Expert Systems With Applications, Vol. 11, No. 2, pp.

109--124, 1996
Published by Elsevier Science lad
Pergamon Printed in Great Britain
0957-4174/96 $15.00+0.00
S0957.4174(96)00038-3

Artificial Intelligence Technologies for Robot Assisted Urban

Search and Rescue

JOHN G. BLITCH

Departmentof Mathand ComputerScience,ColoradoSchoolof Mines,Golden,CO 80401-1887, U.S.A.

Abstract--Structural collapse disasters routinely inspire sympathy not only for victims and their
families, but also for heroic rescue personnel who are faced with a tremendously complex, hazardous and
often frustrating task environment. Military operations and rescue activities in the aftermath of recent
earthquakes and bombings indicate a tremendous need for greater access to denied areas within any
crisis site involving collapsed structures. Recent developments in the remote inspection industry show
great potential for employment of small robotic micro-rover systems in expanded roles for Urban Search
and Rescue.
This paper discusses key issues in the application of robotic systems to Urban Search and Rescue
(USAR) activities and discusses ongoing development of a knowledge-based system for efficient
management of automated search assets. USAR modeling and "micro-bot" employment advantages are
addressed first, followed by a discussion of numerical method shortcomings in the context of search asset
allocation. KNOBSAR is then proposed as an initial expert system prototype designed to interact with
various structural collapse simulation packages and provide advice on search asset allocation to specific
entry points within a crisis site. KNOBSAR structure and design is then illustrated in terms of micro-bot
allocation scenarios to various collapsed structure entry points. The conclusion drawn from Hterature
review, experimentation and personal experience is that A! technologies in the form of robotic platforms
and decision support tools can have a tremendous impact on overall search efficiency for the USAR
community, and represent an important field of study for related military applications.

INTRODUCTION Navarro, 1995). A recent survey of developments in

mobile robot technology, however, revealed that a
THiS PAPEREXAMINESthe use of miniature mobile robots variety of small, lightweight and remarkably robust
(micro-bots) for expanded access to confined spaces and micro-rover systems have become commercially availa-
denied areas within collapsed structures. Because such ble for application to the challenging USAR task domain
systems come in a wide variety of design configurations, (Angle et al., 1995; Blitch, 1995).
extensive analysis is required to take full advantage of Practical experience in actual employment of these
their various employment capabilities. Since the USAR systems for search and rescue operations, however, is
community lacks experience in micro-bot platform virtually non-existent. The closest pool of related
employment, knowledge-based systems are proposed as expertise is most probably found in the power production
a method to assist in efficient management of these and and management industry, where various micro-rover
other valuable search assets. platforms are used for remote inspection of containment
The first practical proposal for use of miniature robotic vessels and processing systems. Angle indicates that
systems in Urban Search and Rescue activities took place small robotic systems can be invaluable for conducting
during victim recovery operations in the aftermath of the inspections of critical nodes which happen to be located
Oklahoma City bombing (Blitch, 1995). Earlier con- in hazardous or confined areas, and discusses AI-based
sideration of mobile robots for USAR work involved behavioral control as a method to achieve more reliable
systems which were too bulky and unreliable for results (Angle et al., 1995).
effective employment in structural collapse situations. With the assumption that micro-rover technology is
Although these early systems proved to be somewhat eventually applied to the USAR task domain, we suggest
immune to hazardous conditions, they offered little the need to capitalize on related expertise within the
advantage over fully protected firefighters, and proved to remote inspection industry, and distribute advice to crisis
be impractical in demonstration scenarios (Mills & site managers in the form of a Knowledge-Based System

109
 !10 J. G. Blitch

(KBS). This prompts us to examine the factors which mount. After shocks and high winds can create further
make micro-bots appropriate for USAR, and consider casualties via falling debris or structural collapse.
key resource management factors which make expert Hygroscopic materials never intended for exposure to
systems desirable as intelligent assistance tools. precipitation absorb fluids and add stress to already
weakened structures. Each step of a 250 lb fireman or
100 lb search dog can set an entire floor crashing down
THE USAR CHALLENGE
upon the levels below, threatening not only other
Disasters which occur in urban areas often result in workers, hut potential survivors as well, This problem is
extended death tolls and horrifying stories of people compounded by the density of sharp, twisted objects
buried alive under massive rubble piles created by which tend to protrude from rubble in all directions. The
collapsed structures. Search activities are impeded by a increased potential for tangled equipment and injured
myriad of hazardous conditions and confined spaces that limbs exacerbates the problem, and requires additional
deny access to voids which may contain survivors. A protective gear for search teams.
wealth of sensory distractors such as random noise, With a variety of complicating factors, such as those
scattered clothing and false odors further exacerbate the encountered after the Oklahoma City bombing, USAR
problem of discriminating between potential survivors, operations are often interrupted or discontinued until an
false readings and actual dead bodies (Beckmann, acceptable factor of safety can be restored to the site.
1991). This restoration task can be quite difficult in its own right
As crisis site managers try to sort through the stream as structural engineers struggle to analyze modified
of incomplete and often contradictory intelligence infor- design characteristics, and compute appropriate shoring
mation developed on site, they are faced with the points. But even with rapid shoring operations and
daunting task of allocating search and extraction assets merciful weather changes, precious time is lost for
from a constantly shifting resource pool in the face of survivors who cling to the hope of imminent rescue.
complicating factors such as weather effects and sub- Such factors weigh heavily in resource allocation
sequent threats from after shocks or secondary bombs. decisions. Crisis site managers must weigh risks at every
Crisis sites are often jammed with floods of media critical node to avoid unnecessary injury to rescuers as
personnel and good Samaritans pledging their support in well as victims in neighboring areas.
any way possible. Although this certainly offers a
refreshing testament to the concept of humanitarian good
will, it further complicates matters for managers who are Confined Workspace
faced with crowd control issues (Borden, 1994; Mills &
Survivors in structural collapse scenarios are most often
Navarro, 1995; O'Connell, 1993).
found in void spaces which are created by partial failure
Although technology has provided the globe with a
of support members. These void spaces vary extensively
wealth of advanced sensors and heavy lift equipment,
in size, location and volume as indicated in the 1989
potential survivors still remain trapped under rubble and
Oakland Freeway collapse (Borden, 1994; Mills &
collapsed structures for many days after a major disaster
Navarro, 1995; O'Connell, 1993). This situation created
strikes. What makes USAR so difficult? The answer lies
voids anywhere from 4 in. to 5 ft in height. Locating and
in analysis of crisis site characteristics, and the con-
searching these void spaces lies at the heart of the USAR
straints they impose on rescue teams. What might
effort to localize survivors. It is absolutely essential for
initially be viewed as a simple dig and sift task
crisis site managers to identify potential voids and
representation quickly builds in complexity as one
survivor densities in order to allocate rescue assets
considers the complicating factors which confront rescue
effectively for work in confined spaces. Unfortunately,
teams who converge on the crisis site in a desperate
human and canine skeletal structures are not conducive
effort to save lives.
to such operations. Their size and joint configurations
restrict maneuverability and dexterity of sensory organs.
Structural Instability
Despite the recent advances in design of structures and
Fire and Hazardous Materials
related materials, the fact remains that the vast majority
of the world's buildings are unable to withstand even The persistent earthquake activity which has occurred all
moderate forces resulting from seismic activity or blast along the Pacific Rim punctuates the drastic effect that
effects. For every "safe" sky scraper such as the World collateral damage can have on disaster sites. Disrupted
Trade Center, there remain thousands of older structures emergency services and public utilities can allow raging
which are destined for the rubble heap in the face of such fires, toxic contamination and electric interference to
destructive influence (Borden, 1994; O'Connell, 1993). continue unchecked for extended periods. Even superbly
As rescue teams search damaged structures, the trained firefighters and rescue workers are unable to
concern for falling debris and worker safety is para- operate in such areas without an adequate supply of
 AI Technologies for Robot Assisted USAR 1I1

