software which was used to render AVI files. Clips
showing a humanoid robot performing the motions
were generated by having the robot perform the mo-
tions and recording them using Digital Video. The
recordings were subsequently cropped to the appro-
priate durations and stored as AVI files using Adobe
Premiere. Examples of the movies can be found at
http://www.dcs.gla.ac.uk/ halej/biomimetics.htm.
4. Experiments
Two experiments were performed that examined
pairwise similarity comparisons of all elements of the
set of movements including motions generated by the
14 different motor control strategies as well as the
captured human movement. In Experiment 1 these
movements were played back on both a humanoid
robot and a computer graphics (CG) character.
Analysis of the data was obtained through multidi-
mensionsal scaling (MDS) (Kruskal and Wish, 1978)
and indicated the presence of outliers that dominated
the MDS solution. In an attempt to more closely in-
spect the fine structure of responses, Experiment 2
examined responses to the CG character when the
outliers were removed.
General methods. In Experiment 1 we exam-
ined pairwise similarity judgements to sets of move-
ments generated by the 15 production techniques.
Excluding self-comparison, a set of 15 elements af-
fords 105 comparisons between pairs. Thus, for each
condition a set of 105 stimuli were constructed. In
Experiment 1 there were 4 sets of 105 pairs corre-
sponding to the combination of 2 modes of presenta-
tion (humanoid or CG character) and 2 movements
((a) or (b)). Experiment 2 used half of these stimuli,
including only movements presented on the CG char-
acter. Pairs of motions for comparison were obtained
by editing together individual video clips so that they
played one after another separated by a brief black
screen. Participants viewed a pair of movements dis-
played on a computer monitor and then gave a rating
of how similar the movements were on a scale of 1
to 10. The MDS algorithim requires dissimilarity
judgements so the similarity ratings were converted
to dissimilarities for Experiment 1. However, for Ex-
periment 2, to avoid this extra conversion we had
participants directly rate how different pairs were.
The particular MDS algorithm used was the IND-
SCAL algorithm (Kruskal and Wish, 1978). It was
chosen since it provides a solution unique up to re-
flections as well as measurements of the variance
accounted for by the underlying dimensions. MDS
works by taking a set of distance measurements be-
tween pairs of items, that for our case were assumed
to be known up to their ordinal ranking, and com-
putes a multidimensional metric representation of
the items. This representation of the items is com-
monly termed the psychological space and gives in-
sight into the relationship among the items. One
important consideration with the use of MDS is that
it doesn’t automatically provide an interpretation of
the underlying dimensions. For this further investi-
gation is required to find properties that can explain
the placement of items along the dimensions.
4-1 Experiment 1 - Pairwise ratings of the
humanoid and CG character
The goal of Experiment 1 was to see what underlying
variables could be found to explain the pattern of
participants’ similarity ratings. In addition, since the
human movement provides a standard of comparison
we can examine results from the viewpoint of how
the other movements are placed around the human
movement.
Design and methods. A total of 20 volunteers
participated in the study with half performing sim-
ilarity ratings on the humanoid while the other half
rated the CG character. For ratings of both the hu-
manoid and CG character, trials were separated into
two blocks of 105 trials based on the two different
movements. Participants performed the two blocks
of trials in a single session separated a by a brief rest.
Results. Similarity judgments were converted
into dissimilarity values and entered into the IND-
SCAL algorithm for MDS. A first pass examined the
number of dimensions appropriate for modeling the
results and found that 2 dimensions was optimal in
the sense that increasing the number of dimensions
did not appreciably increase the variance accounted
for by the solution. In Table 1 we present the re-
sults of the fits of the INDSCAL algorithm, includ-
ing stress, r-squared - amount of variance accounted
for, and the importance measures of the 2 dimen-
sions. The importance of the dimension corresponds
to the amount of variability in the data accounted
for by that particular dimension and are ordered so
that the first dimension corresponds to the dimen-
sion accounting for the larger amount of variance.
The results of the 2D solutions are shown in Fig-
ure 1. Taken together, the results displayed in Table
1 and Figure 1 indicate that although stress and r-
squared measures reveal that a good fit is obtained
to observers’ judgments it appears that this solution
is dominated by the presence of outliers. This is ev-
ident in the plots of movement (a) where for move-
ment (a) production technique MT is very far away
from all the other production techniques for both
the humanoid and the CG character. Technique MT
yields different results most likely because MT op-
timizations over slow movements can produce unex-
pected types of motion (ie pendulous swing for slow
point-point movements). A similar situation is ob-
served for movement (b) where production technique
EPH is far away from other production techniques