Information Technology Reference
In-Depth Information
methods to select the winner. Either a fixed prioritized scheme is defined as in the
subsumption architecture; or a behavior activation level can also be used (action
selection proposed by Maes [46]). An alternative is to use a voting mechanism
[62] where behaviors give votes to a set of actions and the action which obtains
the most votes wins.
An alternative to competitive methods are cooperative strategies, where the
responses of the different behaviors all contribute to the overall response, gener-
ally by means of a weighted vectorial sum [6]. The final output is not the output
of a single behavior but the combination of the outputs of the different active
ones.
Pirjanian [60] makes a slightly different taxonomy. He also points out two
main mechanisms: arbitration or action selection and command fusion. Arbitra-
tion mechanisms are “priority- or state-based” and “winner-takes all”. Command
fusion implies a cooperation among the behaviors in the selection or determi-
nation of the output, whether the final output is a single behavior output or a
vector summation.
2.2
Probabilistic Map-Based Navigation
Among the different approaches to robot control, classical deliberative sense-
plan-act (SPA) control is based on a sequential decomposition of the cognition
process in three basic steps: sensing the environment and matching the sensed
state in a stored centralized world model, planning according to the state, and
acting [56]. In this decomposition of intelligence, the physical agent and the in-
teraction with the environment were put aside. When BB approaches emerged,
they opened the way for the development of the three-layered architectures [29]
that look for the commitment between the two previous strategies. This hy-
brid approach incorporates a layer for basic interaction with the environment
and, therefore, includes embodied and situated agents with interaction capa-
bilities. Like the classical SPA strategy, it continues to rely on a centralized
world model. It is exactly through the use of a centralized world model that the
classical SPA nowadays evolved to the so-called hybrid and the behavior-based
approach. Relying on this global world model, the robot navigation problem is
mainly concerned with answering three questions [45]: Where am I? Where are
the other places regarding me? How can I reach them? These questions give rise
to the three main areas of robot navigation: environment mapping, localization
and planning. Since the late eighties, the two main approaches to spatial mod-
els were metric and topological maps. Elfes [25] introduced Occupancy grids ,
nowadays known as probabilistic approaches, which are metric representations
extended with a certain probabilistic belief of being in a cell [65, 67, 48, 23]. Al-
though different techniques are constantly being developed to acquire the spatial
model[68],theyallmustcopewithincorrect odometry measurements and corre-
spondence problems between data recorded at different time points. At present,
there are robust methods for mapping static environments and those of lim-
ited size, however, mapping unstructured, dynamic or large-scale environments
largely remains an open research problem.
Search WWH ::




Custom Search