water, electrical power and protective clothing. The towing tubes or microphones back down to survivors in
search for survivors often has to wait until collat~i'al order to provide sustainment (in the form of air, water,
threats can be controlled or suppressed enough for rescue medicine, food, communications, etc.) to victims in the
teams to gain access to critical areas. Even when interim between localization and actual extraction. Such
adequate protection and resources are available, the contact can have a tremendous effect on survivors' will to
sensory deprivation caused by thick masks, heavy gloves live and their ability to convey critical information about
and other safety gear can significantly impede a rescue other victims or crisis site characteristics.
team's search process. The emergence of a relatively new industry in remote
inspection platforms has created a variety of small robots
capable of providing extensive survey coverage of many
MINIATURE ROBOTS FOR USAR denied areas. The primary motivation for this industry
The primary advantages realized through employment of comes from the need to inspect containment vessels and
micro-bots as search assets are explored in Blitch (1995). conduits at nuclear power facilities and chemical produc-
To summarize this document, small robotic platforms are tion plants. This has also inspired technology transfer
valuable to the USAR community mainly because they from several related areas, such as NASA's Mars and
can provide access to areas otherwise denied to human Lunar rover programs (Angle et al., 1995). Conse-
and canine search assets due to confined space, structural quently, several families of very capable micro-rover
instability or the presence of hazardous materials (HAZ- systems are now commercially available, ranging in cost
MAT). They achieve this primarily as a result of: from $7k to $18k (see Figs 1--4).
Robotic systems are often proposed as a method by
which to remove humans from dangerous environments.
- - a significantly reduced footprint (both in volume This endeavor, although quite noble and logical from a
and mass) humanitarian standpoint, often meets strong resistance
--relative immunity to hazards from work groups who take great pride in performing
--flexibility of design. dangerous and complex tasks under pressure. Robots can
often be seen as a direct threat to those who depend on
Some systems can also provide polymorphic articulation dangerous work for higher wages and elevated self-
of drive systems and self-illumination as well. Such esteem. Highly publicized failures, such as the Dante
features allow these systems to be: project, have also left many with the impression that
semi-autonomous robotic systems are simply not adap-
tive enough to provide any significant advantage over
--logistically self-sufficient (due to low power con-
well protected human counterparts (Mills & Navarro,
sumption and size)
1995; Miller, 1994; Murphy & Blitch, 1995; Sinden,
------easy to transport or reposition within a crisis site
1990).
--relatively inexpensive to procure and maintain.
The USAR task domain offers an excellent opportu-
nity to explore this issue. By emphasizing micro-bot
When combined with "standard" robot advantages of platforms which can provide access to areas otherwise
24 h operations, relative immunity to fatigue and denied to human or canine search assets, robots are
diversified sensing modalities, micro-bot systems can actually employed as tools by which human workers can
have a tremendous affect on the entire USAR effort. enhance their own performance. The heroic USAR
Even after survivors are located, micro-bots can perform rescue worker who pulls a baby from the rubble pile
critical assistance tasks by returning to an entry point and remains the focus of attention regardless of whether a

FIGURE 1. Inuktun's FIGURE 2. ISR's FIGURE 3. VIT's "Rower FIGURE 4. Inuktun's

"MVGTV". "Pebbles". Standard. "Scissors".
 !12 Z G. Blitch

robot actually located the survivor or not. If anything, computation. By minimizing or maximizing a variety of
micro-bot systems help to further glorify human and cost or utility functions, OR organizations and consult-
canine rescue assets, because they can help find victims ants have enjoyed enormous success in providing
quickly enough to allow their extraction as actual managers around the world with greatly expanded profit
survivors instead of dead bodies. margins (Woolsey & Maurer, 1991). As expected,
As indicated above, damaged structures and rubble military organizations in particular have reaped tremen-
piles create a variety of hazards and obstacles which can dous benefits from such efforts because of the inherent
significantly hamper the ability of human and canine magnitude of their management challenge. Even min-
search crews as they try to gain access and work uscule cost shifts or utility adjustments can have an
effectively in confined spaces. Since robots are not enormous impact on an army's overall efficiency and
"born" with any specific anthropomorphic configuration, combat readiness.
size or mass, we are free to create systems which are As with any problem-solving approach, however,
uniquely suited to USAR tasks. It must be stated from traditional Operations Research methodology (e.g. math-
the outset, however, that even as robots continue to ematical programming) suffers from significant
become more technically advanced and self-sufficient, limitations when confronted with non-numerical prob-
they will never be able to replace the need for human and lems. As Ren and Sheridan point out, such an approach
canine search teams. The medical complexities and dictates the necessity for a relatively hard or determi-
emotional aspects of USAR operations dictate that robot nistic model, that may not be appropriate for a particular
platforms will contribute mainly by offering the rescue task domain (Ren & Sheridan, 1995). Since a sizable
community yet another tool with which to perform their portion of real world decision-making deals with non-
noble and often heroic duties. deterministic environments, the search for crisply
The need for autonomy in USAR oriented micro-bot defined solution sets may not be cost effective or even
systems is established by concern for limitations inherent feasible. This is most certainly the case in the USAR
in tethered or tele-.operations. Traditionally applied arena where uncertainties and subjective evaluations
mobile robot systems, such as those commonly seen in form the basis for a large portion of operational planning
Explosive Ordinance Disposal (EOD) roles, normally activities. Consider the drastically oversimplified USAR
involve control schemes which depend on some sort of scenario in Fig. 5.
direct arrangement (via cable or radio link) between
human and machine. Many references effectively argue
EXAMPLE MATHEMATICAL FORMULATION
the case against these tele-operation control schemes,
FOR USAR
mainly through concern for freedom of movement
constraints, cable failure or entanglement and line of We are faced with an earthquake-induced collapse of a
sight range requirements for radio controlled systems in three story office building in San Francisco, California
non-homogenous environments (Blitch & Murphy, 1996; (where else?). Since the earthquake struck at 6:32 a.m.,
Miller, 1994; Murphy & Blitch, 1995; Sinden, 1990). the building was relatively empty. A structural engineer-
The complex nature of the USAR task domain, ing simulation package has identified three major
however, is probably too demanding to rely solely on a reinforced areas where void spaces are most likely to
single type of control scheme or design. A tethered exist in volumes that could support human life: a nursery,
system, for example, is less vulnerable to vertical drop, cafeteria and vault (see Fig. 5). Personnel unaccounted
and can be tracked or retrieved via cable tension upon for at the time of the earthquake are listed as potential
system failure or incidental collapse of void spaces. survivors. Four different types of micro-bots (with self
Thus, a diverse mix of several reliable systems would be explanatory operating manuals) have been flown in from
most appropriate for USAR. When combined with the around the country to assist in the search effort. We
complicating factors described above, however, this mix restrict the scenario to formulation of a single story
of potential search assets presents crisis site managers representation to promote simplicity and reduce explana-
with a daunting resource allocation task. tion requirements.
We must first analyze the situation and determine a
particular metric to manipulate. Since time is usually the
RESOURCE MANAGEMENT
most important factor in USAR operations, we use
The field of Operations Research (OR) has dealt with minutes of search time as a utility factor to minimize
extensive resource management challenges since its (realizing of course that an actual situation may involve
inception in the aftermath of WWII. Its traditional maximizing other variables such as survivor health,
military flavor has focused primarily on methods priority of extraction, etc.). We then decide upon a set of
designed to handle large quantities of personnel and variables by which to carry out our operation. Binary
material involved in combat support operations. This (on-off: 0-1), integer and real number variables are all
quantitative orientation has pushed OR development in a considered to allow essential flexibility in our modeling
mathematical direction with emphasis on numerical effort.
AI Technologiesfor RobotAssistedUSAR 113

