Page 433 - Introduction to AI Robotics

416

a. 11 Localization and Map Making
b.
Figure 11.17 Example of a.) a global map constructed from previous readings and
b.) a new local observation that must be fit to the global map. Shaded elements in a.
represent possible matches to the shaded element in b.

process. This ignored the cost of feature extraction, however. Feature-based
algorithms were also better able to handle poor initial location estimates. So
if the robot was placed in an office building and told it was facing North
when it was facing South, it would be able to correct that error after it en-
countered one or more gateways.
An important point to remember about localization is that no technique
handles a dynamic environment. If there are people moving about, each lo-
cal update will be different and it may be next to impossible for the robot to
match the past and current observations. If the robot is localizing itself to an
a priori map, it cannot tolerate a large number of discrepancies between the
map and the current state of the real world. For example, furniture shown
in one place on the map but which is actually in another is hard to handle.
Likewise, a hallway in a hospital which is usually clear but suddenly clut-
tered with gurneys and equipment presents a challenge.

11.7.1 Continuous localization and mapping

In order to eliminate the problems with shaft encoders or other propriocep-
tive techniques, current localization methods use exteroception. Exteroceptive
methods involve the robot matching its current perception of the world with
its past observations. Usually the past observations are the map itself. Once
the true position of the robot is known with respect to the map, the current
REGISTRATION perception is then added to the map in a process often called registration.
As seen in Fig. 11.17, matching the current observation to past observations
is not as simple as it sounds. The robot has moved from a to b according to