Content uploaded by Bob Schadenberg
Author content
All content in this area was uploaded by Bob Schadenberg on Aug 24, 2021
Content may be subject to copyright.
36
Predictable Robots for Autistic Children—Variance in
Robot Behaviour, Idiosyncrasies in Autistic Children’s
Characteristics, and Child–Robot Engagement
BOB R. SCHADENBERG and DENNIS REIDSMA, University of Twente
VANESSA EVERS, University of Twente and Nanyang Technological University
DANIEL P. DAVISON, JAMY J. LI, and DIRK K. J. HEYLEN, University of Twente
CARLOS NEVES and PAULO ALVITO, IDMind
JIE SHEN and MAJA PANTIĆ, Imperial College London
BJÖRN W. SCHULLER, University of Augsburg and Imperial College London
NICHOLAS CUMMINS, University of Augsburg and King’s College London
VLAD OLARU and CRISTIAN SMINCHISESCU, Institute of Mathematics
of the Romanian Academy
SNEŽANA BABOVIĆ DIMITRIJEVIĆ and SUNČICA PETROVIĆ, Serbian Society of Autism
AURÉLIE BARANGER, Autism Europe
ALRIA WILLIAMS and ALYSSA M. ALCORN, University College London
ELIZABETH PELLICANO, University College London and Macquarie University
Predictability is important to autistic individuals, and robots have been suggested to meet this need as they
can be programmed to be predictable, as well as elicit social interaction. The eectiveness of robot-assisted
interventions designed for social skill learning presumably depends on the interplay between robot predict-
ability, engagement in learning, and the individual dierences between dierent autistic children. To better
This work was made possible through funding from the European Union’s Horizon 2020 research and innovation program
under grant agreement no: 688835 (DE-ENIGMA).
Authors’ addresses: B. R. Schadenberg, D. Reidsma, D.P. Davison, J. J. Li, and D. K. J. Heylen, University of Twente, Drienerlo-
laan 5, 7522 NB Enschede, The Netherlands; emails: {b.r.schadenberg, d.reidsma, d.p.davison, j.j.li, d.k.j.heylen}@utwente.nl;
V. Evers, University of Twente, Drienerlolaan 5, 7522 NB Enschede, The Netherlands and Nanyang Technological Uni-
versity, 50 Nanyang Ave, Singapore 639798, Singapore; email: v.evers@utwente.nl; C. Neves and P. Alvito, IDMind, Polo
Tecnologico de Lisboa, Lt 1, 1600-546 Lisbon, Portugal; emails: {cneves, palvito}@idmind.pt; J. Shen and M. Pantić, Im-
perial College London, South Kensington Campus Exhibition Rd, South Kensington, London SW7 2AZ, United Kingdom;
emails: {jie.shen07, maja.pantic}@imperial.ac.uk; B. W. Schuller, University of Augsburg, Universitätsstraße 2, 86159 Augs-
burg, Germany and Imperial College London, Exhibition Rd, South Kensington, London SW7 2BX, United Kingdom; email:
schuller@informatik.uni-augsburg.de; N. Cummins, University of Augsburg, Universitätsstraße 2, 86159 Augsburg, Ger-
many and King’s College London, Strand, London WC2R 2LS, United Kingdom; email: nicholas.cummins@informatik.uni-
augsburg.de; V. Olaru and C. Sminchisescu, Institute of Mathematics of the Romanian Academy, Calea Griviţei 21,
Bucharest, Romania; emails: {vlad.olaru, cristian.sminchisescu}@imar.ro; S. B. Dimitrijević and S. Petrović, Serbian Society
of Autism, Gundulicev venac street 38, Belgrade 11000, Serbia; email: suncica.petrovic@yahoo.com; A. Baranger, Autism
Europe, Rue Montoyer 39, Brussels 100, Belgium; email: Aurelie.baranger@autismeurope.org;A. Williams and A. M. Alcorn,
University College London, Gower St, London WC1E 6BT, United Kingdom; emails: {alria.williams, a.alcorn}@ucl.ac.uk; E.
Pellicano, University College London, Gower St, London WC1E 6BT, United Kingdom and Macquarie University, Balaclava
Rd, Macquarie Park NSW 2109, Australia; email: liz.pellicano@mq.edu.au.
This work is licensed under a Creative Commons Attribution International 4.0 License.
© 2021 Copyright held by the owner/author(s).
1073-0516/2021/08-ART36
https://doi.org/10.1145/3468849
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
36:2 B. R. Schadenberg et al.
understand this interplay, we report on a study where 24 autistic children participated in a robot-assisted
intervention. We manipulated the variance in the robot’s behaviour as a way to vary predictability, and
measured the children’s behavioural engagement, visual attention, as well as their individual factors. We
found that the children will continue engaging in the activity behaviourally, but may start to pay less visual
attention over time to activity-relevant locations when the robot is less predictable. Instead, they increas-
ingly start to look away from the activity. Ultimately, this could negatively inuence learning, in particular
for tasks with a visual component. Furthermore, severity of autistic features and expressive language ability
had a signicant impact on behavioural engagement. We consider our results as preliminary evidence that
robot predictability is an important factor for keeping children in a state where learning can occur.
CCS Concepts: • Human-centered computing →Interaction design theory, concepts and paradigms;•
Social and professional topics →People with disabilities;•Computer systems organization →Robotics;
Additional Key Words and Phrases: Predictability, variability, autism spectrum condition, human-robot
interaction, engagement, individual dierences
ACM Reference format:
Bob R. Schadenberg, Dennis Reidsma, Vanessa Evers, Daniel P. Davison, Jamy J. Li, Dirk K. J. Heylen, Car-
los Neves, Paulo Alvito, Jie Shen, Maja Pantić, Björn W. Schuller, Nicholas Cummins, Vlad Olaru, Cristian
Sminchisescu, Snežana Babović Dimitrijević, Sunčica Petrović, Aurélie Baranger, Alria Williams, Alyssa M.
Alcorn, and Elizabeth Pellicano. 2021. Predictable Robots for Autistic Children—Variance in Robot Behaviour,
Idiosyncrasies in Autistic Children’s Characteristics, and Child–Robot Engagement. ACM Trans. Comput.-
Hum. Interact. 28, 5, Article 36 (August 2021), 42 pages.
https://doi.org/10.1145/3468849
1 INTRODUCTION
Autism Spectrum Condition (hereafter referred to as “autism”) is a neurodevelopmental condition
that is characterised by diculties in social communication and interaction (so-called “social fea-
tures”) and by restricted, repetitive behaviour and interests (so-called “non-social features”) [3].
Whether these features are considered disabling for an individual can depend in part on the extent
and nature of support provided by others [97]. This support can include both helping the individual
child or young person to develop skills and strategies (for example, to understand situations and
communicate their needs) and adapting the environment to enable the child to function and learn
within it. In the context of social skill learning, experiencing discomfort due to dealing with un-
predictability is problematic as it prevents children from being in a state where they are ready to
learn. Incorporating a robot in social skill learning might be helpful in that it can provide a highly
predictable manner of learning social skills, as we can systematically control the predictability of
the robot’s behaviour [128]. Indeed, the predictability of a robot is a commonly used argument
for why robots may be promising tools for autism professionals working with autistic children
[e.g., 34,35,43,67,120,141]. Contemporary robot-assisted interventions have also shown that
they can maintain engagement for at least a month, whilst operating autonomously [22,123], and
there are positive indications that such interventions can lead to learning [22,35,123,138] and the
generalisation thereof to dierent contexts [123].
The predictability of robots may make it easier for autistic children to engage in learning in a
robot-assisted intervention and maintain this engagement. However, robots cannot behave fully
predictably as well as provide meaningful learning, as the learning gains may not generalise to the
less predictable world of people [1]. Moreover, a fully predictable robot would perpetuate in repet-
itive behaviour [33], limiting its long-term usefulness, and would also be unable to autonomously
respond to the (unpredictable) dynamics of real-world settings [23]. Both make large scale deploy-
ment of such a robot dicult. Thus, there is a tradeo between making the robot more predictable
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
Predictable Robots for Autistic Children 36:3
and providing the child with meaningful learning content with a tractable robot-assisted interven-
tion, which requires a degree of unpredictability. This unpredictability stems from (a) adding novel
content, which has not yet been learned, (b) making the learned skill’s approximate—at least in
part—what happens in real-world settings with humans, to facilitate generalisability to such highly
unpredictable environments which autistic children need to learn to deal with, and (c) implement-
ing complex robot behaviours (e.g., responsive actions to the environment), for which it may be
dicult to discern the cause of its behaviour and thereby prevent children from learning to pre-
dict the robot’s behaviour. Thus, on the one hand, we want autistic children to be in a state where
learning can occur, where the children show high levels of engagement with the learning material.
This may require very predictable interactions. On the other hand, we want them to learn skills
that are meaningful in the human world, which comes with a degree of unpredictability.A balance
needs to be struck then, where the robot is suciently predictable to maintain engagement while
still providing learning content that is representative of human-human interaction.
What constitutes “suciently predictable” is likely to dier between autistic children as they
are a notoriously heterogeneous group [60]. Some children may be better equipped to deal with
unpredictability than others. For these children, predictability of the robot’s behaviour may not
be as important, and the focus can lie on optimising learning. Furthermore, children’s reactions
to the robot’s unpredictability may be very dierent, and they may nd dierent aspects of un-
predictability problematic [54]. This makes it dicult to generalise ndings on predictability and
autism to autistic children working with robots.
The eectiveness of robot-assisted interventions designed for social skill learning presumably
depends—in part—on the interplay between robot predictability, engagement in learning, and the
individual dierences between dierent autistic children. To better understand this interplay, we
report on a study where autistic children participated in a robot-assisted activity, where we ma-
nipulated the variance in the robot’s behaviour as a way to operationalise predictability, and meas-
ured the children’s individual characteristics as well as their engagement related behaviours. The
robot-assisted activity was developed by the European project “DE-ENIGMA”—which funded the
current study—and revolves around playing with the basics of recognising facial expressions of
emotion. In the remainder of this article, we will rst elaborate on the concepts of engagement
and of predictability in Section 2. This literature forms the basis upon which we base our research
questions and hypothesis (Section 3), and our methods (Section 4). In Section 5, we present the res-
ults of our analysis of the robot’s predictability, and two facets of engagement, namely behavioural
engagement and visual attention. We conclude the article with a discussion on how we interpret
the results of our study and what this means for the interplay between robot predictability, the
two facets of engagement, and the individual dierences in Section 6and conclude on our research
questions in Section 7.
The contribution of this article is threefold. First, we provide a literature-based explication of
predictability as it relates to human–robot interaction (HRI). Second, we have operationalised
the predictability of a robot in a way that it can be assessed in real-life scenarios and provide
measures that can be used to compare the predictability of dierent robots. Finally, we report on
new analyses into how a robot’s predictability inuences two facets of engagement in a dataset of
27 autistic children.
2 BACKGROUND
2.1 Autism, Predictability, and Atypical Predictive Processing
Our senses are constantly dealing with ambiguous sensory information, trying to make sense of
it all. According to Bayesian accounts of perception, the human brain is constantly generating
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
36:4 B. R. Schadenberg et al.
predictions on what sensory information is expected in order to resolve this ambiguity and at-
tempts to minimise the error between the incoming sensory information and the prediction
[7,24,49,63,94]. This is referred to as predictive processing, or predictive coding. When a predic-
tion does not match with the observed sensory information, additional cognitive eort is required
to resolve the mismatch and to identify its cause. When these mismatches happen too often, it
can lead to a person being overwhelmed or anxious [102]. Autistic individuals are believed to fre-
quently experience these mismatches between predicted and observed sensory information due
to a decreased inuence of predictions on perception, which is hypothesised in current Bayesian
accounts of autism to be at the heart of their condition [31,81,104,135], [see 100, for a review].
According to the Bayesian accounts of autism, the non-social features are the result of trying to
maintain a predictable environment through either self-generated sensory information—which
can easily be predicted—or through enacting control over their environment. The social features
of autism, on the other hand, are explained by the highly unpredictable nature of social environ-
ments, which can therefore be dicult to deal with. The Bayesian accounts of autism are not the
only theories that relate prediction diculties with autistic features. For instance, the Empathizing-
Systemizing theory [8,9], posits that the social features of autism arise from a below-average em-
pathy, whereas the non-social features stem from above-average systemizing—identifying lawful
patterns in the information. When faced with information that does not strictly adhere to dis-
cernible laws (i.e., unpredictable environments), such as in social interactions, autistic individuals
struggle to make sense of it. This theory, however, is very descriptive in nature, and does not fully
specify the (dierent) computational mechanisms underlying autism.
A decreased inuence of predictions on the perception in autistic individuals is supported by
various lab studies [6,45,53,80,105,142], although support is not always found [17,83,88,103,140].
Nevertheless, literature has shown that, in practice, oering a more predictable environment is in-
valuable to facilitating learning by providing an environment that puts the child at ease by not
having to deal with the discomfort resulting from unpredictability and requires fewer cognitive
resources to be processed. Current educational practices, therefore emphasise the need for struc-
ture at schools (e.g., through the TEACCH approach [92]), so that autistic children know what to
expect during the day, increasing their engagement in learning [18,87,99]. All things considered,
there is compelling evidence that predictability is important to autistic children.
A construct that may be similar to a preference for predictability—or the intolerance of
unpredictability—of autistic children is that of intolerance of uncertainty (IU). IU is dened
as “the tendency to react negatively on an emotional, cognitive, and behavioral level to uncertain
situations and events” [42], and consists of two components, namely a “desire of predictability”
and “uncertainty paralysis” [13]. Knowing what will happen in the future allows us to increase
the odds of a desirable outcome, or brace for future adversity. However, this requires knowledge
on the probability, timing, and nature of the future event, which is often not available as the future
is intrinsically uncertain. Anticipating the future can therefore induce anxiety, as the uncertainty
reduces our ability to eciently and eectively prepare for the future. Individuals with heightened
levels of anxiety are believed to have an excessive response to uncertainty [56], IU is considered to
be a cognitive vulnerability factor in a broad range of emotional disorders like general anxiety dis-
order, social anxiety, and depression [20,44,64]. IU has also been studied with autistic individuals,
who obtain higher IU scores than the general population [15,21,95]. Furthermore, the higher levels
of IU might explain the high levels of anxiety that are consistently reported in autistic children
[131,144], as IU was found to mediate the relation between autism and anxiety [15,145]. However,
the authors caution that a dierent causal story cannot be ruled out, given the non-experimental
data used for the mediation analysis [15].
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
Predictable Robots for Autistic Children 36:5
2.2 Predictability and Its Operationalisation as Variance in a Robot
Incorporating robotic technology in interventions for autistic children has been proposed to meet
the need for predictability of these children [32,34]. Robots are programmable and can therefore
be designed to be highly predictable. However, regardless of a robot’s programming, it will be dif-
cult to correctly predict the robot’s behaviour when meeting a robot for the rst time. What is
the robot going to do or say rst? When is it going to greet me? How will it greet me? Without
having prior knowledge of the robot, a person is hard pressed to correctly answer these questions.
This example highlights that predicting robot behaviour is a learning process and that the robot
is initially unpredictable. Through interacting with the robot, people become more familiar with
its behaviour, improving their ability to predict its behaviour [40]. However, this requires one to
perceive and learn the structural regularities in the robot’s behaviour—also referred to as invari-
ance detection [51,52]. These structural regularities hold predictive value in that they can be used
to generate predictions regarding the robot’s behaviour in the future. For autistic children, detect-
ing structural regularities (i.e., invariant structures) may be challenging [61]. Varia nce in robot
behaviour can therefore lead to interactions with a robot that are too complex to be learned, and
therefore predicted, or it can take the child longer to learn and predict. Designing a robot to be
highly predictable then might entail programming robot behaviour with low variance such that it
has high structural regularities that can be easily inferred from its behaviour.
In short, we dene the predictability of a robot as an attribute that refers to “the degree in which
a person can quickly and accurately learn to predict a robot’s future behaviour”. For the purpose
of this article, we operationalise the design of predictable robots as decreasing or increasing the
variance in their behaviour, where more variance decreases the predictive value of the robot’s
behaviour and thus decreases its predictability.1
2.2.1 Types of Robot Variance. Variance in robot behaviour can stem from any of the robot’s
expressive modalities, as well as how those modalities are used within the interaction dynamics.
There are several types of variance that we consider in this article, namely variance in speech,
motion, topic, and time:
— S refers to variability in the words that the robot uses (diction) and how
those words are spoken (prosody). Diction variance can be operationalised as using dier-
ent sentence grammars and/or dierent words that have similar meaning. Prosody variance
relates to using dierent prosody and/or dierent voice actors to change intonation, stress,
rhythm, and pitch of the sound. While the use of prerecorded speech can minimise speech
variance, using models for natural language generation to generate dynamic speech may
inuence the variance in the robot’s speech signicantly.
— M refers to variability in the robot’s motions, such as facial expressions
and gestures. A robot could consistently use the same static animation to communicate a
certain intent, which would keep the motion variance low. In contrast, when using models
to generate dynamic motions, such as inverse kinematics, each robot motion can be unique
in its trajectory, increasing variance in the robot’s motions.
— T refers to variability in the timing of robot behaviour, both within ac-
tions (e.g., timing of a motion trajectory) as well as the timing between actions. While this
type of variance is not a type of variance that is actively designed for, temporal delays in
robot behaviour are likely to occur as the result of a lack of computational power, or due to
1For more information on robot predictability, and dierent operationalisation thereof, see Schadenberg et al. [128], or see
Chapter 6 in Schadenberg [124] for a more extensive elaboration.
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
36:6 B. R. Schadenberg et al.
the physics of the robot’s motors, increasing temporal variance. Such temporal delays can
disrupt the interaction of an autistic child [47].
