RoboCup Rescue A Grand Challenge for Multiagent and Intelligent Systems

7/30/2019 RoboCup Rescue A Grand Challenge for Multiagent and Intelligent Systems

http://slidepdf.com/reader/full/robocup-rescue-a-grand-challenge-for-multiagent-and-intelligent-systems 1/14

s Disaster rescue is one of the most serious socialissues th at in volves very large n um bers of hetero-

geneous agents in the hostile environment. The

inten tion of the RoboCup Rescue project is to pro-

mote research and development in this socially

significant domain at various levels, involving

multiagent teamwork coordination , physical

agents for search an d rescue, information infra-

structures, personal digital assistants, a standard

simulator and decision-support systems, evalua-

tion bench m arks for rescue strategies, and robotic

systems that are all integrated into a comprehen-

sive system in t he futu re. For th is effort, which wa s

built on th e success of th e RoboCu p Soccer project,

we will provide forums of t echn ical discussions

and competitive evaluations for researchers andpractitioners. Althou gh th e rescue dom ain is int u-

itively appealing as a large-scale multiagent and

intelligent system domain, analysis has not yet

revealed its domain characteristics. The first

research evaluation meeting will be held at

RoboCup-2001, in conjunction with the Seven-

teenth Internation al Joint Con ference on Artificial

Intelligence (IJCAI-2001), as part of the RoboCup

Rescue Sim ulation League an d RoboCup/ AAAI Res-

cue Robot Co mp etition. In t his article, we present

a detailed analysis of the task domain and eluci-

date characteristics necessary for multiagent and

intelligent system s for th is domain. Then , we pre-

sent an overview of th e RoboCup Rescue project.

At 5:47 AM on 17 Janu ary 1995, an earth-

quake of magnitude 7.2 hit Kobe City,

Japan, killing over 6,432 people; injur-

ing at least 43,800 (recorded in h ospitals); an d

crushing houses belonging to one-fifth of the

city’s 1.5 million people. Some 104,906 build-

ings completely collapsed, and on ly 20 percent

of the city’s buildings were usable after theearthqu ake. Th e damage area con centrated on

a strip of land within 13 miles by 0.6 miles,

and over 2,300,000 people were seriously

affected. The cost for repair o f th e basic in fra-

structure damage exceeded $100 billion, and

total property dam age, including private prop-

erties, well exceeded $1 trillion. The devasta-

tion was at least ten times larger th an t hat of

the 1994 Northridge earthquake that hit that

southern California area. Similar t ragedies

have also taken place in Turkey, Taiwan, and

oth er places on th e globe.

With th e first h it of the earthqu ake, h ouses,

buildings, and other facilities collapsed, androad, railways, and oth er pu blic transportation

systems were totally disrupted. Basic urban

infrastructures, such as electricity, gas, water

sup ply, and sewage system s, were severely

dam aged. Althou gh th e earth quake was devas-

tating, in formation on th e scale of damage was

not immediately transmitted to other parts of

the country because information infrastruc-

tures and personnel to t ransmi t damage

reports were catastrophically damaged so that

th ey were incapable of send ing p recise infor-

mat ion.

Many vict ims were un der col lapsed struc-

tures. Immediately after the earthquake, 285

fires were reported, with 14 large-scale fires,

and many were starting to spread into wider

areas. Firefighting was not effective because

water supply was disrupted, and local reser-

voirs were cracked so th at water leaked out

wi thin h ours . To m ake the s i tua t ion worse ,

roads and open areas that were supposed to

stop th e spread of fires turned in to com bustion

Articles

SPRING 200 1 39

RoboCup RescueA Grand Challenge for

Multiagent and IntelligentSystems

Hiroaki Kitano and Satoshi Tadokoro

Copyrigh t © 2001, American Association for Artificial Int elligence. All rights reserved. 0738-4602-2001 / $2.00

A I M ag az in e V o l u m e 2 2 N u m b e r 1 ( 2 0 0 1 ) ( © A A A I )



structures. Transfer of some of th ese tech n olo-

gies to th e disaster-rescue dom ain at som e poin t

had been our initial plan. Second, disaster res-

cue and soccer have some similar characteris-

tics, such as team work, un certain ty of informa-

tion, and real-time decision, as well as some

largely different characteristics, such as the

number of agents involved, heterogenei ty,

logistic planning, and emergent collaboration.

