ArticlePDF AvailableLiterature Review

Abstract and Figures

One of the many purposes for which social robots are designed is education, and there have been many attempts to systematize their potential in this field. What these attempts have in common is the recognition that learning can be supported in a variety of ways because a learner can be engaged in different activities that foster learning. Up to now, three roles have been proposed when designing these activities for robots: as a teacher or tutor, a learning peer, or a novice. Current research proposes that deciding in favor of one role over another depends on the content or preferred pedagogical form. However, the design of activities changes not only the content of learning, but also the nature of a human–robot social relationship. This is particularly important in language acquisition, which has been recognized as a social endeavor. The following review aims to specify the differences in human–robot social relationships when children learn language through interacting with a social robot. After proposing categories for comparing these different relationships, we review established and more specific, innovative roles that a robot can play in language-learning scenarios. This follows Mead’s (1946) theoretical approach proposing that social roles are performed in interactive acts. These acts are crucial for learning, because not only can they shape the social environment of learning but also engage the learner to different degrees. We specify the degree of engagement by referring to Chi’s (2009) progression of learning activities that range from active, constructive, toward interactive with the latter fostering deeper learning. Taken together, this approach enables us to compare and evaluate different human–robot social relationships that arise when applying a robot in a particular social role.
Content may be subject to copyright.
Social/dialogical roles of social
robots in supporting childrens
learning of language and literacy
A review and analysis of
innovative roles
Katharina J. Rohlng
*, Nicole Altvater-Mackensen
Nathan Caruana
, Rianne van den Berghe
, Barbara Bruno
Nils F. Tolksdorf
and Adriana Hanulíková
Developmental Psycholinguistics, Faculty of Arts and Humanities, Paderborn University, Paderborn,
Developmental Psychology, Psychologisches Institut, Johannes-Gutenberg-Universität
Mainz, English Linguistics, University of Mannheim, Mainz, Germany,
School of Psychological Science,
Macquarie University Centre for Reading, Macquarie University, Sydney, NSW, Australia,
Urban Care &
Education, Windesheim University of Applied Sciences, Almere, Netherlands,
Lausanne, Switzerland,
Language and Cognition, Deutsches Seminar, Albert-Ludwigs-Universität
Freiburg, Freiburg, Germany
One of the many purposes for which social robots are designed is education, and
there have been many attempts to systematize their potential in this eld. What these
attempts have in common is the recognition that learning can be supported in a
variety of ways because a learner can be engaged in different activities that foster
learning. Up to now, three roles have been proposed when designing these activities
for robots: as a teacher or tutor, a learning peer, or a novice. Current research
proposes that deciding in favor of one role over another depends on the content or
preferred pedagogical form. However, the design of activities changes not only the
content of learning, but also the nature of a humanrobot social relationship. This is
particularly important in language acquisition, which has been recognized as a social
endeavor. The following review aims to specify the differences in humanrobot social
relationships when children learn language through interacting with a social robot.
After proposing categories for comparing these different relationships, we review
established and more specic, innovative roles that a robot can play in language-
learning scenarios. This follows Meads (1946) theoretical approach proposing that
social roles are performed in interactive acts. These acts are crucial for learning,
because not only can they shape the social environment of learning but also engage
the learner to different degrees. We specify the degree of engagement by referring to
Chis (2009) progression of learning activities that range from active, constructive,
toward interactive with the latter fostering deeper learning. Taken together, this
approach enables us to compare and evaluate different humanrobot social
relationships that arise when applying a robot in a particular social role.
dialogical roles, humanrobot social relationship, childrobot interaction, language
learning, literacy, social roles
Suna Bensch,
Umeå University, Sweden
Huili Chen,
Massachusetts Institute of Technology,
United States
Jan de Wit,
Tilburg University, Netherlands
Katharina J. Rohlng,
This article was submitted
to Humanoid Robotics,
a section of the journal
Frontiers in Robotics and AI
RECEIVED 17 June 2022
ACCEPTED 19 August 2022
PUBLISHED 05 October 2022
Rohlng KJ, Altvater-Mackensen N,
Caruana N, van den Berghe R, Bruno B,
Tolksdorf NF and Hanulíková A (2022),
Social/dialogical roles of social robots in
supporting childrens learning of
language and literacyA review and
analysis of innovative roles.
Front. Robot. AI 9:971749.
doi: 10.3389/frobt.2022.971749
© 2022 Rohlng, Altvater-Mackensen,
Caruana, van den Berghe, Bruno,
Tolksdorf and Hanulíková. This is an
open-access article distributed under
the terms of the Creative Commons
Attribution License (CC BY). The use,
distribution or reproduction in other
forums is permitted, provided the
original author(s) and the copyright
owner(s) are credited and that the
original publication in this journal is
cited, in accordance with accepted
academic practice. No use, distribution
or reproduction is permitted which does
not comply with these terms.
Frontiers in Robotics and AI frontiersin.org01
TYPE Review
PUBLISHED 05 October 2022
DOI 10.3389/frobt.2022.971749
1 Introduction
A large body of research points to the success of robots designed
for the purpose of education (Mubin et al., 2013;Belpaeme et al., 2018;
Kanero et al., 2018;Lee and Lee, 2022). Existing reviews have
identied a variety of application domains going hand in hand
with different roles for the robots. For learning, Mubin and
colleagues (2013) characterize these roles as levels of involvement
of the robot, and these authors differentiate between a passive robot
that can be used as a tool to be programed and a co-learner(p.3)
that is active, can be more involved, and can inuence the learning
process. Ahmad and colleagues (2021) summarize the social roles that
a robot can fulll in learning as being a teacher or tutor, a learning
peer, or a novice (see Section 2 for more details on established roles).
Roles determine the robots behaviors, but also its responsibility for,
and thus contribution to, the learning process. To fullltheroleofa
teacher, for example, a robot has to initiate the interaction and guide
the learner toward becoming knowledgeable on the taught content.
Ahmad et al. (2021) further propose that the decision for one role over
another depends on the content, the tutor or instructor, the form of
student and the essence of the learning process(p. 295). Hence, the
design of the role that a robot fullls in a learning process clearly has
manifold consequences not only for the robots appearance but also
fortheinteractionandthelearningprocess. Despite these far-reaching
consequences, and even though the literature offers different forms of
designing an interaction with a robot for the purpose of enjoyable (Lin
et al., 2022) and successful learning, little is known about whether and
how these different roles can be designed systematically, let alone how
they differ in shaping a humanrobot social relationship (De Graaf,
2016;Tolksdorf et al., 2020).
In the following review, we aim to specify the differences between
the roles for a robot by proposing categories that make it possible to
perform a systematic comparison. We thereby follow Meads (1946)
theoretical account according to which any interaction brings about a
role. This role is both social and dialogical. It is social, because it
reectsarelationofanindividualtoasocialgroup(Mead, 1946,p.
