THEORY OF COMPANIONS: WHAT CAN THEORETICAL MODELS CONTRIBUTE TO APPLICATIONS AND UNDERSTANDING OF HUMAN-ROBOT INTERACTION?

Abstract

Because it becomes more and more feasible that artificial entities like robots or agents will soon be parts of our daily lives, an essential challenge is to advance the sociability of artifacts. Against this background, a pivotal goal of the Sera project was to develop a theoretical framework for sociable companions as well as for human-artifact interaction. In discussing several levels of sociability from a theoretical point of view, we will critically reflect whether human-companion interaction has to build on basic principles of human-human interaction. Alternative approaches are presented. It is discussed whether a “theory of companions” is necessary and useful and what it should be able to explain and contribute.

INTRODUCTION

Virtual (agents) and embodied (robots) devices that serve as assistive technology, especially to elderly or homebound people, are subject to research and development as can be illustrated by several EU-funded projects covering this issue, such as Companions, LIREC and Semaine. Practical experiences with today's conversational interactive systems, however, show the gap between the utopian “companion” and today's clumsy attempts, as interaction often is disappointing, boring, or completely irritating. Thus, getting people to engage with these artifacts is easy, but keeping them engaged over time is a hard task, because a large number of users refuse to interact with the system again and/or experience aggression, often even culminating in verbal or physical attacks directed towards the system (De Angeli et al. 2005, 2006; Walker et al. 2002). However, as artifacts are supposed to take long-term assistive, coaching, mediating, or educational roles in people's everyday lives, steps have to be taken to enable a blending of these systems into people's lives. To achieve this, the essential challenge is to develop the sociability of artifacts. Within the EU-funded project SERA (Social Engagement with Robots and Agents) we therefore aimed at developing a theoretical framework for sociable companions as well as for human-artifact interaction.

In order to address all aspects of sociability, within this paper we will focus on three different levels (see Table 1). On a micro-level we will address prerequisites for communication by demonstrating in which way theory of mind, perspective taking, and similar abilities enable social interaction. On a meso-level we describe how relationships can be shaped and which factors affect their quality. Finally, on a macro-level we discuss which roles can be helpful frames when trying to shape human-artifact interaction. Therefore, beyond addressing actual interaction and communication, the nature of the relationship and the role of the companion is discussed: we will comment on what we have learned and discussed within the SERA project concerning the question of whether the relationship to the companion should resemble an intimate long-term human-human relationship (e.g., family member, close friend), a non-intimate long-term human-human relationship (e.g., neighbor, mailman) or be based on human-pet relationships.

TABLE 1 Levels of Sociability

Levels of sociability	Corresponding theories
Micro-level: Actual interaction, Prerequisites for communication	– Common ground – Theory of mind – Perspective taking – Shared intentionality
Meso-level: Relationship building	– Need to belong – Prerequisites: attractiveness, reciprocity – Social exchange – dimensions of human-human interaction will play a role (see e.g., dominance, intimacy)
Macro-level: Roles and persona	– assignment of roles by designer versus user

One essential goal of the SERA project is to contribute to the development of theoretical foundations of human-robot/agent social relationships. The basic assumption is that, in order to gain much needed knowledge about long-term, open-ended social relationships between humans and artificial companions, it is important to develop a theoretical framework that will guide data analysis as well as the implementation of the companions’ behaviors. One obvious starting point when developing a theoretical framework that is able to inform architecture and design of companions is to employ theories from psychology, sociology, or philosophy, which certainly focus on human-human interaction. However, several questions can be raised with regard to the usefulness of this approach. For instance, it can be debated to what degree it is useful to conceptualize a theory of companions with reference to human-human-interaction (e.g., common ground, theory of mind). For all of the three levels mentioned above, we will therefore start by describing what is known in the area of human sociability and will subsequently discuss whether this is useful as a starting point for theoretically conceptualizing human-artifact interaction and relationship.

More specifically, it will be asked whether a theoretical framework in the end will be beneficial for deriving guidelines (e.g., for dialogue creation). One of the most important and most consistent results of the SERA project shows that there are large interindividual differences with regard to both, the interaction and relationship building toward an artificial entity. Based on this and derived from frameworks on human communication (e.g., Watzlawick, Beavin, & Jackson, 1967), it will be argued that simple guidelines for message production on the side of the artificial entity will not be helpful. Guidelines will not be able to embrace all important determinants of a specific interaction since human receivers construct the meaning of a message against the background of their own personal background, experiences, moods, etc. What can be learned, however, is that the companion will need basic abilities to understand the users’ reactions to its messages in order to be able to use, for example, repair mechanisms. Also, basic learning of what to avoid when addressing a specific user will be beneficial. In the conclusion it is summarized whether a “theory of companions” is necessary, feasible, and useful, and what it should be able to explain and contribute.

SOCIABILITY ON THREE LEVELS

The final goal of the SERA theoretical framework is to provide a basis for the derivation of guidelines for implementing an artifact's behavior. However, the actual behavior depends on a multitude of aspects, all of which have to be considered when a reasonable architecture for the social behavior of companions interaction should be presented. Sociability is a complex whole that builds on aspects and processes as different as theory of mind and emotionality, situational awareness, and general behavioral patterns. Therefore, the theoretical considerations have to take several levels of sociability into account. Here, we will distinguish among three levels of sociability (micro-level: actual interaction; meso-level: relationship; macro-level: roles). The three dimensions are merely differentiated in order to meaningfully sort different theoretical approaches to describe human-robot interactions. Also, by taking different perspectives we wanted to ensure that all aspects of human-robot interactions will be considered (even if the specific theories presented are certainly a selection). We will first focus on the basic prerequisites for communication which enable actual interaction between humans and artifacts. Here, abilities such as perspective taking, building common ground, theory of mind, and similar approaches are presented. On a second level, we address the relationship between humans and artifacts. Here, we present the mechanisms of human-human relationship building referring to, first, the need to belong (Baumeister and Leary 1995) and attributes such as attractiveness and reciprocity as basic prerequisites for building relationships. Also, patterns for establishing relationships based on equity theory (Walster, Walster, and Berscheid 1978) are described. Additionally, the dimensions of human relationships (for example with regard to dominance) are taken into account. On a third level, we focus on a macro-level that relates to a sociological view of sociability and we discuss the potential roles and personas for artifacts in human-artifact-interaction.