— T refers to variability in the use of dierent actions that address dierent
topics at a particular point in an interaction. For example, at the rst step of an interaction,
increased topic variance could involve a robot saying either a greeting or a comment about
a bird, as opposed to consistently saying a greeting. Types of topic variance related to a
robot can include (1) when a robot responds to its internal state (e.g., battery level, or its
(faulty) perception) which the observer does not know about; (2) when a robot responds to
an external event (e.g., a person walking in on the ongoing session), or item (e.g., a clock)
in the environment; and (3) when a robot responds with an action that is not legible (i.e.,
understandable) to the observer (e.g., randomly saying “beep beep”).
There are other sources of variance (e.g., variance in robot appearance, or morphology), but we
do not utilise these sources in our study for introducing unpredictability. Finally, in addition to
these types of variance, a robot can combine modalities into one action. For example, a robot cap-
able of facial expressions of emotion can more clearly communicate these expressions when they
are combined with short, emotional non-speech expressions [125]. This can also lead to variance
when dierent combinations, intended to express the same multimodal intent, are perceived as
dierent robot actions. For autistic children, processing multimodal information may be more dif-
cult than for typically developing children [25,59,98]. Furthermore, autism professionals report
that presenting multimodal information might cause an information overload [65].
2.2.2 Reducible and Irreducible Unpredictability. While prioritising the minimisation of the
types of robot variance mentioned above can lead to highly predictable robots, doing so severely
restricts the types of applications for robots for autism. An example of such prioritisation of pre-
dictability would be programming it to do the same thing, in exactly the same manner, again and
again, resulting in a very limited repertoire of robot behaviours with the potential for endless re-
petition. For such robots, autistic children may quickly learn to predict their behaviour because
there are only a few behaviours to predict. While such robots may play a role in the broader scope
of robots for autism [39], a degree of unpredictability is unavoidable for most of the robot roles
envisioned by autism professionals [1,66] and researchers [19,37,38,122]. For instance, when
the robot’s role is that of a trainer or educator, where its task is to model, teach, and/or practice
a targeted skill and to provide feedback to the child. When the robot is positioned as a social in-
teraction partner, it will have to perform novel behaviour and in general have a larger range of
behaviours, in order to deal with the complexity, dynamics, and unpredictability of interacting with
a human. The resulting unpredictability of the robot’s behaviour is irreducible, as it is required for
the robot to perform its primary task. That is, even when the robot is carefully designed in terms
of its predictability, irreducible, and unpredictability is unavoidable and does not lead to design
considerations. For a robot that is positioned as a tutor, an example of this would be the unpre-
dictability that can stem from introducing learning content, which is the bare minimum it needs
to do.
There are also technical reasons why a robot is not always behaving predictably, in particular
for robots that behave autonomously, as the technical systems that allow the robot to perceive,
reason, and/or act, can each introduce variance. The robot may respond to what it is perceiving,
which can be faulty, or the observer can be unaware of what the robot is responding to. The
robot’s reasoning may be too complex, or faulty, to be understood by the observer. Or, the intent
the robot tries to convey may simply not be understood by the observer from the robot’s behaviour.
In all of these examples, the resulting robot behaviour is unexpected and was not predicted. As
the resulting unpredictability is (to a certain extent) inherent to using technical systems, we also
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
Predictable Robots for Autistic Children 36:7
consider this to be irreducible unpredictability. In part, this can be solved by utilising a Wizard-of-
Oz paradigm, where a person remotely controls the robot and creates the illusion of interacting
with an autonomous robot, as this reduces the need for certain technical systems. However, for
large-scale and long-term use of robots, the use of a Wizard-of-Oz system is not practical and
requires that the robot can operate autonomously [23,122]. The autism professional using the
robot is too busy paying attention and interacting with the child to also control the robot, and
we deem it unlikely that robot-assisted interventions provide enough value to warrant paying an
additional person to control the robot.
In contrast to irreducible unpredictability, there are also aspects of robot interaction design that
we consider to be reducible. These are instances where adding one feature may improve a certain
aspect of the interaction, but may decrease the robot’s predictability. They are design choices,
rather than a necessity for the robot to perform its primary task. For instance, common design
principles for achieving long-term interactions for typically developing children include using
intelligent tutoring systems to provide the optimal challenge and keep children motivated [110,
126], adding variance to speech to reduce boredom due to endless repetition of a robot’s behaviour
[27,73,77], or personalising speech [85,137]. For a tutoring robot, these features are desirable, but
are not necessarily required for the robot to perform its primary function. In the case of robots
for autistic children, choosing whether to implement such non-essential features that increase
unpredictability is a balancing act. When the robot becomes too unpredictable it may have negative
eects on the interaction and could even negate the intended eect of feature.
2.3 Engagement and Autistic Children Engaging with Robots
The benet of providing a predictable environment for autistic children in which they can learn
is that it may be easier to engage with learning as they do not have to deal with the discomfort
resulting from unpredictability. Engagement is a necessary prerequisite for learning [91], where
higher engagement results in more opportunities for cognitive and social skill learning [48,55].
2.3.1 Engagement Definitions and Measures. Azevedo [5] discuss how the concept of engage-
ment in learning has been (mis)interpreted in various ways in the literature leading to dierent
denitions and meanings of this concept. Within the eld of HRI, engagement is often used as
an outcome measure to say something about the quality and length of a participant’s interaction
with the robot. Engagement is also a concept used for the development of robots that can detect
various stages of the user’s engagement, such as the intention to engage, being engaged, or being
disengaged. An often used denition of engagement in HRI literature is that of Sidner et al. [130],
who dene engagement as “a process by which individuals in an interaction start, maintain and
end their perceived connection to one another”. Other denitions emphasise that engagement is
an aective process formalised as the degree to which an individual wants, or chooses, to engage
with a system [e.g., 14,96]. While there is no consensus on a single denition for engagement,
it is generally viewed as a multi-dimensional concept, including a behavioural, cognitive, and af-
fective component [28,48]. Note that these components of engagement are overlapping and can
sometimes be dicult to disentangle [134].
We are interested in engagement with a robot as it relates to the children’s state in which learn-
ing can occur, since this is what we are trying to achieve with robot-assisted interventions. In such
a context, refers to the child’s participation in learning activities and
involves on-task behaviour. Overall, studies on the engagement in learning of autistic children
mostly relate to behavioural engagement [72], often referred to as “social engagement” when the
task is to engage with another person. As with the denition of engagement, there are also various
approaches to measuring the behavioural engagement of autistic children in their interaction with
a robot. It can be directly assessed through observing the behaviour of the child, annotating the
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
36:8 B. R. Schadenberg et al.
level of behavioural engagement on a macro-behavioural level. For instance, Kim et al. [74]de-
veloped a compliance-based coding scheme for measuring (behavioural) engagement, where the
speed of the autistic child’s reaction to instructions or requests is indicative of the child’s level of
behavioural engagement. In this coding scheme, spontaneous engagement is the highest level of
behavioural engagement, in contrast to a child refusing to comply to the robot or adult’s request
and walking away, which is the lowest level of behavioural engagement. Other studies report on
micro-behavioural interactions, which are used to code the type of engagement, to get a deeper
insight into how autistic children engage [e.g., 78,127]. The choice for a certain measurement of
behavioural engagement appears to be inuenced by the purpose of the study and the type of data
being gathered.
Next, is about the child’s (inferred) interest in the learning activity
and how much the child enjoys it. It is generally assessed through observing the autistic child’s
emotions from which the underlying aect is inferred in terms of valence and/or arousal [e.g.,
74,116,117]. Correctly inferring the aect from the emotional expressions of autistic children can
be dicult, however, as they can produce unique and unusual facial expressions, including blends
of incompatible emotions that are not seen in typically developing children or children with Down
syndrome [148]. Also, the vocal intonation of autistic children can be atypical when expressing
emotions [86,90,101]. Notwithstanding, valence and arousal can be successfully annotated for
autistic children with a sucient agreement between coders [74,117]. Furthermore, using a ma-
chine learning approach trained on data of autistic children, valence and arousal can be detected
using facial, body-pose, and audio features, and heart rate [116].
Lastly, refers to the quantity and quality of the child’s psychological
investment in learning (i.e., use of cognitive eort in order to understand). The cognitive engage-
ment of autistic children is dicult to measure, as the current measures for cognitive engagement
[5] overlap with the other components of engagement [93], or can be too complex to be used by aut-
istic children, such as self-report questionnaires. Task-evoked pupillary responses have long been
associated with attentional engagement [10] and cognitive activity [62,71], as well as emotional
arousal [16], and have been used by researchers to measure the cognitive/aective engagement of
autistic children [e.g., 50]. However, measuring pupil dilation requires carefully controlled experi-
ments and experiment environment to control for other factors that inuence pupil dilation, such
as the pupillary reex to changes in illumination [11]. This makes pupil dilation dicult to use in
real-world settings where the illumination cannot be fully controlled. A concept that is more easily
measured—and is related to cognitive engagement—is that of attention, which is often viewed as
a necessary component for basic forms of engagement to occur [29]. Attention has both a covert
and overt component, where overt visual attention is relying on the gaze xation on a certain loca-
tion, and covert attention involves cognitive processes for paying attention to something without
the movement of the eyes [147]. Indeed, gazing at a particular object is not always indicative of
the person paying attention to that object [107]. Nonetheless, measuring visual attention through
gaze is a commonly used proxy for cognitive engagement in the eld of HRI [4,112], and is also
used with autistic children engaging with robots [e.g., 4,69,70,78,139].
Clearly, the various measures and components of engagement also show overlap. For example,
as Miller [93] noted, gazing at a certain location may also indicate aective engagement, as people
look more at what they like [89]. Altogether, engagement (the concept as a whole) is a fusion of
behavioural, aective, and cognitive components of a person’s involvement with a robot. As each
component of engagement can result in learning, considering all three components together can
provide a richer characterisations of a child’s engagement than any single component. Sometimes
all three components of engagement are combined into one bespoke measure for engagement [e.g.,
68,69,132]. For example, through measuring social signals such as eye gaze, vocalisations, smiles,
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
Predictable Robots for Autistic Children 36:9
spontaneous interactions, and imitation, which can then be converted into an engagement score
by adding one point of engagement for each social signal that is present in a certain segment.
2.3.2 Autistic Children Engaging with Robots. Scassellati et al. [122] discuss that studies on
robots for autism often report positive eects of the robot on the engagement of autistic children
[122], as shown by increases in positive aect [30,76], communication [75,143], and attention
[34,133,139]. Importantly, the engagement that is observed is often social in nature and is
directed not only at the robot, but also at other people near the robot [36,43,46,75,79,114].
The latter is signicant, because from a pedagogical point of view, it does not necessarily matter
whether the child interacts with the robot or whether the robot elicits interaction between the
child and adult, as learning can occur in either case. Robots then are uniquely positioned to assist
autistic children in learning socials skills as they can be both highly predictable in their behaviour,
and therefore deliver learning content in a manner that is easier to process and more comfortable
to them, as well as elicit social interactions with the child through which the skills can be learned.
2.4 Individual Dierences in Autism
Autistic children can vary widely from one another. Autism is a condition that causes atypical-
ities in cognitive, emotional, behavioural, and social functioning; however, such atypicalities are
manifested dierently in both quality and quantity [60]. For instance, some autistic children may
have exceptional intelligence, whereas others have a severe intellectual disability [57]. Some speak
uently, while others never develop spoken language. In light of the Bayesian accounts of autism,
some autistic children should be better equipped to deal with unpredictability than others. As we
are looking to balance the robot’s predictability, what constitutes “suciently predictable” is there-
fore likely to vary between autistic children. Indeed, Goris et al. [54] reported that, when autistic
traits are measured in typically developing adults, these correlate signicantly with preferences for
predictability. Furthermore, autistic individuals’ IU—as we mentioned earlier, a concept that shows
similarities with a desire for predictability — is also known to vary between autistic people [15,21].
Importantly, individual dierences also aect how and to what extent autistic children inter-
act with robot. The types of interaction that autistic children spontaneously initiated in a robot-
assisted activity correlated with individual factors [127]: Those with stronger language ability,
social functioning, and lower autistic features, initiated more functional interactions towards the
robot (e.g., talking to it), in contrast to visual and tactile exploration of the robot’s materials (i.e.,
touching and stroking specic parts of robot, such as its hands). The child’s autistic features have
also been reported to correlate strongly with the level of behavioural engagement of autistic chil-
dren in a robot-assisted activity [78,117].
In summary, children’s reactions to the robot’s unpredictability may be very dierent, and they
may nd dierent aspects of unpredictability problematic. For autistic children who are better
equipped to deal with unpredictability, the predictability of the robot’s behaviour may not be as
important, and the focus can lie on optimising the learning content. In general, the heterogeneity
between autistic children makes it dicult to generalise ndings, but it also stresses that there is no
one-size ts all solution to keeping them engaged in robot-assisted interventions whilst providing
meaningful learning.
3 PROBLEM STATEMENT
To make an informed decision on how to position and design the robot’s behaviour in terms
of its predictability, we need to better understand the interplay between robot predictability,
engagement, and the individual dierences between dierent autistic children. The aim of our
study was to investigate this interplay, where we specically looked at behavioural engagement
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
36:10 B. R. Schadenberg et al.
and visual attention (as a proxy for cognitive engagement) in relation to the robot’s predictability.
We measured these two facets of engagement through manual annotations of observable child
behaviours (see Section 4.7). In our study, autistic children engaged in a robot-assisted activity
that was about the basics of recognising emotional facial expressions. Over the course of four
sessions, the children interacted with a robot that was either low or high in variance of its
behaviour, which was how we operationalised robot predictability (see Section 2.2). We aimed at
addressing the following research questions:
Research question 1: How does variance in the robot’s behaviour aect the autistic child’s
engagement?
— To what extent does variance in the robot’s behaviour aect the autistic child’s behavioural
engagement?
— To what extent does variance in the robot’s behaviour aect the autistic child’s visual
attention?
We hypothesised that initially, in the rst session, there should be no dierence between the low
and high variance conditions, as the robot’s behaviour is novel and still has to be learned. How-
ever, based on the importance of predictability to autistic children, we expected that over ses-
sions the behavioural engagement and visual attention of autistic children should increasingly
diverge between the low-variance robot compared with the high-variance robot in favour of the
low-variance condition. That is, we expected an interaction eect between robot predictability
on behavioural engagement and on visual attention over sessions, but no main eects. For visual
attention, this means that we expected the children to increasingly look less towards the robot
(the source of the unpredictability) and more to non-activity-related locations that provided little
unpredicted sensory input, such as the walls.
Our second research question is related to the individual dierences between autistic children.
Research question 2: How do individual dierences between autistic children inuence their
engagement in the activity in relation to the variance in the robot’s behaviour?
— To what extent do autistic children’s autistic features moderate the relation between the
two facets of engagement and robot predictability?
— To what extent do autistic children’s expressive language ability moderate the relation
between the two facets of engagement and robot predictability?
— To what extent do autistic children’s IU moderate the relation between the two facets of
engagement and robot predictability?
Based on the preliminary ndings of Goris et al. [54], Rudovic et al. [117], and Schadenberg et al.
[127], we hypothesised that autistic children with higher autistic features should be less behaviour-
ally engaged and pay less visually attention to activity-relevant locations (main eects), and their
behavioural engagement should be more strongly and negatively aected than those with lower
autistic features (interaction eect). Our hypothesis was similar for autistic children with lower
expressive language ability. For autistic children who were more sensitive to unpredictability, as
measured through their IU, we expected that they should respond more strongly to more variance
in the robot’s behaviour (interaction eect).
This study is complementary to another study which will be published elsewhere (hereafter
referred to as the complementary study) and is the rst study that assesses the claim on the benets
of robots being highly predictable. That complementary study is about micro behavioural analysis
of autistic children in light of robot predictability. In the study reported in the current article, we
used the same study design and data collection as the other study, but focused on dierent research
questions and conducted dierent analyses to address those questions.
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
Predictable Robots for Autistic Children 36:11
Table 1. Participant Characteristics
Condition
Low-variance High-variance
n(sex) 12 (5 female) 12 (3 female)
Age (years:months)
M(SD) 8:8.92 (1:8.70) 8:4.42 (1:8.49)
Range 6:10–11:7 6:10–11:4
ADOS-2aCalibrated Severity Score
M(SD) 6.08 (1.78) 6.25 (1.60)
Range 4–10 4–10
CARS-2b
M(SD) 27.83 (4.09) 29.17 (6.56)
Range 20.5–33.0 21.5–38.5
SCQc
M(SD) 22.25 (8.72) 25.50 (5.79)
Range 8–37 17–33
Bespoke Scale of Expressive Language
M(SD) 2.92 (0.79) 2.42 (1.24)
Range 2–4 0–4
IUSC-Sd
M(SD) 2.64 (1.04) 3.19 (1.16)
Range 1.42–4.00 1.00–4.92
aADOS-2 [84].
bCARS-2 [129].
cSCQ [118].
dIntolerance of Uncertainty Scale for Children—Simplied (IUSC-S).
4 MATERIALS AND METHODS
4.1 Participants
Autistic children from the United Kingdom were recruited from a special education institution in
the Greater London area. In total, 27 children were recruited of whom 24 (8 girls) were included in
the analysis. These were autistic children with limited spoken communication and high support
needs. For the three children who were excluded from the analysis, participation was discontinued
due to administrative error (1 girl), or due to elevated anxiety during the session (2 boys).2All in-
cluded participants had previously received an independent clinical diagnosis of autism according
the ICD-10 [146], DSM-IV-TR [2], or DSM-V [3]. All children were assessed by the Autism Dia-
gnostic Observation Scale-second edition (ADOS-2) [ADOS-2, 84], the Childhood Autism
Rating Scale-second edition (CARS-2) [CARS-2, 129], the Social Communication Question-
naire (SCQ)[SCQ,118], and a bespoke scale of expressive language ability. In addition to receiving
a clinical diagnosis of autism, all children scored above the autism cuto on ADOS-2 (4 or higher).
