Evaluation of Reactive Architectures

Evaluation of Reactive Architectures

As seen by the follow-corridor example, the two styles of architectures are
very similar in philosophy and the types of results that they can achieve.
Essentially, they are equivalent.
In terms of support for modularity, both decompose the actions and perceptions
needed to performa task into behaviors, although there is some disagreement
over the level of abstraction of a behavior. Subsumption seems to
favor a composition suited for a hardware implementation, while potential
fields methods have nice properties for a software-oriented system.
The niche targetability is also high for both, assuming that the task can
be performed by reflexive behaviors. Indeed, the use of direct perception
emphasizes that reactive robots are truly constructed to fill a niche.
The issue of whether these architectures show an ease of portability to
other domains is more open. Reactive systems are limited to applications
which can be accomplished with reflexive behaviors. They cannot be transferred
to domains where the robot needs to do planning, reasoning about
resource allocation, etc. (this led to the Hybrid Paradigm to be described in
Ch. 7). In practice, very few of the subsumption levels can be ported to new
applications of navigating in an environment without some changes. The
different applications create layers which need to subsume the lower layers
differently. The potential fields methodology performs a bit better in that the

designer can create a library of behaviors and schemas to choose from, with
no implicit reliance on a lower layer.
Neither architecture presents systems which could be called genuinely robust.
The layering of subsumption imparts some graceful degradation if an
upper level is destroyed, but it has no mechanisms to notice that a degradation
has occurred. The finite statemechanisms in the docking behavior show
some resilience, but again, only for situations which can be anticipated and
incorporated into the state diagram. As with animals, a reactive robot will
also do something consistent with its perception of the world, but not always
the right thing.
Summary
Under the Reactive Paradigm, systems are composed of behaviors, which
tightly couple sensing and acting. The organization of the Reactive Paradigm
is SENSE-ACT or S-A, with no PLAN component. Sensing in the Reactive
Paradigm is local to each behavior, or behavior-specific. Each behavior has
direct access to one or more sensors independently of the other behaviors. A
behavior may create and use its own internal world representation, but there
is no global world model as with the Hierarchical Paradigm. As a result,
reactive systems are the fastest executing robotic systems possible.
There are four major characteristics of robots constructed under the Reactive
Paradigm. Behaviors serve as the basic building blocks for robot actions,
even though different designers may have different definitions of what a behavior
entails. As a consequence of using behaviors, the overall behavior
of the robot is emergent. Only local, behavior-specific sensing is permitted.
The use of explicit representations in perceptual processing, even locally, is
avoided in most reactive systems. Explicit representations of the world are
often referred to as maintaining the state of the world internally, or internal
state. Instead, reactive behaviors rely on the world to maintain state (as exemplified
by the gripper controlling whether the robot was looking for soda
cans or for the recycling bin). Animal models are often cited as a basis for
a behavior or the architecture. Behaviors and groups of behaviors which
were inspired by or simulate animal behavior are often considered desirable
and more interesting than hacks. Finally, reactive systems exhibit good
software engineering principles due to the “programming by behavior” approach.
Reactive systems are inherently modular from a software design
perspective. Behaviors can be tested independently, since the overall behav-

ior is emergent. More complex behaviors may be constructed from primitive
behaviors, or from mixing and matching perceptual and motor components.
This supports good software engineering practices, especially low coupling
and high cohesion.
The subsumption architecture is a popular reactive system. Behaviors are
purely reflexive and may not use memory. Behaviors are arranged in layers
of competence, where the lower levels encapsulate more general abilities.
The coordination of layers is done by higher layers, which have more specific
goal-directed behaviors, subsuming lower layers. Behaviors within a layer
are coordinated by finite state automata, and can be readily implemented in
hardware.
Potential fields methodologies are another popular reactive system. Behaviors
in potential field systems must be implemented as potential fields.
All active behaviors contribute a vector; the vectors are summed to produce
a resultant direction and magnitude for travel. Pfields provide a continuous
representation, which is easier to visualize than rule encoding, and are
continuous. The fields can be readily implemented in software, and parameterized
for flexibility and reuse. The vector summation effect formalizes how
to combine behaviors, eliminating issues in how to design behaviors for subsumption.
The fields are often extensible to three dimensions, adding to the
re-usability and portability. In the example in this chapter, behaviors using
potential fields were able to encapsulate several layers in subsumption into
a set of concurrent peer behaviors with no layers. Ch. 5 will give examples
of how to sequence, or assemble, behaviors into more abstract behaviors.
Despite the differences, subsumption and potential fields appear to be
largely equivalent in practice. Both provide support for modularity and
niche targetability. The ease of portability to other domains is relative to
the complexity of the changes in the task and environment. Neither style
of architecture explicitly addresses robustness, although in theory, if only a
higher layer of a subsumption system failed, the lower layers should ensure
robot survivability.