164). For example, by being a tutor, a person has to teach a learner and
is considered to be more knowledgeable than the learner. This relation
shapes attitudes and expectations. These become observable in
interactive acts in the form of verbal and nonverbal
communicative behaviors directed toward the others. This is why
a social role is also dialogical. In other words, whereas attitudes and
expectations are socially motivated, they are accessible through
communication. When interacting with others, the performance of
interactive acts is inuenced by the role. For example, a tutor will
provide an explanation, whereas a learner can ask questions.
Extending what is known about the roles and the way they shape
interactive acts, this review aims to differentiate the abilities that serve
these roles. This extension contributes a framework for the design of
social robots that should raise awareness among scientists, developers,
and users as to what kind of capabilities need to be implemented for
what kind of interaction to serve what kind of educational purpose.
Accordingly, in Table 1, we differentiate the perceptual,cognitive,
and dialogicalabilities that need to be implemented in a robot in
order to fulll a particular role: Whereas perceptualabilities enable a
robot to perceive specic communicative signals, cognitiveabilities
can be implemented in a variety of ways leading to different levels and
complexities in processing the perceived information. Finally, with
dialogicalabilities, robots are able to engage with a social partner.
In addition to our claim that roles shape an interaction, we
consider learning processes to require a particular awareness of roles.
This is because the interactive acts performed are crucial: Not only can
they shape the social environment of learning, but they also engage
the learner to a different degree (Chi, 2009).Thedegreeofengaging
the learner can be specied more clearly when following Chi (2009),
knowledge. Through analyzing the acts performed by tutors and
learners in detail and across many studies, Chi (2009, p. 73)
formulated a conceptual framework for differentiating learning
activities in terms of observable overt activities and underlying
learning processes.Within this framework, activities can be
differentiated into active (doing some physical movements while
learning), constructive (producing additional output with contents
that go beyond the given information), and interactive (participating
in a dialogue characterized by exchange and co-construction of
follow-up activities such as defending a position, elaborating, etc.).
These overt activities differ with respect to the underlying learning
processes (Chi, 2009,p. 77): In active activities, attending processes are
elicited. Their function is to activate existing knowledge or to store
new information. In constructive activities, new knowledge inferences
or integrations are elicited. With these kinds of activities, it is
necessary to organize ones own knowledge in order to gain
coherence. Finally, interactive activities elicit creating processes
(p. 77) in which it is necessary to incorporate a partners
contribution. Based on studies comparing the different forms of
activities, Chi (2009) suggests that for deeper learning to occur,
interactive activities are required. To the best of our knowledge,
activities initiated by social robots have barely been analyzed with
respect to what cognitive processes they elicit and how they might
thus foster learning. In Table 2, we describe different forms of
activities that can guide the design of a robot depending on what
kind of interaction (and thus other forms of learning) is intended.
Following Meads (1946) suggestion that a particular role should be
seen as reecting ones position within a group, we critically reect on
how these activities change the social context and might shape the role
of the other group members involved. For example, when a robot can
provide an individual treatment for a childs limited vocabulary, a
fostering this area of language. Another example could be that the
presence of a social robot fullling a particular role opens up new
possibilities for how other children can be involved.
In the following, we rst review the most common roles that a
social robot fulllsincurrentresearch(Section 2). Adding to the
existing roles, we further review innovative roles that a robot can fulll
for the purpose of promoting language learning and literacy (Section
3). First, the selection of the innovative roles is driven by the intention
Frontiers in Robotics and AI frontiersin.org02
Rohlng et al. 10.3389/frobt.2022.971749
to extend the already well-established roles of a robot as a learner,
tutor, or peer in RALL (robots assisting language learning). We opted
for roles that can be specied in overt dialogical acts. Instead of
characterizing a robot as a tutor, we propose acts with which a robot
can assist a tutor or a teacherthat is, assisting (Section 3.1)and
supporting (Section 3.2). Tackling on novel aspects of assistance and
support, we review possibilities of assisting also children. With regard
to supporting activities, the review focuses on socioemotional aspects
that have barely been considered in the design of robots so far. Second,
our selection of the roles for the review was motivated by aspects from
developmental studies. These aspects are reected in four roles that
strengthen the childs own engagement and learning strategies:
TABLE 1 Denition of innovative roles and references to the existing literature.
Role Denition
of the role
Required abilities of the robot to fulll the role Studies on
learning and
Perceptual Cognitive Dialogical
Robot offers socioemotional
support to alleviate anxiety
and promote engagement
during learning
To perceive, at a minimum,
whether the child is engaged
in the learning activity. This
will allow the robot to
appropriately at key points
in the activity, or to identify
when additional
encouragement is needed if
the child has become
distracted/disengaged. It
would be additionally
benecial, but not essential,
for robots to perceive mood
states in children
To be capable of modeling
the optimal frequency of
feedback and animacy
during a learning activity.
This will enable robots to
provide enough feedback
to facilitate engagement
without the risk of this
form of feedback or
animation form becoming
a distraction. This is
important, because robots
serving this role are not
directly supporting the
learning processes and so
should not be the childs
primary focus of attention
To initiate dialogue that can
demonstrate the robots
sentience and own
engagement in the learning
activity (to promote task
engagement and trust in
robot competence).
Dialogue could also offer
when childs engagement in
task stalls. Verbal
adaptation of robot speech
(e.g., volume, pitch) can also
improve rapport, the social
context, comfort, and
Caruana et al. (2022);
Chen et al. (2020)
Assisting Robot offers socioemotional
assistance to learner (learner
assistant) or takes over tasks
from teacher (teaching
To perceive and record
learner behavior
To interpret and assess
behavior in terms of task
performance; to provide
appropriate input
To initiate and/or maintain
dialogue; to offer
encouragement and/or help;
to provide feedback
Alemi et al. (2015);
Deublein et al. (2018);
Engwall and Lopes (2022);
Hong et al. (2016);Hsiao
et al. (2015)
Prompting Robot invites the learner to
use language expressively
To understand what the
learner is saying. It would be
additionally benecial if it
could detect pronunciation
errors in the learners
speech and provide
To interpret learners
speech; to provide
appropriate input; to
explicitly invite the learner
to speak
To initiate and/or maintain
dialogue; to offer
encouragement and/or help;
to provide feedback on
communicative skills in
general or on pronunciation
Lin et al. (2022)
Role playing Robot acts out a certain role
by collaboratively
negotiating the plot and
To recognize nonverbal
social signals such as facial
expressions, gestures,
posture, pointing, eye
To be trained to produce
proactive and socially
appropriate behavior
To respond adequately and
to predict what type of
interaction and behavior it
should evoke from a child in
a given role play, it should
provide hints and
encouragement in a specic
Ali et al. (2019);Lee et al.
incorrect behavior
Robot acts as an error-prone
tutee for the human tutor
To understand the
corrections given by the
human tutor verbally, via
gestures, and/or via a
mediation tool
To improve on its
proportionally and
according to the received
guidance, without ever
surpassing the
performance of the human
To establish a cooperation
cycle with the human tutor
to converge on a
performance deemed
satisfactory by the tutor
Tanaka and Matsuzoe
Robot initiates an
interaction/dialogue in
which communication
either becomes the focus of
the communicative activity,
is manipulated, or is
reected upon
To perceive the childs
utterances or to deliberately
initiate certain peculiarities
(e.g., longer response times)
in its perception that can be
reected upon
To model childs actions
and infers knowledge
related to task; to suggests
appropriate lessons for
To initiate/elicit a specic
type of dialogue and engage
in an exchange about the
communicative situation
Horwath et al. (2018);
Hsiao et al. (2015);
Ramachandran et al.