The participants’ characteristics can be viewed in Table 1.
This study was reviewed and approved by the ethics committee of University College London,
Institute of Education (REC 1175). For all children, parental consent was obtained prior to their
2Both were in the high-variance condition (see Section 4.3). For one boy, the activity was inducing too much anxiety—he
did not get past the introduction in session one. The other boy repeatedly needed to be calmed in the second session, and
eventually refused to go on. It was then decided to stop the experiment for this child.
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
36:12 B. R. Schadenberg et al.
Fig. 1. Robokind’s humanoid robot R25 called “Zeno” or “Milo”.
participation in our study. Our experimental protocols followed the ethical standards laid down in
the 1964 Declaration of Helsinki.
4.2 Materials
The robot-assisted activity that was developed for the DE-ENIGMA project consisted of a social
robot, a tablet, and one laptop. The social robot that was used is Robokind’s humanoid robot R25
called “Zeno” (see Figure 1). It has ve degrees of freedom in its face and two in its neck, making
it capable of expressing various, recognisable, facial expressions of emotion [119,125]. The robot
also showed a number of bodily gestures, such as waving, cheering, or dancing using its the seven
degree’s of freedom in its body. A 9.7-inch Android tablet was used by the participant to provide
answers to the tasks by selecting (part of) a picture of the robot, or choosing what action the ro-
bot should do. In turn, the robot would autonomously respond to the participant’s choice. The
interaction was recorded through four high-resolution webcams, and one wide-angle webcam, all
with audio. In addition, high quality audio recordings were obtained through a dedicated micro-
phone and 3D video recordings from a Microsoft Kinect. Note that we only used the data from the
webcams for our analyses.
4.3 Experiment Design
We used the data collected in the complementary study. We will briey summarise the experiment
design here. The between-participants study involved two conditions, where the independent
variable was the robot’s behavioural variance, which was either low or high. The children were
randomly assigned to one of the two condition and remained in this condition throughout the
experiment. In the low-variance condition, the robot’s behaviour showed minimal variance. This
means that the verbal and physical behaviour for a certain action was always the same. For ex-
ample, the robot would always say “Hi, my name is Zeno” and wave with its right arm as a way of
greeting the child. In contrast, in the high-variance condition, we implemented speech, motion, tem-
poral, and topic variance, to increase the behavioural variance of all of the robot’s actions. Thus,
the robot displayed behavioural variance throughout the interaction. In this condition, each of the
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
Predictable Robots for Autistic Children 36:13
robot’s actions had four variations that diered the aforementioned types of variance, excluding
the additional actions that were designed to introduce topic variance. This was done as follows:
— S : variability in the words that the robot uses (diction) and how those words
are spoken (prosody).
— M : variability in patterns in the movements of the robot’s face and arms
during emotional facial expressions and robot actions.
— T : variability in time osets between the child pressing a button and the
robot responding to the button press.
— T : the use of dierent actions that address dierent topics at a particular point
in an interaction. Note that in the low-variance condition, the robot actions for topic variance
were responsive actions in that the robot responded to an event (e.g., hearing a noise). In
the high-variance condition, there would be no observable event that explained the robot’s
action.
The variant actions have been designed to display unimodal variance compared with their invari-
ant counterparts. That is, for each robot action, the action can show variance on either speech or
motion—not both modalities. An intent may translate to a combination of both speech and motion
actions, but these will always be shown sequentially as two unimodal actions, and will not be com-
bined to create a multimodal stimulus. For instance, when greeting the child, the robot would rst
say “Hello, my name is Zeno”, which was followed by a wave. We consider this as a single robot
action. For examples of behaviour variants, see Table Ain Appendix A.
The choice for which variant to display for an action was determined through an algorithm. The
action variant was chosen semi-randomly, where it would pick one of the four variants, exclud-
ing the antecedent variant for that action (if any). The latter was to prevent variants to be shown
twice in a row. For actions with topic variance, a dierent selection mechanism needed to be used,
as the variance for these actions related to whether the action is contingent or non-contingent
on (internal or external) events in the environment. We therefore opted to use the Wizard-of-Oz
paradigm, where another experimenter was controlling the robot (the wizard), without being vis-
ible to the child. The wizard was responsible for selecting the topic variant actions. To standardise
the number of topic variance actions, the wizard would be notied through the wizard’s control
interface when it was time for the robot to perform a topic variant/invariant action. At this time,
the wizard would look for one of the topic variant actions that was congruent with the condition
the child was in. In the low-variance condition, the wizard would select topic variant actions that
were a response to an observable event. For instance, the wizard could have the robot say "what
was that noise?" in response to noise outside of the experiment room. For the high-variance con-
dition, the wizard would ensure that the event the robot would respond to was not perceivable. In
this case, the wizard would have the robot respond to noise when there was no noise to be heard.
4.4 Experimental Setup
The study took place at the children’s school, in one of the oces. This room was converted into
an experiment room. A picture of the experimental setup can be seen in Figure 2. The robot stood
on a table facing the child and acted partly autonomously and was partly controlled by the wizard.
This person was sitting in the same room behind a room divider and could view the interaction
through a webcam. The wizard was responsible for selecting the correct task within the activity,
selecting the topic variance actions, and for responding to any unscripted interactions through
using a preset of robot actions such as having the robot say “no”, “yes”, and “I don’t know”. Within
each task, the robot behaved autonomously, although the wizard could interrupt these behaviour
at any time when the situation demanded it. The child would sit in front of the robot and next
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
36:14 B. R. Schadenberg et al.
Fig. 2. The experimental setup. The school’s sta member who accompanied the child would sit behind
the child in the corner of the room. The red circles outline the camera’s that were used for annotating our
engagement measures, which also included the camera from which this image was taken.
to the adult who was presenting the DE-ENIGMA activity. The recording equipment was placed
behind and next to the table on which the robot stood. All computers and laptops were also placed
behind the room divider and were managed by another researcher.
4.5 The DE-ENIGMA Activity
As we mentioned in Section 3, this study is complementary to another study. We use the same
procedure and data collection that is presented in there. For clarity, we describe the DE-ENIGMA
activity here as well, and the experiment procedure in the next subsection.
The participating autistic children engaged in a robot-assisted activity that was developed by the
DE-ENIGMA project which revolves around playing with the basics of recognising facial expres-
sion (see [82] on how the activity was developed and assessed). In the current study, the activity
was merely used to provide an interaction and structure the interactions between the child, robot,
and the adult to ensure that the interactions in the two experimental conditions are comparable
and oer similar opportunities for interaction. The activity itself was the same for both conditions.
The DE-ENIGMA activity starts with an introduction where the robot says, hello. Next, the
robot would display various behaviours to attracts the child’s attention, and to let the child get
more familiar with the various motions and sounds the robot makes. For instance, the robot would
show various facial expressions, do a dance, or play a nursery rhyme. After this introductory phase,
the main body of the activity starts, namely playing several/all of the DE-ENIGMA games.
There were four games (see Figure 7in Appendix Bfor a ow diagram and images of each of
the DE-ENIGMA games), where each game builds on the previous game in complexity. In the rst
game, the children could explore the facial features (mouth, eyes, and eyebrows) of the robot. The
tablet would show the robot’s face and highlight the three facial features. When the child touches
any of the features, the robot would move the features and label them. This way, children could
freely explore the facial features that were of interest to them. In the second game, the robot would
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
Predictable Robots for Autistic Children 36:15
prompt the child to nd one of the features and point them out using the tablet. When the child
provided an incorrect answer, the robot would prompt them to try again. For correct answers,
the robot would provide positive feedback and move the prompted facial feature. The third game
was similar to the rst game, only the child would instead explore facial expressions of emotions
(happiness, sadness, anger, and fear). The robot would explain what and how the robot’s facial
features move for the selected emotion. Finally, for the fourth and nal game, the robot would
prompt the child to identify one of the four emotions in a similar fashion as the second game.
Here the tablet would show four images of the robot, each with a facial expression of emotion.
When the child had played with the robot for around 15 minutes, the robot would prompt the
adult to end the session. After doing so, the robot would say goodbye.
Each game consisted of four opportunities for the child to make a selection, excluding instances
where the child was prompted by the robot again. After the four opportunities the game was over,
the tablet would display a blank screen, and the wizard would need to select the next course of
action. There were also several “choice points” during the activity, where the child could choose a
game or certain robot behaviours. These choice points allowed the adult to defer any requests for
one of the games or certain behaviours to a choice point, rather than denying the child’s request
or interrupting the ongoing game. Moreover, by structuring the choice points and limiting them
to two minutes, it allowed us to control it as much as possible with respect to following the same
programme of content and the same order for each child. During the choice point, the child could
choose from any content that they had already experienced (i.e., both games and generic robot
actions). The child could make the choice through a “choice board”, which contained icons of the
available options. The icons can be added or removed from the choice board, as each was stuck on
the board with velcro. The board and the icons on it were managed by the adult.
The games were designed to allow autistic children with limited receptive or limited expressive
language to complete the games. For instance, the robot would use simple language where each
speech action would consist of only a few words, and the tablet allowed the children to respond
to the robot non-verbally.
4.6 Experiment Procedure
The autistic children would engage in the robot-assisted activity individually, once per day, for
four to ve sessions. The fth session was only applicable to children who did not nish the activ-
ity in four sessions. Each session lasted around 15–20 minutes. The sessions were scheduled on
consecutive school days as much as possible. Due to weekends and the children’s schedule some
variability was inevitable. For the low-variance condition there was an average of 0.36 days in
between sessions (SD = 0.68), and for the high-variance condition this was 0.64 days (SD = 1.22).
Each child was assigned to one of three adults who would lead the sessions. The adult assigned
to the child would also remain with that child for each of the child’s sessions. The adult was tasked
with augmenting the robot’s instructions, supporting the children in using the tablet, giving feed-
back, and responding to their communicative overtures. Theirs was a supportive role, as the robot
delivered the majority of the instructions and feedback. The children were often accompanied by
a school sta member, who would sit in the back of the experiment room. They were asked not to
participate in the activity unless they thought there was an issue such as when the child showed
anxiety.
The content of each day of participation was scheduled to be delivered in a specic order and
was identical across both conditions. When the children met the robot for the rst time, the robot
was covered by a blanket when they walked in the experiment room. The session would start with
the robot being uncovered, introducing itself, and showing what movements it could do, some
facial expressions, and what it sounded like. The goal here was to let the child get comfortable
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
36:16 B. R. Schadenberg et al.
with the robot, what it looked like, and how it behaved. If the child liked any of the actions, they
could be repeated by the wizard. When the child appeared comfortable, the adult would suggest to
transition to the main activity—which the wizard then started—nishing the introduction. In sub-
sequent sessions, the introduction was similar but shorter. After the introduction, the DE-ENIGMA
activity described above would start. During the activity, the children would engage in the games
that were set for that day, as well as in “choice points”, where the child could choose one of the
games or specic robot behaviours they liked. As described in Schadenberg et al. [127], autistic
children often make requests to the robot and denying the child’s request could disrupt the inter-
action. These choice points allowed the adult to defer any requests for one of the games or certain
behaviours to a choice point, rather than denying the child’s request or interrupting the ongoing
game. Moreover, by structuring the choice points and limiting them to two minutes, it allowed us
to control it as much as possible with respect to following the same programme of content and the
same order for each child. During the choice point, the child could choose from any content that
they had already experienced. The child could make the choice through a “choice board”, which
contained icons of the available options. This board and the icons on it were managed by the adult,
and the wizard executed the child’s requests. For children who would or could not choose, the
adult would rst prompt the child by suggesting an activity or a robot action. If the child still did
not express any preference, the adult ask the robot what to do next, upon which the wizard would
select an activity. When the session was over, the robot would say goodbye, and the child went
back to class.
4.7 Measures
4.7.1 Engagement Measures. We chose to operationalise and measure the autistic children’s en-
gagement in terms of behavioural engagement and visual attention as these can be annotated as
patterns of manifest content—child behaviours that are directly observable [108]. This is in contrast
to scoring engagement holistically as projective latent content [108], as this would require the coder
to employ subjective interpretations of the meaning of the behaviour, which is dicult without be-
ing familiar with how the participating children generally behave and understanding the meaning
of their sometimes atypical and idiosyncratic behaviours.
Behavioural engagement. We measured behavioural engagement through expert annotations,
using the coding scheme described in Table 2on segments of 5 seconds. With this coding scheme,
we make the distinction between autistic children being behaviourally engaged or disengaged with
the HRI using a 5-point ordinal scale that denotes the amount of behavioural engagement.
While we believe that stimming behaviour—a self-stimulatory behaviour marked by a repetitive
action or movement of the body—or dgeting can be indicative of behavioural disengagement, it
was problematic to annotate. Stimming behaviour provides sensory input for one of the senses, pre-
venting the child from using this modality for engaging in the activity. However, whilst stimming,
the child can still engage in the activity—and potentially learn—through using other modalities.
For example, a child may be rubbing their hands while speaking to the adult or robot about the
activity. Because we do not distinguish between modalities for annotating behavioural engage-
ment, we decided to code all stimming behaviours that did not prevent the child from engaging
in the activity and interaction were coded as passive. When the stimming was all consuming and
prevented the child from engaging with the activity, we coded it as disengagement.
Visual attention. The experiment procedure and the activity did not allow us to directly meas-
ure cognitive engagement, which is why we opted to measure this through annotating the chil-
dren’s visual attention—a proxy for cognitive engagement. Visual attention relates to the extent
the participant paid overt visual attention to the ongoing activity and interaction, and tells us
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
Predictable Robots for Autistic Children 36:17
Table 2. Coding Scheme for Annotating the Observed Behavioural Engagement of an Autistic Child
Level Meaning Description Example
−2 Fully behaviourally
disengaged
The child exhibits
non-task related behaviour
formostorallofthe
segment.
Child is playing around with
objects that does not involve
the task or interaction with the
robotoradult.Childis
engaging in stimming
behaviours that prevent the
child from partaking in the task.
Child indicates wanting to stop
with the interaction, e.g.
through asking whether the
game or session is nished, or
turning the tablet to the adult
prior to the game’s conclusion.
Talking with the adult about
something unrelated to the
activity or the ongoing triadic
interaction, such as children
saying they are hungry.
−1 Partly behaviourally
disengaged
The child exhibits
non-task related behaviour
forsomeofthesegment.
0 Passive Child does not
behaviourally engage with
the task, nor does the child
show engagement in other
activities.
Child is seemingly listening to
the adult or robot, but does not
communicate back. Child is
looking at the task material, but
does not physically interact
with it. Also, echolalic and
undirected vocalisations were
are included on this level, as
well as stimming behaviours
that do not prevent the child
from partaking in the task.
Covering ears due to auditory
sensitivity.
1 Partly behaviourally
engaged
The child behaviourally
engages with the task
through interacting with
the adult, with the robot
directly or through the
tablet, or other task
materials for some of the
segment.
(Continued)
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
36:18 B. R. Schadenberg et al.
Table 2. Continued
Level Meaning Description Example
2 Fully behaviourally
engaged
The child behaviourally
engages with the task
through interacting with
the adult, with the robot
directly or through the
tablet, or other task
materials for most or all of
the segment.
Pressing a button on the tablet,
requesting facial expressions of
the robot, choosing an activity,
talking with the adult about the
robot, dancing with the robot.
Also includes non-verbal
communication with adult,
such as sharing enjoyment, or
social references after the robot
did something.
more on what or whom the children were engaged with. This was also measured through expert
annotations by coding the children’s gaze direction, but on segments of 2.5 seconds. From these
annotations, we calculated the percentage of time spend looking at a certain direction for each
session.
The coding scheme that we used included the following gaze directions: “robot”, “tablet”, “activ-
ity materials”, “adult”, “school sta member”, “elsewhere”, or “mixed”. The latter was used to an-
notate instances where the child did not focus their gaze at any one point. The activity materials
refers to any materials that were used in the activity, which primarily was the visual choice board
for selecting the next task. And elsewhere refers to any gaze direction that was not in any of the
other categories and thus was not related to the activity. Gaze direction such as the walls, the oor,
the cameras, or the room divider. In contrast, the locations “robot”, “tablet”, “activity materials”,
and “adult” can be considered “activity-relevant locations”. For our analysis of visual attention, we
specically look at the gaze directions towards the robot, and elsewhere, as our hypotheses relate
to these locations. The other locations were annotated to enable better interpretation of the data.
The annotations were done at two levels. The rst (primary) level represents the gaze direction
where the child spent looking for the majority of the segment (1.25 seconds or more). The second
(secondary) level was optional and could be used to annotate a secondary gaze direction during
the segment which lasted at least 0.75 seconds, but no more than 1.25 seconds. This excludes brief
glances, where the child’s gaze would not stay on one direction.
4.7.2 Individual Factors. The complementary study also measured several individual dier-
ences between the autistic children. In the current study, we use a subset of those measures, namely
the CARS-2 for measuring autistic features, a bespoke scale of expressive language, and an adapted
version of the IUSC-parent report form [IUSC, 26].
The IUSC questionnaire was adapted from the complementary study to accommodate to chil-
dren with limited spoken language, as we believed for many of them the IUSC questions were
too dicult for parents to answer about their child. In the remainder of this article, we refer to
this adapted version of the parent report of the IUSC as the IUSC-S. The adapted questions of the
IUSC-S can be seen in Appendix C,Table7, as well as principal component analysis of this adap-
ted questionnaire. Based on this analysis, we excluded three questions for computing the IUSC-S
scores.
The CARS-2 [129] is a 15-item autism screening and diagnostic tool and was administered to
obtain a general measure of characteristics of autism. It was completed based on direct behaviour
observation by a professional as well as reports from parents, teachers, or caretakers. The measure
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
Predictable Robots for Autistic Children 36:19
was completed by the adult who worked with the specic child. The total score on the CARS-2
reects the severity of autistic features with scores of 15.0–29.5 indicating minimal-to-no evidence,
30.0–36.5 is mild-to-moderate severity, and 37.0 and higher is severe autistic features.
