Maze Experiment #2 Methods
Maze Experiment #2
A complete description of this experiment was published in Presence 7(2), 129-143. A
postscript file of the manuscript can be downloaded here Figures are available here. Or view the .pdf file here:
Method
Subjects. A total of 125 people (61 men and 64 women) participated
in the
experiment. 82 of the participants were undergraduates between the ages of 18 and 40
enrolled in an introductory psychology course at the University of Washington. These
people
participated in the experiment in return for extra credit in their class. For the
four weeks of the study during which the Psychology Department’s human subject pool
was unavailable, the remaining 43 participants were recruited through an advertisement
in the campus paper. These recruited participants were between the ages of 18 and 50
and were paid an hourly rate for their participation.
Materials and apparatus.
The real world environment was a 14’ x 18’ maze with movable 7’ tall black curtains,
configured as shown in Figure M3-1. The curtains hung from a rectangular grid of
cables each spaced two feet apart. The ceiling of the maze extended approximately
eight feet above this grid, allowing ambient light into the maze. At four locations,
a large stuffed animal was suspended from the grid, four and a half feet from the
ground, along with a large cardboard numeral. The numerals helped indicate to the
subject the correct path to take through the maze as well as making communication
about the maze’s locations easier. In this way, the numerals and stuffed animals
served as prominent landmarks in the maze.
Figure M3-1
For the virtual portions of the experiment, the room described above was modeled using
WorldUp by Sense8 Corp on a Pentium Pro 200 with an Oxygen 102 graphics accelerator
card (see fig M3-2). At seven locations in the virtual maze we placed red
directional arrows at eye-level. The arrows indicated the correct route to take
through the maze and they also helped to make the space more interpretable for
inexperienced users. A Thrustmaster PFCS joystick provided four degrees of freedom
of navigational control (three degrees of translation plus the ability to pan one’s
viewpoint). Subjects in the immersed condition used a VR4 HMD from Virtual Research
and had additional navigational (and viewpoint) control with a six degree of freedom
head tracker (Polhemus Fastrak). Depending on the location of the user in the VE,
the virtual scenes for both desktop and immersed conditions were rendered between 8.0
and 11.1 frames per second with a mean of 10.0. In addition to the virtual maze room,
a large maze-like virtual environment was created in which participants learned the
rudiments of navigation in a virtual environment. The computer on which the virtual
portions of the experiment were conducted was in the same room (though in a separate
area) as the actual real world maze.
Figure M3-2
Procedure All participants were initially given the Guilford
Zimmerman
standardized
test of spatial orientation ability. They were then given a short task that
familiarized them with wearing a blindfold while walking around a practice maze. This
task gave the experimenter the opportunity to correct the participants of habits such
as walking too slowly and taking too small steps while blindfolded.
(Pilot studies
had shown that a short practice session with a blindfold helped to reduce the variance
in scores during the testing phase of the experiment.) The experimenter then
reconfigured the maze to the standard configuration (see figure M3-1) and allowed
the participant exposure to a version the maze according to their experimental
condition. Twenty participants were randomly assigned to each of the following six
exposure conditions:
1. Blind: Participants in this group were given no
exposure to the mazeroom.
2. Real: For each trial, participants in this group were
given one minute
in
which to explore freely the real world maze. At the beginning of the first trial, the
experimenter pointed out the appropriate route between each object and thereafter,
participants were given no information or advice but were allowed to wander through
the maze on their own.
3. Map: At each trial, participants in this group were
shown a map of the
maze and were asked to study it for one minute. At the beginning of the first study
session, the experimenter oriented the map for the participant and pointed out the
correct route to take through the maze.
4. VR-Desk: This group was given two minutes of exposure
to a virtual
replica
of the maze at each trial. (Previous experiments had shown that one minute of VE
exposure was not enough time to allow a person to navigate through the maze. A two
minute exposure period allowed most participants to navigate through the entire maze
on their first trial.) These participants were seated 24 inches from a 21-inch color
monitor which rendered the virtual scenes at 800 x 600 resolution (true color, 60 Hz
refresh). As with the real world group, participants were initially instructed which
way to go so that they could get to each location in order. The arrows in the virtual
maze also provided path information. The participants’ motion and viewpoint in the
virtual environment were controlled by the user with a joystick.
5. VR-Immersive: At the beginning of each trial, this
group was given two
minutes of exposure to the same virtual maze and were given the same advice as the
other groups on the route to take through the maze; however, they experienced it with
a VR4 head-mounted display (742 x 230 resolution, 60 degree field of view) and a six
degree of freedom tracker. These participants also controlled their motion and gaze
with the joystick.
6. VR-Long Immersive: This group was identical to the
immersive group;
however, at each trial, they were allowed five minutes of exposure time to the virtual
maze.
