Page 161 - Designing Autonomous Mobile Robots : Inside the Mindo f an Intellegent Machine

Chapter 10

Therefore, uncertainty should never be reduced below the magnitude of the current
correction or the minimum uncertainty threshold, whichever is larger.

Learning to be accurately uncertain

If our uncertainty estimates are too pessimistic, they will potentially allow bad data
to be accepted into the position estimates. If the uncertainty estimates are too opti-
mistic, then they may cause true navigational corrections not to be believed. Thus,
the better the uncertainty model, the tighter our filter, and the better the navigation
should be.

Unfortunately, things change. We may find the perfect parameters for uncertainty at
installation, only to find them less accurate months or years later. Gears wear, carpet
wears, road conditions change, etc. For this reason, it is very useful to learn the proper
error factors as the robot drives. This learning can most easily be accomplished by a
program that statistically studies the operation of the robot and the fixes it accepts.
If these statistics indicate that corrections are beginning to regularly approach the
magnitude of the uncertainty, it may be useful to learn new error factors.

Uses of uncertainty

The importance of the uncertainty estimate cannot be overstated. In the coming chap-
ters, we will see that this estimate is crucial in acquiring and processing navigation
data. It is also important in controlling the behavior of the robot in such a way as to
assure safe operation.

144