Thus, invest igat ion of the two domains pro-

vides a deeper understanding of the essence of

mu ltiagent collaboration and auton om ous real-

tim e system s. Th ird, disaster rescue alon e is seri-

ous en ough to just i fy ini t iat ion of th e large-

scale project , and we bel ieve that the

comm un ity and project-man agemen t approach

th at h as been successful in RoboC up Soccer can

bring about rapid progress in research in this

new field.

AI and robotics research can already make

immediate contribut ions within the reach of

current t echn ologies and a nu mber of long-

range contributions with future technologies:First, there is a potential need for simulating

and understanding the opt imal or near-opti-

mal search-and-rescue strategy for a large-scale

disaster. The aston ishin g fact is that t h ere is n o

s imu lator th a t can carry out com prehensive

simu lation of large-scale disasters. Th erefore,

developm ent of th e sim ulator i tself can be a

major contribution. However, more advanced

simula t ion wi th autonomous agents in the

simu lated en viron m ents m igh t m ake a far larg-

er contribution . As illustrated in th e previous

sec t ion, the numbers of inc idents and the

un certainty in m aking confident decisions is

far beyond current human capability. Provid-ing simu lators and a decision -assistance system

would significant ly improve the qual i ty of

decisions and the understanding of the possi-

ble simulation about t o un fold. Such a simu la-

tor and decision-assistance system should be

an integra ted s imula t ion of proper t ies and

infrastructure damage, fire proliferation, ref-

ugee movement , and other factors involved,

and a group of agents should be deployed to

exam ine t h e efficacy of specific search-an d-res-

cue strategies. In the future, such a simulator

should be l in ked to mob i le comm un ica t ion

and data-acquisition systems in th e field. Th is

simulation possibly involves 1,000 to 10,000

agen ts and events.

Second, there are immediate needs and

opportuni t ies for researchers to contribute

search-and-rescue strategies by bui lding a

group of robots that work as a team, each of

wh ich is specialized in specific sensing an d

m obility, as well as inform ation system s th at

can gath er, transm it, and receive information

path ways because debris from th e col lapsed

wood h om es was exposed to air.

Paramedic and rescue teams h ad a h ard time

trying to get to the disaster site because of the

disrupt ion caused b y dam aged roadways, col-

lapsed buildings, and the flood of refugees.

Also, there was insufficient and inaccurate

information about wh ere these teams sho uld

be. Aerial surveillance finally provided a view

of the overal l si tuat ion, but there was no

detailed information about t h e specifics of the

situat ion on the ground. The problem with the

hel icopters and the other aircraft is that the

noise they create hampers ground operations

in locating victims un der th e collapsed h ouses

because the faint sounds that victims create is

th eir on ly source of location information .

The lesson that we learned from the Kobe,

Turkey, and Taiwan earthquakes was the seri-

ous need for a robust , dynamic, intel l igent

plann ing system for search-and -rescue op era-

t ions and for powerful human-machine sys-

tem s, includ ing robo ts and d igital assistan ts, tocope with the ch anging situat ion and to best

save people. The scale of the disaster and the

speed at wh ich the system m ust chan ge accord-

ing to th e situation is far beyond h um an-based

mission planning and results in an extremely

hostile condition that to this point has made

hu m an search and rescue ineffective.

A Grand Challenge

In this article, we describe RoboCup Rescue

(Kitano 2000; Tadokoro et al. 2000; Kitano et

al. 1999) as a grand challenge project for the

AI, multiagent systems, and robotics com mu -n ities. The aim of RoboCu p Rescue is to d evel-

op a series of tech n ologies th at can actual ly

save people in the case of large-scale disasters

and to actual ly operate such systems world-

wide. Although RoboCup Soccer was claimed

to be th e lan dm ark project (Kitano et al. 1997),

RoboC up Rescue is designed as a grand chal-

lenge pro ject to d irectly attack a socially signif-

icant problem. Once accomplished, it will be

th e one of th e largest con tributions that th e AI

and robot ics communi t ies can make for

mankind.