Decision Variables time factor described above. This leads to a minimization

of:
Binary objects (0,1 indicates "on" or "off')

R~=robot i is on (= I), available to search, or off (=0),

unavailable D]j Dl2 Di3
[R,EP,VS,(~ )] + [R,EP,VS2(~ )] + [R,~P, VS3(~ )]+
EPj=entrypoint j is on (= l), selected for exploration,
or off (=0), waiting selection
VSk=voidspacej is on (= 1), selected for exploration,
or off (= 0), waiting selection.

Evaluation~costqualifiers
Dj,=distance from entry point j to void space k
SPi=speedof search in (fl/m in) based on robot i
mobility and sensor configuration
DI2 DI3
[R~EP,V$,(~ )]+[R2EP,V$,(-~ )]+[R:EP,VS,(~ )1+
Dj__~k=search time in minutes.
s/',

Objective Function
Brat
Since our objective is to minimize search time, we would [R~FP.VS,( -~ )] +
represent a global allocation effort as a series of robot-
entry point-void space combinations multiplied by the

INITIAL USAR STRUCTURAL COLLAPSE SCENARIO

Pre -Collapse
~ Nursery
x
o
o
Q
m
"o
g
o
.J

v E.P.4

/
H/W - 1.5' x O.S'
Oist to tit void ',85'
Wet = 0
Current Asset?
Post -Collapse L~,.s Scissors
H/W - 4.1' x 35'
E.P.2 DIIt to tgt valid - 70'
Wet = 0
H/W • 5.5' x 0.8' Currant A m t ?
Dist to tot void • 270' Human
E.P.5
Wet = 1Asset?
Current ~ 4/
Pebbles p H/W : 3.S' x 2.5 °
Dist to tilt void • 250
Wat-O
Current Asset?
Dog

H/W • 0 . 7 ' X 0,5' EPI ) E.P, II

DIst tO tilt void = 40'
Wet • 1 HNV • 0.4 ' x .65 '
RowerASSet?
Current ~ ~ Dlst to tgt void - 30'
Wet • 0, Hot Wires,
Cummt Asset?
MGTV

FIGURE 5. Simplified USAR scenario showing initial "best guess" robot allocation.
114 Z G. Blitch

or to be more succinct: Analysis of this formulation example reveals several

problems with the math programming approach to this
type of problem in general. Identification of descriptor
Minimize: ~ ~ ~ R~EPjVSk(~ )
i=Oj=O k=O variables, for example, can be difficult for characteristics
which are not precisely defined in the problem statement.
where Our asset search speed variable, for example, can be
quite ambiguous depending on client interpretation.
n=the number of robots with potential for allocation Does search speed translate to simple ground speed, or
m=the number of entry points available and are other environmental factors involved? What is the
t=the number of target void spaces. relative importance or weight applied to these factors?
Such subjective evaluation criteria force us to consider a
Constraints variety of time and effort intensive techniques to
translate real word characteristics into some sort of
Feasibility constraints are now added to force micro-bot numerical representation. This numerical representation
selection and check void dimensions versus robot may also incur some rather restrictive assumptions
characteristics. If micro-bot 2 is too wide to fit into entry required for even an initial attempt at feasible solutions.
point 3, for example, any solution with micro-bot 2 For our current situation we make the following
assigned to entry point 3 is infeasible. assumptions.

Constraint variables Constraint Assumptions.

Hi=height of robot i
(1) Each target void space is reachable from each entry
W~=width of robot i
point.
/-//=height of entry pointj
(2) Feasibility is based on physical dimensions.
Wj=width of entry point j
(3) Search time is based on a micro-bot's search speed
times distance required.
Constraint equations

Rj +R2 +R3.... +Rn >-- 1 Although such assumptions are normally appropriate
{at least one robot chosen} and valid for well-defined numerical optimization prob-
EP~ + EP2+ EP3.... + ESm~ 1 lems, they pose a significant threat to the accuracy of our
{at least one entry point chosen} USAR modeling effort. As the situation evolves, and
VSI + VS2 + VS3.... VSt~ I crisis site managers receive more detailed information
{at least one void space chosen} about the crisis site, we become more aware of other
(EPflj) - (RiHi)... >--0 complicating environmental factors as indicated in Fig.
{robot height less than EP height: VEP>O, R>0} 6. Incorporation of these factors into our model and
(EPjWj) - (giwi)...>-~O encoding their impact on the overall USAR effort
{robot height less than EP width: VEP>0, R>0} requires an extensive reformulation effort for application
of mathematical programming techniques.
This brings us to an initial formulation which looks Our original assignment of the Scissors platform to
like: EP 4 and Pebbles to EP 2 was initially feasible, for
example, because EP 2 was relatively dry. This solution
Ojk ) was also appealing from a more subjective standpoint
Minimize: ,=o
~" j=o
~ k=o
~ RiEPjVSk( ~ because Pebbles' semi-autonomous control structure
seemed more appropriate for EP 2's 270 ft target
Subject to: distance. This allocation would save wear and tear on
Scissors' tethered apparatus and reduce the likelihood of
RI+R2+R3....+R,>--1 entanglement over long distances.
{at least one robot chosen } Environment conditions and crisis site characteristics
EP~ +EP2+EP3....+ESm>--I change rapidly, however, and subsequent analysis now
{at least one entry point chosen } reveals that EP 2 is flooding due to rain and/or broken
VSI + VS~ + VS3.... VS,>- I plumbing. We suddenly realize that Pebbles is probably
{at least one void space chosen } not waterproof at all, and that MGTV is marginal. The
(EP:J'Ij) - (RiH~)...>--O Pebbles operator also loses his radio image downlink
{robot height less than EP height: VEP>0, R>0} within 2 rain of its employment because of the thickness
(EPjW~) - (R,W/)... of the rubble pile in the vicinity of EP 2. Structural
-->0 {robot height less than EP width: VEP>O, engineers then inform us that EP 4 is steeply canted,
R>0} EP 1 has a corrosive liquid leaking all over the place,
A! Technologiesfor Robot Assisted USAR 115
EVOLVED USAR STRUCTURAL COLLAPSE SCENARIO
(shows results of crisis site intelligence updates) v.,
HIW • 1.S'x 0.8'
Diet to tOt void -115'
E.P. $ Sulv. Prob. - 0 . 6
HIW • 4.1' X 3.5' Wet - 0
Dist to tilt void ', 70' Grade - Canted
SUN. Prob. '+ 0.7
Wet • 0 Thin walls I rubbh)'
Glade - Steep CurrentAsset?
Light • 7 Pebbles
E.P.2 Current Asset?
H/W " 5.5' x 0.8' Human
Dlst to tot void - 270'
Surv. Prob. : 0.3
Wet - 6
Grade • Fist E.P.5
Hight • 0
Current Asset? EP 3 EP 4 HIW = 3.5 'x2.5 '
Dist to tot void - 250'
Scissors
~SUlV. Prob. • 0.7
EP2 Wet- 0
Grade ,~gradual
Light • e
CurrentAsset?
Dog
EP 1 .....
E.P.t
H/W : 0.7' x 0.5'
Dist to tilt void : 4 0 ' ~ E.P. II
Surv. Prob - 0 . 3 H/W '~ 0.4 ' x .65 '
Wet - 5 DISI to tot void - 30'
Grade = Steep SUlV. Prob. • 0.8
Corrosive Liquid EP 6 Wet- 0
Light =, 0 Glide - grodual
Current Asset? Light,, 2
Hot Wires, Twisted,
Rower
Ckatorad
CumentAsset?
MGTV
F I G U R E 6. S u b s q u e n t U S A R scenario showing environment f l u c t u a t i o n s .

