IEEE Robotics & Automation Magazine - June 2019 - 55

Children's performance of the task was assessed by measuring their performance based on task-solving accuracy
(e.g., accuracy during the imitation task, correct gazing during the joint-attention task, and appropriate pauses during the
turn-taking task).

positive impact on the clinical ASD symptoms, with children
in both groups who had completed the final assessment
reporting lower ADOS severity scores at the end of the treatment, i.e., t(5) = 3.50, p = 0.017 in SHT, and t(4) = 3.25, p =
0.031 in RET.

First Phase
As stated previously, during the first phase of the experiments we used RAT robots remotely controlled under a WoZ
setup. Results from these experiments were used as a basis
for the developing the supervised autonomous system. During this phase, 11 participants with ASDs between the ages of
three and five years were recruited from the Autism Transylvanian Association.
After completing this first phase, we obtained mixed
results. These results were different depending on the task at
hand. During the turn-taking task, the WoZ setup seemed
to achieve gains as good as or even better than under SHT
conditions, especially for children with low skill levels.
Regarding joint attention, RAT and SHT yielded similar
outcomes for the majority of participants. Specific to the
RAT intervention, the results also suggest that the level of
prompting offered by the robot mediator has a direct impact
on the performance of ASD children, with more prompts
resulting in improved performance [24]. For the imitation
task, most of the children previously demonstrated good
performances in baseline sessions, and the RAT condition
did not enhance these skills.

Therapists' Attitudes Toward the System
We conducted an interview to determine therapists' attitudes
toward the DREAM system. Four therapists that have been
working with the system for an average of six months were
interviewed via email. During the interview, we used
open-ended questions and
a short usability survey. A
screenshot of the GUI was
Research in the ethics
used to elicit memories
about their experiences
of social robotics has
with the system.
The DREAM system
increased significantly
was generally appreciated
by the therapists (n = 4).
as it pertains to health care
According to the questionnaire results (a fiveand children.
point Likert scale), the
therapists showed positive
attitudes toward the system, i.e., useful (M = 4.1, Min = 4.0,
and Max = 4.1), satisfying (M = 3.6, Min = 2.4, and Max =
4.0), easy to use (M = 4.2, Min = 4.2, and Max = 4.4), and easy
to learn (M = 3.8, Min = 2.0, and Max = 5.0). They expressed
that the interface of the GUI is easy to use and helps them
deliver an intervention that is both attractive and effective.
According to the interview results, the automatic detection of
behaviors was useful in treating children with ASDs because it
reduced the therapists' potential burden of intervention. They
also found the system to be safe and acceptable. Yet, some
improvements are needed; e.g., increasing the accuracy of recognition, reducing the technical complexity of the system,
and simplifying the GUI. Regarding the possibility of using
the system for other types of therapies, they suggested having
customized GUIs for different applications.

Second Phase
In the second phase of the experimental investigations, we
compare the efficacy of RET and SHT using a randomized
clinical trial design [30]. For this purpose, 27 children were
recruited to participate from different organizations and institutions (most of which are located in Cluj-Napoca, Romania)
that provide educational and/or psychotherapeutic services to
children with autism. To date, 21 of the participants have
completed the full protocol.
The last-observation-carried-forward scores indicated
that both groups showed signs of improvement within a significant time effect, i.e., Wilks' lambda = 0.62, F(8,16) = 3.19,
p = 0.023, and h 2p = .62, with no significant group or interaction effects. A univariate analysis indicated that scores have
improved for imitation (i.e., F(1,25) = 21.79, p <0.001, and
2
h p = .47) and for each of the turn-taking tasks (i.e., sharing
information about what one likes, F(1,25) = 4.50, p = 0.044,
and h 2p = .15); completing a series of figures following a pattern, F(1,25) = 10.22, p = 0.004, and h 2p = .29; and categorizing items, F(1,25) = 11.61, p = 0.002, and h 2p = .32, but not
for joint attention, where baseline differences favoring the
SHT group were observed, F(1,23) = 6.66, p = 0.017, and
2
h p = .23. However, posttest differences between groups were
not significant for joint attention, even when controlling for
baseline scores. In future studies, this outcome will be carefully monitored and more children and sessions may change
it in the expected direction. Both interventions also had a

Ethical Perspective
Lately, research in the ethics of social robotics has increased
significantly as it pertains to health care and children [25].
What are the specific problems raised by autonomous
interaction with mentally disabled children? How can we protect children from exploitation? What if the robot gets the
behavior assessment of the child wrong? How and when does
the therapist need to overrule the behavior of the robot when
needed? These questions all raise important ethical concerns.
Within the DREAM project, we have conducted several studies to explore these and other ethical issues.
In one of the studies, Coeckelbergh et al. [9] attempted to
understand the opinions of parents and therapists about the
appropriateness and benefits of social robots being used in
therapy for children with ASDs. An important finding was
the high acceptability of these robots for helping children with
JUNE 2019

IEEE ROBOTICS & AUTOMATION MAGAZINE

IEEE Robotics & Automation Magazine - June 2019

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - June 2019