Behavior-Based Indoor Navigation - Ambient Intelligence for Scientific Discovery - page 264

consist of determining and maintaining a trajectory to the goal [47]. The main
question to be answered for navigation is not Where am I? but How do I reach
the goal? and the answer does not always require knowing the initial position.
Therefore, the main abilities the agent needs in order to navigate are to move
around and to identify goals. Neither a centralized world model nor the position
of the robot with respect to this model needs to be maintained. Biorobotics, de-
fined as the intersection between robotics andbiology,isanemergentfieldwhose
aim is for biology to contribute to robotics and vice versa [72, 52, 10]. Different
authors [71, 47] classify biomimetic navigation behaviors into two main groups:
1. Local navigation strategies are local control mechanisms that allow the agent
to choose actions based only on its current sensory input. There are four
strategies that fall in that group: search, path integration, taxis and goal
orientation.
2. Way-finding methods are responsible for driving the agent to goals out of the
agent's perceptual range that require recognition of different places and re-
lations among them. They also rely on local strategies. Perception-triggered
response, topological navigation and terrain inspection are the three main
way-finding strategies mentioned in order of complexity.
The BB approach to robot navigation relies on the idea that the control problem
is better assessed by bottom-up design and incremental addition of light-weight
processes, called behaviors, where each one is responsible for reading its own
inputs and sensors, and deciding the adequate motor actions. There is no cen-
tralized world model and data from multiple sensor do not need to be merged
to match the current system state in the stored model. The motor responses of
the several behavioral modules must be somehow coordinated in order to obtain
valid intelligent behavior. As mentioned before, way-finding methods rely on lo-
cal navigation strategies. How these local strategies are coordinated is a matter
of study known as motor fusion in BB robotics, opposed to the well known data
fusion process needed to model data information 1 . The aim is to match sub-
sets of available data with motor decisions; outputs of all the active decisions
somehow merge to obtain the final actions. In this case there is no semantic
interpretation of the data if there is no behavior emergence.
Within this field, action coordination mechanisms can be classified into two
main branches: competitive and cooperative [5]. In competitive or “winner takes
all” strategies, active behaviors compete to reach the actuators and only the
output of the winner has an effect on the robot's behavior. There are different
1 Multi-sensor data fusion seeks to combine information from multiple sources to
achieve inferences that are not feasible from a single sensor. This task is not trivial
as sensor outputs often have overlaps and conflicts, their location is usually highly
distributed, and their configuration very dynamic. In addition, their performance
can vary with time. There are diverse techniques to implement data fusion models.
Most of them are based on classical statistics and have a mathematical background
which guarantee soundness. Hypothesis Tests, Principal Components Analysis and
Transformation, Kalman Filtering, Particle Filtering, etc. are widely used data fusion
methods. See [73, 63, 19] for good reviews of sensor data fusion.