Available via license: CC BY 4.0
Content may be subject to copyright.
1
Preliminary results of a therapeutic lab for promoting autonomies in autistic children
Cristina(Gena,(Rossana(Damiano,(Claudio(Mattutino,(Alessandro(Mazzei,(Stefania(Brighenti(
Dipartimento(di(Informatica,(Università(di(Torino,(cristina.gena@unito.it((
Matteo(Nazzario(,(Valeria(Ricci(
Intesa(Sanpaolo(Innovation(Center,(corso(Inghilterra(3,(Turin,(10138,(Italy(
Camilla(Quarato,(Cesare(Pecone,(Giuseppe(Piccinni(
Jumple(srl,(Via(Isonzo,(55/2,(Casalecchio(di(Reno,(Bologna,(40033,(Italy(
Federica(Liscio,(Loredana(Mazzotta,(Andrea(Meirone,(Francesco(Petriglia(
Fondazione(PAIDEIA(Onlus,(Italy(
This extended abstract describes the preliminary quantitative and qualitative results coming from a therapeutic
laboratory focused on the use of the Pepper robot to promote autonomies and functional acquisitions in highly
functioning (Asperger) children with autism. The laboratory
1
started in February 2021 and lasted until June 2021, and
the weekly meetings lasted two hours and were led by one or two therapists (educators, speech therapists,
psychologists, etc.), helped by 2-3 trainee master students. The participants recruited were four highly functioning
(Asperger) children, aged between 11 and 13 years. There have been in total 16 lab sessions, all recorded by a fixed
camera, in addition to the Pepper’s 2D cameras. Furthermore, trainees filled out evaluation forms provided by
psychotherapists, noting the children autonomy’s progress in a diary with the helping of rating scales [1]. These notes
were then reworked to draw up shared reports, reflecting on the behavior’s evolution and progress of the children
meeting by meeting.
The setting of the lab was an elegant apartment furnished as a real home in the city center. Each meeting had a
similar structure: 1) welcome in the apartment; 2) social moment: dialogue with the robot; 3) moment of snack
preparation; 4) moment of post-snack dialogue; 5) final feedback and goodbye.
The snack preparation was one of the most stimulating moments for the children, dedicated to the preparation, in the
kitchen or directly on the dining room table, of some increasingly complex snacks. The group was led both by Pepper,
instructed to organize and coordinate the activity, and by the therapists, ready to intervene when required.
The goal of the activity was to gradually mitigate the therapist’s aid, so that therapists could only make suggestions
from time to time. Pepper, with the help of the video–modeling [5] encouraged the participants to schematically
organize themselves, listing the ingredients, illustrating the procedures with images, animations, and videos, and giving
the children time to manage the preparation, as well as the possibility of reviewing the steps. The activity gave good
results: the children appreciated the help of Pepper, showing good levels of increasing autonomy, even if, sometimes,
the difficulties in maintaining a high concentration affected the scores reported by the trainees.
During the social moment, Pepper was conversing with children. On the one hand, the robot responded to their
curiosity about itself, and, on the other, guiding a dialogue called “making friends”, in which the robot attempted to
establish a link with each participant, according to their previously declared interests, as advocated in the design of
1
Ethical approval for this study was obtained from the bioethical committee of the University of Turin, with approval number: 0664572
2
social and educational robots [2],[3], also target to autistic users [4]. Indeed, from time to time, Pepper re-proposed
the topics children previously liked most. In the first case, the scenario envisaged that the children took their turns
facing the robot, waiting for it to catch their gaze and listen, and then ask it any question. In the second case, the robot
called the children one by one and began to talk with them on a previously liked topic (e.g., music, video games, etc.),
which followed a script manually updated week by week, with the robot trying to guide the conversation. However,
both the attempts led to unsatisfactory results, often arousing frustration among participants.