The bespoke scale of expressive language measures the spoken language ability of the child. The
adult who gave the sessions rated the child’s expressive language after the last session. This scale
ranged from 0–4, where 0 means “no words”, 1 is “some vocalisations or word approximations”,
2 is “single words”, 3 is “simple sentences” (two to three words), and 4 is “more complex speech,
including complex sentences”. The bespoke score reects the level of expressive language that was
generally used by the child during the sessions.
4.8 Annotation Procedures
As only a few children participated in ve sessions, we only annotated the videos from the rst
four sessions. These sessions were annotated in terms of behavioural engagement levels and visual
attention locations using the ELAN transcription software,3developed by the Max Planck Institute
for Psycholinguistics in Nijmegen, the Netherlands.
4.8.1 Behavioural Engagement Annotation. To determine whether a child is engaged or disen-
gaged requires taking the situational context into account. For example, the child looking away
may be part of the current activity, a response to the adult, or indicative of disengaging from the
interaction. To preserve the situational context, we annotated 1 minute out of every 2 minutes,
excluding only instances where there were technical problems with the system. We started the
annotations from the moment the adult said hello to the robot and ended when the robot had
said goodbye. This resulted in the annotation of 1,660 minutes of video recording, of which 848
minutes were in the low-variance condition and 812 minutes were in the high-variance condition.
During initial testing of the coding scheme for behavioural engagement, we noted that the aut-
istic children sometimes had brief behavioural disengagement episodes of a couple of seconds. We
therefore, divided each minute into segments of 5 seconds, similar to Kim et al. [74] and Simpson
et al. [132].
A single main coder annotated all the recordings, amounting to 9,947 segments. To calculate the
inter-rater reliability of these annotations, a second coder annotated a randomly selected session
for each participant. This amounted to 25% of the recordings being dual-coded, which also con-
tained 25% of all the segments. To determine the agreement between the two coders, Cohen’s κ
statistic was used. There was good agreement between the two coders for behavioural engagement
(Cohen’s κ= .72, 95% CI [.70, .75], p<.001).
To gain insight into the nature of the coder disagreements, we inspected the confusion matrix for
behavioural engagement (see Table 9in Appendix D). The main coder annotated more instances
of disengagement (approximately 14%), which were coded as passive by the secondary coder. Note
that the main coder was not blind to conditions, but the secondary coder was. Given the agreement
between the coders, we do not think it likely that this inuenced our results. All in all, we deem
these values high enough to continue our analysis on the basis of the main coder.
4.8.2 Visual Aention Annotation. For visual attention, we annotated the same minutes as for
behavioural engagement. However, annotating segments of 5 seconds proved too long, as there
were often more than three gaze directions. This made it dicult to code with our annotation
scheme, which accounted for two gaze directions. The segments for visual attention therefore
lasted 2.5 seconds to ensure that there were generally fewer than three gaze directions per segment.
3https://tla.mpi.nl/tools/tla-tools/elan/.
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
36:20 B. R. Schadenberg et al.
Again, a single main coder annotated all the recordings and second coder annotated a randomly
selected session for each participant. The main and secondary coder were the same annotators
as for the annotation of behavioural engagement. For visual attention, this amounted to a total
of 19,855 segments. The dual-coding resulted into 25% of the segments being dual-coded. There
was very good agreement for the primary gaze direction annotations (Cohen’s κ= .87, 95% CI [.86,
.89], p<.001) and good agreement for the additional, secondary gaze direction annotations on the
(Cohen’s κ= .69, 95% CI [.66, .73], p<.001). We deem these values high enough to continue our
analysis on the basis of the main coder. Furthermore, inspection of the confusion matrices (see
Tables 10 and 11 in Appendix D) showed no biases on the coder disagreements for one of the two
coders.
4.9 Data Analysis
For the analysis of both behavioural engagement and visual attention, we used growth models
with a maximum likelihood estimation method. These were modelled using the statistical program
“R” [109], version 3.6.3, with the “nmle” package [106], and analysed with two-tailed tests and a
95% condence level. Growth models allow us to take the multi-level nature into account, where
the behavioural engagement scores/visual attention locations for each of the sessions are level-1
variables, the child is a level-2 variable, and the adult leading the session a level-3 variable. The
adults were randomly assigned to the children regardless of the condition that was assigned to
the child. The adult is therefore a crossed eect. For the visual attention analysis, we removed
annotations where the child looked at the school sta member, as they were not always present
in each session. Additionally, we removed annotations coded as “mixed”, due to the uncertainty
regarding the direction of the child’s gaze.
As recommended by Raudenbush and Bryk [111], we rst make a “basic” model and then add in
variables as appropriate. In this article, we consider the model that includes random coecients,
the condition, and the session (time) as the basic (growth) model. We expected that the children will
have dierent intercepts, as some children will be behaviourally more engaged than other children,
or prefer looking at a certain gaze location more. Furthermore, we also expected that the slopes will
be dierent between children, where one child loses interest faster than other children, resulting
in dierence in behavioural engagement and visual attention. The random coecients account for
these random eects in our models. Next, we further explored to what extent the individual factors
improved the model t of the basic conditional growth model. Finally, we investigated whether the
dierent adults leading the sessions inuenced the two facets of engagement of the children by
adding the adult as a factor.
4.10 Manipulation Check
The study protocol had a exible activity selection and session duration so as to support each
autistic child’s individual preferences. This means that the variability introduced by the robot
diered per session and per participant. In turn, this means that we cannot be certain that the
robot displayed high amount of variance in the high-variance condition, and vice versa for the
low-variance condition. To check to what extent our manipulation of robot variance succeeded,
we calculated the following variables:
(1) A . These should be similar between condi-
tions, and serves as a baseline to put items (2) and (3) into perspective.
(2) A . These are unique within ases-
sion. Of the total number of robot actions, the high-variance conditions should have a more
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
Predictable Robots for Autistic Children 36:21
unique robot actions per minute. This reects the larger pool of unique actions that were
implemented in the high-variance condition.
(3) A . These are actions that the child
has not seen before in either the current or previous sessions. While the content of the
interaction diered per session, there are robot actions that are used throughout the sessions,
such as giving praise. Therefore, this number should drop over sessions in both conditions,
but the high-variance condition should introduce more novel actions per minute than the
low-variance condition in all sessions.
(4) T . This number reects
how often the robot displayed a certain robot action, either in the current session or previous
sessions. The higher the number, the more opportunities the child had to learn to predict this
action. In the high-variance condition, this large amount of unique actions should result in
fewer repetitions per action than in the low-variance condition.
For the manipulation check, we consider all variants of robot actions as being unique actions. For
the variables that were calculated per minute, we excluded the time when there was a technical
diculty. To assess to what extent there is a dierence between the two conditions on each of these
four items, we conducted four mixed ANOVAs, where the condition is a between-subject variable
and session a within-subject variable. Furthermore, we assessed each outlier and consider to re-
move them from the analysis. This was done by placing the outlier in context of what happened
during the session as well as relating the value of the outlier to the values of the other condition.
5 RESULTS
5.1 Manipulation Check
On average the children engaged in the DE-ENIGMA activity for 16 minutes and 13 seconds (SD
= 3 min, 1 s) in the low-variance condition. For the high-variance condition, the average was 15
minutes and 39 seconds (SD = 1 min, 23 s). There was no signicant dierence in the time spent
in the activity between the conditions (F(1, 22) = 3.30, p= .08, η2
p= .13, η2
p90% CI[.00, .34]), over
sessions (F(3, 66) = 2.55, p= .06, η2
p= .10, η2
p90% CI[.00, .20]), nor an interaction eect (F(3, 66) =
0.22, p= .88, η2
p= .01, η2
p90% CI[.00, .03]). Note that the variables reported below are all normalised
to account for dierences between children and sessions in the time that the session lasted.
The extent to which there was variance in the robot’s behaviour per session can be seen in
Figure 3for both conditions. The average number of robot actions per minute for the low-variance
condition was 5.43 (SD = 0.55) and 5.36 (SD = 0.82) for the high-variance condition. There was no
signicant dierence in the average number of robot action per minute over sessions (F(3, 66) =
0.58, p= .632, η2
p= .03, η2
p90% CI[.00, .07]), between conditions (F(1, 22) = 0.09, p= .763, η2
p<
.01, η2
p90% CI[.00, .12]), or an interaction eect (F(3, 66) = 0.91, p= .441, η2
p= .04, η2
p90% CI[.00,
.10]). As expected, there was a signicant dierence between conditions in the average number of
unique robot actions per minute (F(1, 22) = 136.78, p<.001, η2
p= .86, η2
p90% CI[.74, .90]), where this
number was signicantly higher in the high-variance condition (M = 3.80, SD = 0.63) than in the
low-variance condition (M = 1.80, SD = 0.30). Similarly, the dierence between conditions for the
average number of novel robot action per minute was signicant (F(1, 22) = 363.44, p<.001, η2
p= .94,
η2
p90% CI[.89, .96]). The robot in the high-variance condition displayed a higher number of novel
actions per minute than in the low-variance condition for each of the four sessions. For the average
number of repetitions per action there was also a signicant dierence between the conditions (F(1,
22) = 406.46, p<.001,η2
p= .95, η2
p90% CI[.90, .96]). For each session, the high-variance condition had
fewer repetitions per action than in the low-variance condition. Based on these tests, we conclude
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
36:22 B. R. Schadenberg et al.
Fig. 3. Boxplots that show to what extent variance in the robot’s behaviour was achieved. (A) shows the
average number of actions the robot performed per minute, (B) shows the average number of unique actions
performed by the robot actions per minute within a session, (C) shows the average number of actions that the
robot performed for the first time per minute, and (D) shows the cumulative average number of repetitions
per robot action.
that the robot’s variance was signicantly dierent between the conditions, as well as suciently
large given the reported eect sizes.
While we consider the manipulation of the extent of variance in the robot’s behaviour successful,
there were four sessions of three children where the variance of the robot’s behaviour was more
akin to that of the other condition. In the low-variance condition, the rst session of one child
was only partially recorded in our logles due to technical issues. The data loss resulted in a
shorter session recording, which in turn resulted in a higher number of unique and novel actions
per minute, as well as fewer repetitions per action. Nonetheless, it was a regular session where
the child was engaged for most of the time. It is therefore likely that the system performed as
intended and produced low variance in the robot’s behaviour. In the high-variance condition, two
children had sessions where the robot showed few unique actions. For one child, this was the case
in sessions 3 and 4, while for the other it was only for session 3. Note that for session 3, these
instances are not statistical outliers and therefore are not shown as such in Figure 3. For all three
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
Predictable Robots for Autistic Children 36:23
Table 3. Model Fit Measures for Each Model as They Increase in Complexity, as well as Chi-Square
Statistic with 1 Degree of Freedom and Statistical Significance
Model df AIC BIC Log-Likelihood χ2p
Basic growth models
(A) Conditional 9 26.46 49.54 −4.23 - -
(B) Conditional with quadratic trend 9 21.46 44.54 −1.73 - -
(C) Three-level 12 27.45 58.22 −1.72 0.01 .999
M B with covariate
(D) CARS-2 score 10 13.86 39.50 3.07 9.60 .002
(E) Exp. lang. score 10 14.87 40.52 2.56 8.59 .003
(F) IUSC-S score 10 23.43 49.08 −1.72 0.03 .867
M D with another covariate
(G) CARS-2 score and exp. lang. score 11 12.05 40.26 4.98 3.81 .051
M G with interaction between covariate and condition
(H) CARS-2 score * condition 12 14.05 44.82 4.98 <0.01 .961
(I) Exp. lang. score * condition 12 11.08 41.85 6.46 2.97 .085
(J) IUSC-S score * condition 13 15.14 48.48 5.43 0.91 .634
Each model is compared with the rst model with lower complexity as reected by the model’s degrees of freedom.
For the models with one or two covariates, the models are compared, respectively, to M B and M D. The
models with an additional interaction eect between the covariate and the condition are compared with M G.
sessions, the child was disengaged for most of the session, limiting the progression through the
content. In turn, the robot only displayed a limited repertoire of its behaviours, which increased
the chance of displaying actions that had already been displayed before. Given that this happened
in the last two sessions, we cannot exclude that this happened due to their experiences in the rst
and second session. Moreover, the mean number of novel actions per minute and repetitions per
action are in line with the high-variance condition. Therefore, we include these sessions in the
main analysis.
5.2 Behavioural Engagement
The parameter estimates for each of the multi-level models are presented in Table 3. First, we tted
a basic conditional growth model (M A) with random slopes and intercept. As xed eects,
this model contains the session, condition, and an interaction eect between the two. For the
covariance structure, we used a rst-order autoregressive covariance structure. To further explore
the trend of behavioural engagement over sessions, we tted a quadratic and cubic trend instead
of the linear trend. The quadratic trend best tted the change in behavioural engagement over
sessions (M B). Next, we investigated whether accounting for the adult who was leading the
session improved the model (M C). This required a three-level model, where the adult is a
crossed eect. Compared with M B, the three-level model did not signicantly improve the
model t (χ2(1) = 0.01, p= .999). Thus, while the children received their sessions from one of three
adults, this does not explain the variance in their behavioural engagement.
Next, we investigated whether individual dierences in characteristics of the children could
explain the variance in behavioural engagement. We took M B, as it had the best model t,
and investigated whether the CARS-2 score, expressive language score, or IUSC-S score, improved
the model t. Adding the CARS-2 score as covariate signicantly improved M B (χ2(1) = 9.60,
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
36:24 B. R. Schadenberg et al.
Table 4. Parameter Estimates for the Growth Model on Behavioural Engagement with the CARS-2 Score
(M D), Bespoke Expressive Language (Exp. Lang.) Score (M E), or both CARS-2 and Bespoke
Expressive Language Score as Covariate (M G)
M D: CARS-2 M E: Exp. Lang. M G: CARS-2 + Exp. Lang.
Variable b (SE) 95% CI b (SE) 95% CI b (SE) 95% CI
Fixed eects
Intercept 1.59 (0.33) 0.95, 1.59 −0.09 (0.20) −0.47, 0.29 0.89 (0.41) 0.09, 1.69
Session2−0.01 (0.01) −0.03, 0.00 −0.01 (0.01) −0.03, 0.00 −0.01 (0.01) −0.03, 0.00
Condition −0.09 (0.14) −0.37, 0.19 −0.08 (0.13) −0.34, 0.18 −0.05 (0.12) −0.31, 0.20
Ses2:Cond −0.01 (0.01) −0.03, 0.01 −0.01 (0.01) −0.03, 0.01 −0.01 (0.01) −0.03, 0.01
CARS-2 −0.04 (0.01) −0.06, −0.02 - - −0.03 (0.01) −0.05, −0.01
Exp. Lang. - - 0.20 (0.06) 0.07, 0.32 0.13 (0.06) 0.01, 0.25
SD 95% CI SD 95% CI SD 95% CI
Random eects
Intercept 0.41 0.30, 0.56 0.33 0.24, 0.46 0.35 0.25, 0.50
Session 0.12 0.09, 0.17 0.12 0.09, 0.17 0.12 0.09, 0.17
M D: Behavioural engagement ∼Session2* Condition + CARS-2 + (Session |Participant).
M E: Behavioural engagement ∼Session2* Condition + Exp. Lang. + (Session |Participant).
M G: Behavioural engagement ∼Session2* Condition + CARS-2 + Exp. Lang. + (Session |Participant).
The statistically signicant parameter estimates for the xed eects are in bold, excluding the intercept.
p= .002). Similarly, the expressive language score for expressive language signicantly improved
M B (χ2(1) = 8.59, p= .003). Adding the IUSC-S did not signicantly improve M B (χ2(1)
= 0.03, p= .867).
Based on the model t of the models reported above, and whether they signicantly improved
the model t, we consider M D with the CARS-2 score as covariate, and M E, which has
the expressive language score as covariate, as the models that best explain the variance in behavi-
oural engagement. However, to understand to what extent the CARS-2 score and the expressive
language score explain the same variance in behavioural engagement, we tted a model using both
covariates (M G). This did not signicantly improve the model t compared to the best model
with only one covariate (M D). Furthermore, while this model has a better t than M D,
it is also more complex, as reected by a higher BIC.
To investigate whether the covariates actually moderated the eect of the condition (robot pre-
dictability) on behavioural engagement, we further added an interaction eect between the cov-
ariate and the condition to M G. For the CARS-2 score, this did not signicantly improve the
modeltofMG(χ2(1) <0.01, p= .961). Nor was the model t improved when adding an
interaction eect for the expressive language score (χ2(1) = 2.97, p= .085), or the IUSC-S score
(χ2(1) = 0.91, p= .634).
The model parameters of all M D, E, and G can be seen in Table 4. For M D, the CARS-
2 score was signicant (t(21) = −3.52, p= .002). The condition was not signicant (t(21) = −0.64,
p= .529), nor was session2(t(70) = −1.78, p= .079), or the interaction between the condition and
session2(t(70) = −1.18, p= .243). The relationship between the two conditions and behavioural
engagement showed signicant variance in intercepts across the children. In addition, the slopes
signicantly varied across children, and the slopes and intercepts were negatively and signicantly
correlated (r=−.73, 95% CI[−.88, −.45]).
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
Predictable Robots for Autistic Children 36:25
Fig. 4. The predicted values (marginal eects) for behavioural engagement of growth M G, presented
in Table 4.
The parameter estimates for M E, including the expressive language score, show a sim-
ilar trend to M D. The covariate, expressive language score, was signicant (t(21) = 3.25,
p= .004). But neither the condition (t(21) = −0.64, p= .530), session2(t(70) = −1.78, p= .079),
nor the interaction eect between the condition and session2was signicant (t(70) = −1.18, p=
.243). The random intercept and slopes showed a signicant and negative correlation (r=−.52,
95% CI[−.78, −.12]).