We launched RoboCup Rescue within the

context of RoboCup activities for several rea-

sons: First, RoboCup itself was originally

designed t o con tribute to th e n ext-generation

social and indu strial infrastructure by p roviding

innovative technologies. Soccer was chosen

because it in volves man y issues (such as mu lti-

agen t col laborat ion, real-t ime p lann ing, and

integrated sensing) that are essential for tech-

n ologies in futu re indu stries and social infra-

The aim of

RoboCup

Rescue is to

develop aseries of

technologies

that can

actually save

people in th e

case of large-

scale disasters

and to

actuallyoperate such

systems

worldwide.

Articles

40 AI MAGAZINE



on the s i tua t ion and the proper course of

action . For example, to find victims un der the

debris, one robot m igh t h ave hexapot legs and

walk over the debris and insert a rod with a

microphone and a micro-CCD camera in

between debris. An oth er robot m igh t be small

enough to go into the debris with a CCD and

an in frared cam era. These robo ts would collab-

oratively cover the space under the debris. To

efficient ly and completely cover the space,

teamwork would be an essent ial factor. The

number of agents involved in this scenario

could perhaps be less than 20, a l though i t

depends on the size of the single site.

Third, personal digital assistants (PDAs) can

significantly empower search-and-rescue oper-

ations. Such a system must be robust enoughso th at it is operation al in various disaster situ-

at ions. When a certain intel l igence is to be

added to each such device, it h as the poten tial

to chan ge the way a rescue operation might b e

carried ou t. Not on ly personal devices but also

certain in formation infrastructures embedded

in the environment can significantly improve

the efficiency of the operation. For example,

distributin g passive data tags on bu ildin g floors

with respect ive floor numbers can assist in

locating possible victims after a building col-

lapse by using probes to fin d th e distribution of

floor n um bers in th e debris. A n um ber of tech-

nologies can be used and invented within AI,as well as in m ore general tech n ologies, for sav-

ing people

The RoboCup Rescue Project

The RoboCup Rescue project is designed to

maximize the contribution to the society and

at tain high throughp ut in research. From the

research perspective, RoboCup Rescue is

designed to (1) ensure smooth transfer of tech-

no logies inven ted th rough RoboCup activity to

a socially significan t real-world d om ain, (2)establish a dom ain th at comp lem ents features

th at are missing in soccer, and (3) examin e fun -

dam ental principles of teamwork an d real-time

mul t iagent sys tems by having mul t iple

dom ains with certain com mo nalities. Domain

characteristics of soccer, rescue, an d chess are

illustrated in table 1.

RoboCup Rescue consists of four major pro-

jec ts (figu re 1): (1 ) sim ulation , (2 ) robotics a n d

infrastructure, (3) integration, and (4) opera-

tion. Th e simulation project involves developing

com preh en sive disaster- rescue simu lation sys-

tems th at can evolve into deployable real-tim e

decision -supp ort system s and investigating th e

best search-an d-rescue strategies usin g th e sim-

ulator with autonomous and nonautonomous

rescue agen ts. Th e robotics and infrastructure pro-

ject involves developing deployable robotics,

digital assistants, and information infrastruc-

tures that can significant ly improve search-

and-rescue operat ions in real disasters. The

integration project involves integratin g th e com-

prehensive simulat ion system, robotics, and

digital assistan t system s for even tual full-scale

deploymen t. Th e operation project involves step-

wise deployment of the sys tem in the rea l

world.It is impo rtant th at both th e simulation pro-

jec t an d th e robo tics an d in fra st ru ct ure pro jec t

are carried out in parallel. There are a number

of issues that each project can independently

work on and contribute to society. Although

our fin al goal is to bring such techn ologies in to

reality an d actu ally save people, we need bu ilt-

in mechanisms to ensure the sus ta inable

progress of quality research an d d eployable sys-

Articles

SPRING 200 1 41

Rescue Soccer Ch ess

Number of Agents 100 or m ore 11 per team —

Agents in the Team Het ero gen eo us Ho mo gen eo us —

Logistics Major Issue No No

Long-Term Planning Major Issue Less Em ph asized In volved

Em ergent Collaboration Major Issue No No

Hostility En viron men t Oppon en t Players Oppon en t

Real Time Secon ds-Min utes Millisecon ds No

Information Access Very Bad Reason ably Good Perfect

Representation Hybrid Non sym bolic Sym bolic

Control Distributed,Semicentral

Distributed Cen tral

Table 1. Features of Rescue, Soccer, and Chess.



of victims; and h ospital operations (figure 2).

