IEEE Robotics & Automation Magazine - June 2019 - 91

Ball
House
Tree

Butterfly
Two Houses
V-Shape

3
Baseline Intervention
1
0.8 Subcomposition
0.6
0.4
0.2
0
1
2
1
2
1
Symmetry
0.8
0.6
0.4
0.2
0
1
2
1
2
3
1
0.8 Assembly
0.6
0.4
0.2
0
1
2
1
2
3
4
5
Trial
(a)
Independent Correct
Student 3

Challenge Areas
Based on open-ended responses, student performance, and experimenter observations, several areas
that are challenging for designing an intelligent
instruction system were also identified. We believe
the following lessons-learned observations would be
informative toward future work in this area:
● Prompt design is a critical determining factor in
successful instruction. Intrusiveness, verbiage,
and interaction modalities should be designed
with a domain expert.
● Incorporating multiple levels of interaction detail
increases efficiency and trust. Different modes of
interaction, e.g., controlling verbosity in repetitive interactions, can decrease cognitive load and
negate the appearance of a lack of intelligence,
thereby increasing trust.
● Accurate perception is critical. Vision and speech
recognition errors that result in incorrect feedback from the instructor could be harmful to the

JUNE 2019

Figure 14. Student 3: (a) The IRI geometric assembly skill results; here, three skills were taught, with three puzzles per skill (see Figure 9). Nonincreasing performance generally represents
errors, because when using the SMP and backward-chaining method, the number of possible correct independent steps increases until the student is asked to assemble the entire figure
independently. (b) The ARI copy machine, student account, and geometric reasoning skill results. Students are given an instruction phase after baseline, where they must perform the skill
perfectly, followed by an independent phase, where they control their own prompts.

4
3
2
1
100
80
60
40
20
0
100
80
60
40
20
0
100
80
60
40
20
0
Diamond
Hexagon
Trapezoid

Percent Correct

Efficiency of Self-Directed AR Instruction
It is likely that the self-directed nature of the AR
instruction system is responsible for the rapid learning in the AR experiments. We found that providing
the participants with the ability to control their own
prompts, combined with the intelligent AR environment, likely contributed to increased engagement
and efficiency in learning the tasks. In this way,
because students were able to take control of their
own learning experience, they were able to learn
more efficiently. A genuine level of satisfaction with
the experience on the part of the students was
observed, which could also be attributed to the efficiency of the experience.

(b)

Geometric Reasoning
9 10 11 12 13 14 15 16 17 18 19 20 21 22

Student Account

Copy Machine
Independent
Instruction
Baseline

with a robot before instruction, the students showed
greater willingness to work with the IRI again. They
also trusted the robot, were willing to obey the
robot's instructions, and found the experience positive overall. Perhaps unsurprisingly, student 3, who
had the most difficulty, also gave the experience the
lowest scores for how easy it was to learn and how
much she trusted the robot. Regarding the mixed
ratings of the usefulness of the robot's gestures, one
possible explanation for the lower ratings of students 1 and 3 is that, despite giving very positive
feedback overall, student 1's higher affinity meant
she had little need for the IRI's gestures, whereas
student 3 felt more frustrated by the longer time to
acquire the skill, as reflected in her score on the ease
of learning question and as discussed previously.
In open-ended responses, students highlighted
the benefits they perceived in working with the IRIs,
including improved pace, patient repetition, and
relieving the burden on human teachers.

IEEE ROBOTICS & AUTOMATION MAGAZINE

IEEE Robotics & Automation Magazine - June 2019

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