For every level, the implications of modeling human-artifact-interactions and relationships based on the example of human-human interactions are discussed. As an alternative conceptualization, human-pet interactions and relationships are considered on every level.

Microlevel: Prerequisites for Communication

We only slowly begin to understand the massive abilities that humans need for communication. What seems to be easy for every human being is in fact based on numerous different abilities that cannot be easily understood, made explicit as rules, nor be implemented. Recent theories, models, and assumptions from social psychology, cognitive psychology, and developmental psychology show that communication is possible only when both interaction partners either know or refer to the same concepts. This is why robots and agents need a representation of users, their social and cultural backgrounds, and examples of interaction situations and contexts in order to be sociable. What is meant here can aptly be illustrated by Wittgenstein's statement, “If a lion could talk we would not understand it,” referring to the fact that it is useless to implement the ability of natural speech in robots while they are unable to understand concepts that are naturally shared by humans and are taken for granted in communicative interactions. Models on human perspective taking, common ground, theories of mind, etc. all share the view that sender and receiver have fundamental similarities since they share human processing with regard to needs, thoughts, emotions, etc. This also enables the speaker to design messages to be appropriate to what he assumes to be the knowledge of the recipient. In the following, the different approaches are summarized and in a joint conclusion their importance for each form of communication is highlighted.

Perspective Taking

Social perspective taking, i.e., understanding the feelings, thoughts, and motivations of others, is an essential social skill that has been stressed by many researchers. Starting out from a social psychological point of view on the process of perspective taking, Krauss and Fussell (1991) claim that in communication, the fundamental role of knowing what others know is axiomatic and that this has been widely acknowledged (Bakhtin 1981; Clark 1985; Clark and Marshall 1981; Graumann and Herrmann 1989; Mead 1934).

The importance of perspective taking (or role taking, point-of-view appreciation, see Nickerson 1999) has also been stressed by Baldwin (1906), Kohlberg (1969), and Rommeveit (1974). According to their views, the failure to take another's perspective into account can be the basis for misunderstandings and dispute. In this respect, a prerequisite for successful communication is that the message is tailored to the knowledge of the recipient (Krauss and Fussel 1991).

Although this ability has been stressed as an important prerequisite for communication, Krauss and Fussell (1991) state that remarkably little discussion of the process by which communicators take into account the perspective of others can be observed. They review several studies that indicate that speakers indeed take their addressee's knowledge and perspectives into account when they formulate messages. In doing this, the accuracy of people's assessments of others’ knowledge is fairly high, but they seem to be biased in the direction of their own knowledge (see also Nickerson 1999).

Krauss and Fussell (1991) conclude that people's assumptions about other's knowledge are necessarily tentative and best thought of as hypotheses that need to be evaluated and modified over time. Perspective taking is thus not only characterized by theories on what others know; in face-to-face interactions one might additionally use conversational resources such as the possibility to receive feedback on their own assumptions with regard to another's knowledge: “During the course of the interaction, each participant's apparent understandings and failures to understand the partner's messages provide feedback about the appropriateness of the assumptions on which these messages are based” (Krauss and Fussell 1991, p. 21).

In their “interaction adaptation theory” that not only focuses on verbal interaction but also nonverbal mimicry and reciprocity, Burgoon and White (1997) stress the importance of this mutual “online” adaptation and joint construction of the other's knowledge even further: “All message production, and especially that in interpersonal conversation, implicitly begins with an alignment toward the message recipient and the predisposition to calibrate one's messages to the characteristics of the target (as well as the topic, occasion, and setting). Put differently, adaptation is an intrinsic feature of all communication, and, as such, carries with it the implication that to fully understand message production requires knowing the extent to which message content and form are influenced by, and jointly constructed with, co-interactants” (Burgoon and White 1997, p. 282).

By now the user is often the one who tailors his/her messages to the robot or agent and not the other way round, for example, users repeat themselves more slowly or answer in a much simpler way than they would in human-human communication due to the fact that the system has technical shortcomings. However, this is in most cases not sufficient as even basic concepts and—more important—contexts are not shared. Anyway, the aim of companions is not to force humans to adapt to the system, but to design a system with which humans can interact naturally.

Common Ground

The theory of common ground has been proposed as one important aspect of using language. Drawing on Stalnaker (1978) and Karttunen and Peters (1975), Clark (1992) describes common ground as the joint basis for communication: “Two people's common ground is, in effect, the sum of their mutual, common, or joint knowledge, beliefs, and suppositions” (Clark 1992, p. 93). He assumes common ground to be a sine qua non for everything humans do with others: to coordinate and communicate with others, humans have to appeal to their current common ground. This implies that in case there is no common ground, no communication or understanding, respectively, would take place.

Thus, it can be assumed that there should be an initial common ground in each conversation that can be broadened during the interaction. The most obvious starting point in terms of communal common ground is human nature. As an example, Clark (1992) points out that if a sound is audible to someone, he will assume that it is audible to the other as well. Moreover, he explains that people take the same facts of biology for granted and that everyone assumes certain social facts (people use language, live in groups, have names). In most cases, in other words, when interacting with people from the same cultural area, even cultural facts, norms, and procedures can be taken as being part of the common ground. Moreover, humans use the knowledge of cultural communities (e.g., when categorizing people by nationality, profession, etc.) to infer what information members of that community might have. By basic grounding principles during the interaction, humans try to assert the common ground by finding evidence of (joint) category membership, for example.

Furthermore, personal common ground during interactions is built on joint perceptual experiences and joint actions. Here, people try to ground what they do together. Similar to the approach by Krauss and Fussell (1991, see above), it is further assumed that the actual conversation can be used for preventing discrepancies. Humans have verbal and nonverbal strategies to discover and repair situations in which the mutual knowledge is misinterpreted. “Contributors present signals to respondents, and then contributors and respondents work together to reach the mutual belief that the signals have been understood well enough for current purposes” (Clark 1992, p. 252).

In fact, Clark (1992) lists several grounding principles that permit the location of common ground as well as the establishment of mutual knowledge. One example is the linguistic copresence heuristic that entails that anything that has been said during the course of the conversation is known to both. Another general principle that humans rely on is the principle of closure: human agents performing an action require evidence that they have succeeded in performing it. Within human-human interaction, this is relevant in the sense of joint closure: the participants of a joint action try to establish the mutual belief that they have succeeded well enough for current purposes (e.g., shake hands). This might be accomplished by providing each other with subtle feedback. This need for feedback and closure has also been described with regard to human-computer interaction: both telephone buttons that do not beep when pressed or a display that does not change when an action has been taken, are confusing (Norman 1988).