Surprisingly, there is no com preh ensive sim -

ulator for d isaster and rescue operation s. We

consider that the development of a compre-

hen sive sim ulator th at en ables the simu lation

of a m ultiagent rescue operation con tributes to

the quality and effectiveness of an actual res-

cue operation in th e long run . By developing

an integrated and comprehen sive sim ulator for

large-scale disaster rescue, a number of differ-

ent strategies and t actics can be comp ared to

best save people and property. With the

progress of multiagent systems research, tech-

nologies and methodologies invented can

actually be applied to a real search-and -rescue

system that commands fielded personnel and

robots.

Simu lator Architectu re

The simulator should be able to combine vari-

ous domain-specific simulations, such as fire,

building collapse, refugee m ovemen t, and t raf-

fic and p resent th em as a coherent scene. Thecurrent version of the simu lator arch itecture is

shown in figure 3.

A kernel of the simu lator com bines all infor-

mation and updates the status. Several do-

main-specific simulators are connected to the

kernel. Information on an entire disaster field

is stored in a geographic inform ation system. A

nu m ber of agents are deployed in th is simula-

t ion envi ronment to tes t the s t rengths and

weaknesses of the search-and-rescue strategies.

Present ing com plex disaster in formation is a

m ajor challenge in such a sim ulation system .

The current version of the simulator is

equipped with a two-dimensional si tuationmon itor (figure 4), a three-dimen sional view of

collapsed h ouses (figure 5), a th ree-dim ension -

al monitor (figure 6), and a sophisticated lay-

ered presentation system (figure 7).

The bu ilding and hou sing dam age simulator

simulates the degree of damage to buildings

and hou ses. When detailed sim ulation is to be

performed, this simulation needs to be made

block by block, with increased detail for specif-

ic landmarks.

Th e fire simulator simulates the occurrence

of fires and how t hey m ight spread over time.

The comp osition types of buildings and weath -

er factors will be incorporated. The spreading

patterns of fires need to reflect th e collapse of

buildings and th e effectiveness of fire fighting.

Existing sim ulators m odel th is process as a sto-

chastic process of propagating heat and catch-

in g fire with thresholded functions over static

terrain. They do not incorporate damage to

buildings and related fire-fightin g efforts.

The l ife-l ine damage simulator simulates

tem s. Th ree basic principles guide th e research

and d evelopm ent of th e projects: (1) interop er-

ability, (2) open -en dedn ess, an d (3) best prac-

tice.

Int eroperability: Software, robots, and oth er

equipmen t th at comp ly with RoboCup Rescue

standards should h ave a guaranteed level of

interoperabil ity, ensuring th at robots in one

region of the world can be deployed to save

people in oth er region s m ost effectively.

Open-ended system: The en tire system will beopen end ed, so that any new module and tech-

no logies can be p lugged in easily, and th e sys-

tem can be scalable.

Best-practice configurat ion: The system sho uld

be th e collection of th e best-available mo dules

and technologies. The selection of each mod-

ule and techn ology will be made com petitively.

These principles ensure that the results of

the research an d development are always up to

date and that we have a consistent collection

of th e best-available techn ologies. At t he same

time, successful standard form ation guaran tees

a competitive evolution of the overall system

as well as world-wide operational capabilities.

Simulation Project

Disaster simulation requires the integration of

the various aspects of disaster, including fire;

hou sing an d building damage; disruption of

roads, electricity, water supply, gas, and other

infrastructures; movem ent of refugees; status

Articles

42 AI MAGAZINE

Simulation

Project

Robo tics &

Infrastructure

Project

Integration Project

Operation Project

Figure 1. Projects in RoboCup Rescue.



damage to roads, electricity, water supply, gas,

and other infrastructures. These damage simu-

lat ions need to be t ightly coupled with thebuilding and hou sing d amage simulator. Cur-

rent ly, a few sim ulators can predict road block-

age with 70- to 80-percent accuracy (Takahash i

et al. 1998); th ese simu lators, however, are no t

yet coupled with other simulators.

The vict im m odeling simu lator represents

victim s and refugees, critical comp on ent s of a

simulation. Depending on the type of disaster,

the location of victim s, and the m agnitude of

the disaster, the physical and mental damage

th at victim s suffer differs drastically. Such a

simulator needs to reflect what the changing

difference an d u rgency of th e victim s is as well

as how to rescue them. Factors to consider

include th e t ime fram e of the operation, th e

kind of paramedic first aids, hospitals that vic-

t ims can be transported to, and the types of

equipmen t and the expertise of the paramedic

teams.