(2018);Spaulding et al.
Frontiers in Robotics and AI frontiersin.org03
Rohlng et al. 10.3389/frobt.2022.971749
prompting (Section 3.3) that follows caregiversintuitive behaviors,
role playing (Section 3.4), purposefully making mistakes (Section 3.5),
and encouraging metatalk (Section 3.6).Theorderofthepresented
roles can be seen as a progression from tutor- to child-oriented, thus,
increasingly empowering children in their participation that is crucial
for learning (Chi, 2009).
We describe and review the established and innovative roles by
rst specifying different requirements imposed on robotsperceptive,
cognitive, and dialogical abilities. We argue that the established roles
are too coarse-grained to further specify all the abilities needed. In
contrast, the abilities can be further specied for the innovative roles
(Table 1 provides a summary). Furthermore, we analyze the potential
of the innovative roles to evoke interactive acts (summarized in
Table 2) that we evaluate according to a taxonomy proposed by
Chi (2009). Finally (also in Table 2), we analyze what impact the
innovative roles have on the social environment and how they make
an engagement of others possible or necessary. Using this review,
researchers in robotics and developers of social robots can compare
and evaluate different dialogical roles in order to have a better basis for
a decision on the robots desired educational effect in language and
literacy learning.
2 Established dialogical roles that a
social robot fullls
2.1 Social robot as tutor
Certainly, the most common role that a robot can fulll in a
childrobot interaction for language learning and literacy is to
serve as a knowledge resource (e.g., Movellan et al., 2009;Vogt
et al., 2017;2019 in the rst learning session). A meta-analysis
by Belpaeme and colleagues (2018) revealed that in 86% of
studies on robots for education, the robot was designed to serve
TABLE 2 Potential of innovative roles with respect to engagement and involvement of others.
Role Learning/Engagement Shaping the social
context: Involvement of
Active Constructive Interactive
Periodic prompts, generic feedback,
or progress updates (e.g., duration
or proportion of task completed)
Prompting children to reect on
learning success/achievements at
appropriate points during the
learning activity
Sharing the learning activity,
including asking clarication
questions, making errors to elicit
feedback, and establishing a social
context that is nonjudgmental
Likely to be most effective if able to
deploy autonomously safely and
effectively, without the need for an
adult to control it. This is because the
key benet of support robots is that
they can offer a social agent that can
co-experience the activity but is less
likely to be perceived as judgmental/
Assisting Providing or highlighting input Prompting learner to act on
material or to engage in task
Providing feedback, co-solving a
Shift the role of the teacher from
instructor to moderator of learning
Prompting Periodic prompts, generating
questions or topic suggestions
Engaging the learner in problem-
solving discussions (e.g., riddles);
prompts should be fruitful in new
Engaging in discussions (e.g.,
developing a stance toward a
Likely to be most effective if learners
have some previous language
knowledge, to engage in more
elaborate conversations with the
robot. Could work both one-on-one
and in small groups
Role playing Generating questions, topic
suggestions, asking for help
Constructing common ground Negotiating and altering meaning,
exchanging roles
Possible involvement of other
children to simultaneously engage in
peer interactions as well as
robotchild interactions. Adults
might be necessary in case of
younger children
incorrect behavior
Asking for feedback and guidance
in the task
Engaging the learner in reection
and explanations
Sharing the learning activity
including asking clarication
questions, making errors to elicit
feedback, and establishing a social
context that is nonjudgmental
Likely to be most effective if able to
deploy autonomously, safely, and
effectively, without the need for an
adult to control. Possible
involvement of other learners in
collaborative tutoring of the robot to
allow for engaging peer interactions
alongside the robotchild
Providing or highlighting some
kind of linguistic context such as a
narrative, word, sentence
structures, or communicative
practices that can be talked about
Engaging a child in reection and
discussion about the linguistic
context (with a peer)
Providing feedback and co-
constructing new knowledge by
scaffolding the child; e.g., asking
the child to explain a subject to the
Possible involvement of a caregiver
of the child by discussing/reecting
on the linguistic utterances of the
robot or the interaction/turn taking
with the robot
Frontiers in Robotics and AI frontiersin.org04
Rohlng et al. 10.3389/frobt.2022.971749
the role of a tutor or a teacher. Accordingly, a robot acts as a
tutor and a more knowledgeable partner to foster the
acquisition of new knowledge and skills(Chen et al., 2020,
p.3).Chen et al. (2020) rightly point out that the learner gains
from an interaction with a skilled tutor because of the guidance
and scaffolding provided in the learning process. For this, tutors
need to ne-tune to the learners skills in a way that is
temporarily assisting learners to achieve new skills or levels
of understanding they would not reach on their own(Schodde
et al., 2019,p. 1). First approaches to a tutoring robot that can
scaffold the learners behavior are being developed (Schodde
et al., 2019;Cumbal, 2022). They require highly nested abilities
from a robot, and their depiction goes beyond the categories
proposed in Table 1. More specically, they require not only a
model of the learners perceptual, cognitive, and linguistic skills but
also a model of the task. For scaffolding, both models do not just
need to be combined. Verbal guidance also needs to be derived from
the partner model and designed in a way that includes nonverbal
behaviorsandcanadjusttochildrens linguistic capabilities
(Norman et al., 2021;Rohlng et al., 2021).
Although existing studies speak to the high potential of robots for
teaching language and literacy, Kanero et al., 2018 summarize critical
points regarding this research: One is the lack of evidence (or control)
that robots are more effective in the long term than other options. A
further critical point relates to the exposure of children to technology
rather than to human relationships (Sharkey, 2016). In this respect,
Sharkey (2016, p. 295) concludes that there are good reasons not to
encourage fully edged robot teachers.
2.2 Social robot as peer
A group of scholars (Kory Westlund et al., 2016;Belpaeme et al.,
2018) suggest that the decision for one role over another should be
based on whether the interaction with the robot is perceived as being
fun while, at the same time, effective to achieve the learning goal. The
suggested framing(Belpaeme et al., 2018,p. 331) corresponds to
research in developmental studies revealing that children tend to treat
a robot as a peer and social actor (Tanaka and Matsuzoe, 2012;
Breazeal et al., 2016;Oranc and Küntay, 2020). This preference is likely
to lead to acceptability of a robot (Belpaeme et al., 2018)butbears
some limitations, because it is especially pronounced in young
children (Shin and Kim, 2007).
The role of a peer is motivated by its higher acceptability in
children, and it is often contrasted with that of a tutor to
emphasize the companionship (Admoni and Scassellati, 2014).