Theory of Mind

The term “theory of mind” was coined by Premack and Woodruff (1978) as they referred to the “ability—which may or may not be unique to human beings—to explain and predict the actions, both of oneself, and of other intelligent agents” (Carruthers and Smith 1996, p. 19). Theory of mind (ToM) is the ability to see other entities as intentional agents, whose behaviors are influenced by states, beliefs, desires, etc. and the knowledge that other humans wish, feel, know, or believe something (Premack and Premack 1995; Premack and Woodruff 1978; Whiten 1991). In recent years, ToM has been discussed as a basic prerequisite for human-human interaction and various terms have been established: mentalizing (take another person's mental perspective and predict what they can know (Frith and Frith 2003), mind reading (Baron-Cohen 1995) and intentional stance (Dennett 1987) all basically refer to the same ability. Dennett (1987), for example, states that attributing mental states to a complex system (such as a human being) is by far the easiest way of understanding it, in other words, coming up with an explanation of the complex system's behavior and predicting what it will do next. The ability is seen as an important and innate part of human nature that is crucial for all aspects of our everyday social life and our natural way of understanding the social environment: in line with this, Sperber (1993) states that “attribution of mental states is to humans as echolocation is to the bat” (cf. Baron-Cohen 1995, p. 4). Also, Toby and Cosmides (1995) stress the function and innateness of the ability: “We are ‘mindreaders’ by nature, building interpretations of the mental events of others and feeling our constructions as sharply as the physical objects we touch. Humans evolved this ability because, as members of an intensely social, cooperative, and competitive species, our ancestors’ lives depended on how well they could infer what was on one another's minds” (Toby and Cosmides 1995, p. XIII).

Indeed, theory of mind has been discussed as a prerequisite for communication between human interactants: Although “mindreading,” of course, does not allow for a 100% correct prediction of mental states, it provides a general orientation on other people's processes and a prediction of the effects of communication. Baron-Cohen (1995) thus sums up: “A … reason why mind reading is useful, and thus why it may have evolved, is the way in which it allows us to make sense of communication … A range of theorists—Grice (1957). Sperber and Wilson (1986), Austin (1962)—have argued that when we hear someone say something …, aside from decoding the referent of each word (computing its semantics and syntax), the key thing we do as we search for the meaning of the words is to imagine what the speaker's communicative intention might be. That is, we ask ourselves ‘What does he mean?’ Here the word ‘mean’ essentially boils down to ‘intend me to understand'” (p. 27). Hence, in decoding speech, humans go beyond the words we hear or read and hypothesize about the speaker's mental states. Similarly, Frith and Frith (2003) refer to Grice (1957) and his idea that a successful understanding of an utterance depends upon perceiving the intention of the speaker as well as to Sperber and Wilson's (1995) theory of relevance. They conclude that pragmatics of speech rely on mentalizing and that in many real-life cases the understanding of an utterance cannot be based solely on the meanings of the individual words (semantics) or upon the grammar by which they are connected (syntax).

The functioning of a theory of mind module has so far been merely described on a global level. While during the first years it was mainly analyzed whether the ToM ability is unique to humans and when it develops in children, recently the neurological bases of ToM have drawn interest (David et al. 2006; Vogeley et al. 2001; Shamay-Tsoory et al. 2006; Shamay-Tsoory et al. 2005). In parallel, models become more elaborated. Recently, cognitive based ToM (assumption concerning cognitive states) and affective based ToM (assumption concerning the other person's emotion) have been differentiated (Shamay-Tsoory et al. 2006; Shamay-Tsoory et al. 2005; Kalbe et al. 2007). Another distinction has been made between theory-theory and simulation-theory (Carruthers and Smith 1996) that basically disagree on the basis of a theory of mind. “Theory-theorists” suggest that the ability to explain and predict behavior is based on a folk psychological theory of the structure and functioning of the mind that might either be innate, learned individually, or acquired through a process of enculturation (Gopnik 1993; Wellman 1990; see Carruthers and Smith 1996). The simulationist view (Heal 1986; Harris 1989; Gordon 1986), on the other hand, holds that “what lies at the root of our mature mind-reading abilities is not any sort of theory, but rather an ability to project ourselves imaginatively into another person's perspective, simulating their mental activity with our own” (Carruthers and Smith 1996, p. 3). Thus, Humphrey (1984) argues that humans mind read by using their own experiences of introspection as a simulation of another's mental states. Gordon (1986) also holds a radical simulationism view, stating that human's concepts of mental states are acquired through a process of simulation, without subjects needing to have introspective access to their own mental states as such. With regard to the origin of the ToM ability, the different theoretical perspectives also hold diverging views: while theory-theorists partly propose that the theory is learned on the basis of experience (Gopnik 1996; Astington 1996; see also Frith and Frith, 2003) and partly assume that folk-psychology is embodied in an innate theory of mind module, the ability to simulate is always conceptualized as an innate genetic endowment. But not all authors see the different views as mutually exclusive: Heal (1996) and Perner (1996) try to close the gap between the different approaches and argue for a simulation/theory mix.

One of the most sophisticated models of ToM is presented by Baron-Cohen (1995) who explains the child's development of mind reading. He proposes an eye-direction detector and an intentionality detector that provide input to a shared-attention mechanism. While these allow the ability to read behaviors in terms of volitional mental states and to read eye direction in terms of perceptual mental states, the ToM module is needed as a way of representing the set of epistemic mental states and to tie together the volitional, perceptual, and epistemic mental state into a coherent understanding of how mental states and actions are related.

Consequences for Human-Artifact Interaction

In sum, it might be stated that the different models on common ground, perspective taking, and ToM show major similarities with regard to the fact that all propose that humans possess a direct but implicit knowledge of other humans (be it via simulation or via a learned or innate theory on fellow humans) that forms a starting point for mutual comprehension. Building on this, the dialogue can be used to clarify and broaden mutual knowledge by means of grounding processes. Concepts that can be seen as common human knowledge that can be built on are fairly simple. They include, for example, that humans have a limited cognitive capacity and cannot do several tasks at once; that humans have to eat several times a day; or that someone who does not immediately know the answer to a question shows that he/she is thinking about it.