EP 1, 2 and 6 are almost completely dark and that EP 6 (EP.,Hj)- (R~H~)->O

is heavily cluttered with several hot wires which they axe {robot height less than EP height: VEP>O,
unable shut off from the vault's emergency power supply R>0}
(see Fig. 6). While these new conditions cause a series of (EPjWj) - (R,W~)~O
rather frantic phone calls to robot manufacturers to {robot width less than EP width: VEP>0,
discuss unanticipated system vulnerabilities and capabil- R>0}
ities, our immediate concern is how to re-allocate s2eejw'rj - 2_>0
resources to reflect these changes. {if robot 2 is on with EPj, then EPj
In addition to adding new variables to reflect wetness, wetness--<2VEP > 0 }
gradient, lighting and hazards, we must also figure out R , E P jWTj - 6_>0
how to encode and reflect their relative importance to our {if robot 1 is on with EPj, then EPj
final solution. Upon successful acquisition of new wetness -<6VEP>0}
system performance data and extended discussion of RsEP~Hj - 1>0
void attribute representation, we arrive at the reformula- {if robot 3 is on with EPj, then EPj electrical
tion indicated below: hazard is off, e.g. -<lVEP>0}
R2EPjCHj- 1->0
{if robot 2 is on with EPj, then EPj chemical hazard
Minimize: Y. ~. ~ R,EPyS~(
i=0 j=0 k=0
~) is off, e.g. -<lVEP>0}

where
Subject to: C~j=cost in speed degradation (0 < C~< 1) for robot i
due to obstacle clutter in entry point j
R~+R2+Rs....+R,->I WTj=wetness factor (integer) for entry point j from
{at least one robot chosen } 0 = dry to 10 = flooded
EP2+EPs....+ESm>--I EHi=electricity hazard factor (integer) for entry point
{at least one entry point chosen} j from 0=low to 10=high
+ VS2 + VS3.... VSt>>-I CHj=corrosion hazard factor (integer) for entry point
{at least one void space chosen} j from 0=low to 10=high.
116 J. G. Blitch

After adding variables and changing integer values, our tions (Wu & Coppins, 1981).
computer most likely chugs through another set of Another significant concern for this particular problem
multiple IP/LP iterations and finally spits out a new involves computational complexity. The use of mixed
allocation listing for our beloved micro-bots. While integer formulation is necessary to accurately model the
reading this list to a skeptical technical search team scenario with certain objects "turned" on or off, and real
leader, however, the Colorado USAR TF commander number variables used as evaluator tools for these actors.
arrives with new information. She informs us that Woolsey and Maurer (1991) point out however, that
debriefings from survivors who made it out before the mixed integer formulations are notoriously difficult and
building collapsed indicate a high probability of survi- computationally expensive because they tend to involve
vors (including 2 pregnant females) in the nursery area. recursive Linear Programming (LP) oscillations which
The "chief" has designated this as the highest priority of continue somewhat randomly until integer values are
all target void spaces. He wants us to get our most obtained. This creates a tedious, unpredictable and
effective micro-bot systems down to EP 3 or EP 4 to help resource intensive solution cycle that can extend from to
out the human rescue team which got stuck in a crevice. hours to days even on large "super computer" platforms.
The technical search team leader poses the tongue in Although continued progress in parallel computing and
cheek question: "Now which robot is best for this advanced algorithm development methods may offer
situation?!" After several groans and shrugs, we start the some relief to such challenges in the future, access to
whole analysis and reformulation process over again to related hardware and software products will probably not
reflect some sort of priority variable in our objective be forthcoming in the USAR community for quite some
function such as: Pk=void space priority (integer 0--10) time.
reflects survivor probability expected in a particular void In spite of their widely enjoyed success in a variety of
space (via scent detection, noise, crisis site intelligence, OR application arenas, therefore, the numerical program-
key leader intuition, etc.) which prompts increased ming approach suffers from extensive limitations when
priority of localization/extraction for that area due to: applied to non-deterministic task domains which are not
crisply defined or dynamically oriented. Since the USAR
(1) Suspected victim health condition
environment involves subjective decisions made in the
(2) Ease of extraction
face of fluctuating criteria, these limitations are encoun-
(3) Victims' knowledge of crisis site/other potential
tered often. This examination of USAR related factors,
survivor locations.
therefore, indicates that current mathematical program-
ming techniques are inadequate for time-sensitive
The formulation exercise illustrated above reflects the assistance in search asset resource allocation, and
dynamic nature of this particular task domain and inspires us to consider methods that are more symbolic in
presents a significant obstacle to effective application of nature.
math programming methods. Since crisis site character-
istics are in a constant state of flux, variable coefficients RELATED AI TECHNIQUES
and constraints are likely to undergo radical adjustments
Although Artificial Intelligence represents a relatively
regularly. Weather effects or after shocks that alter void
new field of computer science, its basic concepts are well
dimensions and hazard characteristics, for example, can founded in the historical origin of human thought. Many
make previous micro-bot allocation solutions inefficient
chose to regard human cognition in terms of domain
or even infeasible. Discovery of new voids, imposition of specificity. Harmon and Sawyer, for example, separate
search priorities, or receipt of more detailed intelligence "shallow" reasoning methodologies which are domain
will likewise affect solution development through addi- specific, from "deep" reasoning techniques that tend to
tion or deletion of new variables. be more generalized in nature (Harmon & Sawyer,
A wealth of experience in effective OR application to 1990). The shallow approach is further described as
challenging problems can provide senior consultants heuristic oriented and experience dependent, while deep
with a powerful capacity for clever constraint formula- reasoning involves the application of principles and
tions. Even with this expanded ability to manipulate
axioms to new situations (see Fig. 7). The generalized
variables, however, the extended effort required to
nature of the math programming approach discussed
encode and then explain the validity of a particular
above prompts the continued examination of deep
numeric representation to key personnel can significantly reasoning methods for USAR resource allocation.
reduce a client's confidence in the final solution.
Although Wu and Coppins indicate that detailed sensitiv-
State Space Search
ity analysis and goal programming techniques can
accommodate these variations to a certain degree, the An initial attempt at application of AI technologies to
vast majority of crisis site fluctuations will require an this task domain would most likely involve some sort of
entire reformulation of the problem with significant time state search. The first problem encountered with this
lost in analysis and encoding of new variable representa- methodology involves the complexity of the state
AI Technologies for Robot Assisted USAR I17
Domain and
Performance Theodes