According to Belpaeme et al.s (2018) meta-analysis, robots were
designed to fulll the role of a peer or a novice only in 9% of the
studies investigating robots for educational purposes. The
characteristic of a peer interaction is that both partners are
responsible for knowledge construction (Chi, 2009). Thus,
they both need to be engaged. The motivation for the rst
robot fullling the role of a peer was to establish a
relationship between the robot and the children (Kanda et al.,
2004). The authors considered the robots ability to identify and
recognize its peers as a prerequisite for this relationship that could only
evolve over time. The ability to identify and recognize is based on
perceptual and cognitive skills.Fordialogicskills,Baxterand
colleagues (2017) found that personalized robots are a better
design to engage children in a learning interaction. With
personalization,the authors refer to adaptation of non-verbal
behavior, personable language content, and alignment to task
performance(Baxter et al., 2017,p.4).Clearly,thedesignofa
robot as a peer demands all abilities to be responsive to the partner in
order to achieve cooperation. However, in current research, beyond its
emphasis on engagement, it is not exactly specied what kind of
activities the role should bring about to foster what kind of learning
process. In Section 3, we therefore suggest that the coarse-grained
demands on cooperation can be broken down into some specics.
2.3 Social robot as learner
What is established in current research is that a robot can
fulll the role of a less knowledgeable partner, and a learner. A
meta-analysis of robots in education has suggested that a robot in
this role can support skill consolidation and mastery (Belpaeme
et al., 2018). In fact, some studies have investigated whether
particular roles of a robotas much enjoyment and engagement
as they may provideare benecial for learning. Tanaka and
Matsuzoe (2012), for example, put children into a teaching role
and investigated learning-by-teaching effects when Japanese
children taught some English words to a care-receiving
robot(p. 78). To leverage the childs knowledge to the level
of an expert in an interaction is clearly engaging but might also
foster empathy and compassion (Lee et al., 2019).
Tanaka and Matsuzoe (2012) found that a robot fullling the
learners role can foster English verb learning in Japanese
children. Whereas these studies add a lot to the growing
possibilities of applying a robot, they lack a clear comparison
to other potential roles. In addition, for future research, it
remains an open question whether the learning effect is
relying on the teaching and caring activity that requires a
tutor to adapt to the learner or on the knowledge
construction ability of the learner that is feedbacked to the
tutor. For the role of an active learner, it is necessary to
display an increase of learning by, for example, responses that
are substantive and meaningful(Chi, 2009,p. 82). The
implementation of such responses reecting cognitive and/or
dialogical processing is lacking in the current design of robotic
1 We thank a reviewer for this idea.
Frontiers in Robotics and AI frontiersin.org05
Rohlng et al. 10.3389/frobt.2022.971749
To summarize, the established roles offer a coarse-grained
differentiation into three roles (see Figure 1). However, as
highlighted in Section 2.1, these roles demand nested skills
that currently cannot be implemented in a satisfactory
manner in a robot. With the innovative dialogical roles in
Section 3, we propose an extension to the established roles.
They make some aspects of the roles feasible and specify the
required capabilities in Table 1. At the same time, they are
innovative, because they uncover abilities that further extend
the established categories (see Figure 1) toward the social
competence (on a horizontal axis), thereby bringing in more
uid domain knowledge (on a vertical axis). Thus, they offer
novel forms of education for the purpose of language learning
and literacy, even though the roles proposed below do not
exhaust the possibilities that the two dimensions in Figure 1
can yield. By focusing on what abilities foster what kind of
activities that can be applied in learning interaction, we add
to the current literature that offers little discussion on whether
and how the different roles can be designed systematically.
3 Innovative dialogical roles that a
social robot can fulll
3.1 Using a social robot to assist learning
A role that seems implicit in most applications of social
robots is that of assistant in service (Cõaićet al., 2019), care
(Broekens et al., 2009), or education (Belpaeme et al., 2018). In
the role of an assistant supporting language learning and
literacy development, robots can be a tutor, novice or peer
(Van den Berghe et al., 2019;Neumann, 2020). Yet, the
specic potential or contribution of the assistant role per se
has rarely been addressed in detail. In educational settings,
assistance can be considered in (at least) two ways: focusing
either on assisting the teacher to support instruction or on
assisting the learner to solve a given task. Both of these roles
arguably entail unique opportunities and challenges, and we
will elaborate on them below.
Considering the teacher-assisting role, the robot often acts
as a tutor providing learning material in a classroom setting
allowing the teacher to focus on student performance rather
than on instruction (Alemi et al., 2015;Hong et al., 2016).
Hence, the robot assists the teacher by taking over specic
instructional tasks, thereby freeing educational resources. In
classroom settings, the robot could further collect, process,
and monitor data on learning performance and progress that
might inform subsequent teaching on either the group or
individual level. Considering such personalization, robots
have the potential to scaffold learning by providing
learning content that is tailored to individual abilities
(Gordon and Breazeal, 2015). The use of social robots in
et al., 2021) can be understood in a similar vein. Further
functions that teaching assistant robots may take over include
the role of an invigilator who supervises student actions
(Ahmad et al., 2021), as a native speaker in second-
language learning that models language behavior and elicits
language production (Han et al., 2012;Spaulding et al., 2018),
or as catalyst that facilitates the interaction between learners
in a setting in which the teacher is not actively involved
(Engwall and Lopes, 2022). In all these settings, the robot
may be perceived as less intimidating and judgmental from the
learnersperspective as well as more motivating and engaging
than the human teacher (see Sections 2.2 and 3.2).
Importantly, because the robot might reduce the
educational and cognitive load of teachers, their assistance
makes it possible to shift the role of the teacher from
instructor to moderator of learning. In Chis (2009)
taxonomy, the teaching assistant might, thus, allow the
teacher to build more constructive and interactive learning
Considering the learner-assisting role, the robot can
provide direct support when the learner encounters
difculties by providing information to help solve the
problem at hand (Crossman et al., 2018). For instance, the
robot can use locational or verbal cues to focus the learners
attention on relevant material in the input (Hemminghaus
and Koop, 2017). Robots might also support learning
indirectly by giving feedback on task performance (Gordon
and Breazeal, 2015;Hsiao et al., 2015), by mirroring learner
Summary of the various social/dialogical roles presented in
this review. On the left, the established roles differ with respect to
the domain knowledge. They can be further specied in the
innovative roles that differ in terms of social competence.
Frontiers in Robotics and AI frontiersin.org06
Rohlng et al. 10.3389/frobt.2022.971749
behavior or providing a different perspective (Wood et al.,
2017), by increasing motivation and engagement (Deublein
et al., 2018), or by focusing and redirecting attention to the
task when the learner drifts off. In an embodied way, a robot
might further compensate for impairments by, for example,
reading out text for a visually impaired learner and
representing or connecting a learner who cannot be
physically present in a class (Newhartetal.,2016). In all
these scenarios, the robot fullls a leveraging role and enables
the learners to succeed in a task that they might fail without
the robots assistance.