Within the EU-funded project SERA, in which this theoretical framework was developed, data on long-term interactions with a robot installed in the household were collected. The robot was a revised version of the rabbit-shaped robot Nabaztag that was able to lead a dialogue on fitness-related issues. Especially in the least advanced form of the robot, at the beginning of the project, we observed that the participants’ goals were often misinterpreted—which is partly due to the fact that the Nabaztag has no concept of the subjects’ (humans’) goals and needs, respectively. From the data, we selected a specific scenario where the participant's intention to leave the house was misinterpreted, and we derived implications for an alternative course of the interaction. In the given situation, the subject enters the room wearing a scarf labeled ‘Liverpool.’ Then, the participant puts his/her jacket aside and takes his/her keys, which lie in front of the Nabaztag. The dialogue given below follows (Video: it1_p1_POct03_1246.mov):

Nabaztag: … It looks like you're going swimming. Please could you press the video button on your way past? Have a good time.

Subject: Not going swimming.

Nabaztag: Recording on.

Subject: No I'm not going swimming,that has been abandoned.

[No more reactions of the Nabaztag]

Every human being, in circumstances similar to this situation, could process a variety of cues indicating that the subject is not going swimming. An obvious visual aspect a human would acknowledge is that the subject is not carrying a sports bag. Although the Nabaztag has no visual perception and is thus not able to process this visual cue, a variety of other conversational and contextual information is available. For instance, it might not be the usual time to do sports (contextual information). For humans this is the usual time to have lunch instead. A part of humans’ communal common ground is that human beings have the physical need to eat at least three times a day. If the Nabaztag had at least basic communal common ground it could use this information to infer the subject's actual needs and goals, respectively.

What is missing in this scenario is not only a representation of the user's wishes and beliefs in the Nabaztag, but also general knowledge about human nature. More precisely, the Nabaztag lacks common ground and a theory of mind. If at least basic communal common ground were available, the course of the interaction could be as follows:

Nabaztag: It looks like you're going swimming. Please could you press the video button on your way first?

Subject: Not going swimming.

Nabaztag: Then you would probably go to lunch?

Subject: Yes I am.

Nabaztag: Recording on. Have a good time.

Subject: See you later.

The obvious consequence of these considerations is thus to try to implement common ground, perspective taking, and theory-of-mind-like abilities, including the agent's “awareness” of its own abilities and the basic knowledge about the human interaction partner. Therefore, a user model is needed that incorporates global knowledge on human needs and states. As described by all the models summarized above, humanness enables humans to directly infer states and thereby communication effects on other people—either by simulating or by an in-built or learned theory on other humans. The agent has neither: if it has a mind at all, it has no theory of its own mind and thus no possibility to simulate or impute its knowledge on others. Also, no agent so far has a complete user module that can be compared to a theory of mind in the sense of the theory-theory approach. Thus, a large number of small aspects that are taken for granted in human-human communication and would thus never be made explicit (e.g., feedback rules in turn taking) are not present in human-agent communication.

Clark's (1992) statement that it is hard to exaggerate the number and variety of basic concepts we take as common ground for everyone especially shows that it will be difficult to compile the knowledge on human nature that we rely on in everyday communication—even if we tried. Also, to simply implement rules or knowledge will probably not be sufficient: as Frith and Frith (2003) aptly state, mere knowledge will not be enough to successfully mentalize: “The bottom line of the idea of mentalizing is that we predict what other individuals will do in a given situation from their desires, their knowledge, and their beliefs, and not from the actual state of the world” (Frith and Frith 2003, p. 6). Nevertheless, a few researchers have begun to consider theory of mind as a fruitful concept that should be implemented in agents (Peters 2006), robots (Breazeal et al. 2004), or multi-agent systems (Marsella and Pynadath 2005).

However, a word of caution has to be issued: the requirement that a companion needs ToM capabilities addresses two challenges that have plagued AI for decades: the so-called “common sense problem” and the user-modeling problem. The project SERA does not pretend to be able to solve them once and for all time. Instead of postulating generalized and extensive ToM capabilities, then, it seems more useful to “demystify” them by first categorizing them into different types and then analyzing individually which requirements can be met with reasonable effort and which alternatives or workarounds can be found to compensate for others. Admittedly, this process involves a considerably reduced view of ToM capabilities as compared to human-human communication. We think that such a simplification is justified by the goal of achieving any kind of ToM capabilities in the system, in particular if the alternative is to give up any effort in the face of overwhelming complexity.

For a clearer view of the mentioned ToM capabilities, it is helpful to distinguish them by their properties:

1.	general: the “common ground” that can be assumed to be shared by all users, background
2.	individual: assumptions about the individual user and specific context
3.	static: assumptions of the system that will not change over time
4.	dynamic: what changes over time, what has to be learned or otherwise acquired at runtime

These properties provide us with a matrix of four types of capabilities, exemplified in the following matrix (see Table 2):

TABLE 2 Matrix of ToM Capabilities

	General	Individual
Static	time of day, human biorhythm widely shared interests, e.g., weather, news common language and conversation mechanisms (e.g., backchanneling) intentional stance (the basic assumption that the user is an autonomous agent) cultural practices, normative behavior, e.g., “taking the key” means “going out”, affective meanings assumptions about the target group assumptions about the respective social roles (status/power distribution, distance/closeness) assumptions about relationships	user's name, gender, age user's habits, specific biorhythm user's agenda, regular or scheduled activities user's interests, hobbies, preferences user's social network user's personality, personal style
Dynamic	Ageing of usersPassing of time (e.g., seasons)Relationship evolution, e.g., increase in familiarity	user's moods and needs (current, past, and prospective) user's plans and goals changes in social network (new contacts etc.) change in human-companion relationship user's change of habits, schedule, interests etc. Interaction history

It may be noted that some capabilities have both general and individual, and static and dynamic components, for example, relationship building and maintenance:

	general, static: assumptions about social role and status of the system and the user, respectively
	general, dynamic: assumptions about the (normal, expected) evolution of relationships and changes of status
	individual, static: the user's prevalent model of the system (e.g., device-like, pet-like), user's personality traits and relationship preferences
	individual, dynamic: building and maintaining the specific relationship.

