Page 255 - Mechatronics for Safety, Security and Dependability in a New Era

Ch48-I044963.fm Page 239 Tuesday, August 1, 2006 4:04 PM
Tuesday, August
Page 239
1, 2006
4:04 PM
Ch48-I044963.fm
239
239

Temporal filter Fm
(a) Feature extractors (b) Tasks in which the robot learned feature extractors

Figure 2: Feature extractors which are generated in the past tasks

EXPERIMENT
Experimental Setting

We used a small mobile robot which is about 40 cm high and has a camera with a fixed orientation to
look ahead at the floor. The task is to move along a given path to a destination. The size of I o and I f
in pixels is 64 x 54 and that of I c is 8 x 6. We defined the dimension of a substate as m = 1. The
robot can move at a translational speed v and a steering speed a> independently. To reduce the
computation cost, we discretized the state and action space and calculated the probabilities. We set the
histoiy length to h = 10 . The thresholds are set to H lh = 0.4 and C lh = 0.8h.

The robot was given three feature extractors shown in Figure 2. Fs,Fc, and Fm are generated in
tasks A, B, and C, respectively. The feature extractors have the following characteristics.
• 3 x 3 spatial filter Fs: This type of filter calculates sum of weighted brightness values of nine
neighbouring pixels. The generated filter emphasizes and inhibits horizontal edge.
• Color filter Fc : This type of filter calculates sum of weighted red, green, and blue. The generated
filter inhibits red.
• Spatial filter Fm : This type of filter calculates sum of weighted past five images. The generated
filter emphasizes the current image and inhibits the past image.

Feature Extractor Selection
The task given to the robot is shown in Figure 3 (a). The robot moves to the front of the door and waits
for it to open. It moves to the destination after the door opens. The environment is the same as that of
tasks A, B, and C. We gave three episodes of successful instances (L = 234,254,233). After learning,
the robot divided all instances into 13 subsets using ISODATA algorithm.

Figure 3(b) shows the learned behavior and Figure 3(c) shows the selected feature extractors at each
time step. The selected feature extractors differ depending on the situation. The average number of
selected feature extractors per step is 1.57. Figure 3(d) shows the selected subsets of instances at each
step. When the robot could not choose an action from the selected subsets because of low reliability, it
used all instances to decide again. 9) in the figure shows the step when the robot could choose an
action from the selected subsets. It is verified that the robot accomplishes the task while selecting