Domain ~ ~
Dependent ~ Surface
Facts Knowledge

q~.~,\~Os~ Shallow Reasoning AI:

~~ t~°~'
KBS and Industrial Robotics

CompiledKnowledge
No
Knowledge
~ ~ Leer.,_
~-"m8 tro~ 8
Deep Reasoning Ah
Neural Networks and Autonomous
I:ID~--~--- ~¢h°°lo o- Robotics
_,_ ,,v~r la....~ - "na aOo,.
Domain - rclAt,.,._ ~ -=
independent "t"IPL~'_o ~ Deep
Definitions
. . . . . . . . .
-~:, ~1~,_ ~
"a/O~,/. ,.~""--~
Knowledge

' a's

General Theories
FIGURE 7. General reasoning development diagram.

representations themselves. Since search assets will need when considering the addition of symbols or values for
to be located in three-dimensional space and associated each combination of entry point, void space and micro-
with a wide variety of crisis site characteristics, an bet. With the consideration of essential descriptive
effective state representation is bound to be quite large. factors that impact on allocation decisions (wetness,
Consider the following allocation state representation hazards, weather, lighting, etc.) the state representation
which is based on the variables previously established in requirement itself becomes difficult and unwieldy. Even
our Integer Programming formulation: with the development of simplifying assumptions and
efficient management of this enormous state space search
Initial state:. requirement, computational resource limitations inspire
our exploration of other methods to simplify the
problem. The pursuit of knowledge for the reduction of
state space search requirements proceeds along divergent
Ri=robot located at {x,y,z}, with characteristics
paths as indicated in Fig. 7. We maintain the momentum
{a,b,c ..... z} established in consideration of numerical programming
EPi=entry point located at {x,y,z }, with characteristics
with continued examination of the deep reasoning
{a,b,c ..... z} paradigm.
VSk=void space located at {x,y,z}, with characteristics
{a,b,c ..... z}
Tjk= time for robot i to search from EPj to VSk.
DeepReasoningThroughNeuralNetworks
The study of artificial reasoning has progressed largely
where through creation of mechanical devices since the time
man first counted with pebbles. As such, many of AI's
x,y,z...represent object locations in three dimensions
more interesting features have had to wait for concurrent
progress in computer hardware and microprocessor
a,b,c...represent void descriptors such as noisy, wet,
development in order to create effective applications.
dark, contaminated, twisted.., etc.
The connectionist approach, for example, suffered from
a sequence of ambitious false starts and political setbacks
Goal state:
until more effective platforms and marketing could spark
[Ri,EPj, VSt, combination with smallest T~] new interest in the form of back propagation neural
networks (Russel & Norvig, 1995).
This representation will require tremendous expansion By using combined weights between randomly acti-
 118 J. G. Blitch

vated nodes in an effort to mimic human thought, reasoning in non-deterministic environments because it
Artificial Neural Networks (ANNs) have proven to be a avoids the necessity for a rigidly defined problem
remarkably capable decision support tool, especially for formulation. By matching rule preconditions, operations
problems involving pattern recognition. Such systems and postconditions with successful sequences that have
process task requirements and input by comparing occurred in the past, a problem's state space is tremen-
cumulative weight arrangements with desired outcome, dously reduced (albeit with a corresponding loss of
in order to create an error representation. The error is generalization). This avoids the need for the tedious and
then propagated back through its node layers to modify resource intensive algorithmic processing required by the
connection strengths in an effort to learn a more deep reasoning approach in our integer programming
appropriate response. Further refinement of inherent example. It also reacts much more effectively to dynamic
decision surfaces allows the ANN approach to develop problem situations, because reformulation of an entire
deep reasoning capabilities that can often be generalized problem structure is no longer necessary.
to other task domains without a significant loss of Rule-based inference also allows for symbolic imple-
effectiveness. mentation which can represent outcomes (or
As a numerically oriented and generalized process, an post-conditions) in clear text verbiage as opposed to
ANN shares several of the features mentioned above numerical interpretation. This allows for development of
with math programming. Zak et al. point out advantages explanation subsystems which can provide a clear
gained through the use of ANN models for classic LP indication of where a particular reasoning sequence went
problems (Zak et al., 1995). The proposed models wrong as opposed to numerical methods which just crash
compete for early convergence time based on various or chug along forever in pursuit of NP complete
error computation functions and network design charac- solutions based on a certain arrangement of variables.
teristics. These efforts are certainly worthwhile in pursuit The corresponding modification or addition of a new rule
of increased computational performance, but do little to structure is therefore greatly simplified, and the system
mediate the other limitations associated with determi- develops an expanded capability to handle more diverse
nistic requirements indicated earlier. A neural network situations over time. In any case, such procedures are
decision support tool would still require, for example, a bound to be more flexible in the face of fluctuating
well-defined problem representation with relatively crisp environment features than reformulation in mathematical
input data. Interpretation of network output values also terms. The symbolic, clear text representation of recom-
incurs the same client translation necessity indicated for mended solutions also has tremendous appeal for a client
math programming methods. Although connection base that operates in the chaotic USAR environment,
weights can be modified fairly easily, an applied neural thus it possesses a clear advantage over numerical
network would most likely require frequent retraining methods.
and/or structural adjustments (architecture, number of It is evident that the shallow reasoning approach offers
nodes, layers, etc.) in order to respond to the fluctuating some distinct advantages over other methods for
USAR task domain and site characteristics. Although resource management in dynamic environments. Its
ANN analytic characteristics make them a good candi- associated cost in terms of generalization, however, can
date for hybrid system development, they are be significant if care is not taken to limit applications to
inappropriate for application as a stand alone USAR narrow domain sets. Systems which employ shallow
resource management tool. reasoning tend to be somewhat brittle if applied to areas
outside the specialized environment from which their
APPLICATION OF EXPERT SYSTEMS rules were extracted. In this sense, an expert system
would lack the common sense necessary to realize, for
The search for intelligent decision support assistance
example, that it just recommended a micro-bot for
now continues along the shallow reasoning path with
cardio-pulmonary surgery when fed medically-oriented
consideration of a more heuristic-oriented approach.
input parameters. While this is a significant limitation for
Since operations conducted in the USAR environment
many large scale problem sets, it is probably of minimal
are based extensively on experience gained from pre-
concern for crisis site managers who by nature maintain
vious disasters, shallow reasoning methods are initially
a tightly controlled focus on their own specialized search
more appealing than generalized algorithms. The follow-
and rescue task domain.
ing perspective on related expert system development is
undertaken with the intent to find appropriate models for
development of KNOBSAR, a resource allocation assis- Hybrid Systems
tance tool for USAR crisis site managers.
Continued progress in the area of overall systems
development has also led to a new breed of hybrid
Shallow Reasoning
systems which attempt to combine the effects of multiple
With the application of a heuristic-oriented rule base, an problem solving methodologies in a single application
expert system approach also allows for more effective package. Neural networks, for example, are often
 Ai Technologies for Robot Assisted USAR 119