The distinction of these capabilities makes it possible to analyze them and to design for them individually:

1.	General, static components are usually “built in” and even implicit in the design. These assumptions are present on practically every level, from the language used to the decision to use the key hook as a “meaningful object,” or when it is appropriate to launch the “good morning” dialogue. Designers practically cannot help themselves using the numerous resources they share with their users. When a system is designed for portability between target groups, applications, and cultures, however, it becomes important (and turns out challenging) to make them explicit and to “package” them. The challenge here is that the amount of general knowledge humans have and use is intimidating: “normal” everyday behavior relies on life-long learning and experiences. The methods to deal with these problems include: limiting user expectations (by appearance of the device, by self-disclosure) restricting domains (in our case, the domain of fitness and activity) task-oriented interaction (in our case, the task of supporting the user's monitoring of his/her physical activities and fitness) tricks, e.g., pattern matching and general encouraging feedback in the ELIZA tradition as used in chatbots; by changing topics back to the domains and tasks covered by the system, by back-channeling that invokes understanding, and so forth.
2.	General, dynamic components can also be “hard-coded,” but on a different, higher level. They involve change or substitution of behaviors (e.g., form of address) which can be built in from the start but is triggered by events such as date or the number of past interactions. They could also be triggered by remote intervention (e.g., via the Internet).
3.	Individual, static components can be either built in or user configured. In our case, the subject's activity plan is elicited in pre-test interviews and is built in from the beginning. A different way to acquire components that still does not involve machine learning would be a more open system that gives the user the possibility to change his/her profile, agenda, contacts, and preferences, via a different and technically simpler modality than that used for interaction, for example, a screen-based graphical user interface (GUI), where the complexity can range from questionnaire to an authoring environment.
4.	Individual, dynamic components are undoubtedly the biggest challenge. Two capabilities can be distinguished: learning and adaptation. By learning capabilities we mean the acquisition of knowledge and consequent permanent change of behavior. In other words: prior assumptions have to be overwritten, new knowledge has to be added, (e.g., a new contact or a change of habits). This corresponds to long-term memory. By adaptation we mean the temporary change of behavior triggered by perception or interaction, (e.g., the user's mood or health). It also includes short-term memory, in other words, building a history of the current interaction (e.g., knowing which topics have been covered) or of several interactions during the day.

The latter aspects are especially similar to what has already been suggested in 2003 in an approach by Fong et al. They also state that that a meaningful interaction between human and social robot is only possible when the robot is able to perceive the world as humans do and can have human-oriented perception. Also, they posit that a lack of common ground can lead to problems in communication. Socially situated learning is proposed as a possibility to resolve these problems, since learning can be used to transfer information. With regard to the differences in sensing and perception of human and robot (agent), learning becomes crucial to overcome the problem of the different views of the world.

Alternative Approaches: Human-Dog Interaction

Shortcomings and advantages of the human communication paradigm have implicitly already been mentioned. The major advantage is, of course, that humans will not have to adapt in any way when they want to communicate with robots or virtual agents. On the other hand, the major shortcoming is, as depicted above, that crucial abilities for human-like communication cannot easily be implemented. Thus, the question has to be asked whether human-like communication is actually necessary—all the more if we do not plan to have relationships with robots that resemble the relationships to our partner or children (see discussion below). Alternatively, we can ask whether it would be sufficient to provide an artificial entity with the communicative abilities of, for example, a dog. Also, it can be asked whether it would be feasible to develop a radically new and innovative form to model communication between humans and artifacts – one that draws neither on human-human communication nor on human-animal communication.

The answer might depend on the level of interaction one is targeting. Here, we would argue that with regard to a very basic level of communication (e.g., the concrete dialogue) the basic prerequisites of human communication have to be modeled in order to guarantee a basic mutual understanding. This might already be true when communication is not even intended to be verbal but, for instance, is based on gestures. In this line, Tomasello (2008) impressively shows that some human communicative acts cannot be understood by primates since they lack a cooperative mindset and therefore are not able to understand a gesture that is meant as altruistic information for the interaction partner. The ability that enables humans to produce and understand these kinds of gestures or communicative acts is termed “shared intentionality.” Without this ability, it will not be possible to understand gestures and other communicative signals —especially when these are issued in a cooperative context.

Because these abilities and the corresponding communicative acts operate on an involuntary, automatic level, and are ubiquitous (as humans are inherently social and focused on communication and cooperation), humans will not easily be able to suppress these mechanisms of communication to learn new forms. Humans will certainly be able to learn ten sentences that will work in a communication with robots/agents but this (1) will not be satisfying, especially when long-term relationships should be built and (2) will only work for aspects that humans have to learn in addition to their usual communication but not work for things humans have to omit when interacting with robots/agents. With regard to the latter, there are numerous examples from the SERA data collection showing that humans—although they are consciously well aware that the robot does not understand anything they say—make extensive use of communication (e.g., lengthily correct the robot's assumptions) that will not be processed by the robot.

Nevertheless, human-dog communication has long been proposed as a framework that might be helpful to model human-robot/agent interaction (Dautenhahn 2004; Dautenhahn and Billard 1999). When human-dog communication was first mentioned in the debate on the design of human-robot interaction, it was assumed that dogs learn to communicate with humans during their ontogenesis (Dautenhahn and Billard 1999). This was seen as a potential model for robots that could be built to also learn to adapt to a human and his orher individual peculiarity. However, current research shows that dogs are able to communicate with humans (even better than apes and primates can) not because of their learning abilities in ontogeny, but due to the fact that they have been subject to a human-steered evolution over thousands of years (Tomasello 2008). Similarly, Miklósi (2009) argues: “Researchers assume that over these years there has been both an unconscious and a conscious selection for dogs with enhanced communicative abilities which fitted the particular task. Thus the ease at which (in a normal case) a pet dog develops a communicative relationship with the owner or the whole social group (‘family’), is neither accidental nor is it destitute of some evolutionary predispositions.” Indeed, dogs have several abilities that facilitate smooth interaction with humans: they are able to initiate communicative interactions, rely on visual human gestures, and recognize simple forms of visual (joint) attention (Miklósi 2009). It has been shown that if dogs face an unsolvable task or situation they utilize both gazing and gaze alternation in order to initiate a communicative interaction with humans (Miklósi et al. 2000; 2003). Given these results and insights it has to be noted that for implementing actual communicative acts in robots and agents, the dog-human-interaction model does not provide a more fruitful basis compared with human-human interaction, because dogs also have basic abilities (in this case acquired by human-steered selection) that will not be easily isolated and implemented. Also, it has to be noted that humans themselves in interactions with humans do not use any other interaction patterns as when they interact with fellow dogs. They all rely on the same principles of joint attention, theory of mind, deictic gestures, shared intentionality, and other basic prerequisites for communication. They also talk to their dogs although the dogs probably do not understand the utterances. The only difference might be that their expectations are lower than the expectations they have when approaching a fellow human or—maybe—a robot or agent with human-like appearance. This, however, is not an aspect of the micro-level but rather of the meso-level in terms of the relationship that is built between human and robot—and which will be commented upon in the following paragraph.