Articles

SPRING 2001 43

Disaster Simu lator Disaster Preven tion / Search an d Rescue Agent sTraffic Flow Model

Traffic Ma pDoctor Agent

Comm ander Agent

Rescuer Agen t

Robot Agent

Fire Fighter Agen t

Police Agen t

Victim Agen t

Back Sup por ter Agen t

Fire Spread Mod el

Mass Behavior Model

Building Collapse Map

Structu ral Vibration Model

Tsunami Model

Seism ic Motion ModelFluid Dynam ics Model

Liquefaction Mo del

Geographic

Information

System

Mission Critical

Man-Machine

Interface

AR Presen tation of Inform ation

Victims • Mass Media • Rescuers

Fire Fight ers • Police • Volunteers

Doctors • Commanders

Disaster Contro l Center

Prim e Minister • PresidentMayor • Politicians

Behavior Com m and

Transmission

Virtu al DisasterExperience

Workers • Students • Children

Sh opp ing Malls • Museums

Real-World

Interface

Disaster Inform ation

CollectionVictims • Mass Media

Action Command Transmission

Traffic Signa ls • Evacuation Signals

Gas Valves • Water Gates

Electricity Controls • Rescue Robot s

Army

• Virtual Experience an d Training

• Con dition ing of Optim al Action in Disaster

• Action Simulation of Parties of Rescue,

• Fire Fighters, and Back Supporters

• Simu lation of Disaster Con trol Strategy

• Decision Supp ort at Urgent Disaster

• En ligh ten m ent o f Disaster Prevent ion• Generic Disaster Prediction

• Urban Plan n ing for Disaster Prevention

• Real-Tim e Mon itoring of Dam age

• Synch ron ous Simu lation Using Real Data

• Efficient Generalization of Rescue Parties

Disaster Informat ion Collection

Seismometers

Tsunam i Meters

Video Cam eras

Reconn aissance Parties

Mobile Telecom m un icators

Combustion Model

Mass Psychology

R e al Di s a s t e r

S of t w ar e W o

r l d

H um an W or d

P l u g g e d I n B y

R e s e a r c h e r s W

o r l d w i d e

Intern ation al In terdisciplinary

Coo perative Research

Figure 2. The Overall Concept for the Simulation Project.



important because massive refugee marching

to escape from a disaster site seriously blocks

rescue t raf fic. It is possible that some artificial

life type of approach can provide reasonable

sim ulation of th is aspect. The sim ulation, h ow-

ever, needs to b e interactive because of chan g-

ing terrain. Collapsed buildin gs and roads and

police road blockades can con tinu ously chan ge

th e terrain, and t h e effects of these chan ges on

refugees need to be simu lated.

The data-collection an d -visualization simu-

lator sim ulates the d ata collection of a situa-

tion in a d isaster site; the visualization of com-

plex information is a major issue for such

system s to h ave practical value. Data collection

is not a major problem if we use the simulator

Victim m odeling also has a serious time-sen-

sitive element. Certain casualties need to be

taken care of within a short period of time, and

if a paramedic team arrives late to the scene,

the n ature of the operation could be very dif-

ferent . If certain casualties are less serious, th eir

priority would be lower, and rescue would be

directed toward vict ims with more urgent

needs. Saving such victims can become time

critical as victim att rition gets serious, an d ot h -

er urgent factors are revealed.

The refugee behavior simulator simulates a

large number of people who are trying to

escape from a disaster site, trying to find a

secure place, and searching for family and

friend s. Th e m ovemen t of a refugee is critically

Articles

44 AI MAGAZINE

GIS

GIS In terface

KERNEL

Fire

Stations

Hospitals

Police Statio n s

Disaster Sim 1

Disaster Sim n

Interagent Protocol

Monitoring Viewer

Civilian Agent s

Professional Agents

Building In formation

Road In formation

Planner Information Fire Brigades

Police Agen ts

Army Agents

Medical Agents

Figure 3. An Architecture of the Simulator.



for offl ine simulation during training and

research. However, if we wan t to use th e simu-

lation system as a decision-support tool and

assist decision m akers, th e m eth od of appro-

priately collecting and consolidating informa-

tion is a major issue. No such effort is under-

way a t t he momen t , and t he number o f

research op portu n ities is vast.

