Page 354 - Introduction to AI Robotics

337
9.4 Associative Methods
the vehicle is in OR1, the landmarks are found in the same order if the ro-
bot always turns in the same direction (either always clockwise or always
counter-clockwise): building-mountain-tower. If the vehicle is in OR2,the
landmarks will be tree-building-intersection.
An interesting attribute of orientation regions and landmark pair bound-
aries is that the vehicle can directly perceive when it has entered a new
orientation region. For example, as the robot moves from OR1 to OR2,the
building-tower landmark pair boundary is in front of it, then is on both sides
of the vehicle, and is finally behind it. The transition is a perceptual event,
and does not require knowledge about the distance of the landmarks, just
that the relationship of the landmarks to the robot has just changed.
The use of orientation regions allows the robot to create an outdoor topo-
logical map as it explores the world, or to localize itself (coarsely) to a metric
map. The robot does not have to be able to estimate the range to any of the
landmarks.
How the robot navigates inside an orientation region has the flavor of vi-
sual homing. The robot may need to retrace its path as precisely as possible
through an orientation region. (A military robot might need to return with-
out being seen.) Without knowing the range to the landmarks bounding the
orientation region, the robot is helpless. But if it has to remember the angles
to each landmark every n minutes, it can move to follow the angles. A set of
angles remembered at a point along the path is called a viewframe.
One amusing aspect of the viewframe approach is that it assumed the ro-
bot had cameras literally in the back of its head. Unfortunately the ALV
vehicle did not; all the sensors faced forward. Daryl Lawton tells the story
of trying to convince the union driver of the vehicle to stop every 10 feet
and do a 360 turn so he could record the viewframes. After much plead-
ing, the driver finally honored the request, though it wasn’t clear if he ever
understood why the crazy scientist wanted a high-tech truck capable of au-
tonomous navigation to go in circles every 10 feet!
Associative methods are interesting because of the tight coupling of sens-
ing to homing. The image signature and viewframe concepts do not require
the robot to recognize explicitly what a landmark is, only that it is percep-
tually stable and distinguishable for the region of interest. Unfortunately
associative methods require massive storage and are brittle in the presence
of a dynamic world where landmarks may be suddenly occluded or change.
Of course, this is more problematic for indoor environments than outdoor
ones.