Meso-Level

On the meso-level, we will discuss models for relationship building of humans and artifacts. Here, we will comment on a human attribute that will facilitate the relationship building also with artificial entities: the need to belong. Additionally, several factors for the decision to engage in relationship building are named. Also, theories describing the prerequisites for the establishment and maintenance of relationships are summarized as well as basic dimensions of human relationships described.

The Need to Belong

When seen from the part of the user, we can conclude that relationship building has a good chance of taking place as long as basic prerequisites on the part of the robot/agent are met. This is due to the fact that humans have been shown to possess a need to build relationships. This has been termed the “need to belong” (Baumeister and Leary 1995). In their article on belongingness, Baumeister and Leary (1995) suggest that “human beings are fundamentally and pervasively motivated by a need to belong, that is, by a strong desire to form and maintain enduring interpersonal attachments” (p. 522). This human motivation has, according to Baumeister and Leary, “multiple links to cognitive processes, emotional patterns, behavioral responses, and health and well-being” (p. 522). Consequently, all of us are interested in having warm and positive relationships and making and maintaining friendships as key conditions for happiness (Berscheid 1985; Berscheid and Peplau 1983; Berscheid and Reis 1998). As social beings, our need to belong leads us to regularly seek the company of other people. This tendency is also referred to as affiliation. According to these findings, affiliation is a natural phenomenon, a deep-rooted pillar of human existence. The need to belong is the basis for the social orientation of human beings. In order to satisfy this need, we seek the company of others: we build groups (e.g., families, cliques) and are interested in the others’ lives, we help each other and join clubs just because the satisfaction of the need to affiliate makes us happy (see also Cacioppo and Patrick 2008; Ryan and Deci 2000). In line with this, Kappas (2005) aptly stated that humans are “free monadic radicals.” They will try to bond and affiliate with anything that is interactive and has basic social cues such as, for example, speech (see Reeves and Nass 1996; Nass and Moon 2000).

With regard to the design of the robot, one would not derive that an artificial entity itself will need a need to belong, but that one will have to be careful when deciding which attributes to include so that the humans’ need to belong will also be transferred to artificial entities. However, it is also a fact that when there is choice, humans will not just bond with any entity but that there are factors that influence who is attractive and whom we choose for the establishment of a relationship (see Aronson, Wilson and Akert 2007).

Factors for Attraction and Establishment of a Relationship

It can be assumed that humans will draw on similar criteria, as they experience in human-human encounters, when deciding whether they would like to interact again with a robot. Besides propinquity (see mere exposure effect; Zajonc 1968) and similarity (Berscheid and Reis 1998; McPherson, Smith-Lovin, and Cook 2001), it is of course (physical) attractiveness that plays an important role. Here, the finding, “what is beautiful is good” (Dion, Berscheid, and Walster, 1972), in the sense that attractive people are also rated positively in other aspects can also be assumed to be true for robots. An additional factor influencing relationship building is reciprocal liking: since all humans like to be liked, we are attracted to others who behave as if they like us. No matter if the signals are nonverbal or verbal, whether we like a person or not depends our judgment about the extent to which the other person likes us (Berscheid and Walster 1978; Kenny 1994; Kenny and La Voie 1982; Kubitschek and Hallinan 1998). Liking can even compensate the absence of similarity (Gold, Ryckman, and Mosley 1984). As Curtis and Miller (1986) demonstrated, reciprocal liking might as well be the result of a self-fulfilling prophecy. People who believed they were liked by their counterpart, generally behaved more likeable and were at the same time liked more than the participants who believed they were disliked. In line with this, design guidelines for artificial entities should follow from knowledge about reciprocal liking. The robot should give its user the impression that it likes him or her and appreciates his or her presence since this increases the likeability of the system, as long as this is authentically implemented. Depending on the setting, this may well be realized with the help of ingratiation (i.e., by praising the user). In contrast to people with a positive or moderate self-concept, however, people with a negative self-concept tend not to respond to the friendly behavior of others and will accordingly provoke negative reactions affirming their negative self-concept (Swann, Stein-Seroussiand, and McNulty 1992). Therefore, again, it is important to not rely too much on seemingly simple, straightforward rules that are derived.

Theories on Social Exchange and Equity

A theory that brings together the different determinants of attraction discussed so far is the social exchange theory (Homans 1961; Blau 1964; Thibaut and Kelley 1959). Assuming that relationships are comparable to a marketplace where costs and benefits are exchanged according to economic principles, this theory is described as “the idea that people's feelings about a relationship depend on their perception of the rewards and costs of the relation, in the kind of relationship they deserve, and their chances of having a better relationship with someone else” (Aronson, Wilson, and Akert 2007, p. 319).

A person's level of satisfaction in a relationship is determined by the comparison level (Kelley and Thibaut 1978; Thibaut and Kelley 1959). This expectation about the outcome of rewards and punishments the person is likely to receive in a relationship is established through experience gained in a number of previous encounters and results in a high or a low comparison level. Furthermore, the level of satisfaction also depends on the evaluation of the comparison level for alternatives, in other words, the assumption on what one would receive in an alternative relationship or the perception of how likely one could find an alternative partner to replace the old relationship. As results of research on social exchange theory has shown, costs borne and rewards earned in a relationship are important to both friends and romantic couples, and influence how satisfied people feel about a relationship (Bui, Peplau, and Hill 1996; Le and Agnew 2003; Rusbult 1983; Rusbult, Martz, and Agnew 1998).

Resulting from criticism accusing social exchange theory for neglecting equity or fairness as an essential variable in social relations, the so called equity theory was proposed. It assumes that people are concerned about equitable relationships in which the contribution of rewards and costs made by the partners are roughly equal (Homans 1961; Walster, Walster, and Berscheid 1978). According to the proponents of equity theory, relational satisfaction arises from equitable relationships. Compared to inequitable relationships, in which the partners feel uneasy about the perceived imbalance, equitable relationships are the happiest and most stable relations.

Further, with regard to social exchange in close relations, the investment model has been developed. It suggests that in long-term relationships, not only the level of satisfaction with a relationship regarding rewards and costs, comparison level, and the comparison level for alternatives play a role but also the perception of what has been invested that would be lost by ending the relationship (Rusbult 1983). Thus, in order to be able to predict the duration of an intimate relationship one has to know about these additional determining factors.

