Page 74 - Introduction to AI Robotics

57
2.4 Representative Architectures
NHC has several advantages. It differs from Strips in that it interleaves
planning and acting. The robot comes up with a plan, starts executing it,
then changes it if the world is different than it expected. Notice that the de-
composition is inherently hierarchical in intelligence and scope. The Mission
Planner is “smarter” than the Navigator, who is smarter than the Pilot. The
Mission Planner is responsible for a higher level of abstraction then the Nav-
igator, etc. We will see that other architectures, both in the Hierarchical and
Hybrid paradigms, will make use of the NHC organization.
One disadvantage of the NHC decomposition of the planning function is
that it is appropriate only for navigation tasks. The division of responsibili-
ties seems less helpful, or clear, for tasks such as picking up a box, rather than
just moving over to it. The role of a Pilot in controlling end-effectors is not
clear. At the time of its initial development, NHC was never implemented
and tested on a real mobile robot; hardware costs during the Hierarchical
period forced most roboticists to work in simulation.

2.4.2 NIST RCS

Jim Albus at the National Bureau of Standards (later renamed the National
Institute of Standards and Technology or NIST) anticipated the need for intel-
ligent industrial manipulators, even as engineering and AI researchers were
splitting into two groups. He saw that one of the major obstacles in apply-
ing AI to manufacturing robots was that there were no common terms, no
common set of design standards. This made industry and equipment man-
ufacturers leery of AI, for fear of buying an expensive robot that would not
be compatible with robots purchased in the future. He developed a very de-
tailed architecture called the Real-time Control System (RCS) Architecture to
serve as a guide for manufacturers who wanted to add more intelligence to
their robots. RCS used NHC in its design, as shown in Fig. 2.7.
SENSE activities are grouped into a set of modules under the heading
of sensory perception. The output of the sensors is passed off to the world
modeling module which constructs a global map using information in its
associated knowledge database about the sensors and any domain know-
ledge (e.g., the robot is operating underwater). This organization is similar
to NHC. The main difference is that the sensory perception module intro-
duces a useful preprocessing step between the sensor and the fusion into a
world model. As will be seen in Ch. 6, sensor preprocessing is often referred
to as feature extraction.