Prior to the maze exposure, all of the participants in the virtual conditions had been
given between 30 and 75 minutes of instruction and training in how to use the input
devices efficiently. A virtual practice world was used in which the elements of
navigation with the joystick (and tracker) were trained and practiced. After learning
the basic navigation skills, these participants were timed on a "virtual obstacle
course" that required extensive use of the elements of navigation and concentrated on
those that would be important for navigating through the maze room. Subjects were not
allowed to proceed with the experiment until they could complete the obstacle course
in under four minutes. Training time for participants thus varied depending on their
abilities. All but four people were able to complete the obstacle course in less than
four minutes. It was clear to the experimenter that the four participants who did not
complete the obstacle course had difficulty physically moving the joystick because of
its relatively high spring tension. These four people were randomly re-assigned to
one of the three non-virtual conditions.
After encountering either a virtual, a real, or a map version of the maze, subjects
were blindfolded and escorted to the beginning of the real world maze. They were then
instructed to touch each stuffed animal in order, as quickly as possible, while
minimizing the number of times that they hit the walls of the maze. As participants
went through the maze blindfolded, the experimenter timed them and counted how many
times they touched the walls. Participants were informed and continually reminded of
how they were being scored, and were asked to do their best to minimize their time and
touches of the walls (or "bump count"). This process of exposure to the maze followed
by a blindfolded walk-through task was repeated six times.
After the sixth exposure to the maze, the experimenter gave the
participant a
distracting task while he altered two of the curtains in the maze (see Figure M3-3).
This new configuration was identical to the one on which subjects had been trained
except that two of the possible three paths between the first and third stuffed
animals were now blocked. In the new configuration, both the most familiar and the
most direct path were blocked. The participant was instructed that his or her task
was no longer to touch each animal. Rather, the task was to go as quickly as possible
from the first stuffed animal directly to the third stuffed animal. When the subjects
discovered that the typical path between the first and third animals had been blocked,
they were forced to rely on their mental representation of the maze and integrate the
piece-meal knowledge they had acquired to that point. We refer to this task as the
"integration task." The experimenter recorded how long the participants took to
complete the integration task, how many times they touched the maze walls, and the
route(s) that they attempted to take.
Figure M3-3
Finally, participants were given a 30 question true/false test in which they
identified whether a given map of the room correctly represented a portion of the
maze. False items on this test (see figure M3-4a) were wrong because they either
showed an incorrect route between locations (see figure M3-4b) or an incorrect
relative position of locations (see figure 5b). Though both types of items required
configurational knowledge, we refer to the items showing possible routes between
object locations as ‘route items.’ Those items which showed only the relative
location of points we refer to as ‘survey items.’
Figure M3-4a
Figure M3-4b
Results
Across all trials, the partial correlation (controlling for subject) between the
participant’s time through the maze and their bump count was quite high (r(832) = .94,
p < .0005). Because of this correlation, all of the subsequent analyses will be
univariate, treating time through the maze as the dependent variable. None of the
results we obtain differ substantially when we consider bump count as an additional
(or sole) measure.
Figure M3-5 illustrates the effect of repeated exposures on time through the maze
for each experimental group. Differences in transfer of spatial knowledge between the
non-blind conditions were evaluated using a repeated measures fixed effects ANOVA,
treating trial number (1 - 6) as a within subject independent variable, experimental
condition (map, real, desk, short immersion, and long immersion) and gender as between
subject independent variables. Time through the maze was the dependent variable. In
summary, this analysis yielded a significant two-way interaction between condition and
trial (F(20, 286) = 1.84, p = .017) and significant main effects of trial (F(5, 86) =
31.73, p < .0005), condition (F(4, 90) = 2.73, p = .034) and gender (F(1, 90) = 18.75,
p < .0005). All other effects were not significant. We defer analysis of the gender
related effects to the section on individual differences.
Figure M3-5
Not surprisingly, subjects’ performance in all conditions improved steadily over
trials -- the main effect of trial was significant (F(5, 86) = 31.73, p < .0005).
More importantly, the rate of this improvement depended on the type of training the
participant had received, probably because participants in the different conditions
were converging towards the same asymptote. The interaction of trial and condition
was significant (F(20, 286) = 1.84, p = .017).
