Page 284 - Introduction to AI Robotics

267
7.4 Managerial Architectures
which form the behavior schema. The schemas themselves can consist of as-
semblages of primitive schemas, coordinated by finite state machines. Sche-
mas can share information, if necessary, through links established by the
Motor Schema Manager. Behaviors are not restricted to being purely reflex-
ive; behavior specific knowledge, representations, and memory is permitted
within the schemas. The motor schemas, however, are restricted to potential
fields.
The fifth subsystem, Homeostatic Control, falls into a gray area between
deliberation and reaction. The purpose of Homeostatic control is to modify
the relationship between behaviors by changing the gains as a function of the
“health” of the robot or other constraints. As an example, consider a plan-
etary rover operating on a rocky planet. The robot is tasked to physically
remove rock samples from various locations around the planet and deliver
them to a return vehicle. The return vehicle has a fixed launch date; it will
blast off, returning to Earth on a set day and time no matter what. Now, the
rover may be provided with default gains on its behaviors which produce
a conservative behavior. It may stay two meters away from each obstacle,
giving itself a wide margin of error. At the beginning of the mission, such a
conservative overall behavior appears reasonable. Now consider what hap-
pens towards the time when the return vehicle is set to launch. If the robot is
near the return vehicle, it should be willing to shave corners and reduce the
margin by which it avoids obstacles in order to ensure delivery. The robot
should be willing to perform the equivalent of sacrificing its own existence
for the sake of the mission.
The issue becomes how to do homeostatic. Many aspects of AuRA are
motivated by biology, and homeostatic control is no exception. Rather than
put a module in the deliberative portion to explicitly reason about how to
change the overall behavior of the robot, biology suggests that animals sub-
consciously modify their behaviors all the time in response to internal needs.
For example, an animal who needs food becomes increasingly focused on
finding food. Human behavior changes in response to insulin. Fortunately,
changing the emergent behavior is straightforward in a potential fields rep-
resentation of behaviors, since the output vector produced by each behavior
can be scaled by a gain term. Returning to the case of the planetary rover
rushing to make the last delivery, the gain on the move-to-goal behavior at-
tracting the rover to the return vehicle should start going up, while the gain
on the avoid-obstacle behavior should start going down as a function of time
to launch.