The marginal means for behavioural engagement, estimated by M G, can be seen in Figure 4.
M G, which includes the CARS-2 and expressive language score as covariates, shows that
both covariates signicantly contribute to predicting the variance in the behavioural engagement
of the children. For the CARS-2 the parameter estimate is −0.03 (t(20) = −2.51, p= .021). This
means that the model estimates that autistic children who scored higher on the CARS-2 were less
behaviourally engaged. For the expressive language score, the parameter estimate is 0.13 (t(20) =
2.18, p= .042), which means that autistic children with more complex expressive language, with
scores ranging from 0 to 4, were also more behaviourally engaged than those with less complex
expressive language. The model shows that the condition was not signicant, nor did the condition
interact with the sessions.
5.3 Visual Aention
The visual attention over the sessions can be seen in Figure 5. As our hypotheses only relate to
the “robot” and “elsewhere” gaze direction, we will not report on the other annotated gaze direc-
tions. Again, we tted multi-level models, using random slopes, random intercepts, and rst-order
autoregressive covariance structure, to model the participants’ visual attention with the robot and
their visual attention elsewhere. The parameter estimates for the conditional growth model on
visual attention towards the robot can be seen in Table 5. The relationship between the robot’s
variance and visual attention with the robot showed signicant variance in intercepts across the
children, but the slopes were non-signicant. We investigated whether accounting for the dier-
ences between children, as measured by the individual factors, improved the model. In contrast
to the previous results on behavioural engagement, this was not the case for any of the measures.
The model that best ts the data is therefore a conditional growth model with random intercepts.
The visual attention towards the robot signicantly decreased over sessions (t(70) = −2.91, p=
.005). This could be due to other factors that inuence visual attention, such as a novelty eect
that wears o, or boredom. There was no eect of condition (t(22) = 0.56, p= .579), nor was there
an interaction eect between condition and session (t(70) = −1.24, p= .219).
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
36:26 B. R. Schadenberg et al.
Fig. 5. The (unmodelled) percentage of time a participant spent looking at each of the annotated directions
for each of the sessions (visual aention) and each of the conditions.
The model parameters for the children’s visual attention elsewhere canalsobeseeninTable5.
We found no signicant variance in the slopes across children. The variance in intercepts did signi-
cantly dier. Again, accounting for dierences on the autism-specic measures did not improve
the model. Therefore, the data was best described by a conditional growth model with random in-
tercept. The visual attention elsewhere did not signicantly decrease over sessions (t(70) = −0.12,
p= .908), nor was there a signicant dierence between conditions (t(22) = −1.75, p= .095). There
was, however, a signicant interaction eect between the condition and session (t(70) = 3.90, p<
.001). In the high-variance condition, children looked increasingly “elsewhere” over sessions com-
pared with the low-variance condition. Thus, in contrast to the results for behavioural engagement,
this shows an impact of robot predictability on the children’s visual attention—which is indicative
of engagement—to the robot-assisted activity.
6 DISCUSSION
The goal of our study was to investigate the interplay between robot predictability, the behavi-
oural engagement and visual attention to the activity, and the idiosyncrasies therein between
autistic children. To that end, we manipulated the variance in the robot’s behaviour as a way to
operationalise predictability, and measured the children’s behavioural engagement and visual
attention in a robot-assisted activity, as well as individual factors.
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
Predictable Robots for Autistic Children 36:27
Table 5. Parameter Estimates for the Conditional Growth Models for Visual Aention
towards the Robot and Elsewhere
T E
Variable b (SE) 95% CI b (SE) 95% CI
Fixed eects
Intercept 0.46 (0.04) 0.38, 0.55 0.22 (0.04) 0.14, 0.29
Session −0.03 (0.01) −0.05, −0.01 −0.00 (0.01) −0.02, 0.02
Condition 0.03 (0.06) −0.09, 0.15 −0.10 (0.05) −0.21, 0.00
Session:Condition −0.02 (0.01) −0.05, 0.01 0.05 (0.01) 0.03, 0.08
SD 95% CI SD 95% CI
Random eects
Intercept 0.10 0.07, 0.14 0.10 0.08, 0.15
V ∼Session * Condition + (1 |Participant).
V ∼Session * Condition + (1 |Participant).
The statistically signicant parameter estimates for the xed eects are in bold, excluding the
intercept.
Fig. 6. The predicted values (marginal eects) for the visual aention towards the robot (A) and elsewhere
(B) of the growth models presented in Table 5.
We found that for the less predictable robot, autistic children paid less visual attention to activity-
relevant locations as the sessions progressed. Rather, the children started to pay more attention
to locations where nothing moves (annotated as “elsewhere”), such as the walls, the cameras, or
the room divider. In contrast to our predictions, however, the robot’s predictability did not impact
on behavioural engagement. The children continued to engage with the robot-assisted activity re-
gardless of the robot’s predictability. We did nd, though, that individual dierences in children’s
background characteristics were related to their behavioural engagement. Higher autistic features
were related to less behavioural engagement of autistic children, while the children’s expressive
language ability was related to greater behavioural engagement. Visual attention, on the other
hand, was not inuenced by any of the individual dierences measured. Finally, we found no evid-
ence for a relation between the autistic children’s IU and their response to the robot’s predictability
in terms of their behavioural engagement and visual attention.
In conclusion, it appears that the children continued engaging in the robot-assisted activity, but
started to pay less visual attention over time to the activity-relevant locations when the robot was
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
36:28 B. R. Schadenberg et al.
less predictable. Instead, the children started to pay more attention to locations that visually do not
change (i.e., no changes in sensory information), such as the walls, or the oor—they look away
from the activity. We believe this might be a coping strategy to minimise sensory input and deal
with anxiety resulting from the inability to learn to accurately predict the robot’s actions. Similarly
to how stimming can be used by autistic children to generate predictable sensory information to
deal with an overload of unpredictability [135]. For learning, paying less visual attention to the
activity-relevant locations is problematic as it indicates that the child is less strongly engaged
with learning tasks. In particular, this may impact long-term use of robot-assisted interventions,
as in our study this eect got stronger over sessions. We may also expect that eventually the
children may start to be less behaviourally engaged, when the encouragement from the adult starts
to fail in motivating the child to engage with the robot and deal with the resulting unpredictability.
However, as we did not measure the children’s learning, we cannot draw a rm conclusion about
whether the diminished visual attention also impacts the learning.
The result that individual dierences inuence the behavioural engagement of autistic children
is in line with previous studies that found correlations between autistic features, measured through
the CARS-2, and behavioural engagement [117,127]. Similarly, Kostrubiec and Kruck [78]found
correlations between autistic features, measured via the SCQ, and the proportion of prosocial be-
haviours in a robot-assisted intervention. Not only do individual dierences seem to inuence to
what degree autistic children are behaviourally engaged in activities with robots, they are also
correlated with how they behaviourally engage [127]. Researchers often report increased engage-
ment, increased levels of attention, and novel social behaviours when incorporating a robot in the
interaction [38,113,121,122]. For such studies, it is then important to account for the individual
dierences and report these. But the relationship between the children’s individual dierences and
behavioural engagement also raises a new question: Do the children’s individual dierences only
inuence their behavioural engagement directly, or do they moderate the relationship between the
eect of the robot on behavioural engagement. Understanding this relationship would allow us to
better determine which autistic children may benet in particular from such robot activities or in-
terventions. Lastly, these ndings also suggest that the way autistic children interact with, or in the
presence of, robots could potentially be indicative of their autistic features. This warrants further
research, but might be particularly interesting for robots that are used for diagnosing autism.
Whether certain individual factors also moderate the eect of robot predictability on engage-
ment remains unanswered by this study, as we found no evidence that this was the case. In par-
ticular, we had expected that those higher in IU would also be more strongly aected by the ro-
bot’s predictability, given the similarity between IU and unpredictability. While other studies did
nd relationships between IU and autistic features, such as the presence of sensory sensitivities
[95], repetitive motor behaviours, and insistence on sameness [145], we found no results in our
study that support a relationship between IU and predictability. Similarly, in an earlier study with
(typically developing) adults where we manipulated predictability only in terms of topic variance,
we found no relationship either between IU and the social perception of the robot in terms of
warmth, competence, and discomfort [128]. While unpredictability and uncertainty are often used
interchangeably, suggesting conceptual similarity [56], our results suggest that greater caution is
warranted when doing so. In the current study, we dened and manipulated robot predictability
in terms of making it more or less dicult to learn to predict the robot’s behaviour. However, the
questionnaires on IU are more about uncertainty regarding events further into the future than the
robot’s next action(s) would be. The children knew that they were going to interact with “Zeno the
robot” at a certain point during the day, that they would then play several games with the robot,
and then go back to class—a xed structure. In that sense, the children could predict the robot’s
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
Predictable Robots for Autistic Children 36:29
behaviour on a more abstract level (i.e., the robot will perform actions related to the games), but not
on precisely when, where, and how those robot actions would appear. Possibly, IU relates to an in-
tolerance for situations without a clear structure on what is going to happen in a longer time span,
resulting in the unpredictability of future events. Predictability, as we dened and manipulated it,
related more to predicting sensory information in the immediate future. This distinction may also
be relevant when considering the role of predictability and uncertainty in the non-social features
of ASC, in particular in insistence of sameness (e.g., inexible adherence to routines, or ritualised
patterns) given that both IU and insistence of sameness refer to liking things to be predictable and
a dislike of change [21].
To our knowledge, this is the rst study that has operationalised and manipulated predictabil-
ity in a real-world setting and showed how the extent of unpredictability can be quantied. The
metrics in our manipulation check can be used to compare the degree of unpredictability between
studies using robots. In general, robotic technology is uniquely position for investigating predict-
ability in autistic children, as they allow us to carefully manipulate its predictability (unlike with
humans), but they also elicit social interactions in autistic children. There is some preliminary
work in trying to teach autistic children to deal with unpredictability [e.g., 58,115]. To this end,
robots may be particularly useful in that its unpredictability can be carefully increased in both in-
tensity as well as in predicting dierent aspects of the environment (e.g., predicting robot actions,
or predicting future events).
6.1 Limitations
In our study, we manipulated the predictability of the robot’s behaviour through its variance. This
ought to have made it more dicult to learn to predict the robot’s behaviour. Through our manip-
ulation check, we concluded that this manipulation was successful. However, possibly even in the
high-variance condition, the robot’s behaviour was not problematically unpredictable, as is the un-
predictability of human behaviour for example. In practice, robots can be more unpredictable than
we could manipulate in our study, as we needed to keep the conditions comparable and avoid con-
founding factors. In the future, robots will become more sophisticated and capable of levels closer
to humanlike behaviour. In turn, so too does their ability to be unpredictable. Thus, robot predict-
ability could aect the engagement of autistic children more strongly when they become more
sophisticated and humanlike. In our study, we looked at the ability to predict the robot’s actions,
which is not the same as the extent which one perceives the robot to be predictable (attributed predict-
ability) [128]. Even though a robot’s behaviour is more dicult to predict, it can still be considered
to be more predictable [41,128]. Current measures for attributed predictability are too complex to
be used for autistic children in our study, as they rely heavily on language and intellectual abil-
ity. When new measures become available, it would be interesting to assess the autistic children’s
attributed predictability in relation to dierent levels of variance in the robot’s behaviour.
Note that there are several sources of variance that could not be controlled fully in our study.
First, there is inherent variation in robot motion due to limits in reproducibility of motion by the
robot’s stepper motors, which may dier slightly with each operation. Second, we were not able
to control for the aective quality and the attractiveness for engagement of each specic variant
we chose, despite the fact that we selected ones with similar verbal and/or motion qualities. For
example, children may have experienced more negative aect when the robot used the specic
term “Time for dancing”, which was present only in the high-variance condition. Additionally, for
some sessions, technical diculties resulted in unintended behaviours and form of behaviours,
introducing some variability. These instances were not annotated, but the eects could possibly
carry over to later in the session.
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
36:30 B. R. Schadenberg et al.
Finally, we also note that, as with most studies that concern robots and autistic children (see
[12,38]), our sample size was relatively low. This can negatively inuence the generalisability of
our results and prevents us from drawing strong conclusions. It is therefore important to take the
reported margins of error into account when interpreting our results.
7 CONCLUSION
Predictability is important to autistic individuals, and robots have been suggested to meet this
need as they can be programmed to be predictable, as well as to elicit social interaction. However,
little was known about the interplay between robot predictability, engagement in learning, and the
individual dierences between autistic children. Here, we systematically manipulated the robot’s
predictability, and measured the behavioural and visual attention of the autistic children. Addi-
tionally, we also measured several individual factors, including the children’s autistic features,
expressive language ability, and IU. We found that the children will continue engaging in the
activity behaviourally, but start to pay less visual attention over time to activity-relevant loca-
tions when the robot is less predictable. Instead, they increasingly start to look away from the
activity. Ultimately, this could negatively inuence learning. In particular for tasks with a visual
component, where paying less visual attention leads to fewer opportunities for learning. Further-
more, we found that the severity of autistic features and expressive language ability had a sig-
nicant impact on behavioural engagement. This nding is relevant for robots used for diagnos-
ing autism and raises the question whether individual dierences only directly inuence behavi-
oural engagement or whether they moderate the eect of a robot on the children’s behavioural
engagement.
We consider our results as preliminary evidence that robot predictability is an important
factor for keeping children in a state where learning can occur. In particular, in long-term in-
teractions with many sessions, our results indicate that the trend of paying less visual atten-
tion increases over time. As individual dierences between autistic children were shown to
have a signicant impact on behavioural engagement, future studies should therefore be care-
ful to account for these dierences. Finally, our study indicates that predictability can be stud-
ied in real-life scenarios with real stimuli, rather than articial stimuli in lab settings. We also
showed how the degree of predictability can be quantied in a way that it can be used as a
manipulation check to display the degree of unpredictability between conditions in a real-world
setting.
Once the engagement of the children with the robot has been further claried, future research
should consider looking into the aective component of engagement to investigate whether a
more predictable robot is more enjoyable to the children. Additionally, future research is needed
to determine whether “higher quality” engagement also leads to more and faster learning. After
all, increased engagement alone is insucient to justify robots to assist in interventions for aut-
istic children when it does not also lead to increased learning. In our study, we opted for a
holistic approach to increasing the robot’s unpredictability by implementing several types of
variance. Future research should examine to what extent each types of variance inuences aut-
istic children. This would allow us to more carefully take the robot’s predictability into account
when designing its behaviour. Lastly, the goal of our study was to investigate how we should
design robots to best aid autistic children in learning, specically focusing on their need for pre-
dictable environments whilst taking the children’s idiosyncrasies into account. Future research
could also investigate the role of a robot’s predictability in engaging typically-developing chil-
dren in learning, in order to assess whether or not predictability is uniquely important to autistic
children.
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
Predictable Robots for Autistic Children 36:31
APPENDICES
A EXAMPLES OF THE VARIATIONS OF THE ROBOT’S BEHAVIOURS
Table 6. Some Examples of the Variations for Certain Robot Actions
Description Variant one
(Default)
Variant two Variant three Variant four
Introduction actions
Greeting “Hi, my name is
Zeno.” + wave
with right arm
“Hello. I am
Zeno the robot.”
+ wave with left
arm
“I am Zeno.
Hello!” + wave
with right arm
“Hi, I am Zeno. +
wave with left
arm”
Game actions
Prompt eyes “Find eyes.” “Choosing eyes.” “Where are
eyes?”
“Find my eyes.”
Resolve eyes “You found my
eyes!” + eyes
open then close
“You found
them!” + eyes
left then right
“That’s right” +
eyes open then
close
“You found...
eyes!” + eyes left
then right
Generic actions
Praise “Good job!” “Well done!” “Good working!” “Excellent!”
Unsure “I don’t know” “Sorry, don’t
know”
“Maybe?” “Not sure”
Topic variant actions
Low-variance condition High-variance condition
Description Triggering event Action Triggering event Action
Respond to noise There was noise “What’s that
noise?”
There was no
noise
“What’s that
noise?”
Respond to
child’s presence
Child is out of
view of the robot
“Where are
you?”
Child is sitting
in front of the
robot
“Where are
you?”
Respond to WiFi
connection lost
- - Robot loses WiFi
connection
(unobservable
internal event)
“Reconnecting.
Reconnecting.
Success.”
In the low-variance condition, the robot would only perform the default variant of an action. In the high-variance
condition, the robot would select one of the four variants of an action. For the topic variant actions, there is only one
variant. These actions were triggered by the wizard, who made sure that the actions were a response to an observable
event in the low-variance condition. In the high-variance condition, there was no observable event that could explain
the action. There were also more types of topic variant actions in the high-variance condition, as some of those actions
were plausible responses to internal events. The child had no way of observing these internal events, thus they only
occurred in the high-variance condition.
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
36:32 B. R. Schadenberg et al.
B FLOW DIAGRAM OF THE DE-ENIGMA GAMES
Fig. 7. Flow diagram of the four DE-ENIGMA games. First, the robot or adult would prompt the child with
a question or to explore the options that were shown on the tablet. The child could then select one of the
options, which was highlighted with the green square, or one of the four images in the Yellow Game. Aer
selecting an option, the highlights on the tablet are removed. The robot then responded to the child’s option
by labelling the chosen option and moving the respective facial feature, or displaying the facial expression.
In case of the Blue and Yellow Game, the robot would also evaluate the child’s answer and provide posit-
ive feedback. Each of the sequences shown were repeated four times for each game (with dierent facial
features/facial expressions).