Early Learning: effects of immersion and maps
Table M3-1 shows mean times through the maze on the first two trials for each
experimental group. Participants who were allowed only one minute of exposure to the
real maze were able to traverse it blindfolded much faster on the first two trials
than those participants in the other conditions. On average, subjects in all VR
conditions performed worse in the initial trials (M = 270.51 s) than those people in
either the real world (M = 163.32 s) or map (M = 242.88 s) conditions. By the second
trial, only the group that was given a much longer training time in the immersive VE
was able to outperform participants trained on the map (Mlong immerse = 122.05 s; Mmap
= 191.70 s). The lag in performance for the participants who trained in most VE
conditions is partly responsible for the significance of the trial by condition
interaction. Statistical comparisons between immersed and non-immersed VE groups and
between VE training in general and map training are not significant over the first two
trials. The only significant Helmert contrast comparing group differences on the
first two trials is the one that compares the real world group with all other
non-blind groups. The difference between the real world group’s mean time and that of
the other non-blind groups over the first two trials was, with 95% confidence,
estimated to lie within the interval (14.76, 286.85).
Table M3-1
Later learning: asymptotic performance and the effect of long
immersion
Figure M3-6 illustrates the mean times for each group on the blindfold task after
the sixth training session. By the sixth trial, participants in the long immersive
condition outperformed those in the real world training group (Mlong immerse = 40.95;
Mreal = 56.5 seconds), although this difference is not significant. Participants who
trained in the other conditions converged on a somewhat worse performance. The
contrast comparing mean times for the real and long immersed condition with the two
other VE conditions is significant (t(43) = 3.22, p = .002).
Figure M3-6
The convergence of performance between the real and virtual groups by the sixth
trial cannot be attributed to the learning of the environment that occurs while the
subject is blindfolded. By the sixth trial, those subjects in the blind condition are
still performing significantly worse than the other groups. The contrast comparing
the times of the blind group on the sixth trial with those of the other groups is
highly significant (t(19) = 4.55, p = .0002).
Representation differences
Differences in mental representations after the sixth trial were measured by combining
the results of the integration task and the true-false questionnaire. The integration
task forced participants to access their mental representation of the maze by blocking
off the familiar path from one maze location to the other. In addition to recording
the time to complete the integration task and the number of times the participant ran
into the maze walls, the experimenter also recorded which alternate routes the subject
attempted to take. Two statistics were derived from the true-false questionnaires:
the total percent correct, and a "survey score" which was calculated by subtracting
the number of correct ‘route’ items (see figure M3-4a) from the number of correct
‘survey’ items (see figure M3-4b).
Representation differences were tested using a fixed effect MANOVA with gender and
experimental condition (real world, map, desktop, immersive, and long immersive) as
between subjects independent variables. Time on the integration task, bumps in the
maze on the integration task, whether the participant initially attempted the shortest
route, percent correct on the true-false test, and survey score on the true false test
were included as dependent variables. The analysis revealed significant main effects
of both condition (F(20, 253) = 1.90, p = .013) and gender (F(5, 76) = 2.58, p =
.033). The interaction of condition and gender was not significant. The most
sensitive predictor of group differences in spatial representations was whether the
participant initially attempted to take the shortest route in the integration task.
Follow-up univariate analysis on this variable revealed a highly significant effect of
training condition (F(4, 90) = 3.76, p = .007). The difference in means between these
groups is due primarily to the lower scores of the immersive conditions. On average,
participants in the immersed conditions chose to take the shortest route on the
integration task only 35% of the time, whereas, people in the other conditions
averaged 63%. This difference is significant (t(79) = 2.90, p = .005).
Individual differences
Most of our dependent measures varied reliably with gender. On average, in all
non-blind experimental groups, men outperformed women at the blindfold task (F(1, 90)
= 18.75, p < .0005). A gender effect was particularly strong for women who trained in
the three VE conditions. This trend is illustrated in figure M3-7. VE-trained women
performed significantly worse than men in the VE conditions (the 95% confidence
interval for the mean difference between VE men and VE women [assuming equal
variances] was 225.21 +/- 141.45 s ). They also performed significantly worse than
women trained in the real world (the 95% CI for the mean difference between these
groups [assuming equal variance] was 271.65 +/- 173.24 s). Moreover, there was not a
significant difference between women and men who trained in the real world.
Figure M3-7
Men also outperformed women on four of the five measures of spatial representation
(all but the survey score from the true-false questionnaire), and the MANOVA conducted
on our representation measures showed a significant effect of gender (F(5, 76) = 2.58,
p = .033). T-tests confirmed that men took less time to complete the integration task
(t(65) = 2.70, p = .009), touched the walls of the maze less frequently (t(65) = 2.40,
p = .020), and scored higher on the true-false test (t(91) = 2.39, p = .019).
The Guilford Zimmerman test of spatial orientation was moderately predictive of a
participant’s overall performance on the true-false test (r(95) = .44, p = < .0005);
however, it was not predictive of the survey score derived from the true-false test,
nor was it predictive of any of the behavioral measures of spatial knowledge.
[ return to top ]