Visualization aspects are also im portan t. The

amount of information that needs to be dis-

played and conveyed to decision makers and

people in th e field is tremend ous. It would n ot

be useful for the systems to display al l the

information th at would be available. People

would be unable to quickly comprehend al l

this information. There is a need for selective

display and a new m ethod of providing m ean-

ingful information that is suited to comman-

ders, rescue personnel, and civilians. The

research on this area has just started (Shinjoh

et al. 1999).

Open Evolution of Simu lator

The simulator is designed to be scalable in

terms of functions and the size of simulation

events. New simulator functions can be added

by specialized plug-in simu lation m odules with

an interface protocol that complies with th e

defined stan dard. If an yon e provides a specific

simulation module that has a higher perfor-

Art icles

SPRING 2001 45

Figure 4. A Two-Dimensional Situation Monitor View.

Figure 5. A Simulator View of Collapsed Houses.



m ance rating than the existing m odule, it will

be possible for a better module to replace an

existing module, so that the simulator systems

are always at their best con figuration.

The important issue is to make resources

open and available to researchers, so that any-

one who wants to participate can make their

contribution from an y of a num ber of aspects.

Thus, i t is important that protocols and the

overall architecture are well agreed on at th e

start and improved on based on ongoing dis-

cussions and t he developm ental progress of the

technologies.

Industrial sectors can contribute by provid-

ing the way in wh ich their systems can be con-

n ected to th e sim ulator, as well as making th eir

codes available, so t hat th e RoboCu p Rescue

simulator can be the hub of all rescue-related

m odules. In th e long term, com panies will be

able to make sure that th eir products are con-

sistent with th e de facto open standard for all

th e disaster-rescue simu lation an d d ecision -

support systems. Already, there are local gov-ernm ents (mostly in Japan for n ow) that are

working on rescue strategy and u rban plan nin g

using th e RoboCu p Rescue simu lator.

Search-and-Rescue Strategy

Finding out the best search-and-rescue strate-

gies for large-scale disasters involves state-of-

the-art plannin g and m ultiagent research, such

as teamwork (Tambe an d Zhang 1998), plan-

n ing un der uncerta in ty (Pol lack and Zhang

1998), resource boun dedness (Russel 1995),

hierarchical planning (Kambhampati et al .

1998), and real-time planning.Just like the RoboCup Soccer simulation

league, the RoboCup Rescue simulation project

provides a simu lator and a set of scenario so th at

researchers interested in rescue strategy will be

able to investigate an d evaluate th eir approach.

In conjunction with several government sec-

tors, researchers who are interested in disaster

simulation efforts will provide realistic scenarios

of disasters for a few cities in th e world to th ose

who are interested in rescue operations. Any

number of researchers can work on the same

scenario to com pare the ad vantages and disad-

vant ages of each rescue strategy.

There are a nu m ber of issues to be investigat-

ed. Although we cannot possibly create an

exhau stive list of research issues, th e followin g

m ajor issues can be addressed.

Multiagent Planning

This dom ain can involve plann ing and execu-

tion monitoring for more than 10,000 agents

with different physical and informational

Art icles

46 AI MAGAZINE

Figure 6. A Three-Dimensional Situation Monitor View.

Figure 7. A Layered Display of Complex Disaster Information.



capabilities un der a dynam ically ch anging h os-

tile environ m ent . Plann ers can work in several

ways: First, th ey can coordinat e global strategy

with several subordinate planners. Flexible

hierarchical planning is essential to achieving

large-scale plann ing. Second , plann ing is done

in a distributed manner because disaster sites

are typically spread over wide areas, and com -

munication between them is l imited. Each

local plannin g system , however, m ust comm u-

nicate with the ot her plann ers, and th e global

planner must synchronize activi t ies and

update th e status to achieve an op timal strate-

gy. Third, contrary to the notion of a global

planning system, there is a need to establish

techn ologies that effect totally distributed an d

asynchronous planning when no clear global

planning center exists, or the expected global

planning site is destroyed. Whether the most

ef ficient strategy planning can be achieved by

global plannin g with subordinate plann ers or

totally distributed plann ing is a major research

issue. The entire planning system needs to beflexible enough t o dynam ically reformulate the

plann ing strategy from distributed plannin g to

h ierarchical plann ing an d vice versa.