To fulll the teacher- as well as the learner-assisting role,
the robot needs to understand the learning task at hand and its
demands. It has to assess the learnerscompetence, estimate
their potential abilities, and monitor progress. In addition to
registering overt task performance, this requires a robot to
interpret multimodal cues (eye gaze, body posture, etc.) in
order to infer the learners cognitive and emotional state and
to react to situations in which the learner might struggle with a
given task. To assist the learner in solving a task, the robot
further needs to be able to provide appropriate cues via
gestures, speech, or other means of directing attention and
providing information (see also Vogt et al., 2017). Critically,
to scaffold learning optimally, the robot needs to know not
only how to help in a given context but also when its help is
required and then balance its assistance accordingly. If the
robot is too helpful this might increase short-term success but
reduce the opportunity for long-term learning (for a
discussion of the assistance problem, see Koedinger et al.,
3.2 Using a social robot to offer
socioemotional support to children
An emerging role for educational robots is to support
children by either engaging and motivating them during a
learning activity or by alleviating anxiety associated with
learning a new or challenging skill. A robot can serve this role
exclusively or in conjunction with other roles of being a tutor or a
peer (see Section 3.1). Unlike other education robot roles,
supportive robots can promote learning indirectly by
optimizing the childs socioemotional context to promote
engagement in (and mitigate avoidance of) learning tasks.
This could well be of signicant value for children who have
learning, attention, or literacy difculties. Such supportive roles
have been examined widely in healthcare settings to improve
treatment adherence for chronic conditions (e.g., asthma)
(Ferrante et al., 2021) and to improve mental health outcomes
for isolated and inrm children (Jeong et al., 2015;Rossi et al.,
2022). Crossman et al. (2018) exposed 87 children to a stressful
task. Those who interacted with a social robot experienced
greater reductions in stress on subjective state anxiety and
salivary cortisol measures in comparison to children from two
control conditions, in which either the robot was turned off or
not present. However, the broader potential for social robots to
simultaneously support socioemotional and education outcomes
in children remains unexplored. A recent mixed-methods study
by Caruana et al., (2022) explored the potential for three different
robots (NAO, MiRo, and Cozmo) to support childrens
engagement in reading. Whereas only one robot (NAO) had
the capacity to engage in social dialogue, all robots demonstrated
the potential to support childrens engagement by responding
with nonspeech vocalizations, sounds, and movements (e.g.,
grunting, head shaking, and tail wagging). During in-depth
interviews, most children reported that the robots offered a
welcomed, engaging, calming,and nonjudgmental social
context for reading. For this reason, many children expressed
a preference for reading a difcult book to a robot than to either
their teacher, or themselves alone. This preference was most
prevalent among children who interacted with NAO. Further,
children directly associated the dialogical (in)abilities of their
chosen robot as a signal of its intelligence and sentience, and
thus its capacity to comprehend the learning activity and assist
the child if needed. As such, whilst supportive robots need not
serve in dialogical capacities (e.g., Paro the nonverbal
zoomorphic robot), social dialogue may help robots to
assert their social presence and their capacity to fully co-
experience the learning activity. Further, features of robot
speech can shape a supportive learning contextparticularly
when robots can adapt/entrain to features of a human
interlocutors speech (Kumar et al., 2010;Gulz et al., 2011).
For example, Lubold et al., (2018) observed middle school
children teaching a robot to solve ratio problems. Children
experienced greater learning and social rapport when the
robot engaged in social dialogue and adapted to the childs
pitch, compared to robots who engaged in social dialogue
thus conclude that dialogical and supportive robots offer great
promise for education interventions. However, fullling their
full potential will require autonomous robots that can
accurately and rapidly perceive, comprehend, and respond
to child speech in the absence of adult operators. Such needs
currently outstrip the capabilities of most robots and speech
recognitionproduction systems.
The studies presented above suggest that the mere
presence of supportiverobots has the potential to
improve childrens socioemotional state during difcult
learning tasks (see Caruana et al., 2022, for a discussion).
However, larger gains are likely to be seen if robots can
actively motivate children or change the way children
evaluate a learning activity and their capacity to complete
it. In this respect, consider a child experiencing reading
difculty and associated anxiety. A support robot may
promote active engagement simply by offering periodic
prompts that encourage the child to continue reading or
Frontiers in Robotics and AI frontiersin.org07
Rohlng et al. 10.3389/frobt.2022.971749
relay the amount of time or the number of pages already read
in a session. Such capabilities would be easy to automate. It is
important to highlight that for such a support, robots must be
able to rapidly sense and recognize when a child is making
errors and thus stops reading or becomes anxious. Then, they
could offer sensitive encouragement and reinforcement when
the child needs it, and thus encourage the child to persist. This
could be accompanied by questions prompting the child to
reect on their reading success: That was a hard one, but you
read it! How do you feel?Another future direction is to
develop robots that can support reading engagement through
multiple dialogical roles (e.g., a supportive co-learner) because
they can also make errors themselves to elicit corrective
feedback from the child (see Section 3.5,Grimminger and
Rohlng, 2017). They could also ask spontaneous
comprehension questionsa successful method that is
known from dialogical reading (Blewitt et al., 2009)to
promote and check the childs attention while framing the
robot as engaged, competent, but also un-intimidatingly
apprehension. Indeed, recent work has shown that robots
that can adaptively move between the roles of tutor or
peer/novice can maximally support childrenslearningand
emotional needs during vocabulary acquisition (Chen et al.,
2020). This again demonstrates that the social roles robots can
adopt during learning interventions do not need to be xed or
discrete; and, indeed, dynamic, and adaptive education roles
are likely to best position social robots as socioemotional
supports for children engaging in learning activities. We
will come to this point within the Discussion (see also
Figure 1).
3.3 Using a social robot to nudge or
prompt childrens communicative
Social robots are particularly useful compared to other types
of technology to prompt or nudge otherscommunicative
behavior. They often have a humanoid or animal-like
appearance, and this increases the tendency to
anthropomorphize them (Mubin et al., 2013), thus, making it
more likely that people will speak to them. Lin et al. (2022)
discussed oral interactions between learners and robots in their
recent RALL review. One of their ndings was that interactive
oral tasks (such as engaging in dialogues with robots) are often
used in foreign-language-learning classesbecause robots
provide learners with the opportunity to engage in dialogues
with a nativespeakerand these activities are aimed more at
practicing communicative skills than improving grammatical
accuracy. Robots are used for such communicative activities
for a reason: Learners are often less anxious about engaging in
dialogues with social robots than peer learners or teachers,
because they feel less judged and less afraid of making
mistakes (Alemi et al., 2015; see also Section 3.2). The robot
may thus serve as a middle ground between the benets of a tutor
(high-quality language input and feedback) and less-anxiety-
inducing environment of practicing with a peer. This does not
mean, however, that robots can be used only in elaborate
conversational classes with language learners who have some
prociency in the language that they speak with the robot. Robots
can also be used to prompt novice learners to use expressive
language. For example, children in Vogt et al. (2019) had little
prior knowledge of English and were invited by the robot to
repeat English target words. This study investigated a long-term
effect of learning by testing childrens recall experimentally, but it
did not address the question of whether children were motivated
to use the learned words in their everyday context. In this sense,
to prompt somebody means to provide an impulse that is taken
up not only in a context that requires or elicits itas is the case in
an experimental session requiring active learningbut also in
other contexts. The exploitation of the impulse in other contexts
would reect the constructive learning (see Table 2).
