IEEE Robotics & Automation Magazine - June 2019 - 54

WoZ
SPARC
AL

WoZ

SPARC
AL

Workload

Performance

AL
SPARC

Autonomy

WoZ

Figure 7. The expected ideal behaviors of SPARC over time compared with WoZ and autonomous learning (AL), based on the
therapist's workload, the robot's performance, and autonomy.

In case the action autonomously decided by the robot is not
proper, the therapist can deny the suggested action and manually select a more appropriate one. We have proposed a learning-from-demonstration method called supervised progressively
autonomous robot competencies (SPARC), so that the robot can
learn from the manual actions of the therapist and improve its
suggested actions for the next interaction [21]. As shown in
Figure 7, SPARC aims at maintaining high level of performance
throughout the interaction (e.g., in WoZ) while ensuring a light
workload for the therapist (i.e., autonomous learning).
The self-monitoring system attempts to overcome possible
technical and ethical limitations. This system currently acts as
a logging mechanism and is connected using the therapist's
supervisory interface. The therapist can overrule the robot's
proposed actions via the GUI. In future applications and
based on a set of rules, it would act as an alarm system that is
triggered when the robot detects technical limitations and
ethical issues. This system also provides recorded data (e.g., a
child's performance and the robot's operation) for therapists
and engineers to evaluate the efficacy of a RET system.
Clinical Experiments and Results
From a clinical perspective, this project seeks to determine
how much RET can improve joint attention, imitation, and
turn-taking skills in ASD children as well as how the gains
obtained within these interactions compare to standard interventions. Therefore, the clinical experiments were divided
into two phases: one using RAT robots under a WoZ system,

Figure 8. The intervention platform used in DREAM. A child sits
in front of a robot and an interactive screen.

IEEE ROBOTICS & AUTOMATION MAGAZINE

JUNE 2019

and another using RET within a supervised autonomous system. Both phases have been compared to standard human
treatment (SHT) conditions.
The experiments were conducted using a classical singlecase alternative treatment design. Children participated in six
to eight baseline sessions, followed by eight SHT sessions and
eight WoZ or RET sessions. Within the baseline sessions, the
child interacts with a human partner who does not offer any
feedback regarding the child's performance. The purpose of
these sessions is to identify the initial level of skills and their
variability before the child receives any of the two interventions
(i.e., SHT or RET, where either the human or robotic partners
give feedback that is based on the child's performance).
The conditions were randomized to mitigate the ordering
effect. After the baseline sessions, the order for each intervention session (either SHT or RAT/RET) was established based
on a random schedule that contained a random sequence
indicating which session should be performed next. The
schedule was different for each child.
Before the baseline session, we used the Autism Diagnostic
Observation Schedule (ADOS) instrument [22] to confirm
children's diagnosis of autism and assess which were their
social and communication abilities. We also employed ADOS
as a measurement tool to quantify-before and after interventions-the differences in the scores.
After the initial ADOS measurements were taken and the
baseline session completed, children interacted with either a
robot or a human, with an additional person always acting as
a mediator between the child and interaction partner. The
tasks to be tested were implemented following a discrete trial
format, i.e., within a highly structured environment, the
behaviors broken into discrete subskills, and a child taught to
respond to explicit prompting (e.g., "Do you like me?").
We employed the humanoid robot NAO [29] to assess our
hypothesis. For certain tasks, we used the electronic Sand
Tray therapy kit [23], a 26-in capacitive touchscreen and associated control server where images can be manipulated by
dragging (on the side of the human partner) or simulated
dragging (on the side of the robot partner). Moreover, an
intervention table was designed to capture sensory information (shown in Figure 8) by employing three RGB cameras
and two Kinect sensors.

IEEE Robotics & Automation Magazine - June 2019

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