included as an external or internal interface element to of essential experience pools caused through unantici-
provide effective analysis of input data for expert s~i~m pated workforce reti~ment, migration, personal injury or
processing. The inherent divergence of the deep and process bottlenecks. The intent was to either capture an
shallow reasoning approaches however, has been an expert's implicit knowledge (usually in the form of
obstacle to effective integration of such modules within generalized heuristics and case-based algorithms) prior
a single system. to his departure, or disseminate it to other apprentice
Ren and Sheridan, however, propose the use of fuzzy experts during periods of his unavailability (Harmon &
logic interface transformations to effectively navigate Sawyer, 1990).
this reasoning divergence barrier (Ren & Sheridan, In KBS development, the expert's knowledge is
1995). Their decision framework, called ADAUS for typically "captured" through the creation of an inference
Adaptive Decision Aiding Under Soft Boundaries, net composed of relatively simple if-then rules. These
makes use of both deterministic and non-deterministic rules are chained together to produce a somewhat
subsystems in an effort to provide the best of both shallow reasoning effect for a narrow task domain. The
representation strategies. The Classic Mathematical Pro- system processes input along the sequence of rules;
gramming (CMP) provides a "hard" or deterministic eventually arriving either at an appropriate answer, or a
component that is combined with a capacity for "softer" subsequent demand for additional input corresponding to
reasoning in the symbolic-oriented Linguistic Knowl- unsatisfied rules. In this manner, a KBS can be said to
edge Base (LKB). Integration of their respective efforts filter out relatively simple problems which follow
are provided through application of fuzzy constraints and established rules from those which demand a more
an Adaptive Calibration Module (ACM). extensive reasoning process. As the KBS operator is
Such an ambitious combination of diverse reasoning called upon to react to unsolved problems or suspicious
power would offer tremendous potential for development solutions, he conducts "how" and "why" traces through
of flexible and robust resource allocation tools that are the inference net to determine which rules require
appropriate for dynamic and complex environments. The additional input, modification or augmentation by condi-
USAR community, however, is in need of immediate tional subrules (Harmon & Sawyer, 1990).
assistance in resource allocation procedures, and cannot Validation and verification testing of these systems
afford to wait for further development of such innovative revealed an extensive capability for relief of critical
techniques. Thus, we propose the development of a expert bottlenecks. They also provided a residual training
simplified expert system prototype for immediate effect for expert apprentices because the KBSs could
resource allocation assistance, with an eye for future actually explain the how and why of each step they
evolution of related hybrid systems. follow enroute to answers. This explanatory character-
istic emphasizes one of the most significant advantages a
KNOWLEDGE-BASED METHODOLOGY KBS has over conventional computer programs: it
actually explains its reasoning process, and can be
The primary motivation for the development of expert modified directly by a user without extensive computer
systems came from a desire to distribute human expertise programming experience (Harmon & Sawyer, 1990;
that had been "captured" by a knowledge base in an Russel & Norvig, 1995).
effort to relieve overworked key personnel, or provide Other KBSs simply seek to ease the workload on
guidance in situations where they became unavailable. overtaxed experts who inadvertently create process
As one of the original founders of KBS theory, bottlenecks and reduce overall production. With the
Feigenbaum proposed that what really made a person recent expansion of laptop computing power, however,
"smart" in a productive sense, was the actual specialized KBSs can now be used to actually transport expertise to
knowledge he possessed, regardless of the methodology locations currently denied to an actual human expert.
or ease with which it was acquired (Feigenbaum et al., These factors all serve to expand KBS flexibility, and
1988; Barr et al., 1989). A pilot, for example, is effective make expert systems particularly useful in disruptive
at guiding his aircraft because of his acquired knowledge environments.
of how its control surfaces will make it respond in a
variety of aerodynamic conditions. Specialized knowl-
edge makes the pilot "smart" (e.g. effective) enough to KBS APPLICABILITY
fly safely--regardless of whether he took 6 months or 6 A KBS approach was chosen for the USAR asset
years to learn this particular skill. Hence, "the power of allocation problem for several reasons. First, the need for
an AI program is proportional to its knowledge rather rapid dissemination of knowledge oriented expertise is
than its reasoning ability". well established in the USAR community. Expert
The use of Knowledge-Based Systems to capture and availability in the USAR environment is not only limited
disseminate human expertise is most commonly seen in by bottlenecked processing and prolonged work shifts,
the industrial and financial communities. They were but by disruption of transportation and communication
primarily designed to compensate for the unavailability networks as well. The capability of a KBS to explain its
 120 J. G. Blitch

rule-based solutions, and filter out all but the most indicated earlier.
complex problems in a user friendly manner can Effective development of a KBS prototype depends
significantly reduce the time required to allocate upon simplicity and narrow focus in its initial design.
resources while also minimizing fatigue and bottleneck Focus is provided by limiting potential search assets to
effects for crisis site managers. the robotic systems listed above, and simplicity is sought
Second, the majority of USAR procedures are based through pursuit of a relatively straightforward rule-based
on heuristics that are implicitly developed through years system. This rule-based approach is further defined to
of experience and training as opposed to the strict use backward chaining for its inference sequence,
application of statistical formulas and rigid algorithms. because of its goal-driven nature. Since the system has
The complexities of structural collapse prediction fac- only four possible outcomes (recommend one of the
tors, combined with human behavior characteristics three available micro-bot systems to a void space or none
make rapid solutions from strict application of deep at all) the problem structure falls into the goal-driven
reasoning methods highly unlikely. Rescuers need the paradigm as indicated by Harmon and Sawyer (1990).
rapid, approximate solutions that are easily explained Hence the KNOBSAR prototype is designed with
and modified by Knowledge-Based Systems in a backward chaining in mind.
dynamic environment.
Third, the amount of intelligence data that must be Basic Structure
sorted, analyzed and prioritized in USAR scenarios
A schematic of KNOBSAR's general structure is shown
dictates that computer-assisted database management of
in Fig. 8. The system receives information about
some sort will eventually be required on site. Database
essential crisis site characteristics through routine inter-
management is an inherently critical part of KBS
face with a number of integrated databases which will
employment, because of its dependence on correlated
most likely store projected internal void characteristics
data, and its associated rule development and modifica-
from a variety of simulation packages and on-site
tion.
intelligence sources. As queries for advice about a
These factors have been recognized by a variety of
particular crisis site entry point are processed, the user
expert system designers in related asset allocation task
interface consults with the knowledge base to determine
domains. Amirkhanian and Baker (1992), for example,
applicable rules. The inference engine then "chains"
applied results of earlier work by Alkass and Harris
through these rules to arrive at a best guess for the
(1988) to develop an expert system for more efficient
optimal micro-bot to assign to that particular entry point.
allocation of earth-moving equipment in construction
Exceptions or partial solutions (e.g. allocation choice
projects. Although Aldershot (1994) refers to a variety of
narrowed to two or more micro-bots) are described to the
optimization oriented expert system applications, their
user through the explanation subsystem. This subsystem
use for allocation of miniature robots or other poly-
not only provides a suggested solution, but also the chain
morphic systems remains an unexplored area of
of reasoning by which it was obtained. In this manner,
research.
the user gains confidence in the system's allocation
process, and can readily correct logic deficiencies
KNOBSAR PROTOTYPE DEVELOPMENT through addition or modification of rule sequences.
By processing allocation queries in this manner,
Harmon and Sawyer separate KB systems into three
KNOBSAR actually provides a filtering effect for crisis
basic types: pure induction, simple rule-based and hybrid
managers who are overwhelmed with resource allocation
systems (Harmon & Sawyer, 1990). Induction systems
issues. Since it strains out solutions which can be
are valuable in the derivation of unspecified rules from a
obtained through simple rule chaining, key leadership
high density of accurate and complete data, which are
figures are left to concentrate on the more difficult
then used to reason about more general scenarios. Since
problems and exception cases which demand their
we have already established that the USAR task domain
personal attention. Guidance for exception handling is
is inaccessible and incomplete, this approach is inap-
incorporated into the system's knowledge base so that
propriate for our needs. Rule-based systems on the other
solutions are "grown" for future situations of a similar
hand, attempt to extract or derive rules from actual
nature. These characteristics illustrate KNOBSAR's con-
experts in the field through a series of interviews or other
tribution to greater efficiency in USAR asset allocation.
knowledge acquisition techniques. Such an approach
offers the best option for KNOBSAR development since
Assumptions
most crisis site managers primarily follow internally
constructed heuristics which are derived from a wealth of KNOBSAR's initial development is based on some
hard won experience. Hybrid systems, which offer a important assumptions. First, each of the assets is
combination of the two techniques, as well as potential considered in a binary integer fashion. Thus, degraded or
incorporation of other deep reasoning strategies, are partial asset capabilities are not currently factored into
rejected here in the interest of simplicity and speed as the allocation process. Second, the certainty of input data
AI Technologies for Robot Assisted USAR 121