Facing all the tasks in which a robot prompts children in their
behavior, it becomes apparent that speech recognition
technology plays an important role in their design. Yet,
although advancements are being made, current speech
technology is still limited in recognizing speech, especially that
of young children (Kennedy et al., 2017). This is probably one of
the reasons for why only a few studies with novice learners invite
the learner to use language expressively. Most studies include
games in which learners can respond in other ways such as
selecting an answer on a tablet (de Wit et al., 2018). If robots are
to effectively prompt children to use language expressively, they
should be able to recognize what children are saying. It would be
even better if they could detect language input with such accuracy
that they could correct pronunciation errors. In that case, robots
could provide learners with not only a less-stressful environment
in which to practice a language but also feedback on how to
improve their pronunciation. High-functioning speech
recognition, however, is not the full story. The ability to
initiate and maintain a dialogue is still a challenge for social
robots along with the ability to adjust expressions for a particular
content (Conti et al., 2020).
3.4 Using a social robot in role plays
The term role playrefers to a scenario in which children
imagine and act out a certain role by collaboratively negotiating
the plot. Scenarios can be exibly adjusted to create reality or
ction. They provide the context for negotiating and altering
meaning and thus allow for constructing common ground.
Common ground refers to the shared knowledge between
interlocutors that is critical in communication and that can be
enriched through processing and accumulating new information
Frontiers in Robotics and AI frontiersin.org08
Rohlng et al. 10.3389/frobt.2022.971749
in communicative interactions (Clark, 2015). Role plays have
been shown to enhance metacommunicative skills and to provide
adequate language learning support in the preschool years (e.g.,
Andresen 2005) as well as in the classroom context of foreign-
language learning (Raz, 1985;Al-Arishi, 1994). It is therefore
reasonable to ask whether a social robot could successfully
accomplish such role plays or role-playing games. For
example, the robot could take the role of a shopkeeper that
interacts with a child customer, the role of a police ofcer
identifying a shoplifter, or the role of a repair service that
needs help with technical tools in order to x a bike. The
robot could serve as a model for a child who, when switching
roles, could imitate behavioral and linguistic patterns (see also
Carpenter et al., 2005) that are appropriate in a given scenario.
The number of studies on this issue is heavily limited and covers
mainly adult participants. This is true for shop scenarios in which
a robot provides a service through natural interactions with adult
speakers (e.g., Lee et al., 2011;Liu et al., 2018); language cafés in
which learners engage in small talk with a robot who takes up
different roles such as interviewer, narrator, facilitator, or
interlocutor (Engwall et al., 2021); and in a role-playing
scenario inspired by Game of Thrones to teach a novel
language (Ali et al., 2019).
Role plays go beyond the more common conversational
classes discussed in Section 2, because they are based on task-
based language teaching (Ellis, 2003) and require building a
social relationship between the participants in a given
scenario. This relationship is achieved by producing proactive
and socially appropriate behavior. Whereas the high social
appropriateness of predicted and proactive behaviors (Liu
et al., 2018) as well as replies to social contexts that are
coordinated and timely (Belpaeme et al., 2018) are currently
not implemented in robots, this could foster learning of role-
related procedures. For example, the robot needs to predict what
type of interaction and behavior it should evoke from a child in a
given role play and to respond adequately. It should also adjust its
behavior exibly when switching roles in a given scenario. Again,
the recognition of nonverbal social signals such as facial
expressions, gestures, and posture is needed for the right
interpretation of the situation and for appropriate replies
providing hints and encouragement in a specic role.
In summary, role plays could foster vocabulary acquisition
and the development of communicative competence. Following
the task-based language teaching approach (Ellis, 2003), a child
or a second-language learner could practice meaningful real-life
verbal skills by establishing a social relationship with the robot
and solving a task or pretending to do so. Finally, role plays
provide an excellent opportunity for what are considered to be
the best forms of teaching: activating knowledge, boosting
meaning negotiations, and the co-construction of knowledge
(Chi, 2009). Such activities could be particularly relevant for
learners from diverse cultural backgrounds, who could practice
conversations in the new language before having them with
actual speakers of that language. Future studies could examine
the effectiveness of a robot in meaningful role-play interactions
and learning gains as a function of the design, functionalities of
the robot, and most importantly the individual characteristics of
the child interacting with the robot.
3.5 Using an incorrect social robot to
promote reection and error correction
Since the late 1960s, when researchers made the surprising
discovery that, in a peer tutoring setting, peer tutors progressed
more than their own tutees (Cloward, 1967), literature has
extensively investigated the benets of learning-by-teaching
interactions (Duran, 2017). Learning with the expectation to
teach was found to promote the identication of key content
elements and their organization in a meaningful representation
(Benware and Deci, 1984), whereas learning and explaining
was found to allow for more persistent learning gains than
learning with the expectation to teach (Fiorella and Mayer,
2013). Explaining to others provides the tutor with more and
better opportunities to recognize own areas for improvement,
reorganize their own knowledge, and repair their own errors by
exercising their metacognitive skills (Duran, 2017).
In an effort to provide learners with more opportunities for
engaging in learning-by-teachinginteractions, teachable virtual
agents (Biswas et al., 2005) and teachable robots (Tanaka and
Matsuzoe, 2012;Hood et al., 2015;Walker et al., 2016) have been
developed. Within the latter category, an intriguing research
avenue is to explore the effect that purposefully designed
incorrect behaviors of a robot have on the engagement and
learning of its human tutor. With numerous studies in the
eld of humanrobot interaction supporting the fact that
faulty robots are consistently perceived as more likeable than
their infallible alternatives (Ragni et al., 2016;Mirnig et al., 2017),
it seems plausible to argue that incorrect
teachable robots further
reduce the learnersanxiety, making them feel even less judged
and afraid of making mistakes (Alemi et al., 2015). Childrens
promptness at adopting a care-taking attitude toward social
robots showing weaknesses (Tanaka and Matsuzoe, 2012)
seems particularly apt; and, in addition, childrens motivation
and engagement as tutors speaks to the protégé effect,which
seems to especially benet lower-achieving students (Chase et al.,
2009). Finally, incorrect behaviors displayed by the robot can 1)
be the trigger for a correction spontaneously provided by the
child (thus providing a natural framing for a learning-by-
2 We propose the terms incorrect behaviorand incorrect teachable
robotsto denote, respectively, a robot behavior purposefully designed
to include mistakes and the social robot exhibiting such behavior in the
context of a learning-by-teachinginteraction. We use this term to
distinguish from faulty robots,whose mistakes are not necessarily
planned for, while still emphasizing the nonoptimality of their actions.
Frontiers in Robotics and AI frontiersin.org09
Rohlng et al. 10.3389/frobt.2022.971749
teachinginteraction), and 2) via a careful design of the robots
behavior, highlight specic errors and thus nudge the child
towards specic corrections. These last two benets, uniquely
brought by incorrect teachable robots, have been investigated by
two studies.
