Page 76 - Introduction to AI Robotics

59
2.4 Representative Architectures
The Value Judgment module provides most of the functionality associated
with the PLAN activity: it plans, then simulates the plans to ensure they will
work. Then, as with Shakey, the Planner hands off the plan to another mod-
ule, Behavior Generation, which converts the plans into actions that the robot
can actually perform (ACT). Notice that the Behavior Generation module is
similar to the Pilot in NHC, but there appears to be less focus on navigation
tasks. The term “behavior” will be used by Reactive and Hybrid Delibera-
tive/Reactive architectures. (This use of “behavior” in RCS is a bit of retrofit,
as Albus and his colleagues at NIST have attempted to incorporate new ad-
vances. The integration of all sensing into a global world model for planning
and acting keeps RCS a Hierarchical architecture.) There is another module,
operator interface, which is not shown which allows a human to “observe”
and debug what a program constructed with the architecture is doing.
The standard was adapted by many government agencies, such as NASA
and the US Bureau of Mines, who were contracting with universities and
companies to build robot prototypes. RCS serves as a blueprint for saying:
“here’s the types of sensors I want, and they’ll be fused by this module into a
global map, etc.” The architecture was considered too detailed and restrictive
when it was initially developed by most AI researchers, who continued de-
velopment of new architectures and paradigms on their own. Fig. 2.8 shows
three of the diverse mobile robots that have used RCS.
A close inspection of the NHC and RCS architectures suggests that they
are well suited for semi-autonomous control. The human operator could
provide the world model (via eyes and brain), decide the mission, decom-
pose it into a plan, and then into actions. The lower level controller (robot)
would carry out the actions. As robotics advanced, the robot could replace
more functions and “move up” the autonomy hierarchy. For example, tak-
ing over the pilot’s responsibilities; the human could instruct the robot to
stay on the road until the first left turn. As AI advanced, the human would
only have to serve as the Mission Planner: “go to the White House.” And so
on. Albus noted this and worked with JPL to develop a version of RCS for
NASREM teleoperating a robot arm in space. This is called the NASREM architecture
and is still in use today.

2.4.3 Evaluation of hierarchical architectures
Recall from Part I that there are four criteria for evaluating an architecture:
support for modularity, niche targetability, ease of portability to other do-
mains, and robustness. NHC and RCS both provide some guidelines in how