Clark and Mills (1993) differentiate between exchange relationships that usually appear among new acquaintances and communal relationships, typically found among close friends, members of a family, or romantic partners. Exchange relationships are characterized by the need for equity (i.e., the need for equal contributions by the parties in terms of costs and rewards). In relationships governed by the equity norm, people keep track of the contributions and feel exploited when they perceive an imbalance between their own contribution and the benefits they receive from the relationship. Communal relationships, in contrast, are more characterized by responding to the needs of partners, family members, or friends without expecting to be paid back (Clark 1984, 1986; Clark and Mills 1993; Mills and Clark 1982, 1994, 2001; Vaananen et al. 2005).

Thus, equity in long-term relations operates in slightly different ways. Although in new and casual encounters, we expect an immediate compensation of a contribution, a rigid tit-for-tat rule, we apply looser give-and-take rules the longer and closer we get to know a person (Kollack, Blumstein, and Schwartz 1994; Laursen and Hartup 2002; Vaananen et al. 2005).

According to these models, not only an initial balance of costs and rewards in the beginning of robot-human relationships has to be considered and catered for, but it has also to be taken into account that these short-term relationships that follow specific rules move on to become long-term relationships that follow different rules at a certain point. Thus, the robot/agent has to be of use to the user, so that the user might at least initially feel a balance in the relation. A user's feeling of a balance between contributions and rewards from the interaction with a robot is important for the maintenance of the relationship in the beginning. This can be achieved when the agent/robot is able to effectively help with everyday tasks such as reminding of appointments, providing the weather forecast, or receiving and announcing messages. After this initial phase in which a give-and-take rule is applied, the user possibly perceives his/her relationship toward the agent/robot as a communal one, so that equal contributions become less important. Potentially, users feel a strong bond with their robot, so they do not consider or reject alternatives and might feel bad about ending the relationship. Besides these features that generally can be implemented once before the interaction starts, a specific model of the user and the common “history” of user and robot will be needed in order to render ongoing communication and relationship management and development successful and satisfying.

The question that arises is whether humans tend to compare a relationship with an artificial entity with the cost and rewards invested in “real” human-human relationships, or if other rules are applied. Also, it has to be considered, to what kind of relationships the relationship with a robot/agent is compared. Comparing the robot with a child may probably trigger other feelings and reactions and imply other expectations compared with, say, a pet. People invest much money in the best food and the best medical care for their pets. The fact that many people have intense relationships with their dogs, cats, or birds—although these animals can neither speak nor have any concept of human communication—suggests that the emotional rewards people perceive seem to outweigh the costs they invest. Unlike these animals, robots are no living creatures, they are not warm and do not (at the moment) make the impression of acting autonomously. However, our data in the SERA project show that people are influenced by a robot's presence, at least; they feel that there is “something.” A further difference in the relationship between humans and robot is the emotional component involved in ending a relationship. Humans would probably not have the impression that they let their robot down when they leave the house or ignore it, etc. This is also true for ending the relationship. If future research, however, shows that humans build bonds that will lead them to feel sorry for the ending of the relationship with a robot/agent, of course ethical questions will have to be discussed. This was already brought up by Sproull et al. (1996) concerning the relationships people might build with computers: “Many people want computers to be responsive to people. But do we also want people to be responsive to computers?” (p. 119).

Basic Dimensions of Human Relationships

As was similarly shown with regard to communicative acts on the micro-level, we can assume that a human will not use fundamentally different patterns in a relationship with a robot compared with a relationship with a fellow human. One can differentiate basic patterns and dimensions for the perception of a relationship that will not differ from those that are applied in human-human relationships. Most prominent patterns and dimensions are intimacy (as discussed above, for example, with regard to the need to belong) and dominance.

While the concept of “Need to Belong” covers one dimension of interpersonal relationships, there are, of course, several other dimensions to it. Not only warmth, friendliness, and closeness play a role in bonding and belonging but also hierarchy and status (Burgoon and Dunbar 2000). Metaphorically, humans seek the place where they belong not only “horizontally” (= peers, friends, equals) but also “vertically” (= hierarchical, super/subordinated). There are power differences among family and clique members, and belonging to a family/clique involves not only relationships of closeness, but also having a secure place in the hierarchy. Equal to human-human relationships, this might as well be relevant to a human-companion relationship that is likely to have status/power differences. This second dimension also plays a role with regard to, for example, the implementation of (im)politeness, which has to do primarily with status threat/respect (“vertical” distance).

Conclusions with Regard to Relationship Building with Robots

In conclusion, it can be said that similar to what has been shown on the micro-level, with regard to specific communicative acts, here it also becomes apparent that a human being does not leave his/her usual view on the world. Just as the establishment of relationships with fellow humans depends on the question whether and to what extent the individual's need to belong is already satisfied, this will also be true for the establishment of relationships with robots and agents. Likewise, robots and agents will not be chosen as potential companions based on fundamentally different criteria than those that apply for fellow humans. Also, in a relationship with robots/agents, the same equity mechanisms will apply as in human-human relationships, and people will certainly strive for a balance between gains and costs. Also, the dimensions people use to perceive their relationships in cannot be radically new dimensions but will, for example, target the aspect of dominance and hierarchical structures. Thus, the human does not leave his/her world just because she/he is interacting with a robot. However, the readiness to apply all these aspects is of course not independent from the design of the agent/robot. If the entity looks human, people will involuntarily be more inclined to draw on their usual relationship experiences than when it is a blinking blue ball.

Macro-Level

On the macro-level, we will discuss the role of the robot companion in a more general sense. Here, we will rely on the very few studies and texts that have been presented in robot/agent research and will summarize the discussion that we had on this topic during the SERA project. Therefore, this paragraph rather addresses less empirical and more normative aspects of companions.

One of the few studies in this realm has been presented by Dautenhan (2004). She explored people's perceptions and attitudes toward the idea of a future robot companion for the home. Results (of only 28 participants) indicated that a large proportion of participants were in favor of a robot companion and saw its potential role as being an assistant, machine, or servant. However, only a few expressed the wish that the robot companion might be a friend. Household tasks were preferred to child/animal care tasks.

Discussions in the SERA project also centered around the notion that robots and agents should not be designed in such a way that they take the role of the partner or child. Just as Fernaeus (2009) suggests a framework model of technology as a resource for human action and defines technology as something that can assist with certain tasks, it was concluded that the robot's persona and task should not be too close to what we know from intimate human relationships.

