Making Robot ’ s A itudes Predictable : A Stereotype Content Model for Human-Robot Interaction in Groups

Stereotypes play a key role both in human and in robot perception. As such, these may play an important role in predicting behavior towards the targets of said stereotypes. In this paper, we argue for the usefulness of exploring how stable dimensions of stereotype content, more specically warmth and competence, apply to HRI. We propose to do so by examining the impact of these characteristics, displayed by robots, on the emotions, behaviors and future intention of participants to interact and work with robots. We chose these two stereotypical dimensions given that research has demonstrated their underlying and ubiquitous inuence on people perception and responses. Moreover, we decided to explore this issue in the context of small group interactions due to the ever-common nature of this type of social arrangements in people’s everyday life.


INTRODUCTION
People's perception about themselves and others is in uenced by a multitude of factors.One of those factors is the stereotypes individuals develop and apply to make sense of the world.Cognitive psychology theories argue that stereotypes are be er categorized in terms of implicit personality theory, as they constitute heuristics for stimuli interpretation that due to their general acceptance can be considered as part of a process of social cognition [4].us, these constitute part of a larger set of knowledge structures that guide interpretation of other's people behavior and are relevant predictors of own behavior [2], [3], [9].Moreover, research in Human-Robot Interaction (henceforth, HRI) has demonstrated the presence of this type of social categorization (for example, based on gender [11]), drawing a ention to the question of how di erent dimensions of stereotype content can impact the way robots are perceived, and thus, responded to.Answering this question can yield important ndings to take into account during the robot conception design by contributing to the creation of more socially e ective robots (by providing people a consistent form of social information about the robot that they can take into account to predict its behavior and thus, adjust their expectations about it and act accordingly).us, if we consider stereotypes can function as social cues in the context of HRI, this can spare the user from the e ort of looking for further information during the interaction, thereby increasing the level of transparency and intuitiveness of this interaction.

Externalizing individual attitudes in groups of human-robot interactions
e present work is done under the AMIGOS project [1] which aims to contribute to the literature by considering the roles of emotions and adaptation in Human-Robot-Interaction, in the context of group interactions.Our aim is to create a data-driven model for group interactions that allows robots in that context, to adjust their behavior according to the situational characteristics and the preferences of each individual set of users.us, our goal is to endow the robot with the ability to generate context-adaptive responses using interactive machine learning techniques.In order to do so, we use a card-game scenario in which two participants engage in an entertaining task with two robots (for more details, see [1]), and analyze how mixed group behave towards one another in the role of partners and opponents.us, our goal is to endow the robot with the ability to generate context-adaptive responses as robots and humans interact in small groups.As such, making explicit certain types of a itudes and behaviors in robots is very important for a natural communication to emerge.Signi cant work has been done in making expressions of robots legible and predictable, in particular at the motion level [8] [13].Other work has pointed out the need for exaggeration of actions and features in robots, to make them more natural [19].Here we argue that a itudes can be made more salient in a robot by leveraging the power that stereotypes have in communication.

Stereotypes in Human and Robot Interaction
People are social creatures who tend to a ribute human-like characteristics to a broad range of non-human elements.is might include random pa erns [16], virtual agents [15] and robots [20].
In the case of robot perception, studies have already demonstrated the important role of stereotypes in robot's trait evaluation.For example, Eyssel and Hegel compared two robots displaying di erent gender facial characteristics [10].ese authors found that the male robot was perceived as more agentic, whereas the female robot was perceived as being more communal.Furthermore, using a voice gender manipulation, that could be either synthesized robot-like or human-like, Eyssel and colleagues [12] also found that participants tended to evaluate the same-gender robot more positively across a large range of social dimensions, suggesting the existence of some sort of projection mechanism.Moreover, when robots' are stereotyped, it appears that the stereotypes that are associated with them are congruent with those that occur in Human to Human Interaction (HHI).For example, the aforementioned typical assumed gender-role stereotypes have been also been consistently veri ed in HRI [20], hinting at the existence of some level of extrapolation of people's stereotypes about other people to stereotypes about what other social actors, in this instance, robots, do or are supposed to do in a certain situation.

e Stereotype Content Model: Why does Competence Matter
e term stereotypes implies a gestalt view of people perception, suggesting the notion that some traits can be more central than others, in organizing our perception of other people [20].According to the Stereotype Content Model (hereina er, SCM) [14]; [7] warmth and competence are the two main stable content dimensions of stereotypes.ese are central to group stereotypes and have been linked to speci c emotional and behavioral outcomes [7] (see g 1).SCM considers two levels of competence and warmth (low and high).e combination of those two, can be associated with a set of traits that are considered more or less socially desirable and, thus can elicit di erent emotional and behavioral responses [14].Social robots have not yet been perfected to the point where a certain degree of incompetence (i.e.failure to adjust and choose a proper response) is not to be expected.In fact, social situations, especially those involving more than one human, can present very challenging environments that make it hard for a robot to identify the relevant characteristics to take into account to form a response that is adequate.Furthermore, a high level of competence, might not always be desirable as it can evoke machine-like associated stereotypes (i.e.computers/machines don't fail ) or because a high level of taskorientation might be perceived as a threat [17].Research in HHI demonstrates consistent group stereotypes associated with perceptions of di erent levels of competence and warmth [14], that can result in di erent behavioral approaches.
Goals and Hypothesis.Although some studies have already focused on the competence and warmth dimensions in robot perception (e.g.: [5], we believe that a more systematic understanding of how di erent levels of competence and warmth displayed by the robot intertwine to mold an overall impression of the robot, is still missing.An examination of how these di erences in perception a ect the emotional and behavioral responses towards the robot is also lacking.In this context, we expect participants to display similar emotional responses to those observed in HHI, towards robots displaying di erent levels of warmth and competence.(see g.1).We also aim to examine the moderating role of trait congruency in the future intention to work with robots.Furthermore, we Figure 1: Content dimensions of SCM: ER stands for expected emotional response, whereas BR stands for expected behavioral response [14] also expect participants to display a higher level of future intention to work with the robot that displays similar characteristics to him/herself.

ONGOING WORK
Social interactions are complex phenomena, that might include different forms of interpersonal communication and social messaging.
ese social interactions become increasingly more complex if one looks at them from a group interaction standpoint, as this adds complexity to the analysis.
However, as groups of humans and robots might emerge in a near future it becomes relevant to consider how they act and communicate in groups.Answering the question we present in this paper will allow for the development of robots that can interact both according to the situation and to the type of emotional or behavioral response it wishes to evoke. is also helps the development of robots that can evoke consistent mental models and social categorizations from humans and that, in turn, will make robots' behaviors and intentions appear more predictable from the user perspective.
To address these questions, we are currently working on an entertaining interaction card-game scenario , where two human participants are required to play with two robots [6], displaying high and low levels of warmth and competence. is task has been used in previous studies of the AMIGOS project and details regarding the experimental design and task can be found elsewhere [1], [17] .Warmth will be manipulated through the u erances spoken by the robot whereas competence will be manipulated through the game-solving algorithm implemented in each robot (for more details, see [18]).us, our main goal is to understand the behavioral and emotional implications of di erent judgments of warmth and competence levels and how these might a ect future intention to work with robots.