Abstract
The Subsumption Architecture is a special case of behavior based
control for robotics. Behavioral modules are added as "layers" with each
layer performing a complete behavior. Higher level behaviors override
lower level ones by taking control of their effectors or manipulating
their internal states. The control layers are built up out of finite
state machines connected by links that act essentially like wires. To
test this architecture on a reasonably complex problem, a prototype
airplane controller was developed. This controller flies a simulated
aircraft from take-off to landing and was run on a "C" based
implementation of the subsumption architecture. Several lessons where
learned from this effort. The subsumption architecture as currently
defined is not sufficiently modular. A clean interface between different
behaviors would be desirable. And finally, a more general relationship
than strict hierarchy between high level and low level modules is
required. None of these problems is insoluble within the behavior based
approach but all must be solved if realistic problems are to be dealt
with. Some candidate solutions are given.
Abstract
The recent increase in the use of range images may suggest the
revision of some of the techniques developed for intensity images so that
they adapt to range images more effectively. An important topic is image
registration. A scheme is developed to register range images in an
environment where distinctive features are scarce. When each image
overlaps with several other images, the registration must also be
performed at the global level. This is particularly challenging because
of the possibility of bending and compression in some forms of range
images (i.e., the relative position of data points on the image reference
surface may be inaccurate). The "primitives" used for local registration
are contours of constant range, which are extracted from data and are
represented by means of a modified chain code method. All "best" matches
of pairs of contours are considered tentative until their "geometrical"
implications are evaluated and a consistent majority has emerged. To do
global registration, a cost function is constructed and minimized. Terms
contributing to the cost include violation of local matches as well as
compression and bending in range images. In cases where there is no
appreciable compression and bending in the images, the proposed global
scheme could improve the quality of local registration by enforcing
consistency among them. In particular, we have implemented this scheme to
map the floor of the ocean, where the range data is obtained by a
multibeam echosounder system installed aboard a sailing ship producing
multiple overlapping range images. The system that we have developed is
the first automated system for correctly registered mapping of the ocean
floor; it is efficient and robust.
Abstract
A new kind of feature extraction operator for range images is
introduced that facilitates object recognition in several ways. It
consists of three points in 3-space fixed at the vertices of an
equilateral triangle and one or more curves, called test curves, fixed in
the reference frame of the triangle. This mathematical structure is then
moved as a rigid body until the vertices all lie on the surface of some
range image or modeled object. The point(s) of intersection of the test
curve(s) and the surface are used to define local shape features which are
invariant under rigid motions. These features can be used to
automatically find distinctive regions at which to begin recognition, to
rapidly screen candidate modeled objects for a match, and to speed pruning
in the generation of interpretation trees. Tripod operators are
applicable to all 3-D shapes, and reduce the need for specialized feature
detectors.
Abstract
The tripod operator is a class of feature extraction operators for
range images which facilitate the recognition and localization of objects.
It consists of three points in 3-space fixed at the vertices of an
equilateral triangle and several curves, called test curves, fixed in the
reference frame of the triangle. This mathematical structure is then
moved as a rigid body until the three vertices lie on the surface of some
range image or modeled object. The point (s) of intersection of the test
curve(s) and the surface are used to define local shape features which are
invariant under rigid motions. These features can be used to
automatically find distinctive regions at which to begin recognition, to
rapidly screen candidate objects for a match, and to speed pruning in the
generation of interpretation trees. Tripod operators are applicable to
all 3-D shapes, and reduce the need for specialized feature detectors. A
key property is that they can be moved on the surface of an object in only
three DOF (like a surveyor's tripod on the ground). Consequently, only a
3-dimensional manifold of feature space points can be generated, for any
number of test curves. Thus, objects can be represented compactly, and in
a form allowing fast matching. They are used here to characterize objects
by generating a cloud of points in feature space for each object by random
placement of the operator. Then new feature measurements are made by
operator placements in a range image containing one of those objects.
Using a simple nearest-neighbor approach, we determine which objects are
rejected and which remain as recognition candidates. Experiments were
performed using this approach in order to measure the discriminating power
of tripod operators.
Abstract
Techniques are described for the automatic diagnosis of primarily
analog systems. These results arose from several years of work at NRL in
this area, along with a fully implemented research prototype diagnosis
system, FIS (Fault Isolation System). Key features are a local
qualitative causal model of replaceable module behavior, the absence of
the single fault assumption, a rigorous probabilistic treatment of fault
probabilities, dynamic best test selection based on heuristics or entropy,
and efficient algorithms for computing the probability and the entropy of
Boolean expressions.
Intelligent Decision Aids
Intelligent M4 Systems
Machine Learning
Neural Networks
Back to NCARAI Library