Behavior-Based Indoor Navigation - Ambient Intelligence for Scientific Discovery - page 273

trained MLP network tell us the adequate parameter values for each distance. To
do so, we placed the robot at different positions and considered a valid tilt angle
if there existed a zoom range for which the MLP was able to robustly identify
the panel. Although the collected data did not show a precise linear relationship
among the parameters, we approximated the curve using a linear function with
bounded minimum and maximum values. Figure 4 shows the ideal relationship
between the distance to the wall and the tilt angle ( g ( x )), the collected data and
the selected linear approximation ( f ( x )).
With the dynamic tilt adjustment behavior running on the robot, we collected
data for the zoom values in a similar manner. Letting the robot select the tilt
angle and changing the zoom value using uniform intervals, we captured the
MLP results for each zoom value and the sonar measured distance to the wall.
Using, as centroid, the mean value of the acceptable zoom range for each
measured distance, we interpolated the curve using a B-spline, a mathemat-
ical, parametric definition of a series of connected geometric curve segments,
belonging to the spline curve family extensively used in the CAD industry and
more widely in computing for 3D geometry generation and modeling [58]. Fig-
ure 5 shows the obtained curve. Moreover, the selection of middle points of valid
zoom ranges makes final behavior more robust as it allows a confidence interval
for noisy readings (see [8] for more details).

3.4 Behavior Coordination for Task Achievement

In previous sections we showed how the robot is able to perform collision-free
navigation in a privileged compass orientation and dynamic landmark detection.
In order to acquire favorable data from the environment (i.e detect emergency
exit panels from different viewpoints) the dynamic landmark detection system
was improved with an active camera head control.

tilt (deg) sonar/tilt relation

65

g(x) = atan(1650/x)

60

55

50

45

40

f(x) = 68.8 − 0.01861*x

35

30

25

collected data

20

15

500 1000 1500 2000 2500 3000

dist (mm)

Fig. 4. Sonar measured distance versus tilt angle.