Behavior-Based Indoor Navigation - Ambient Intelligence for Scientific Discovery - page 266

Self-localization consists of determining the robot's position within the model
starting from an initial unknown state. There are several approaches that try to
solve the localization task. Markov localization is a method based on an extension
of Hidden Markov models that maintain a belief over the robot global config-
uration space. It is based on the so named Markov assumption or static world
assumption that presumes current robot sensor readings rely only on current po-
sition. Markov localization is a passive probabilistic process that calculates the
probability the robot is at each state as a function of the acquired sensor model
and the cinematic model [18, 27, 57]. Due to the high requirements of updating
and maintaining the probability density function over the whole set of states, a
number of authors proposed localization methods based on particle filters, where
the a posteriori belief is represented by a set of particles together with an as-
sociated weighting factor of each particle, a discrete subset of the probability
distribution [61, 28, 70].
A classic approach for generating maps is based on Kalman filters [37].
Kalman filter-based mapping algorithms are often referred to as SLAM algo-
rithms. SLAM (Simultaneous Localization And Mapping) or CML (Concurrent
Mapping and Localization) is not a solution itself, but a problem concerned with
building the map while jointly computing the robot's localization [68]. The cou-
pling of these two tasks should relieve the correspondence problem [44], which
is hard to solve when mapping and localization are tackled separately. It is per-
formedintwomainsteps:anexploration phase to reach different places and
location revisiting for consistency that can also drive the robot to new unvisited
locations [74, 33].
Planning is a traditional field of Artificial Intelligence (AI) and many algo-
rithms and techniques that cope with problems as different in nature to robotics
and manipulators as graphics animation or non-invasive surgery have been devel-
oped [41, 38, 42, 66]. However, planning in dynamic environments often requires
re-planning due to changes in the robot-environment state [14].

2.3 Role of Machine Learning in Navigation

One of the characteristics of intelligent behavior is the capacity of adaptation.
The ability to learn about the environment has long been considered an impor-
tant characteristic of artificial intelligent systems. A number of mobile robots are
able to learn from their navigational experience. Automatic learning paradigms
are involved in the so called Machine Learning area, where a large number of
models are defined. Mobile Robotics widely apply Machine Learning techniques
for navigation [24]. Learning can be used for improving the interaction of the
robot-environment system [12].
Over the last few years, many algorithms have been developed for very diverse
problems, including natural language understanding, control, face recognition,
etc. Learning may improve the whole system performance in different forms: con-
cept generalization from multiple examples, past experience reuse, new concept
discovering for environmental natural landmark learning and, of course, behavior
implementation and coordination.