IEEE Robotics & Automation Magazine - June 2019 - 17

of each infant limb at 128 Hz using the quaternion filter supplied by the Opal sensor software development kit. The
acceleration information was provided to the robot to drive
its response behavior.
During the portions of the study involving contingent
rewards, the robot's reward behavior was activated when
the infant's right leg moved above a resultant acceleration
threshold. This acceleration threshold was determined
based on our previous work [20]; we found a threshold of
3 m/s 2 to be appropriate for detecting leg movements in
the root-sum-of-squares signal from the three filtered
accelerometer axes. During the robot's reward behaviors,
the infant movement perception system paused and did not
monitor for new acceleration over the threshold until the
robot finished executing its reward condition behavior.
Because of the manner in which data were collected, any
infant movements above the acceleration threshold that
occurred during the robot reward kicks (each approximately 1.2 s in duration) were not considered in our later contingent learning analysis.

H2: Imitation
Infants will tend to imitate the NAO robot by kicking the ball in front of them with a knee-extension
movement. Past work in the mirror neuron literature [10] suggests that an infant may naturally imitate a similarly sized humanoid robot. This is part of
the motivation for using a robot in this work and
for selecting the NAO robot in particular. At the
same time, imitation is not as well understood as
contingent learning; therefore, this hypothesis
is exploratory.

Hypotheses
This study tested three main hypotheses (H).

Data Collection
During the study, we tracked the following aspects of the
infant's state: physical responses (i.e., motion measured by
inertial sensors), visual responses (i.e., eye gaze), and
behavioral state responses (i.e., infant-alertness level). We
collected physical response data from the arm- and legworn Opal inertial sensors and gaze direction data from
the head-worn eye tracker. Three RGB cameras and a
Microsoft Kinect One RGB-D camera captured front, side,
and face views of the infant for post hoc assessment of
infant behavior.

H1: Learning the Contingency
The majority of infants will learn the contingency, i.e.,
they will move their right leg above the acceleration
threshold at a rate of 1.5 times or greater in one or more
of the contingency phases than is done during the initial
baseline. This hypothesis aligns with findings from past
work such as [5], where infants learned to move in a particular way to gain contingent rewards.

H3: Parent Surveys
The parents' perception of infant behavioral state and
contingency learning will provide additional insights
beyond the quantitative and qualitative data collected
from the infants. Because the parents are present in the
lab and able to observe the infant's behavior, parental insights may augment our understanding of the
child's experience.

Table 2. The frequency of infant motion above the acceleration threshold during study phases.
Infant

Baseline
(Moves/Min)

Learner
Threshold
(Moves/Min)

Movement Reward
Activity
(Moves/Min)

Movement and
Lights Reward
Activity (Moves/Min)

Movement and
Sound Reward
Activity (Moves/Min)

Extinction
(Moves/Min)

TD2

11.62

15.75

18.37

TD3

10.87

7.87

8.25

TD4

7.5

25.12

16.5

4.5

TD5

11.62

14.62

TD6

17.5

26.25

25.87

16.12

10.5

TD7

14.5

21.75

14.62

18.37

9.5

TD8

7.5

11.25

26.25

16.12

16.5

TD9

4.5

6.75

13.87

5.25

13.5

10.5

TD10

7.5

18.37

12.75

17.25

9.5

TD11

10.5

15.75

14.62

25.5

14.5

TD12

34.12

33.75

TD13

3.37

7.12

14.25

Note: The third column shows the learning threshold, i.e., the leg-movement rate required to conclude that the infant has learned the reward. Shaded
headings match the subsequent plot color coding. Gray boxes indicate phases during which contingent reward learning occurred.

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