Page 116 - Introduction to AI Robotics

99
3.7 Summary
Ch. 5 will provide a case study of a robot which was programmed to fol-
low white lines in a path-following competition by using the affordance of
white. It was distracted off course by the white shoes of a judge. Fortunately
that design flaw was compensated for when the robot got back on course by
reacting to a row of white dandelions in seed.
Robots introduce other challenges not so critical in animals. One of the
most problematic attributes of the Reactive Paradigm, Ch. 4, is that roboti-
cists have no real mechanism for completely predicting emergent behaviors.
Since a psychologist can’t predict with perfect certainty what a human will
do under a stressful situation, it seems reasonable that a roboticist using prin-
ciples of human intelligence wouldn’t be able to predict what a robot would
do either. However, robotics end-users (military, NASA, nuclear industry)
have been reluctant to accept robots without a guarantee of what it will do
in critical situations.

3.7 Summary

A behavior is the fundamental element of biological intelligence, and will
serve as the fundamental component of intelligence in most robot systems.
BEHAVIOR A behavior is defined as a mapping of sensory inputs to a pattern of motor
actions which then are used to achieve a task. Innate Releasing Mechanisms
are one model of how intelligence is organized. IRMs model intelligence at
Level 2 of a computational theory, describing the process but not the imple-
mentation. In IRM, releasers activate a behavior. A releaser may be either
an internal state (motivation) and/or an environmental stimulus. Unfortu-
nately, IRMs do not make the interactions between concurrent, or potentially
concurrent, behaviors easy to identify or diagram.
Perception in behaviors serves two roles, either as a releaser for a behavior
or as the percept which guides the behavior. The same percept can be used
both as a releaser and a guide; for example, a fish can respond to a lure and
follow it. In addition to the way in which perception is used, there appear to
be two pathways for processing perception. The direct perception pathway
uses affordances: perceivable potentialities for action inherent in the envi-
ronment. Affordances are particularly attractive to roboticists because they
can be extracted without inference, memory, or intermediate representations.
The recognition pathway makes use of memory and global representations
to identify and label specific things in the world.
Important principles which can be extracted from natural intelligence are: