Page 115 - Introduction to AI Robotics

98
3 Biological Foundations of the Reactive Paradigm
In order to simplify sensing, perception should filter sensing and consider
only what is relevant to the behavior (action-oriented perception).
Direct perception (affordances) reduces the computational complexity of
sensing, and permits actions to occur without memory, inference, or in-
terpretation.

Behaviors are independent, but the output from one 1) may be combined
with another to produce a resultant output, or 2) may serve to inhibit
another (competing-cooperating).

Unfortunately, studying natural intelligence does not give a complete pic-
ture of how intelligence works. In particular there are several unresolved
UNRESOLVED ISSUES issues:
How to resolve conflicts between concurrent behaviors? Robots will be re-
quired to perform concurrent tasks; for example, a rescue robot sent in
to evacuate a building will have to navigate hallways while looking for
rooms to examine for people, as well as look for signs of a spreading fire.
Should the designer specify dominant behaviors? Combine? Let conflict-
ing behaviors cancel and have alternative behavior triggered? Indeed, one
of the biggest divisions in robot architectures is how they handle concur-
rent behaviors.
When are explicit knowledge representations and memory necessary? Direct
perception is wonderful in theory, but can a designer be sure that an af-
fordance has not been missed?
How to set up and/or learn new sequences of behaviors? Learning appears to be
a fundamental component of generating complex behaviors in advanced
animals. However, the ethological and cognitive literature is unsure of
the mechanisms for learning.

It is also important to remember that natural intelligence does not map
perfectly onto the needs and realities of programming robots. One major
advantage that animal intelligence has over robotic intelligence is evolution.
Animals evolved in a way that leads to survival of the species. But robots are
expensive and only a small number are built at any given time. Therefore, in-
dividual robots must “survive,” not species. This puts tremendous pressure
on robot designers to get a design right the first time. The lack of evolution-
ary pressures over long periods of time makes robots extremely vulnerable
to design errors introduced by a poor understanding of the robot’s ecology.