~ tructurslCollatpse "~
imuletion& Mapping I
ackeges S
RuteBue Exceptions
Crisis Site Managers

Guidance for Rule Exceptions

rttical Crisis Site Data

\ ~\ntelligence
lC°llected Miorobot AIIooatloe i k -- -- - - Microbot AIIooetion

FIGURE 8. General KNOBSAR design and organization.

is assumed to be 100% for simplicity of operation within dialogue in order to capitalize on shared familiarity with
the knowledge base and inference engine. Thus, stochas- USAR procedures. Clarification of technical issues and
tic methodologies are expected to manage uncertainty as conflict between expert recommendations were handled
efficiently as possible within the simulation and mapping on a case by case basis with subsequent question and
packages indicated as part of KNOBSAR's input parame- answer periods.
ters. Further development of this complex issue is These methods resulted in the development of an
deferred for future versions and prototype modification. object-oriented knowledge representation by establishing
Although these assumptions may place KNOBSAR's object attribute value (OAV) triplets for objects and if-
recommendations somewhat short of "optimum" alloca- then rules for process concepts. The expert interview
tion advice, it is doubtful that the incomplete and process then correlated OAV triplets to arrive at simple
non-deterministic nature of the USAR environment will rules of thumb. These rules, or heuristics, were
ever provide enough information to approach actual sequenced and connected to form an initial inference
optimality. By using derived heuristics the system can network. A working inference engine prototype was then
provide a best guess at optimum advice in spite of such developed using the common generate, test and debug
assumptions, thereby establishing an effective solution procedures applied to an initial inference net. The
from which to progress as uncertainty is diminished. inference engine prototype will subsequently be
expanded through successive iterations of the verifica-
tion-validation-test cycle, in conjunction with expected
Knowledge Acquisition Process
user requirements.
The development of KNOBSAR's knowledge base Many KBSs are oriented on capturing the expertise of
depends primarily on a detailed review of related a single expert before he/she retires, or becomes
literature and interviews with five prominent USAR unavailable for guidance. This system, however, capital-
leaders and four remote inspection system designers izes on a pool of expertise from several different sources
from around the country. The author performed the vast within the USAR and Robotics communities and com-
majority of knowledge acquisition tasks in order to bines it with a variety of Operations Research
satisfy pedagogical objectives and take advantage of his management heuristics described by Woolsey and
familiarity with the task domain. Interviews with USAR Maurer (1991). Conflicts which arise within these
experts were conducted informally with a general topic respective expertise pools are resolved by rule segmenta-
122 J. G. Blitch

tion or referred to FEMA and the Center for Earthquake from the operator or from a variety of data bases. It
Preparedness for further mediation. provides search asset allocation advice in the form of
micro-bat ranking for particular voids. The narrow
System Inputs design approach indicated above prompts us to restrict
our resource allocation options to the four micro-bat
KNOBSAR inputs consist of descriptive information
systems described above. It makes use of the LEVEL 5
received directly through the user interface or imported
OBJECT development tool to capitalize on advantages
from various simulation-oriented databases as indicated
realized in object-oriented modeling combined with
in Figl 8. This information will normally consist of crisis
natural language processing for extended flexibility and
site characteristics such as void space dimensions,
programming simplicity. It also incorporates a visual
contaminants, location indices, etc. The system also
programming package for the design of a powerful, yet
prompts the user for additional inference-onented crite-
user friendly Graphical User Interface (GUI). KNOB-
ria such as essential micro-bot availability data and
SAR's initial design is being developed and tested using
which particular void space to consider for allocation.
an IBM PC laptop computer with a 486 DX 33 processor
Initial input screens will be based on a simplified
to emphasize its field versatility and ease of employment.
Graphical User Interface (GUI) as illustrated in Fig. 9. Such considerations are intended to enhance client
Subsequent development will incorporate user preferred acceptance and interest in further research.
GUI characteristics identified during the verification, System performance will be measured by the time
validation and test cycle. required for a user to obtain a solution and the correlation
of these solution sets with what actual experts would
System Outputs have recommended under similar circumstances. Valida-
Advice which is recommended by the system gets passed tion of the system will be conducted during selected
back to the user as a rank ordered set of micro-bots for training sessions with USAR task forces and will be
allocation to a specific void space. In order to maximize based on expected reductions in overall search time and
an effective user friendly approach, this information will asset allocation mistakes (inefficiencies). As the intended
be combined with the critical reasoning data and client for system development, specific system feedback
guidance prompted by the system input screen(s). As will be provided through demonstration to Mr Pete
with the input data displays however, this information Bakersky, the USAR director for FEMA's region VIII.
will initially take the form of a simplified GUI as
indicated in Fig. 10, and incorporate more elaborate SUMMARY AND MILITARY RESEARCH
features with subsequent development. CONTRIBUTIONS
In examining the USAR challenge, this paper has
Implementation and Testing
explored the application of AI technologies to a complex
KNOBSAR is designed around a user friendly interface and dynamic environment. By illustrating the value of
that accepts various crisis site information either directly semi-autonomous micro-bots as search assets, we hope
IIIHI Ilnl~'l I ]
Die OKi
S e a r c h A s s e t I r r [ b r m a t i o n DisplaTy c,,~-~,,, ~ s , t t,,~,
Aterg)ute Table for ALl SeL~ch AueW
I I 1 us^eta I ,,,. .i,. .,,:

~ .ng. [mathi~dthl ~- i hr. [

Void Attr~ume8 Table

Void ht width )wetness I grade !lighting I clu
I.D. gt.) [ (ft.) rank rank [ r a
I i

i4.1 Ste p .7 (1 |
EP4 ~1.5 !O.e iO [Canted i8 O
EPS 3., )2., 0 [Graduali8 [0 E
r . ' :2 ,i ° m

To have tire btferem~e FJq0ae evaluam poe~ible

choices: Select a Void be the tllde 8here and
click o n the "ALlocate A s s e ~ tl Void" ButtoR

FIGURE 9. KNOBSAR input screen showing data from simulation packages.