C IUSC-S QUESTIONNAIRE
The IUSC-S questionniare can be seen in Table 7. To assess how many dimensions of intolerance
of uncertainty the IUSC-S measured, we conducted a Principle Component Analysis (PCA) on 15
items with varimax rotation. The Kaiser-Meyer-Olkin (KMO) measure veried the sampling ad-
equacy for the analysis (KMO = .65), and all KMO values for individual items were >.55. Given our
sample size of 24, this is sucient, but mediocre. Bartlett’s test of sphericity indicated that correla-
tions between items were suciently large for PCA (χ2(105) = 312.61, p<.001). Two components
had eigenvalues above Kaiser’s criterion of 1 and together explained 70.85% of the variance. Table 8
shows the component loadings for each item, where we highlighted items with loadings of .72 or
higher and with a cross loading dierence greater than .2, based on recommendations from Stevens
[136]. Based on these component loadings, we exclude item 2, 7, and 11 for computing the IUSC-S
scores.
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
Predictable Robots for Autistic Children 36:33
Table 7. The IUSC-S, Parent-Report form, Adapted from the Complementary Study
Questions
1. Uncertainty makes my child’s life intolerable.
2. My child’s mind cannot be relaxed if he/she does not know what will happen tomorrow.
3. Uncertainty makes my child uneasy, anxious, or stressed.
4. Unforeseen events upset my child greatly.
5. It frustrates my child to not have all the information he/she needs in a situation.
6. Uncertainty keeps my child from living a full life.
7. When it’s time to act, uncertainty paralyzes my child.
8. When my child is uncertain he/she cannot function very well.
9. Other children seem to be more certain than my child.
10. Uncertainty makes my child unhappy or sad.
11. My child always wants to know what the future has in store for him/her.
12. My child cannot stand being taken by surprise.
13. Uncertainty keeps my child from sleeping soundly.
14. My child tries to get away from all uncertain situations.
15. The ambiguities of life stress my child.
Table 8. The Component Loadings for the Items on
The IU as Measured by the IUSC-S
Item Component 1 Component 2
Item 6 .90 −.12
Item 4 .87 .20
Item 15 .87 −.09
Item 3 .86 −.25
Item 1 .86 −.12
Item 12 .84 .02
Item 10 .84 −.23
Item 5 .79 −.27
Item 14 .78 .27
Item 13 .77 .17
Item 9 .74 −.17
Item 8 .73 .11
Item 7 .70 .55
Item 11 .58 .58
Item 2 .48 −.63
Eigenvalue 9.16 1.47
% of variance 61.08 9.77
Cronbach’s α.96 .34
The component loadings that meet the recommendations
of Stevens [137] are in bold.
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
36:34 B. R. Schadenberg et al.
D CONFUSION MATRICES FOR BEHAVIOURAL ENGAGEMENT AND
VISUAL AT TENTION
Table 9. Confusion Matrix of the Annotations for
Behavioural Engagement between the Main Coder
and Secondary Coder
Secondary coder
Main coder −2−10 1 2 Total
−2188 47 28 14 4 281
−16105 53 15 1 180
0720706 60 2 795
191779796 57 958
224360198 267
Total 212 193 869 945 262 2,481
The number of annotations that both coders annotated similarly
are in bold.
Table 10. Confusion Matrix of the Primary Annotations for Visual Aention between
the Main Coder and Secondary Coder
Secondary coder
Main coder Robot Tablet Teaching Mats. Adult Assistant Elsewhere Mixed Total
Robot 1,599 23 10 16 0 60 27 1,735
Tablet 19 1,070 5702412
1,137
Teaching Mats. 72 534 50 13 4 565
Adult 12 8 6 373 02819
446
Assistant 10 0 010 24
17
Elsewhere 28 40 11 16 0 904 19 1,018
Mixed 11 7 5 9 0 25 55 112
Total 1,677 1,150 571 426 10 1,056 140 5,030
The number of annotations that both coders annotated similarly are in bold.
Table 11. Confusion Matrix of the Secondary Annotations for Visual Aention between the
Main Coder and Secondary Coder
Secondary coder
Main coder Robot Tablet Teaching Mats. Adult Assistant Elsewhere Mixed Total
Robot 305 16 7 13 1 23 16 381
Tablet 19 122 16076
161
Teaching Mats. 10 1 60 40 5 3 83
Adult 13 7 3 118 07 9
157
Assistant 00 0 0700
7
Elsewhere 17 8 4 11 0 119 10 169
Mixed 64 0 61 3 52 72
Total 370 158 75 158 9 164 96 1,030
The number of annotations that both coders annotated similarly are in bold.
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
Predictable Robots for Autistic Children 36:35
ACKNOWLEDGMENTS
The authors and the DE-ENIGMA project are extremely grateful to the school that hosted this re-
search, and the sta members, children, and families who generously gave their time for the study.
We would also like to thank Meike Berkho, Hannah Viner, Audrey McMillion, Mutluhan Ersoy,
Manual Milling, Alice Baird, Mihai Zanr, Elisabeta Oneata, Vesna Petrović, Bridgette Connell,
Lynn Packwood, and Cristina Fernández Álvarez de Eulate for their help with this study.
REFERENCES
[1] Alyssa M. Alcorn, Eloise Ainger, Vicky Charisi, Stefania Mantinioti, Sunčica Petrović, Bob R. Schadenberg, Teresa
Tavassoli, and Elizabeth Pellicano. 2019. Educators’ views on using humanoid robots with autistic learners in special
education settings in england. Frontiers in Robotics and AI 6, November (Nov. 2019), 1–15. DOI:https://doi.org/10.
3389/frobt.2019.00107
[2] American Psychiatric Association. 2000. Diagnostic and Statistical Manual of Mental Disorders (4th ed.). American
Psychiatric Association, Arlington, VA. DOI:https://doi.org/10.1176/appi.books.9780890423349
[3] American Psychiatric Association. 2013. Diagnostic and Statistical Manual of Mental Disorders (5th e d.). Author, Wash-
ington, DC. DOI:https://doi.org/10.1176/appi.books.9780890425596
[4] Salvatore M. Anzalone, Soane Boucenna, Serena Ivaldi, and Mohamed Chetouani. 2015. Evaluating the engagement
with social robots. International Journal of Social Robotics 7, 4 (Aug. 2015), 465–478. DOI:https://doi.org/10.1007/
s12369-015-0298-7
[5] Roger Azevedo. 2015. Dening and measuring engagement and learning in science: Conceptual, theoretical, method-
ological, and analytical issues. Educational Psychologist 50, 1 (Jan. 2015), 84–94. DOI:https://doi.org/10.1080/00461520.
2015.1004069
[6] Joshua H. Balsters, Matthew A. J. Apps, Dimitris Bolis, Rea Lehner, Louise Gallagher, and Nicole Wenderoth. 2017.
Disrupted prediction errors index social decits in autism spectrum disorder. Brain 140, 1 (Jan. 2017), 235–246.
DOI:https://doi.org/10.1093/brain/aww287
[7] Moshe Bar. 2007. The proactive brain: Using analogies and associations to generate predictions. Trends in Cognitive
Sciences 11, 7 (Jul. 2007), 280–289. DOI:https://doi.org/10.1016/j.tics.2007.05.005
[8] Simon Baron-Cohen. 2002. The extreme male brain theory of autism. Trends in Cognitive Sciences 6, 6 (Jun. 2002),
248–254. DOI:https://doi.org/10.1016/S1364-6613(02)01904-6
[9] Simon Baron-Cohen. 2009. Autism: The empathizing-systemizing (E-S) theory. Annals of the New York Academy of
Sciences 1156, 1 (2009), 68–80. DOI:https://doi.org/10.1111/j.1749-6632.2009.04467.x
[10] Jackson Beatty. 1982. Task-evoked pupillary responses, processing load, and the structure of processing resources.
Psychological Bulletin 91, 2 (1982), 276–292. DOI:https://doi.org/10.1037/0033-2909.91.2.276
[11] Jackson Beatty and Brennis Lucero-Wagoner. 2000. The pupillary system. In Proceedings of the Handbook of Psycho-
physiology (2nd ed.), John T. Cacioppo, Louius G. Tassinary, and Gary G. Berntson (Eds.). Cambridge University
Press, New York, NY, 142–162.
[12] Momotaz Begum, Richard W. Serna, and Holly A. Yanco. 2016. Are robots ready to deliver autism interventions? A
comprehensive review. International Journal of Social Robotics 8, 2 (Mar. 2016), 157–181. DOI:https://doi.org/10.1007/
s12369-016-0346-y
[13] Howard Berenbaum, Keith Bredemeier, and Renee J. Thompson. 2008. Intolerance of uncertainty: Exploring its di-
mensionality and associations with need for cognitive closure, psychopathology, and personality. Journal of Anxiety
Disorders 22, 1 (Jan. 2008), 117–125. DOI:https://doi.org/10.1016/j.janxdis.2007.01.004
[14] Timothy W. Bickmore, Daniel Schulman, and Langxuan Yin. 2010. Maintaining engagement in long-term interven-
tions with relational agents. Applied Articial Intelligence 24, 6 ( Jul. 2010), 648–666. DOI:https://doi.org/10.1080/
08839514.2010.492259
[15] Christina Boulter, Mark H. Freeston, Mikle South, and Jacqui Rodgers. 2014. Intolerance of uncertainty as a frame-
work for understanding anxiety in children and adolescents with autism spectrum disorders. Journal of Autism and
Developmental Disorders 44, 6 (2014), 1391–1402. DOI:https://doi.org/10.1007/s10803-013-2001-x
[16] Margaret M. Bradley, Laura Miccoli, Miguel A. Escrig, and Peter J. Lang. 2008. The pupil as a measure of emotional
arousal and autonomic activation. Psychophysiology 45, 4 (Jul. 2008), 602–607. DOI:https://doi.org/10.1111/j.1469-
8986.2008.00654.x
[17] Ricarda Braukmann, Emma Ward, Roy S. Hessels, Harold Bekkering, Jan K. Buitelaar, and Sabine Hunnius. 2018.
Action prediction in 10-month-old infants at high and low familial risk for Autism Spectrum Disorder. Research in
Autism Spectrum Disorders 49, May 2017 (May 2018), 34–46. DOI:https://doi.org/10.1016/j.rasd.2018.02.004
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
36:36 B. R. Schadenberg et al.
[18] Linley C. Bryan and David L. Gast. 2000. Teaching on-task and on-schedule behaviors to high-functioning children
with autism via picture activity schedules. Journal of Autism and Developmental Disorders 30, 6 (2000), 553–567.
DOI:https://doi.org/10.1023/A:1005687310346
[19] John-John Cabibihan, Hifza Javed, Marcelo Ang, and Sharifah Mariam Aljunied. 2013. Why robots? A survey on the
roles and benets of social robots in the therapy of children with autism. International Journal of Social Robotics 5, 4
(Nov. 2013), 593–618. DOI:https://doi.org/10.1007/s12369-013-0202-2
[20] R. Nicholas Carleton. 2016. Into the unknown: A review and synthesis of contemporary models involving uncertainty.
Journal of Anxiety Disorders 39 (Apr. 2016), 30–43. DOI:https://doi.org/10.1016/j.janxdis.2016.02.007
[21] Paul D. Chamberlain, Jacqui Rodgers, Michael J. Crowley, Sarah E. White, Mark H. Freeston, and Mikle South. 2013.
A potentiated startle study of uncertainty and contextual anxiety in adolescents diagnosed with autism spectrum
disorder. Molecular Autism 4, 1 (2013), 31. DOI:https://doi.org/10.1186/2040-2392-4- 31
[22] Caitlyn Clabaugh, Kartik Mahajan, Shomik Jain, Roxanna Pakkar, David Becerra, Zhonghao Shi, Eric C. Deng, Rhi-
anna Lee, Gisele Ragusa, and Maja J. Matarić. 2019. Long-term personalization of an in-home socially assistive
robot for children with autism spectrum disorders. Frontiers in Robotics and AI 6, November (Nov. 2019), 1–18.
DOI:https://doi.org/10.3389/frobt.2019.00110
[23] Caitlyn Clabaugh and Maja J. Matarić. 2019. Escaping Oz: Autonomy in socially assistive robotics. Annual Review
of Control, Robotics, and Autonomous Systems 2, 1 (May 2019), 33–61. DOI:https://doi.org/10.1146/annurev-control-
060117-104911
[24] Andy Clark. 2013. Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral
and Brain Sciences 36, 3 (Jun. 2013), 181–204. DOI:https://doi.org/10.1017/S0140525X12000477
[25] Olivier Collignon, Geneviève Charbonneau, Frédéric Peters, Marouane Nassim, Maryse Lassonde, Franco Lepore,
Laurent Mottron, and Armando Bertone. 2013. Reduced multisensory facilitation in persons with autism. Cortex 49,
6 (Jun. 2013), 1704–1710. DOI:https://doi.org/10.1016/j.cortex.2012.06.001
[26] Jonathan S. Comer, Amy K. Roy, Jami M. Furr, Kristin Gotimer, Rinad S. Beidas, Michel J. Dugas, and Philip C. Kendall.
2009. The intolerance of uncertainty scale for children: A psychometric evaluation. Psychological Assessment 21, 3
(2009), 402–411. DOI:https://doi.org/10.1037/a0016719
[27] Alexandre Coninx, Paul E. Baxter, Elettra Oleari, Sara Bellini, Bert P.B. Bierman, Olivier A. Blanson Henkemans, Lola
Cañamero, Piero Cosi, Valentin Enescu, Raquel Ros Espinoza, Antoine Hiolle, Rémi Humbert, Bernd Kiefer, Ivana
Kruij-Korbayová, Rosemarijn Looije, Marco Mosconi, Mark A. Neerincx, Giulio Paci, Georgios Patsis, Clara Pozzi,
Francesca Sacchitelli, Hichem Sahli, Alberto Sanna, Giacomo Sommavilla, Fabio Tesser, Yiannis Demiris, and Tony
Belpaeme. 2015. Towards long-term social child-robot interaction: Using multi-activity switching to engage young
users. Journal of Human–Robot Interaction 5, 1 (Aug. 2015), 32–64. DOI:https://doi.org/10.5898/JHRI.5.1.Coninx
[28] James P. Connell and James G. Wellborn. 1991. Competence, autonomy, and relatedness: A motivational analysis
of self-system processes. In Self processes and Development. Lawrence Erlbaum Associates, Inc, Hillsdale, NJ, 43–77.
Retrieved from https://psycnet.apa.org/record/1991-97029-002.
[29] Lee J. Corrigan, Christopher Peters, Dennis Küster, and Ginevra Castellano. 2016. Engagement perception and gen-
eration for social robots and virtual agents. In Toward Robotic Socially Believable Behaving Systems. Anna Esposito
and Lakhmi C. Jain (Eds.), Intelligent Systems Reference Library, Vol. 105, Springer, International Publishing, Cham,
29–51. DOI:https://doi.org/10.1007/978-3-319-31056- 5_4
[30] Cristina A. Costescu, Bram Vanderborght, and Daniel O. David. 2015. Reversal learning task in children with autism
spectrum disorder: A robot-based approach. Journal of Autism and Developmental Disorders 45, 11 (2015), 3715–3725.
DOI:https://doi.org/10.1007/s10803-014-2319-z
[31] Sander Van de Cruys, Kris Evers, Ruth Van der Hallen, Lien Van Eylen, Bart Boets, Lee De-Wit, and Johan Wagemans.
2014. Precise minds in uncertain worlds: Predictive coding in autism. Psychological Review 121, 4 (2014), 649–675.
DOI:https://doi.org/10.1037/a0037665
[32] Kerstin Dautenhahn. 1999. Robots as social actors: Aurora and the case of autism. In Proceedings of the 3rd Cognitive
Technology Conference (CT’99). Vol. 359. M.I.N.D. Lab, San Fransisco, CA, 359–374. Retrieved from https://books.
google.nl/books?id=QM-- tgAACAAJ.
[33] Kerstin Dautenhahn. 2007. Socially intelligent robots: Dimensions of human–robot interaction. Philosophical Trans-
actions of the Royal Society B: Biological Sciences 362, 1480 (Apr. 2007), 679–704. DOI:https://doi.org/10.1098/rstb.2006.
2004
[34] Kerstin Dautenhahn and Iain Werry. 2004. Towards interactive robots in autism therapy: Background, motivation
and challenges. Pragmatics & Cognition 12, 1 (2004), 1–35. DOI:https://doi.org/10.1075/pc.12.1.03dau
[35] Daniel O. David, Cristina A. Costescu, Silviu-Andrei Matu, Aurora Szentagotai, and Anca Dobrean. 2020. Eects of a
robot-enhanced intervention for children with ASD on teaching turn-taking skills. Journal of Educational Computing
Research 58, 1 (Mar. 2020), 29–62. DOI:https://doi.org/10.1177/0735633119830344
[36] Lorenzo Desideri, Marco Negrini, Massimiliano Malavasi, Daniela Tanzini, Aziz Rouame, Maria Cristina Cutrone,
Paola Bonifacci, and Evert-Jan Hoogerwerf. 2018. Using a humanoid robot as a complement to interventions for
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
Predictable Robots for Autistic Children 36:37
children with autism spectrum disorder: A pilot study. Advances in Neurodevelopmental Disorders 2, 3 (2018), 273–
285. DOI:https://doi.org/10.1007/s41252-018-0066-4
[37] Joshua J. Diehl, Charles R. Crowell, Michael Villano, Kristin Wier, Karen Tang, and Laurel D. Riek. 2014. Clin-
ical applications of robots in autism spectrum disorder diagnosis and treatment. In Comprehensive Guide to Aut-
ism. Vinood B. Patel, Victor R. Preedy, and Colin R. Martin (Eds.). Springer, New York, NY, 411–422. DOI:https:
//doi.org/10.1007/978-1-4614-4788- 7_14
[38] Joshua J. Diehl, Lauren M. Schmitt, Michael Villano, and Charles R. Crowell. 2012. The clinical use of robots for
individuals with autism spectrum disorders: A critical review. Research in Autism Spectrum Disorders 6, 1 (Jan. 2012),
249–262. DOI:https://doi.org/10.1016/j.rasd.2011.05.006
[39] Lucy Diep, John-John Cabibihan, and Gregor Wolbring. 2015. Social Robots: Views of special education teachers. In
Proceedings of the 3rd 2015 Workshop on ICTs for improving Patients Rehabilitation Research Techniques.ACM,New
York, NY, 160–163. DOI:https://doi.org/10.1145/2838944.2838983
[40] Anca Dragan and Siddhartha Srinivasa. 2014. Familiarization to robot motion. In Proceedings of the 2014 ACM/IEEE
International Conference on Human–Robot Interaction. ACM, New York, NY, 366–373. DOI:https://doi.org/10.1145/
2559636.2559674
[41] Katherine Driggs-Campbell and Ruzena Bajcsy. 2016. Communicating intent on the road through human-inspired
control schemes. In 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems. IEEE, 3042–3047.