In addition, th e agents involved are h etero-

geneous. There are rescue people an d robo ts,

helicopters and airplanes, ground-based vehi-

c les , and o ther au tonomous and nonau-

tonomous agents.

Real-Tim e–Anytime Planning

Quite obviously, the need for plann ing un der

real-tim e constraint s is critical in t h is dom ain,

given that the si tuation can change quickly

and often for the worse . Coming up wi th agood plan as quickly as possible might not be

good enough —in th is case, a plan agent sh ould

be able to provide a plan on deman d.

Heterogeneous Agent

The plann er should be able to recognize the

capability of each agent and coordinate them

to accomplish certain tasks. It might often be

th e case that th e capability necessary is m issin g

in the agents available on si te. The planner

shou ld be able to access th e poten tial availabil-

i ty of missing features and request that the

capability be added to the agent or that new

agents be added to th e team.

Robust Plan n ing

Much of the information available to agents

and decision makers during a disaster can be

incorrect, partial, and essentially unreliable.

Planning systems should be able to assess the

reliability of in formation and cope with p ossi-

ble errors. A feature that enables the p lann er to

actively request in formation gathering would

be critical. Not on ly inform ation b ut also actual

agents can be involved in accident s or disabled

for various reasons. In addition, because of

un expected h ardships at th e site of a disaster, it

is no t always guaranteed th at plans dispatched

to agent s will be executed. There mu st be ways

of copin g with th e range of un certainty an d be

able to replan an d execute in real time.

Mixed-Initiative Planning

Many t imes, the human rescuers wil l make

their own decisions and execute a search-and-

rescue plan that is not necessarily consistent

with tha t of th e p lanner. The p lann er must

cope with such a complex situation an d n ego-

tiate to find t he best solution. At th e same tim e,

if act ions taken by human rescue personnel

that are incon sistent with th e planner worked

well, certain learnin g and archival capabilities

can transfer such successful actions to other

sites.

Execution Mon itorin g

In such a hostile environment, it is not at all

guaranteed th at comm ands and operators dis-

patched from the plann er will be executed as

expected. The m ajor questions to add ress are

(1) how to gath er information on t he status of

plan execution, (2) how to cope with the d is-

ruption of comm unication, and (3) how to deal

with misinformation when personnel on site

are given incorrect information. This issue of

execution m onitoring is tight ly coupled with

robust plann ing.

Scenario-Based Plann ingIt is total ly incon ceivable that we would be

completely prepared for disasters. There are a

number of search-and-rescue scenarios and

doctrin es, just like in m ilitary plan nin g. There

are political and social decisions that are not

easily incorporated into automated planning,

and rescue strategies devised by responsible

of ficers need to be accom m odated. Thu s, plan-

ners m ust be able to use a given set of plan sce-

narios as well as create th eir own p lan.

Resource-Bounded Planning

Search-and-rescue operations are always con-

strained by resources, such as m aterials, per-

sonnel, and time. For example, the num ber of

robots and personn el that can arrive at a specif-

ic site within a given time limit, is seriously

limited by traf fic, fatigue, an d available agents.

The p lanner m ust be able to take into account

such con straints and p ossibly propose plans to

mitigate such constraints and st i l l maintain

long-range strategies.

Finding out

the best

search-and-

rescue

strategies for

large-scale

disastersinvolves state-

of-the-art

planning and

multiagent

research, such

as teamwork,

planning

under

uncertainty,resource

boundedness,

hierarchical

planning, and

real-time

planning.

Art icles

SPRING 2001 47



n ot battle proven. Altho ugh efforts to imp rove

the level of autonomy will continue, it would

be better if we provided technologies with an

adjustable level of autonomy, so that systems

can be used as teleoperation systems when

autonomous systems are yet technically infea-

sible. With the advancement of technologies,

the system can gradually increase the level of

auton om y when judged appropriate.

