 
Reference source not found. right).

4 Discussion
In the present study, we investigated how someone’s behavior is influenced by the presence of a robotic agent and especially how this influence is moderated by individual 
differences in the attitude towards robots. Our results showed that when playing with 
Cozmo, participants overall performance in the game improved as indicated by a 
smaller number of explosions for joint than individual trials. This effect was clearly 
driven by Cozmo’s actions. When only analyzing trials in which participants successfully stopped the balloon themselves (Human trials), results showed that participants 
stopped the balloon slightly earlier (as evident from faster reaction times) and at smaller 
sizes in the joint compared to the individual trials. Such a difference in the performance 
suggests that the presence of Cozmo triggered the action in joint trials, leading to a 
social facilitation effect exerted by the robot. Interestingly, the social facilitation effect 
occurred despite that the optimal strategy to lose as few points as possible was to react 
as late as possible (i.e. withholding the action). It should be noted that social facilitation 
was not observed in the number of lost points because the relation between stop size 
and actual feedback was not fully linear (see Table 1).
Our results showed that the effect exerted by the social presence of Cozmo varied as 
function of individual attitude towards robots, with higher scores on the FSQ questionnaire being associated with better performance (i.e. later balloon stops and fewer points
lost) in joint trials, and higher scores on the Discomfort subscale of the RoSAS leading to later balloon stops (i.e. slower reaction times) in joint trials. Together the correlation patterns suggest that the more negative someone is regarding robots, the less s/he 
is influenced by the presence of the robot and the less s/he shows social facilitation 
effect. Such a result is in line with the hypothesis of Schellen and Wykowska [11], that 
negative attitudes towards robots might be a moderating factor for social presence effects. Future work should look at how the present effects translate to the use of other robots, especially humanoid robot. Humanoid robots could for example further increase 
social facilitation effects by inducing more social presence. Future studies should also 
account whether trustworthiness toward the robot may affect the decision to intervene 
earlier to prevent losing points.

5 Conclusion
Our results indicate that sharing a task with a robot apparently triggers the execution of 
action, even when it is not beneficial. This should be carefully considered when introducing robots in situations with social or work-related responsibility. In a similar vein, 
our results highlight the importance of carefully examining individual differences in 
the attitude towards robots, since they differently affect how people behave in the presence of a robotic agent.

6 Acknowledgements
This project has received funding from the European Research Council (ERC) under 
the European Union’s Horizon 2020 research and innovation program (grant awarded 
to AW, titled “InStance: Intentional Stance for Social Attunement.” G.A. No: ERC2016-StG-715058).