Page 350 - Introduction to AI Robotics

333
9.4 Associative Methods
rors could use multiple trips between nodes to build up a reasonable metric
map, since most of the errors would average out. Another attractive aspect
of the distinctive place approach is that it supports discovery of new land-
marks as the robot explores an unknown environment. As long as the robot
found something distinctive that it could reliably localize itself to, it could be
put on a topological map. Then as it repeatedly moved to it, the robot could
construct a metric map.
Returning to the discussion of landmarks, it should be noticed that a land-
mark must be unique to a node pair. There can’t be any corners in the real
world that are not on the graph between the nodes or else the robot will
localize itself incorrectly.
The distinctive place approach as originally formulated encountered some
problems when behavior-based roboticists began to apply it to real robots.
One of the most challenging problems was perception. Good distinctive
places are hard to come by; configurations that seemed useful to humans,
like corners, proved difficult to reliably sense and localize against. Features
that were easy to sense often were too numerous in the world, and so were
not locally unique. Another challenge was learning the local control strategy.
As the robot explored an unknown environment, it was easy to imagine that
it could find distinctive places. But how did it learn the appropriate local
control strategy? In an indoor environment, the robot might resort to always
using wall following, even though other behaviors would be better suited.
How would it ever try something different? Another open issue is the prob-
lem of indistinguishable locations. The issue of indistinguishable locations
has also been tackled to some degree by work with probablistic methods,
which will be covered in Ch. 11.

9.4 Associative Methods

ASSOCIATIVE METHODS Associative methods for topological navigation essentially create a behavior
which converts sensor observations into the direction to go to reach a partic-
ular landmark. The underlying assumption is that a location or landmark of
interest for navigation usually has two attributes:

PERCEPTUAL STABILITY 1. perceptual stability: that views of the location that are close together should
look similar

PERCEPTUAL 2. perceptual distinguishability: that views far away should look different.
DISTINGUISHABILITY