DOI:https://doi.org/10.1109/IROS.2016.7759471
[42] Michel J. Dugas, Kristin Buhr, and Robert Ladouceur. 2004. The role of intolerance of uncertainty in etiology and
maintenance. In Generalized Anxiety Disorder: Advances in Research and Practice. R. G. Heimberg, C. L. Turk, & D. S.
Mennin (Eds.), Guilford Press, New York, NY143–163.
[43] Audrey Duquette, François Michaud, and Henri Mercier. 2008. Exploring the use of a mobile robot as an imitation
agent with children with low-functioning autism. Autonomous Robots 24, 2 (Feb. 2008), 147–157. DOI:https://doi.org/
10.1007/s10514-007-9056-5
[44] Danielle A. Einstein. 2014. Extension of the transdiagnostic model to focus on intolerance of uncertainty: A review
of the literature and implications for treatment. Clinical Psychology: Science and Practice 21, 3 (Sep. 2014), 280–300.
DOI:https://doi.org/10.1111/cpsp.12077
[45] Louise Ewing, Elizabeth Pellicano, and Gillian Rhodes. 2013. Atypical updating of face representations with experi-
ence in children with autism. Developmental Science 16, 1 (Jan. 2013), 116–123. DOI:https://doi.org/10.1111/desc.12007
[46] David J. Feil-Seifer and Maja J. Matarić. 2009. Toward socially assistive robotics for augmenting interventions for
children with autism spectrum disorders. In Experimental Robotics,O.Khatib,V.Kumar,andG.J.Pappas(Eds.),
Vol. 54. Springer, Berlin Heidelberg, 201–210. DOI:https://doi.org/10.1007/978-3-642-00196- 3_24
[47] Cindy Ferrara and Suzanne D. Hill. 1980. The responsiveness of autistic children to the predictability of social
and nonsocial toys. Journal of Autism and Developmental Disorders 10, 1 (1980), 51–57. DOI:https://doi.org/10.1007/
BF02408432
[48] Jennifer A. Fredricks, Phyllis C. Blumenfeld, and Alison H. Paris. 2004. School engagement: Potential of the
concept, state of the evidence. Review of Educational Research 74, 1 (Mar. 2004), 59–109. DOI:https://doi.org/10.3102/
00346543074001059
[49] Karl J. Friston. 2005. A theor y of cortical responses. Philosophical Transactions of the Royal Society B: Biological Sciences
360, 1456 (Apr. 2005), 815–836. DOI:https://doi.org/10.1098/rstb.2005.1622
[50] Morgan Frost-Karlsson, Martyna Alexandra Galazka, Christopher Gillberg, Carina Gillberg, Carmela Miniscalco, Eva
Billstedt, Nouchine Hadjikhani, and Jakob Åsberg Johnels. 2019. Social scene perception in autism spectrum disorder:
An eye-tracking and pupillometric study. Journal of Clinical and Experimental Neuropsychology 41, 10 (Nov. 2019),
1024–1032. DOI:https://doi.org/10.1080/13803395.2019.1646214
[51] Eleanor J. Gibson and Anne D. Pick. 2000. Perceptual Learning and Development: An Ecological Approach.Oxford
University Press, New York, NY.
[52] Lakshmi J. Gogate and George Hollich. 2010. Invariance detection within an interactive system: A perceptual gateway
to language development. Psychological Review 117, 2 (2010), 496–516. DOI:https://doi.org/10.1037/a0019049
[53] Judith Goris, Senne Braem, Annabel D. Nijhof, Davide Rigoni, Eliane Deschrijver, Sander Van de Cruys, Jan R. Wi-
ersema, and Marcel Brass. 2018. Sensory prediction errors are less modulated by global context in autism spec-
trum disorder. Biological Psychiatry: Cognitive Neuroscience and Neuroimaging 3, 8 (Aug. 2018), 667–674. DOI:https:
//doi.org/10.1016/j.bpsc.2018.02.003
[54] Judith Goris, Marcel Brass, Charlotte Cambier, Jeroen Delplanque, Jan R. Wiersema, and Senne Braem. 2020. The
relation between preference for predictability and autistic traits. Autism Research 13, 7 (Jul. 2020), 1144–1154.
DOI:https://doi.org/10.1002/aur.2244
[55] Charles R. Greenwood. 1991. Longitudinal analysis of time, engagement, and achievement in at-risk versus non-risk
students. Exceptional Children 57, 6 (May 1991), 521–535. DOI:https://doi.org/10.1177/001440299105700606
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
36:38 B. R. Schadenberg et al.
[56] Dan W. Grupe and Jack B. Nitschke. 2013. Uncertainty and anticipation in anxiety: An integrated neurobiological
and psychological perspective. Nature Reviews Neuroscience 14, 7 (Jul. 2013), 488–501. DOI:https://doi.org/10.1038/
nrn3524
[57] Rebecca Grzadzinski, Marisela Huerta, and Catherine Lord. 2013. DSM-5 and autism spectrum disorders (ASDs): An
opportunity for identifying ASD subtypes. Molecular Autism 4, 1 (2013), 12. DOI:https://doi.org/10.1186/2040-2392-
4-12
[58] Victoria Hallett, Joanne Mueller, Lauren Breese, Megan Hollett, Bryony Beresford, Annie Irvine, Andrew Pickles,
Vicky Slonims, Stephen Scott, Tony Charman, and Emily Simono. 2021. Introducing ‘predictive parenting’: A
feasibility study of a new group parenting intervention targeting emotional and behavioral diculties in chil-
dren with autism spectrum disorder. Journal of Autism and Developmental Disorders 51, 1 (Jan. 2021), 323–333.
DOI:https://doi.org/10.1007/s10803-020-04442-2
[59] Francesca Happé and Uta Frith. 2006. The weak coherence account: Detail-focused cognitive style in autism spectrum
disorders. Journal of Autism and Developmental Disorders 36, 1 (Jan 2006), 5–25. DOI:https://doi.org/10.1007/s10803-
005-0039-0
[60] Francesca Happé, Angelica Ronald, and Robert Plomin. 2006. Time to give up on a single explanation for autism.
Nature Neuroscience 9, 10 (Oct. 2006), 1218–1220. DOI:https://doi.org/10.1038/nn1770
[61] Annika Hellendoorn, Lex Wijnroks, and Paul P. M. Leseman. 2015. Unraveling the nature of autism: Finding order
amid change. Frontiers in Psychology 6 (2015), 359. DOI:https://doi.org/10.3389/fpsyg.2015.00359
[62] E. H. Hess and J. M. Polt. 1964. Pupil size in relation to mental activity during simple problem-solving. Science 143,
3611 (Mar. 1964), 1190–1192. DOI:https://doi.org/10.1126/science.143.3611.1190
[63] Jakob Hohwy. 2013. The Predictive Mind. Oxford University Press, Oxford. DOI:https://doi.org/10.1093/acprof:oso/
9780199682737.001.0001
[64] Ryan Y. Hong and Mike W.L. Cheung. 2015. The Structure of cognitive vulnerabilities to depression and anxiety.
Clinical Psychological Science 3, 6 (Nov. 2015), 892–912. DOI:https://doi.org/10.1177/2167702614553789
[65] Claire A. G. J. Huijnen, Monique A. S. Lexis, Rianne Jansens, and Luc P. de Witte. 2017. How to implement robots in
interventions for children with autism? A co-creation study involving people with autism, parents and professionals.
Journal of Autism and Developmental Disorders 47, 10 (Oct. 2017), 3079–3096. DOI:https://doi.org/10.1007/s10803-017-
3235-9
[66] Claire A. G. J. Huijnen, Monique A. S. Lexis, Rianne Jansens, and Luc P. de Witte. 2019. Roles, strengths and chal-
lenges of using robots in interventions for children with autism spectrum disorder (ASD). Journal of Autism and
Developmental Disorders 49, 1 (Jan. 2019), 11–21. DOI:https://doi.org/10.1007/s10803-018-3683-x
[67] Bibi E. B. M. Huskens, Rianne Verschuur, Jan C. C. Gillesen, Robert Didden, and Emilia I. Barakova. 2013. Pro-
moting question-asking in school-aged children with autism spectrum disorders: Eectiveness of a robot interven-
tion compared to a human-trainer intervention. Developmental Neurorehabilitation 16, 5 (2013), 345–356. DOI:https:
//doi.org/10.3109/17518423.2012.739212
[68] Shomik Jain, Balasubramanian Thiagarajan, Zhonghao Shi, Caitlyn Clabaugh, and Maja J. Matarić. 2020. Modeling
engagement in long-term, in-home socially assistive robot interventions for children with autism spectrum disorders.
Science Robotics 5, 39 (Feb. 2020), eaaz3791. DOI:https://doi.org/10.1126/scirobotics.aaz3791
[69] Hifza Javed, Rachael Burns, Myounghoon Jeon, Ayanna M. Howard, and Chung Hyuk Park. 2019. A robotic frame-
work to facilitate sensory experiences for children with autism spectrum disorder. ACM Transactions on Human–
Robot Interaction 9, 1 (Dec. 2019), 1–26. DOI:https://doi.org/10.1145/3359613
[70] Hifza Javed, WonHyong Lee, and Chung Hyuk Park. 2020. Toward an automated measure of social engagement for
children with autism spectrum disorder—a personalized computational modeling approach. Frontiers in Robotics and
AI 7, April (Apr. 2020), 14. DOI:https://doi.org/10.3389/frobt.2020.00043
[71] D. Kahneman and J. Beatty. 1966. Pupil diameter and load on memory. Science 154, 3756 (Dec. 1966), 1583–1585.
DOI:https://doi.org/10.1126/science.154.3756.1583
[72] Deb Keen. 2009. Engagement of children with autism in learning. Australasian Journal of Special Education 33, 2
(2009), 130–140. DOI:https://doi.org/10.1375/ajse.33.2.130
[73] C.D. Kidd and Cynthia Breazeal. 2008. Robots at home: Understanding long-term human–robot interaction. In Pro-
ceedings of the 2008 IEEE/RSJ International Conference on Intelligent Robots and Systems. IEEE, 3230–3235. DOI:https:
//doi.org/10.1109/IROS.2008.4651113
[74] Elizabeth Kim, Rhea Paul, Frederick Shic, and Brian Scassellati. 2012. Bridging the research Gap: Making HRI useful
to individuals with autism. Journal of Human–Robot Interaction 1, 1 (Aug. 2012), 26–54. DOI:https://doi.org/10.5898/
JHRI.1.1.Kim
[75] Elizabeth S. Kim, Lauren D. Berkovits, Emily P. Bernier, Dan Leyzberg, Frederick Shic, Rhea Paul, and Brian Scassel-
lati. 2013. Social robots as embedded reinforcers of social behavior in children with autism. Journal of Autism and
Developmental Disorders 43, 5 (May 2013), 1038–1049. DOI:https://doi.org/10.1007/s10803-012-1645-2
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
Predictable Robots for Autistic Children 36:39
[76] Elizabeth S. Kim, Christopher M. Daniell, Corinne Makar, Julia Elia, Brian Scassellati, and Frederick Shic. 2015. Po-
tential clinical impact of positive aect in robot interactions for autism intervention. In Proceedings of the 2015
International Conference on Aective Computing and Intelligent Interaction. IEEE, 8–13. DOI:https://doi.org/10.1109/
ACII.2015.7344544
[77] Jacqueline Kory and Cynthia Breazeal. 2014. Storytelling with robots: Learning companions for preschool children’s
language development. In Proceedings of the 23rd IEEE International Symposium on Robot and Human Interactive
Communication. IEEE, 643–648. DOI:https://doi.org/10.1109/ROMAN.2014.6926325
[78] Viviane Kostrubiec and Jeanne Kruck. 2020. Collaborative research project: Developing and testing a robot-assisted
intervention for children with autism. Frontiers in Robotics and AI 7, March (Mar. 2020), 1–16. DOI:https://doi.org/
10.3389/frobt.2020.00037
[79] Hideki Kozima, Marek P. Michalowski, and Cocoro Nakagawa. 2009. Keepon. International Journal of Social Robotics
1, 1 (Jan. 2009), 3–18. DOI:https://doi.org/10.1007/s12369- 008-0009- 8
[80] Rebecca P. Lawson, Christoph Mathys, and Geraint Rees. 2017. Adults with autism overestimate the volatility of the
sensory environment. Nature Neuroscience 20, 9 (Jul. 2017), 1293–1299. DOI:https://doi.org/10.1038/nn.4615
[81] Rebecca P. Lawson, Geraint Rees, and Karl J. Friston. 2014. An aberrant precision account of autism. Frontiers in
Human Neuroscience 8, May (May 2014), 302. DOI:https://doi.org/10.3389/fnhum.2014.00302
[82] Jamy J. Li, Daniel Davison, Alyssa M. Alcorn, Alria Williams, Snezana Babovic Dimitrijevic, Sunčica Petrović, Pau-
line Chevalier, Bob R. Schadenberg, Eloise Ainger, Liz Pellicano, and Vanessa Evers. 2020. Non-participatory user-
centered design of accessible teacher-teleoperated robot and tablets for minimally verbal autistic children. In Pro-
ceedings of the 13th ACM International Conference on PErvasive Technologies Related to Assistive Environments (PETRA
’20). 51–59. DOI:https://doi.org/10.1145/3389189.3393738
[83] Itay Lieder, Vincent Adam, Or Frenkel, Sagi Jae-Dax, Maneesh Sahani, and Merav Ahissar. 2019. Perceptual bias
reveals slow-updating in autism and fast-forgetting in dyslexia. Nature Neuroscience 22, 2 (Feb. 2019), 256–264.
DOI:https://doi.org/10.1038/s41593-018-0308-9
[84] Catherine Lord, Michael Rutter, Pamela C. DiLavore, Susan Risi, Katherine Gotham, and Somer L. Bishop. 2012.
Autism Diagnostic Observation Schedule: ADOS-2. Western Psychological Services, Los Angeles, CA. Retrieved from
https://www.wpspublish.com/store/p/2648/ados-2-autism- diagnostic-observation-schedule- second-edition
[85] Nichola Lubold, Erin Walker, and Heather Pon-Barry. 2016. Eects of voice-adaptation and social dialogue on per-
ceptions of a robotic learning companion. In Proceedings of the 2016 11th ACM/IEEE International Conference on
Human–Robot Interaction. IEEE, 255–262. DOI:https://doi.org/10.1109/HRI.2016.7451760
[86] Hope Macdonald, Michael Rutter, Patricia Howlin, Patricia Rios, Ann Le Conteur, Christopher Evered, and Susan Fol-
stein. 1989. Recognition and expression of emotional cues by autistic and normal adults. Journal of Child Psychology
and Psychiatry 30, 6 (Nov. 1989), 865–877. DOI:https://doi.org/10.1111/j.1469-7610.1989.tb00288.x
[87] Gregory S. MacDu, Patricia J. Krantz, and Lynn E. McClannahan. 1993. Teaching children with autism to use photo-
graphic activity schedules: Maintenance and generalization of complex response chains. Journal of Applied Behavior
Analysis 26, 1 (Mar. 1993), 89–97. DOI:https://doi.org/10.1901/jaba.1993.26-89
[88] Catherine Manning, James Kilner, Louise Neil, Themelis Karaminis, and Elizabeth Pellicano. 2017. Children on the
autism spectrum update their behaviour in response to a volatile environment. Developmental Science 20, 5 (Sep.
2017), e12435. DOI:https://doi.org/10.1111/desc.12435
[89] Lizzie Maughan, Sergei Gutnikov, and Rob Stevens. 2007. Like more, look more. Look more, like more: The evidence
from eye-tracking. Journal of Brand Management 14, 4 (Apr. 2007), 335–342. DOI:https://doi.org/10.1057/palgrave.
bm.2550074
[90] Joanne McCann and Sue Peppé. 2003. Prosody in autism spectrum disorders: a critical review. International Journal of
Language & Communication Disorders 38, 4 (Jan. 2003), 325–350. DOI:https://doi.org/10.1080/1368282031000154204
[91] Linda McCormick, Mar yJ. O. Noonan, and Ronald Heck. 1998. Variables aecting engagement in inclusive preschool
classrooms. Journal of Early Intervention 21, 2 (Jan. 1998), 160–176. DOI:https://doi.org/10.1177/105381519802100208
[92] Gary B. Mesibov and Victoria Shea. 2010. The TEACCH program in the era of evidence-based practice. Journal of
Autism and Developmental Disorders 40, 5 (May 2010), 570–579. DOI:https://doi.org/10.1007/s10803-009-0901- 6
[93] Brian W. Miller. 2015. Using reading times and eye-movements to measure cognitive engagement. Educational Psy-
chologist 50, 1 (Jan. 2015), 31–42. DOI:https://doi.org/10.1080/00461520.2015.1004068
[94] D. Mumford. 1992. On the computational architecture of the neocortex. Biological Cybernetics 66, 3 (Jan. 1992), 241–
251. DOI:https://doi.org/10.1007/BF00198477
[95] Louise Neil, Nora Choque Olsson, and Elizabeth Pellicano. 2016. The relationship between intolerance of uncertainty,
sensory sensitivities, and anxiety in autistic and typically developing children. Journal of Autism and Developmental
Disorders 46, 6 (Jun. 2016), 1962–1973. DOI:https://doi.org/10.1007/s10803-016-2721- 9
[96] Heather L. O’Brien and Elaine G. Toms. 2008. What is user engagement? A conceptual framework for dening user
engagement with technology. Journal of the American Society for Information Science and Technology 59, 6 (Apr. 2008),
938–955. DOI:https://doi.org/10.1002/asi.20801
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
36:40 B. R. Schadenberg et al.
