Model-Oriented Architectures

Model-Oriented Architectures

Both the managerial and state-hierarchy styles of architectures evolved directly
from the Reactive Paradigm. The designers sought to addmore cognitive
functionality to the Reactive Paradigm. As such, managerial and statehierarchy
styles are more bottom-up in flavor, emphasizing behaviors or
skills as the basic building blocks. However, a new influx of researchers
from the traditional AI community, particularly Stanford and SRI, has en-

tered robotics and made a distinct contribution. Their architectures have a
more top-down, symbolic flavor than managerial and state-hierarchies. One
hallmark of these architectures is that they concentrate symbolic manipulation
around a global world model. However, unlike most other Hybrid
architectures, which create a global world model in parallel with behaviorspecific
sensing, this global world model also serves to supply perception to
the behaviors (or behavior equivalents). In this case, the global world model
serves as a virtual sensor.
The use of a single global world model for sensing appears to be a throwback
to the Hierarchical Paradigm, and conceptually it is. However, there
are four practical differences. First, the monolithic global world model is
often less ambitious in scope and more cleverly organized than earlier systems.
The world model is often only interested in labeling regions of the
sensed world with symbols such as: hallway, door, my office, etc. Second,
perceptual processing is often done with distributed processing, so that slow
perceptual routines run asynchronously of faster routines, and the behaviors
have access to the latest information. In effect, the “eavesdropping”
on perception for behaviors is an equivalent form of distributed processing.
Third, sensor errors and uncertainty can be filtered using sensor fusion over
time. This can dramatically improve the performance of the robot. Fourth,
increases in processor speeds and optimizing compilers have mitigated the
processing bottleneck.
Two of the best known model-oriented architectures are the Saphira architecture
developed by Kurt Konolige with numerous others at SRI, and the
Task Control Architecture (TCA) by Reid Simmons which has been extended
to do multi-task planning with the Prodigy system. The Saphira architecture
comes with the ActivMedia Pioneer robots.