Page 111 - Introduction to AI Robotics

94
3 Biological Foundations of the Reactive Paradigm
to use for different environmental conditions. Schema theory is expressive
enough to represent basic concepts like IRMs, plus it supports building new
behaviors out of primitive components. This will be discussed in more detail
in later chapters.
This alternative way of creating a behavior by choosing between alterna-
tive perceptual and motor schemas can be thought of as:

Behavior::Schema
Data environmental_state
Methods choose_PS(environmental_state)
perceptual_schema_1()
perceptual_schema_2()
motor_schema()

Arbib and colleagues did work constructing computer models of visually
guided behaviors in frogs and toads. They used schema theory to represent
RANA COMPUTATRIX the toad’s behavior in computational terms, and called their model rana com-
putatrix (rana is the classification for toads and frogs). The model explained
Ingle’s observations as to what occasionally happens when a toad sees two
flies at once. 33 Toads and frogs can be characterized as responding visually
to either small, moving objects and large, moving objects. Small, moving ob-
jects release the feeding behavior, where the toad orients itself towards the
object (taxis) and then snaps at it. (If the object turns out not to be a fly,
the toad can spit it out.) Large moving objects release the fleeing behavior,
causing the toad to hop away. The feeding behavior can be modeled as a
behavioral schema, or template, shown in Fig. 3.9.
When the toad sees a fly, an instance of the behavior is instantiated; the
toad turns toward that object and snaps at it. Arbib’s group went one level
7
further on the computational theory. They implemented the taxis behavior
as a vector field: rana computatrix would literally feel an attractive force
along the direction of the fly. This direction and intensity (magnitude) was
represented as a vector. The direction indicated where rana had to turn and
the magnitude indicated the strength of snapping. This is shown in Fig. 3.10.
What is particularly interesting is that the rana computatrix program pre-
dicts what Ingle saw in real toads and frogs when they are presented with
two flies simultaneously. In this case, each fly releases a separate instance of
the feeding behavior. Each behavior produces the vector that the toad needs
to turn to in order to snap at that fly, without knowing that the other be-