[97] Mike Oliver. 2013. The social model of disability: Thirty years on. Disability and Society 28, 7 (Oct. 2013), 1024–1026.
DOI:https://doi.org/10.1080/09687599.2013.818773
[98] Alexia Ostrolenk, Vanessa A. Bao, Laurent Mottron, Olivier Collignon, and Armando Bertone. 2019. Reduced multi-
sensory facilitation in adolescents and adults on the autism spectrum. Scientic Reports 9, 1 (Dec. 2019), 11965.
DOI:https://doi.org/10.1038/s41598-019-48413-9
[99] Mark O’Reilly, Je Sigafoos, Giulio Lancioni, Chaturi Edrisinha, and Alonzo Andrews. 2005. An examination of the
eects of a classroom activity schedule on levels of self-injury and engagement for a child with severe autism. Journal
of Autism and Developmental Disorders 35, 3 (Jun. 2005), 305–311. DOI:https://doi.org/10.1007/s10803-005-3294- 1
[100] Colin J. Palmer, Rebecca P. Lawson, and Jakob Hohwy. 2017. Bayesian approaches to autism: Towards volatility,
action, and behavior. Psychological Bulletin 143, 5 (2017), 521–542. DOI:https://doi.org/10.1037/bul0000097
[101] Rhea Paul, Amy Augustyn, Ami Klin, and Fred R. Volkmar. 2005. Perception and production of prosody by speakers
with autism spectrum disorders. Journal of Autism and Developmental Disorders 35, 2 (Apr. 2005), 205–220. DOI:https:
//doi.org/10.1007/s10803-004-1999-1
[102] Martin P. Paulus and Murray B. Stein. 2006. An insular view of anxiety. Biological Psychiatry 60, 4 (Aug. 2006), 383–
387. DOI:https://doi.org/10.1016/j.biopsych.2006.03.042
[103] Philip J. Pell, Isabelle Mareschal, Andrew J. Calder, Elisabeth A. H. von dem Hagen, Colin W.G. Cliord, Simon
Baron-Cohen, and Michael P. Ewbank. 2016. Intact priors for gaze direction in adults with high-functioning autism
spectrum conditions. Molecular Autism 7, 1 (Dec. 2016), 25. DOI:https://doi.org/10.1186/s13229-016-0085-9
[104] Elizabeth Pellicano and David Burr. 2012. When the world becomes ’too real’: A Bayesian explanation of autistic
perception. Trends in Cognitive Sciences 16, 10 (2012), 504–510. DOI:https://doi.org/10.1016/j.tics.2012.08.009
[105] Elizabeth Pellicano, Linda Jeery, David Burr, and Gillian Rhodes. 2007. Abnormal adaptive face-coding mechanisms
in children with autism spectrum disorder. Current Biology 17, 17 (Sep. 2007), 1508–1512. DOI:https://doi.org/10.1016/
j.cub.2007.07.065
[106] Jose Pinheiro, Douglas Bates, Saikat DebRoy, Deepayan Sarkar, and R Core Team. 2020. Nlme: Linear and Nonlinear
Mixed Eects Models. Retrieved from https://CRAN.R-project.org/package=nlme
[107] Michael I. Posner. 1980. Orienting of attention. Quarterly Journal of Experimental Psychology 32, 1 (Feb 1980), 3–25.
DOI:https://doi.org/10.1080/00335558008248231
[108] W. James Potter and Deborah Levine-Donnerstein. 1999. Rethinking validity and reliability in content ana-
lysis. Journal of Applied Communication Research 27, 3 (Aug. 1999), 258–284. DOI:https://doi.org/10.1080/
00909889909365539
[109] R Core Team. 2020. R: A Language and Environment for Statistical Computing. Technical Report. R Foundation for
Statistical Computing, Vienna. Retrieved from http://www.r-project.org/
[110] Aditi Ramachandran, Chien-Ming Huang, and Brian Scassellati. 2017. Give me a break!. In Proceedings of the 2017
ACM/IEEE International Conference on Human–Robot Interaction. ACM, New York, NY, 146–155. DOI:https://doi.org/
10.1145/2909824.3020209
[111] Stephen W. Raudenbush and Anthony S. Bryk. 2002. Hierarchical Linear Models. Applications and Data Analysis
Methods (2nd ed.). Sage Publications, Thousand Oaks, CA.
[112] Charles Rich, Brett Ponsleur, Aaron Holroyd, and Candace L. Sidner. 2010. Recognizing engagement in human–robot
interaction. In Proceeding of the 5th ACM/IEEE International Conference on Human–Robot Interaction (HRI ’10). ACM
Press, New York, NY, 375. DOI:https://doi.org/10.1145/1734454.1734580
[113] Daniel J. Ricks and Mark B. Colton. 2010. Trends and considerations in robot-assisted autism therapy. In Proceedings
of the 2010 IEEE International Conference on Robotics and Automation. IEEE, 4354–4359. DOI:https://doi.org/10.1109/
ROBOT.2010.5509327
[114] Ben Robins, Kerstin Dautenhahn, René te Boekhorst, and Aude Billard. 2005. Robotic assistants in therapy and educa-
tion of children with autism: Can a small humanoid robot help encourage social interaction skills? Universal Access
in the Information Society 4, 2 (Dec. 2005), 105–120. DOI:https://doi.org/10.1007/s10209-005-0116- 3
[115] Jacqui Rodgers, Anna Hodgson, Kerry Shields, Catharine Wright, Emma Honey, and Mark Freeston. 2017. Towards a
treatment for intolerance of uncertainty in young people with autism spectrum disorder: Development of the coping
with uncertainty in everyday situations (CUES©) programme. Journal of Autism and Developmental Disorders 47, 12
(Dec. 2017), 3959–3966. DOI:https://doi.org/10.1007/s10803-016-2924-0
[116] Ognjen Rudovic, Jaeryoung Lee, Miles Dai, Björn Schuller, and Rosalind W. Picard. 2018. Personalized machine learn-
ing for robot perception of aect and engagement in autism therapy. Science Robotics 3, 19 (Jun. 2018), eaao6760.
DOI:https://doi.org/10.1126/scirobotics.aao6760
[117] Ognjen Rudovic, Jaeryoung Lee, Lea Mascarell-Maricic, Björn W. Schuller, and Rosalind W. Picard. 2017. Measuring
engagement in robot-assisted autism therapy: A cross-cultural study. Frontiers in Robotics and AI 4, July (Jul. 2017),
36. DOI:https://doi.org/10.3389/frobt.2017.00036
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
Predictable Robots for Autistic Children 36:41
[118] Michael Rutter, Anthony Bailey, and Cathrine Lord. 2003. The Social Communication Questionnaire: Manual. Western
Psychological Services, Los Angeles, CA. Retrieved from https://www.wpspublish.com/scq-social-communication-
questionnaire
[119] Michelle J. Salvador, Sophia Silver, and Mohammad H. Mahoor. 2015. An emotion recognition comparative study of
autistic and typically-developing children using the zeno robot. In Procedings of the 2015 IEEE International Conference
on Robotics and Automation. IEEE, 6128–6133. DOI:https://doi.org/10.1109/ICRA.2015.7140059
[120] Felippe Sartorato, Leon Przybylowski, and Diana K Sarko. 2017. Improving therapeutic outcomes in autism spectrum
disorders: Enhancing social communication and sensory processing through the use of interactive robots. Journal of
Psychiatric Research 90 (Jul. 2017), 1–11. DOI:https://doi.org/10.1016/j.jpsychires.2017.02.004
[121] Brian Scassellati. 2007. How social robots will help us to diagnose, treat, and understand autism. In Robotics Research.
Sebastian Thrun, Rodney Brooks, and Hugh Durrant-Whyte (Eds.). Springer, Berlin Heidelberg, 552–563. DOI:https:
//doi.org/10.1007/978-3-540-48113- 3_47
[122] Brian Scassellati, Henny Admoni, and Maja J. Matarić. 2012. Robots for use in autism research. Annual Review of
Biomedical Engineering 14, 1 (2012), 275–294. DOI:https://doi.org/10.1146/annurev-bioeng-071811-150036
[123] Brian Scassellati, Laura Boccanfuso, Chien-Ming Huang, Marilena Mademtzi, Meiying Qin, Nicole Salomons, Pamela
Ventola, and Frederick Shic. 2018. Improving social skills in children with ASD using a long-term, in-home social
robot. Science Robotics 3, 21 (8 2018), eaat7544. https://doi.org/10.1126/scirobotics.aat7544
[124] Bob R. Schadenberg. 2021. Robots for Autistic Children: Understanding and Facilitating Predictability for Engagement
in Learning. Ph.D. Dissertation. University of Twente, Enschede. DOI:https://doi.org/10.3990/1.9789036551649
[125] Bob R. Schadenberg, Dirk K. J. Heylen, and Vanessa Evers. 2018. Aect bursts to constrain the meaning of the facial
expressions of the humanoid robot Zeno. In Proceedings of the 1st Workshop on Social Interaction and Multimodal
Expression for Socially Intelligent Robots 30–39. Retrieved from http://ceur-ws.org/Vol-2059/paper4.pdf
[126] Bob R. Schadenberg, Mark A. Neerincx, Fokie Cnossen, and Rosemarijn Looije. 2017. Personalising game diculty
to keep children motivated to play with a social robot: A Bayesian approach. Cognitive Systems Research 43 (2017),
222–231. DOI:https://doi.org/10.1016/j.cogsys.2016.08.003
[127] Bob R. Schadenberg, Dennis Reidsma, Dirk K. J. Heylen, and Vanessa Evers. 2020. Dierences in spontaneous inter-
actions of autistic children in an interaction with an adult and humanoid robot. Frontiers in Robotics and AI 7(Mar.
2020), 19. DOI:https://doi.org/10.3389/frobt.2020.00028
[128] Bob R. Schadenberg, Dennis Reidsma, Dirk K. J. Heylen, and Vanessa Evers. in press. “I see what you did there”:
Understanding people’s social perception of a robot and its predictability. ACM Transactions on Human–Robot Inter-
action 10, 3 (in press), 1–27. DOI:https://doi.org/10.1145/3461534
[129] Eric Schopler, Mary E. Van Bourgondien, Glenna J. Wellman, and Steven R. Love. 2010. The Childhood Autism Rat-
ing Scale (CARS-2). Western Psychological Services, Los Angeles, CA. Retrieved from https://www.wpspublish.com/
store/p/2696/cars-2-childhood-autism- rating-scale-second- edition
[130] Candace L. Sidner, Christopher Lee, Cory D. Kidd, Neal Lesh, and Charles Rich. 2005. Explorations in engagement
for humans and robots. Articial Intelligence 166, 1-2 (Aug. 2005), 140–164. DOI:https://doi.org/10.1016/j.artint.2005.
03.005
[131] Emily Simono, Andrew Pickles, Tony Charman, Susie Chandler, Tom Loucas, and Gillian Baird. 2008. Psychi-
atric disorders in children with autism spectrum disorders: Prevalence, comorbidity, and associated factors in a
population-derived sample. Journal of the American Academy of Child & Adolescent Psychiatry 47, 8 (Aug. 2008),
921–929. DOI:https://doi.org/10.1097/CHI.0b013e318179964f
[132] Kate Simpson, Deb Keen, and Janeen Lamb. 2013. The use of music to engage children with autism in a receptive
labelling task. Research in Autism Spectrum Disorders 7, 12 (Dec. 2013), 1489–1496. DOI:https://doi.org/10.1016/j.rasd.
2013.08.013
[133] Ramona E. Simut, Johan Vanderfaeillie, Andreea Peca, Greet Van de Perre, and Bram Vanderborght. 2016. Children
with autism spectrum disorders make a fruit salad with probo, the social robot: An interaction study. Journal of
Autism and Developmental Disorders 46, 1 (Jan. 2016), 113–126. DOI:https://doi.org/10.1007/s10803-015-2556- 9
[134] Gale M. Sinatra, Benjamin C. Heddy, and Doug Lombardi. 2015. The challenges of dening and measuring student
engagement in science. Educational Psychologist 50, 1 (1 2015), 1–13. https://doi.org/10.1080/00461520.2014.1002924
[135] Pawan Sinha, Margaret M. Kjelgaard, Tapan K. Gandhi, Kleovoulos Tsourides, Annie L. Cardinaux, Dimitrios
Pantazis, Sidney P. Diamond, and Richard M. Held. 2014. Autism as a disorder of prediction. Proceedings of the Na-
tional Academy of Sciences 111, 42 (2014), 15220–15225. DOI:https://doi.org/10.1073/pnas.1416797111
[136] James P. Stevens. 2002. Applied Multivariate Statistics for the Social Sciences (5th ed.). Taylor & Francis, New York,
NY.
[137] Caroline L. van Straten, Iris Smeekens, Emilia I. Barakova, Jerey C. Glennon, Jan K. Buitelaar, and Aoju Chen.
2018. Eects of robots’ intonation and bodily appearance on robot-mediated communicative treatment outcomes for
children with autism spectrum disorder. Personal and Ubiquitous Computing 22, 2 (Apr. 2018), 379–390. DOI:https:
//doi.org/10.1007/s00779-017-1060-y
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
36:42 B. R. Schadenberg et al.
[138] Dag Sverre Syrdal, Kerstin Dautenhahn, Ben Robins, Efstathia Karakosta, and Nan Cannon Jones. 2020. Kaspar in
the wild: Experiences from deploying a small humanoid robot in a nursery school for children with autism. Paladyn,
Journal of Behavioral Robotics 11, 1 (Jul. 2020), 301–326. DOI:https://doi.org/10.1515/pjbr-2020-0019
[139] Adriana Tapus, Andreea Peca, Amir Aly, Cristina A. Pop, Lavinia Jisa, Sebastian Pintea, Alina S. Rusu, and Daniel O.
David. 2012. Children with autism social engagement in interaction with Nao, an imitative robot: A series of single
case experiments. Interaction Studies 13, 3 (2012), 315–347. DOI:https://doi.org/10.1075/is.13.3.01tap
[140] Furtuna G. Tewolde, Dorothy V. M. Bishop, and Catherine Manning. 2018. Visual motion prediction and verbal false
memory performance in autistic children. Autism Research 11, 3 (Mar. 2018), 509–518. DOI:https://doi.org/10.1002/
aur.1915
[141] Serge Thill, Cristina A. Pop, Tony Belpaeme, Tom Ziemke, and Bram Vanderborght. 2012. Robot-assisted therapy
for autism spectrum disorders with (Partially) autonomous control: Challenges and outlook. Paladyn, Journal of
Behavioral Robotics 3, 4 (Jan. 2012), 209–217. DOI:https://doi.org/10.2478/s13230-013-0107-7
[142] Marco Turi, David C. Burr, Roberta Igliozzi, David Aagten-Murphy, Filippo Muratori, and Elizabeth Pellicano. 2015.
Children with autism spectrum disorder show reduced adaptation to number. Proceedings of the National Academy
of Sciences 112, 25 (Jun. 2015), 7868–7872. DOI:https://doi.org/10.1073/pnas.1504099112
[143] Joshua Wainer, Ben Robins, Farshid Amirabdollahian, and Kerstin Dautenhahn. 2014. Using the humanoid robot
KASPAR to autonomously play triadic games and facilitate collaborative play among children with autism. IEEE
Transactions on Autonomous Mental Development 6, 3 (Sep. 2014), 183–199. DOI:https://doi.org/10.1109/TAMD.2014.
2303116
[144] Susan W. White, Donald Oswald, Thomas Ollendick, and Lawrence Scahill. 2009. Anxiety in children and adolescents
with autism spectrum disorders. Clinical Psychology Review 29, 3 (Apr. 2009), 216–229. DOI:https://doi.org/10.1016/
j.cpr.2009.01.003
[145] Sarah Wigham, Jacqui Rodgers, Mikle South, Helen McConachie, and Mark H. Freeston. 2015. The interplay between
sensory processing abnormalities, intolerance of uncertainty, anxiety and restricted and repetitive behaviours in
autism spectrum disorder. Journal of Autism and Developmental Disorders 45, 4 (2015), 943–952. DOI:https://doi.org/
10.1007/s10803-014-2248-x
[146] World Health Organization. 1992. The ICD-10 Classication of Mental and Behavioural Disorders: Clinical Descrip-
tions and Diagnostic Guidelines. World Health Organization, Geneva.Retrieved from https://apps.who.int/iris/handle/
10665/37958
[147] Shu-Chieh Wu and Roger W. Remington. 2003. Characteristics of covert and overt visual orienting: Evidence from
attentional and oculomotor capture. Journal of Experimental Psychology: Human Perception and Performance 29, 5
(2003), 1050–1067. DOI:https://doi.org/10.1037/0096-1523.29.5.1050
[148] Nurit Yirmiya, Connie Kasari, Marian Sigman, and Peter Mundy. 1989. Facial expressions of aect in autistic, mentally
retarded and normal children. Journal of Child Psychology and Psychiatry 30, 5 (Sep. 1989), 725–735. DOI:https://doi.
org/10.1111/j.1469-7610.1989.tb00785.x
Received November 2020; revised March 2021; accepted May 2021
ACM Transactions on Computer-Human Interaction, Vol. 28, No. 5, Article 36. Publication date: August 2021.