In a scenario aiming to teach English verbs to 3- to 6-year-old
Japanese children, a NAO robot was used in a learning-by-
teachinginteraction that envisioned a rst phase in which an
adult teacher taught both the child and the robot, followed by a
phase in which the child could revise the robots initially wrong
understanding and correct it (Tanaka and Matsuzoe, 2012).
Results indicated a higher learning gain compared to a control
group of children not interacting with the robot in both an
immediate and a delayed post-test administered 35 weeks after
the experiment. Interestingly, parents reported that their children
liked the experience of teaching the robot so much that they
continued to play it at homeeven days and weeks after the
experimentthus suggesting that the interaction with the robot
could have promoted their spontaneous learning. In this study,
the robots initially incorrect behavior was used as a trigger and
motivator for the learning-by-teachinginteraction. It would be
interesting to compare such a robot with a traditionalrobot
tutee to verify the effects of the robots mistakes on the childrens
engagement in the interaction and learning.
In another study known as CoWriter (Hood et al., 2015;
Chandra et al., 2017), a NAO robot with poor handwriting skills
was used in a learning-by-teachinginteraction with children to
stimulate their metacognition, empathy, and self-esteem in
addition to their handwriting skills. The interaction consisted
of the following sequence: The child rst selects the letter to help
NAO practice on. The robot then writesit on a digital tablet
(concretely, moving its nger in front of the tablet to follow the
letters trajectory). Finally, it asks for feedback as well as an
example from the child and incorporates this in its next attempt
to write the letter. It continues its attempts until the child deems
the result to be satisfactory. Beside incorporating the childs
feedback, at each iteration the robots letter includes a
deformation in proportion, breaks and/or alignment that
Chandra and colleagues (2017) identied as categories of
mistakes commonly performed by children. Helping the robot
improve its handwriting thus induced the children to reect on
common handwriting mistakes and how to correct them. In
20182019, the CoWriter setup was integrated into the weekly
occupational therapy sessions for a 10-years old boy with a
complex neurodevelopmental disorder including severe
dysgraphia, for whom previous therapies had not led to
noticeable improvements (Gargot et al., 2021). After
20 sessions, the boys handwriting skills had improved
signicantly, and a decrease in avoidance behaviors as well as
better commitment to handwriting practice could be observed.
As a consequence, he could go back to a regular school where he
received special education. Interestingly, although the boy
reected on the intentions behind the robots behavior (It is
not the robot who learns, it is meGargot et al., 2021,p.6)
relatively early in the intervention, he kept tutoring the robot and
engaging with it throughout the sessions.
Unfortunately, the tremendous potential of incorrect
teachable robots, which combines the advantages brought by
learning-by-teachinginteraction with the unique engagement
and metacognition boost provided by the protégé effect inspired
by the robots shortcomings, comes at a high technical cost. Quite
ironically, a good bad-performingrobot is possibly even more
difcult to design than a good well-performingrobot. What
seems particularly difcult is to concurrently and consistently
ensure that 1) the robots incorrect behavior triggers the desired
effects in terms of engagement and metacognition without
generating frustration in the human tutor (Biswas et al.,
2005), 2) the robot improves over time, thus incorporating the
tutors scaffolding behavior and making it interactive (Chi, 2009)
while 3) never surpassing the tutors own competence, which
would negatively impact the tutors self-esteem. Tanaka and
Matsuzoe (2012) circumvented this problem by remotely tele-
operating the robot with a Wizard-of-Oz approach, whereas
Hood et al. (2015) and Chandra et al. (2017) relied on a
dataset of adult handwriting samples to dene shape
deformations to apply to the lettersmodels, thereby ensuring
the robotsbad handwriting.Upon merging the robots own
poor letter with the example provided by the child, errors are
either mitigated (if they do not appear in the example) or
reinforced: This enables the child to see their own mistakes in
the robots handwriting and reect on them. Exporting such a
sophisticated interaction to more complex contexts, possibly
involving social and verbal interaction, is an open research
3.6 Using a social robot to encourage
metatalk in children
A dialogical role that has received little attention in current
implementations of social robots is that of a robot encouraging
metatalk. This term encompasses both the ability to talk about
communication in general (metacommunication) or language
in particular (metalinguistics). In the case of communication,
there can be talk about organizing the ongoing interaction or
about hypothetical conversations; in the case of language, the
talkcanbeaboutitsstructure(Aukrust, 2004). In both cases,
communication becomes the focus of the communicative
activity, is manipulated, or is reected upon. To date, there
has been hardly any focus on this ability in childrobot
interaction studies on language learning, or it is usually
addressed only implicitly as in studies on improving
childrens narrative abilities (e.g., Kory Westlund and
Breazeal, 2019). To reect on communication is an ability
relying on metacognition. Metacognition, in turn, can be
dened as awareness and management of onesown
Frontiers in Robotics and AI frontiersin.org10
Rohlng et al. 10.3389/frobt.2022.971749
thought(Kuhn and Dean, 2004,p. 270) and is driven by
executive control.
Considering metalinguistic abilities as foundational for a
childs communicative competence and literacy outcomes
(Heller, 2014;Stude, 2014), Hsiao et al., (2015) carried out a
study in which a robot interacted with a pair of two kindergarten
children in a book-reading situation in which the children were
requested to reect on sentence patterns and compare words.
The robot was equipped with the ability to produce speech and
sounds as well as to automatically provide emotional responses to
the childrens turns via its integrated touchscreen. Interestingly,
in addition to improving childrens overall reading skills, the
authors reported that opportunities arose for children to share
reections about the meaning of the text, such as one child
explaining the meaning of a word to the other or both discussing
parts over which one of them disagreed (Hsiao et al., 2015,
p. 287). Furthermore, and with regard to the reading abilities
necessary to solve math problems, Ramachandran et al. (2018)
applied a robot that encouraged 11-year-old children to think
alouda metacognitive strategy. Their results indicated positive
effects of a social robot on the studentsengagement and
compliance with the proposed thinking-aloud strategy. The
authors concluded that a robot can support the use of a
metacognitive strategy to enhance problem solving in children.
In addition, recently, Spaulding and colleagues (2021) used a
robot as a learning companion to play language games with
children. It engaged them in activities of spelling and rhyming
words in order to promote their phonological awareness,
i.e., knowledge about sounds in their spoken language
(Spaulding et al., 2021). The contextual environment was
displayed on a tablet, and by taking turns with the child, the
robot performed game tasks and responded to the childs input
on the tablet with socioemotional behaviors. In addition, the
robot modeled the learners behavior and demonstrated
exemplary tasks to the child based on the childs play actions
and state of knowledge (Spaulding et al., 2021,p. 5).
Taken together, these ideas highlight a clear potential for the
use of robots in the dialogical role of initiating metatalk in
childrenalbeit the aforementioned studies focus mainly on
talk about language in particular rather than metaknowledge
about dialogue. Additionally, in terms of Chis (2009) proposed
categories of engagement, further interactive capabilities are
needed. Whereas the approaches presented here can be
considered primarily as actively and constructively eliciting
childrens elaborations and reections that go beyond the
immediate learning content presented, future empirical studies
should move toward an interactive learning. Specically,
interaction scenarios aiming to enable children to talk about
dialogical features such as peculiarities experienced within the
turn-taking with a robot that takes longer to react (Tolksdorf
et al., 2021), could elicit reections on dialogical processes in
general, and turn-taking in particular: Children could reect on
how long a pause between a question and its answer could be and
what reactions a too long pause elicits. These reections could
help children to develop coping strategies for peculiarities or
idiosyncrasies within a dialogue without referring to persons
(Horwath et al., 2018). Overall, there seems to be an inexhaustible
potential for using social robots to support metatalk in children.
