IEEE Robotics & Automation Magazine - December 2012 - 32

Summation of Posture Change Norms
Summation of Posture Change Norms
Summation of Posture Change Norms

20
0.0325
15

3
4
Subject Number
(a)

2.32e-4
1.18e-4

5.19e-4
4.96e-4

0.0873

3
4
Subject Number
(b)

0.0049
0.022
0.033
0.0073

3.05e-7
1.36e-5 4.63e-4
8.27e-8
0.0028
0.0021

3
4
Subject Number
(c)

Figure 8. Mean and standard deviation of the summation of
the posture change norm of test subjects in the (a) baseline,
(b) first learning, and (c) final training experiments. The dark
gray, white, and light gray bars indicate the mean and standard
deviation values during each of the intermediate learning steps
(after every ten trials). In (a), the baseline experiment, only
Subject 2 shows a significant improvement after all trials. In
(b) the first learning experiment, Subjects 2, 4, and 5 show
significant improvements. In (c) the final experiment, the
interaction with the robot improved for all subjects. With each
learning trial, the indicated values decrease, and the movement
of the robot becomes smoother and more synchronized with
that of the subject.

IEEE ROBOTICS & AUTOMATION MAGAZINE

DECEMBER 2012

decreased as the experiment progressed. In the baseline
experiment, only one subject was able to significantly
improve the interactions, where statistical significance is
computed using a t test. None of the other subjects were able
to improve their interactions. In the first experiment, in
which the proposed learning system is used, three subjects
show significant improvement. Finally, in the second learning experiment, all of the subjects showed significant
improvement in their interactions. This indicates that while a
human can adapt to a robot and thus improve their interactions (as in the baseline experiment), this adaptation can be
significantly improved by empowering the robot with learning capabilities (first and second learning experiments). We
also analyzed the maximum values of the posture change
norm during the interaction. Figure 9 shows the change in
the maximum posture change norm during each learning
phase of the baseline experiment and the first learning experiment. No significant difference in the maximum posture
change norm is observed in the baseline experiment. On the
other hand, in the learning experiment, there are large
changes in the maximum posture change norm. For all subjects, the values drastically decrease after learning.
Still, one possible implication from above results cannot be ruled out by the experiments performed so far.
Specifically, it remains unclear how much the learning
system contributes to the improvement of interaction. A
possible argument would be that the observed improvements are due to the long-term habituation and experience with the robot. If this argument is true, then we
should see a similar improvement of interactions as above,
even if we simply repeat the baseline experiment (where
learning is disabled) three times in a row. To investigate
this question, we performed the aforementioned experiment (three times baseline) with all subjects. For the subjects, the experiment looked exactly the same as the other
experiments: the difference was not transparent.
Figure 10 compares the summation of posture change
norm between the first and the third baseline experiment.
In each of the experiments, only one subject made significant improvement during the intermediate learning
steps. On the whole, although for some subjects slight
improvement was visible (notably Subject 5), the results
are not as comprehensive as when learning is enabled.
This means that, while long-term habituation and experience aids the learning process, it is not sufficient for a
general improvement in PHRI.
Discussion
The following observations are based on the results of the
above experiments. First, when learning and adaptation were
only possible on the side of the human caregiver, generally, little or no improvement could be measured. However, even in
this asymmetric learning situation, at least one subject was able
to adapt to the robot so as to significantly improve the interaction quality. This shows the human ability to quickly adapt to
new situations and motor tasks. The second observation is that

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2012