A! Technologiesfor Robot Assisted USAR 123

to inspire their employment as extremely capable, low- much more demanding and elusive problem: especially
cost search assets with a tremendous return ~on in terms of polymorphic platforms and obstacle negotia-
investment potential for future disaster situations. With tion. Thus, the development of knowledge-based
the proposal of the KNOBSAR expert system prototype, resource management systems represents a challenging
we further promote the use of AI methods to compensate and valuable field of study for future military opera-
for the resource management difficulties encountered in tions.
more generalized numerical problem-solving techniques. This paper's analysis of the USAR task domain reveals
An expert system application of specific domain knowl- a wealth of commonality with typical military opera-
edge for more efficient resource allocation also provides tions. Not only is the environment inherently unstable
an excellent modeling fit for structural collapse simula- and hazardous, but the need for rapid allocation of
tion and mapping products because of the synergistic resources by typically exhausted and overwhelmed
effect of their combination. By combining a reasoning leaders is common. The choice to put human or machine
methodology with a shared knowledge representation, in harms way serves to intensify the stressful and
KNOBSAR can maximize a modeling effort's impact by emotional context of such decisions. KNOBSAR illus-
providing valuable advice in a user friendly manner. trates the KBS role as an adaptive filter for tedious and
Military benefits derived from technology transfer of routine management issues, thereby relieving manage-
this research are plentiful. Virtually any discussion of ment bottlenecks and allowing key leaders to concentrate
military operations in the next century will involve on more complex and difficult decisions.
employment of unmanned systems to some degree. The Perhaps the greatest advantage of this approach lies in
same is already true for space exploration, mining and a the domain tailored degree of knowledge control. By
host of other challenging endeavors. Military operations acting as smart mechanical advisors, knowledge-based
in particular will benefit tremendously from the efficient applications like KNOBSAR can be a tremendous asset
management of these unmanned systems. Although the for complex combat decision analysis--without the
industrial robotic community has already accomplished threat of subjugating man to machine logic. Regardless
much in the way of process optimization, effective of the quality or quantity of advice, the human command
allocation of autonomous mobile robots represents a element always retains the final authority for a decision.

Untith~.d
File OK!
7-7-

Microbot Allocation Display -'7

Microbot Specifications Table

model num model name manufacturer! description Tgt. Void Tgt. Void
Feasibility Allocation
Rank

SC 1 Scissors Inuktun Micro-tracks TRUE 2

Pebb 1 Pebbles IS Robotics Semi TRUE 3
ROWER :tovvstdl VIT Sophisticate< FALSE 0

Model [Mvgtv ]

Model# [MVGTV2 C } ID of Target Void

Maaufaetaz~r [Inuktun ] I I
Variable geometry with very small
[)escr@tion payload. Not submersible,

FIGURE 10. KNOBSAR output showing allocation results screen.

 124 J. G. Blitch

As we move forward into the 21st century however, it Blitch J.G. (1995). AI applications for robot assisted urban search and
would be foolish to ignore the tremendous potential rescue. The AI Exchange, 9, 15-20.
value realized in machine-based reasoning tools. Only Biitch, J. G. & Murphy, R. R. (1996). Alternative scenarios for military
deployment of unmanned ground vehicles. Submitted to the 2nd
with continued examination of such technology can we Conference on Robots for Challenging Environments, Albuquerque
ensure progress toward optimal development the noble NM, June 1996.
USAR effort and preservation of freedom in the face of Borden EW. (1994). USAR operations: an overview (Northridge
adversity. Earthquake). Fire Engineering, 147, 92-98.
Feigenbaum, E., McCorduck, P. & Nil, P. H. (1988). The rise of the
expert company. New York, N.Y.: Time Books.
Acknowledgementv--This work is supported by the U.S. Army Harmon, P., & Sawyer, B. (1990). Creating expert systems for business
Artificial Intelligence Center (USAIC), the National Association for and industry. New York: Wiley.
Search and Rescue (NASR) and the 10th Special Forces Group (ABN). Mills, C. & Navarre, M. (1995). Urban search and rescue task force
Special thanks go to Mr Pete Bakersky (FEMA), LTC Steve Woffinden overview. In Proceedings of the 6th International Symposium on
(USAIC) and Mr Chuck Mills (NASAR) for their active research Rescue Dogs, Boulder, CO.
support and co-ordination efforts. This work is also indebted to Dr Miller D.P. (1994). Coping with limited on-board memory and
Robin Murphy (CRIS), Dr R. E. D. Woolsey (CSM), Dr Ruth Manrer communications bandwidth in mobile robot systems. IEEE tronsac-
(CSM-APICS), Dr Jean Bell (CSM) and Bill Tucker for their technical tions on systems, man and cybernetics, 2,4, I 58--72.
guidance and design assistance. Murphy, R. R. & Blitch, J. G. (1995). Multiple autonomous robots for
space mining. In Proceedings 3rd International Symposium on
REFERENCES Mine Mechanization and Automation, Colorado School of Mines,
Golden CO.
Alkass, S. & Harris, E (1988). Expert system for earthmoving O'Connell J.P. (1993). Search tactics and procedures below grade
equipment selection in road construction. Journal of Construction (World Trade Center Bombing). Fire Engineering, 146, 61-66.
Engineering and Management, 114. Ren, C. & Sheridan, T. B. (1995). A hybrid knowledge based system
Aidershot, H. (1994). Expert systems and optimization, John Epton for adaptive decision support. In IEEE Transactions on Systems,
(Ed.). Brookfield, VT.: Ashgate. Man, and Cybernetics, Volume 25, Oct. 1995.
Angle, C., Brooks, R. A. & Jones, J. (1995) Robust, behavior based Russel, S. J. & Norvig, P. (1995). Artificial Intelligence: A Modern
robots for inspection. In Proceedings of the Eighth Annual Idaho Approach. Englewood, N.J.: Prentice Hall.
National Engineering Laboratory Computing Symposium. Idaho Sinden F.W. (1990). The tethered robot problem, The international
Falls, ID. journal of robotics research, 9, I 122-133.
Amirkhanian, S.N. & Baker, N.J. (1992). Expert system for equipment Woolsey, R. E. D. & Maurer, R. (1991). Useful management series,
selection of earth moving operations. Journal of Construction Vol. 1. Golden CO: Colorado School of Mines.
Engineering and Management, 118, 313-318. Wu, N. & Coppins, R. (1981). Linear Programming and Extensions.
Barr, A., Cohen, P. R. & Feigenbanm, E. A. (1989). The handbook of New York, N.Y.: McGraw-Hill.
artificial intelligence. Reading, MA: Addison-Wesley. Zak S.H., Upatising V. & Hui S. (1995). Solving linear programming
Beckmann, U. (1991). Listening devices in heavy search and rescue. problems with neural networks: a comparative study. IEEE
Fire Engineering, 144, 87-90. Transactions on Neural Networks, 6, 94-104.