Page 240 - Introduction to Autonomous Mobile Robots

225
Mobile Robot Localization
(
p il) directly using a model of the robot’s sensor behavior, its position , and the local
l
l
environmental metric map around .
The sensor model must calculate the probability of a specific perceptual measurement
given that its likelihood is justified by known errors of the sonar or laser rangefinder sen-
sors. Three key assumptions are used to construct this sensor model:
1. If an object in the metric map is detected by a range sensor, the measurement error can
be described with a distribution that has a mean at the correct reading.
2. There should always be a nonzero chance that a range sensor will read any measurement
value, even if this measurement disagrees sharply with the environmental geometry.
3. In contrast to the generic error described in (2), there is a specific failure mode in ranging
sensors whereby the signal is absorbed or coherently reflected, causing the sensor’s
range measurement to be maximal. Therefore, there is a local peak in the probability
density distribution at the maximal reading of a range sensor.
By validating these assumptions using empirical sonar trials in multiple environments,
the research group has delivered to Rhino a conservative and powerful sensor model for its
particular sensors.
Figure 5.24 provides a simple 1D example of the grid-based Markov localization algo-
rithm. The robot begins with a flat probability density function for its possible location. In
other words, it initially has no bias regarding position. As the robot encounters first one
door and then a second door, the probability density function over possible positions
becomes first multimodal and finally unimodal and sharply defined. The ability of a
Markov localization system to localize the robot from an initially lost belief state is its key
distinguishing feature.
The resulting robot localization system has been part of a navigation system that has
demonstrated great success both at the University of Bonn (Germany) and at a public
museum in Bonn. This is a challenging application because of the dynamic nature of the
environment, as the robot’s sensors are frequently subject to occlusion due to humans gath-
ering around the robot. Rhino’s ability to function well in this setting is a demonstration of
the power of the Markov localization approach.

Reducing computational complexity: randomized sampling. A great many steps are
taken in real-world implementations such as Rhino in order to effect computational gains.
These are valuable because, with an exact cell decomposition representation and use of raw
sensor values rather than abstraction to features, such a robot has a massive computational
effort associated with each perceptual update.
One class of techniques deserves mention because it can significantly reduce the com-
putational overhead of techniques that employ fixed-cell decomposition representations.