In fact, it could be argued that the problem arose at the roots of design, since both activities were very far from what
we could really define a "conversation", other than a simple transmission of information. Bringing the mind back to the
sociologist Sherry Turkle [6] conversations convey much more than the details of an argument: it is not just a question
of answers, but of what they mean. As the developers did not implement real dialogue autonomy in the robot, Pepper
showed no progress in the interaction, leaving the trainees to take note of the children’s inclinations, and planning,
from one meeting to the next, a new dialogue that considered what emerged.
The results from this experience showed some critical issues to be addressed in future works. Concerning the dialogue
system, at least two related features need to be empowered. On the one side, the dialogue system needs to be
improved in robustness and in precision. The actual conversations show a high degree of expectation from children
about the robot’s knowledge: to fulfill this expectation, one needs to have a correct and precise semantic
representation of the children’s questions in encyclopedic and commonsense domains. A possible improvement could
be based on the construction of an annotated corpus by using a Wizard of Oz approach [7]. In this way, one could train
a machine learning frame based natural language understanding system, starting from the annotation of user intents.
The other improvement concerns the preparation of back-up dialogue strategies that the dialogue system can adopt
in the case of non-comprehensible questions/utterances from the child or sentences not strictly related. An interesting
possibility for building a back-up strategy is using large language-models as BERT [8].
We also defined an ontology for the possible topics of children interests whose classes and properties have been
defined by extracting them from DBpedia
2
. As future work, we will integrate the robot’s dialogue with this knowledge
base to make the robot able to navigate the ontology and reason on it, thus enriching its dialogue strategies.
Focusing on the specific autistic children’s features, we must notice that the autistic functioning distorts the essence
of the conversation. The exchange of utterances does not produce the pleasure of sharing but is functional to obtaining
something more concrete, such as searching for information. If the goal is reaching a typical conversation, the results
could always be unsatisfactory in this context. At the same time, this calls for the collection of conversational data
targeted at this specific group of interactants.
REFERENCES
[1] Federica Cena, Cristina Gena, Pierluigi Grillo, Tsvi Kuflik, Fabiana Vernero & Alan J. Wecker (2017) How scales
influence user rating behaviour in recommender systems, Behaviour & Information Technology, 36:10, 985-
1004, DOI: 10.1080/0144929X.2017.1322145
[2] Cietto, V., Gena, C., Lombardi, I., Mattutino, C., & Vaudano, C. (2018, September). Co-designing with kids an
educational robot. In 2018 IEEE Workshop on Advanced Robotics and its Social Impacts (ARSO) (pp. 139-140). IEEE.
2
https://www.dbpedia.org/
3
[3] Gena, C., Mattutino, C., Perosino, G., Trainito, M., Vaudano, C., & Cellie, D. (2020, May). Design and development
of a social, educational and affective robot. In 2020 IEEE Conference on Evolving and Adaptive Intelligent Systems
(EAIS) (pp. 1-8). IEEE.
[4] Gena, C., Mattutino, C., Brighenti, S., Meirone, A., Petriglia, F., Mazzotta, L., Liscio, F., Nazzario, M., Ricci, V.,
Quarato, C., Pecone, C., & Piccinni, G. (2022). Sugar, Salt & Pepper - Humanoid robotics for autism. IUI
Workshops 2021, ArXiv, abs/2203.07543.
[5] McCoy, K., & Hermansen, E. (2007). Video modeling for individuals with autism: A review of model types and
effects. Education and treatment of children, 183-213.
[6] Sherry Turkle, RECLAIMING CONVERSATION, The Power of Talk in a Digital Age, 436 pp. Penguin Pres, 2016
[7] Riek, L. D. (2012). Wizard of oz studies in hri: a systematic review and new reporting guidelines. Journal of Human-
Robot Interaction, 1(1), 119-136.
[8] Xiaodong Gu, Kang Min Yoo, Jung-Woo Ha, DialogBERT: Discourse-Aware Response Generation via Learning to
Recover and Rank Utterances, AAAI 2021, https://arxiv.org/abs/2012.01775