Therefore, less intimate social roles or personalities were discussed, such as a butler or maid personality, a health adviser or a manager (for a specific part of the user's life). All of these social roles were associated with different capabilities of the system and expectations by the user.

However, from an empirical stance, it can be seen as difficult if not impossible to predefine the role of the robot. In the end, the human user defines the way she/he perceives, communicates with the robot/agent, and which role she/he assigns to the artificial entity (see also the results of the media equation, Reeves and Nass, 1996). Moreover, all of these approaches are very restrictive to only one kind of persona, social role, or personality and thus are not really drawn from life as humans are not limited to one social role but fulfill a variety of roles in daily life. The conclusion was derived that the robot cannot consistently take one social role, and it should not. The starting point for the robot's persona is that of the “companion.” We have to go beyond imitation of single human roles toward a genuine companion identity, which is a collection of different identities. This concept is more comparable to real life where humans also incorporate a variety of social roles and different identities. In rich, long-term, human-human relationships, it is normal to integrate the diversity of social identities of the partners: one may have working relationships with a friend, mothers who act as playmates, couples who are “buddies” in sport or hobby, etc. In consequence, we decided that the artificial entity should be perceived as autonomous and pro-active, and as being of real use. It is thus not fruitful to create “the” perfect persona, but instead to offer opportunities to the user to attribute roles and personality.

LIMITS OF DESIGNING FOR SOCIABILITY

As was discussed throughout this paper, it became apparent on every level of sociability that the potential of the designer to design for sociability is limited and that, in the end, it is the user who defines communication, relationship, and roles. We would nevertheless like to highlight two specific aspects that make it difficult to take design decisions and which challenge the attempt to present a theory of companions that enables deriving rules for the design of robots and agents.

1.

Difficulty to derive design decisions from theory Here, several aspects are causative for this difficulty: on the one hand, there is still a lack of understanding of, for example, human communication. On the other hand, what is known is increasingly complex and cannot be implemented easily (see, e.g., basic prerequisites such as common ground or theory of mind that seem to be simple at first glance, but it takes years to implement just the first aspects, see Breazeal et al. 2004). This leads to the fact that although developers would like to derive the implementations (e.g., dialogue structure) from theory, they end up falling back on their personal experience and skills. And it is, indeed, difficult to derive concrete aspects such as dialogue moves from something as general as theories and assumptions of human communication. What is, however, necessary, is to critically reflect the concrete dialogue, once it has been designed, against the background of the theoretical knowledge. More important, how it is perceived by the user has to be tested empirically. Just as in human-human communication, we will see discrepancies between the sender's communicative goal and the receiver's perception and understanding of the message. This directly leads to the second aspect that makes the task to design human-robot/agent interaction difficult:

2.

Individuality of specific user, idiosyncratic construction of communication Our results, when observing the participants interacting with the robot rabbit in the SERA project, show that people's reactions are utterly individual and idiosyncratic. Although the same functionalities and abilities were implemented, the participants’ handling of and communication with the system differed to a large extent. Employing a qualitative rather than quantitative approach, this fact emerged more clearly than in previous studies. This, of course, leads to the conclusion that it will not be possible to design a system that will be perceived as helpful, efficient, and satisfying by all participants. In this line, Dautenhahn (2004) already suggested that individualized and personalized robot companions will have to be developed. She argues that individualized robots are necessary due to human nature, since people have individual needs, likes and dislikes, preferences, and personalities that a companion would have to adapt to. On the other hand, it must be noted that individual differences will not hinder the development of sensible companions just as the fact that people have individual attributes does not hinder the engineering of cars.

However, this is not to say that there are no fruitful endeavors to advance human-robot interaction or that the depicted difficulties will ultimately obstruct the development of robots and artifacts that do not only lead meaningful conversations with humans but also qualify as companions. Highlighting these difficulties would hopefully lead to greater caution and awareness with regard to these aspects.

CONCLUSION AND IMPLICATIONS FOR FUTURE STUDIES AND METHODOLOGY

The aim of this paper was to discuss whether a theory of companions is needed and what it has to include. More specifically, it was argued that sociability is a complex whole and if we would like to implement it in its complexity, we have to attend to different levels that address the actual communication, the relationship, and the roles that might be assigned. On all levels one might draw on theories from human-human interaction, and indeed it was concluded that there is no real alternative to this. This is due to the fact that humans in their interactions with robots and agents will not stop to employ and expect the communicative mechanisms they are used to. Also, it was shown that the only realistic alternative to human-human communication as a model, human-dog communication, also largely relies on the same mechanisms, since dogs have been adapted to the human communication system. Although there does not seem to be a way to establish a radically different model for human-robot/agent interaction, we would not say that a theory of companions is obsolete and that merely human-human communication should be used as a framework. Although we suggest deriving aspects from human-human interaction, companions need not necessarily mimic human-human relationships. They are devices that satisfy certain needs of their owners and have their uses and functions in the owners’ lives. When they have the function to support the owners’ health, well-being, and independent living, however, they assume a role that goes far beyond that of, say, a vacuum cleaner, and they have to be able to maintain that role over a longer period. In this light, it becomes essential to investigate how long-term relationships are built and re-built on the micro-level of conversational interaction. A theory of companions in the form it has been sketched here, as grounding in human-human communication but being amended by additional aspects, is necessary to guide empirical analyses in the realm. Moreover, the theory of companions should be refined to in future guide not only empirical studies but derive design guidelines (but see discussion above on problems with this approach) and to help explain human reactions toward artificial entities.

Future research should also include qualitative aspects because results of the SERA project showed that these are especially helpful for observing and understanding people's idiosyncratic reactions. However, future research also has to be quantitative, as this opens up the possibility to include additional aspects like tasks that will have to be focused not only within the theory of companions but also in empirical analyses.

In sum, it has been shown that a theory of companions without considering the human user and his/her needs, perceptions, and communication patterns will not be useful. Instead of a theory of companions we therefore need a “Theory of Companion-Human Interaction” in which the human user is taken into account from the very first steps in the design process onward to the empirical analysis of successful long-term relationships between humans and artifacts.

Acknowledgments

This work was supported by funding from the European Community's Seventh Framework Programme [FP7/2007-2013] under grant agreement no. 231868 and project name Social Engagement with Robots and Agents (SERA).