Page 304 - Introduction to AI Robotics

287
7.9 Interleaving Deliberation and Reactive Control
matches in planning and reaction times are no longer a compelling reason to
enforce a strict separation of deliberation and reaction. However, the soft-
ware engineering reason for the partition remains: things which operate on
symbols and global information should be in the deliberator; things which
operate directly on sensors and actuators should be in the reactor.
The above example of Navigation illustrates a top-down interleaving of
deliberation and reaction. The deliberative layer(s) decompose the mission
into finer and finer steps until it arrives at a set of behaviors capable of ac-
complishing the first subgoal. Another example of interleaving deliberation
and reaction is the role of the deliberative layer(s) in supplying the reactor
with expectations and virtual sensors. And it is important to note that delib-
eration can be triggered bottom-up as well.
It is not always the case that the deliberator generates the set of behaviors,
turns them on, and lets them execute until they either complete the subtask
or fail. In the case of sensing, it might be desirable for the deliberator to make
some of the global models being constructed available to the behaviors. For
example, consider a robot vehicle following a road. Now suppose that it is
looking for the large tree, since it is supposed to turn right at the intersec-
tion after the tree. In order to maintain the correct sequence of behaviors,
the deliberator has a global world model where trees are noted. The reactive
behaviors of following a road and avoid an obstacle do not need the explicit
representation of a tree. But what if the tree casts a shadow which might con-
fuse the follow-path behavior and cause it to give the robot incorrect steering
commands? In this case, it would be useful if the information about the pres-
ence of a tree could be absorbed by the behavior. One way to do this is to
permit methods on the world model to act as virtual sensors or perceptual
schema. Then the follow-road behavior could use the virtual sensor, which
tells the behavior when the road boundaries extracted by the vision sensor
are probably distorted, to ignore affected regions in the image that would
normally be observed by the vision sensor. This is an example of how the
SELECTIVE ATTENTION deliberative layer can aid with selective attention, or filtering of perception for
a context.
But the reactor may wish to trigger deliberation under many circumstances.
Most of these cases involve failure or exception handling. Suppose that a
sensor breaks and the perceptual schema cannot produce a percept. The
schema cannot diagnose itself. Therefore, its failure would cause a delib-
erative function, such as the Sensing Manager in SFX, to be triggered. If the
Sensing Manager cannot find a replacement perceptual schema or equivalent
behavior, then it can no longer meet the constraints imposed by the Mission