4 Discussion
In this review, we have proposed innovative roles that a social
robot can fulll in an interaction with a child. This extends
established roles that enable a social robot to act as a tutor, peer/
companion, or novice/learner (Mubin et al., 2013;Ahmad et al.,
2021;Lin et al., 2022) through being less rigid and more uid in
relation to the established role categories (see Figure 1). At the
same time, these roles offer innovative forms of education for
learning language and literacy that further differentiate the
existing categories with respect to domain knowledge and
social competence (see Figure 1). In addition to our review,
we have contributed an analytical framework that can be used to
specify the differences in humanrobot social relationships when
children learn language through interacting with a social robot.
Based on the theoretical account proposed by Mead (1946)
suggesting that social roles shape interactive acts, we identied
the overt acts that we then subjected to a critical evaluation
reecting the different humanrobot social relationships that
arise when applying a robot in a particular social role.
Identifying the skills that enable a robot to perform overt acts
is a crucial step in our analysis of innovative roles. In Table 1,we
differentiated between perceptual, cognitive, and dialogical skills
that are required from a robot to fulll a role. This differentiation
can be seen as problematic, because cognitive and dialogical skills
are intertwined. However, whereas cognitive skills are covert,
dialogical behaviors are overt and easy to assess. In addition,
overt behaviors are based more on pragmatic skills, and they
result in decisions on choosing, for example, what formulation is
appropriate for a situation. While analyzing the established roles
for the skills they require, we realized that they are too coarse-
grained to clearly identify the skills. Summarizing the innovative
roles in Table 1, we have to highlight the fact that the majority of
them requires skills that have still hardly ever been implemented
in robots. Instead, both cognitive and dialogical skills are pre-
programmed, and in most studies, semi-implementation of such
skills is realized by applying the Wizard-of-Oz-method.
After having gained a clearer picture on the robots skills as
prerequisites to a social role, we drew on the abilities specied in
Table 1 to analyze different types of overt activities that a robot
could initiate and thus use to engage with a learner (Table 2). Our
analysis is guided by a taxonomy suggested by Chi (2009) who
differentiated between active, constructive, and interactive
activities. Providing empirical support for deeper learning
being scaffolded more by interactive rather than constructive
or active activities, we specied in Table 2 whether and with what
Frontiers in Robotics and AI frontiersin.org11
Rohlng et al. 10.3389/frobt.2022.971749
behaviors the three types of learning activities can be realized
when the robot is fullling a specic role. It becomes clear that
interactive learningrequires more responsive behavior in a
robot than active learningdoes. It is important to note that this
responsivity goes beyond simple contingency (Cangelosi et al.,
2010) and requires more semantics. In this respect, McGillion
et al. (2013) suggest the term semantic contingencyto highlight
the meaningful action that can follow on from a behavior.
Because, given the current state of the art, robots are not able
to understand the actions of their partners, they cannot select
behavior that is both temporally and semantically contingent
(such as providing not only feedback but appropriate
feedback)behavior that would better relate to the action that
the partner has just performed. Consequently, a form of memory
for the interaction would be necessary to enable a robot to
perform more activities linked to interactive learning. With
such a memory for the history of interaction(Rohlng
et al., 2016,p. 4), a robot could monitor the childs
engagement in order to note deviations, and then provide
hints, encouragement, or reassurance.
When reecting on the educational potential of the
innovative social roles, we noted that most of them (see
Sections 3.13.5) require a robot to react to childrens
multimodal communicative behavior. Interestingly, despite
plenty of evidence suggesting that a robot needs to react
contingently and multimodally when interacting with children
(Belpaeme et al., 2018), the dialogue situation is often mediated
via the use of a tablet, and this restricts the childs behavior to
some choices on a screen (Vogt et al., 2017). From a technical
point of view, this compensates for the robots inability to react to
the variability in childrens communicative behavior. Yet,
considering the design of interaction, this compensation
clearly limits the richness of information. For example,
Tolksdorf and Mertens (2020) found that children take much
longer pauses when answering to a robot in which they make use
of gestures and gaze rather than verbal behavior. Current robots
cannot yet cope with such reactions, often resulting in interaction
breaking down (Rohlng et al., 2021). What robots thus currently
lack is a system that takes full advantage of the robots body
behaving contingently and adapting to both individual
behavioral patterns and childrens speech (Kennedy et al.,
2017). As long as such a dialogue system is not implemented,
robotic technology will lack the interaction capabilities crucial for
communication with children.
Finally, it is important to emphasize that not only can a robot
appear in one role or the other, but that it can also change roles
just like children do in their (learning) interactions. A change in
dialogical rolesfor example from being tutee to a tutoris
likely to boost learning in children by providing them with both
perspectives. As discussed in Section 3.5, tutoring can be a more
valuable learning experience than being tutored; the question of
whether fullling one role and then changing to the other could
advance learning even more has barely been addressed in
research so far. This pertains to both studies on childrobot
interaction as well as studies on child development. Among the
former, Chen et al. (2020) recently proposed a robot capable of a
role adaptation. Furthermore, these authors demonstrated that
an adaptive role in which a robot shifts between tutor and peer/
novice can lead to both learning and emotional support (see also
Lee et al., 2019). Hence, role adaptation clearly brings about new
possibilities to enhance learning processes. Whereas we are not
aware of further studies systematically investigating the
advantage of a role change for language and literacy learning,
there is some discussion on the phenomenon of role reversal.
More specically, in studies with a gray parrot acquiring labels,
Pepperberg (2002) argued that for the animal to learn, it was
important to observe the roles of a tutor and tutee being changed
and thus modeled independently from the persons. In
developmental studies, Carpenter and colleagues (2005) have
argued that role reversal is important for children to recognize
the reciprocity of linguistic symbols. A role reversal is dened by
the child performing an action toward an adult in the same way
that the adult performed it toward him or her(Carpenter et al.,
2005,p. 254). This enables the child to learn that there is a
reciprocal substitution between demonstrator and learner(p.
255). In their study, Carpenter et al. (2005) found positive
correlations between the amount of role reversal imitation and
the childrens comprehension and production of pronouns.
These studies make it plausible to argue that childrens
learning can be advanced by a role reversal. For the design of
a robot, however, it is a challenge to provide the capabilities
required for a particular role, in addition to being able to reverse
the interaction protocol. Chen et al. (2020) solved this by
reinforcement learning and concluded that an adaptive,
reciprocal peer is more engaging, interesting, and fun for
children.However, a ne-grained analysis of the learning
activities would be necessary to further evaluate the exchange,
to determine whether the obtained effect is driven by more