Another issue for rescue robots is that in

m any cases, they will h ave to be used for other

purposes as well. Disasters do not always take

place. Having dedicated robots and intelligent

systems only for disasters that might happen

only once in 30 years is not practical from an

engineering or an economic point of view. Of

course, certain numbers of robots can be spe-

cially designed for rescue an d u sed for fire and

other small-scale disasters that happen more

frequently. The system has to be used in daily

life so that it is justified economically, and we

can be sure that it will be function al at the tim e

of the disaster.AI and robo tics can con tribute to th e search-

and-rescue problem in a number of different

areas. With reference to actual robots and

PDAs, rather than high-level planning,

m echan ical aspects of robots, the robots’ phys-

ical performance, and concrete multirobot col-

laboration m ust be investigated. The following

issues, amon g man y, need to be add ressed:

First are robust robot systems that can go

into disaster si tes, which means moving

around o n an un even surface and in a h ostile

environment.

Second are multiple sensing systems with

flexible con figurations, so that sensors can beplaced and extended into debris to capture

sign s of victim s. Th is task can be accom plish ed

not only by single robots but also by multiro-

bot teams. Intelligent systems are also needed

to autom atically, or semiautom atically, carry

out sensor placements and self-positioning.

Third are m ult iple-robot systems t hat can

collaborate as a team with h um an p ersonn el to

rescue vict ims. Such robots should work

together to remove debris, so that elaborate

plannin g and com mu nication are unn ecessary.

Fourth is a human-robot interface that

enables rescue personnel to control a team of

robots for m ore effective operation s.

Fifth is an interface with strategic planning

system s to coordin ate local operation s with

global operations.

Most of th ese issues are still open problem s

for AI and robotics; th us, th is domain is a rich

source of insp i rat ion th a t can p romote n ot

only rescue-related robotics and AI but also

more general technologies. Already, there is

Data-Collection Agents an d Plan n ing

One of the major problems in a disaster is, as

repeated many times already, the dif ficulty in

getting accurate information in a reasonable

amou nt of t ime. There is a need for explici t

planning and a specific agent for data collec-

tion, which is tied to other agents and plan-

ners, so th at th e necessary information can be

obtain ed faster and m ore reliably. This issue istotally open and has n ot yet been investigated.

Robotics and

Infrastructure Project

The robotics and infrastructure project focuses

on th e developm ent of robots, digital assis-

tants, and infrastructures to enhance search-

and-rescue operations with AI, robotics, and

information technologies.

The most im portant task in search and res-

cue is to find victims under the debris. Unless

we identify th e location of victims, rescue op er-at ions cannot take place. Thus, the ini t ial

emphasis of the project might focus on,

althou gh n ot exclusively, establishin g the t ech-

nologies for highly effective and practical sys-

tems for victim search. Concurrently, effort

will be made to achieve robotic systems that

work with human rescue personnel to rescue

victim s at th e disaster site.

In ei ther si tuation, robots need to work

together as a multiagent robotics team as well

as with hum an rescue personnel and comm an-

ders. Assum ing th at robo ts, digital devices, and

human rescue personnel are al l t reated as

agents, we have a massive multiagent roboticsproblem. Various kinds of robot can be used

and form t eams at various levels. Som e robots

might be very small , and hundreds of them

will be deployed to go in to sm all spaces with in

the debris to find victims. Other robots might

be b ig , wi th a human on board to remove

debris so th at victims can be rescued.

At th e tim e of a d isaster, person-power is at

an absolute sh ortage. It is not feasible that two

or more persons are needed to operate one

robot at the disaster si te unless the robot is

extremely ef ficient for victim search. It is pre-

ferred that m ultiple robots be operated by on e

person; thus, robots must have an adjustable

level of autonomy. Robots should have auton-

omy because they can search for victims with-

out h um an h elp. However, it sho uld be realized

that a disaster si te is very complex, and i t

would be difficult to build fully autonomous

systems that can accomplish a task without

any h um an intervention . Also, rescue of ficials

would generally resist technologies that were

The most

important

task in search

and rescue is

to find victims

under the

debris.

Art icles

48 AI MAGAZINE



some pion eering research in th is area (Murph y

et al. 2000). Some robots and multiple-robot

teams th at h ave been d eveloped for other pur-

poses, including multisegment robots for pipe

inspection s at GMD (Scholl et al. 1999), can be

applied for rescue purp oses.

Integration

The in tegra t ion of s im ula t ion , informat ion

infras t ruc ture , and robot ics systems i s the

ultim ate techn ical goal of the p roject. Figure

8 gives an abstract view of how this integra-

tion will take place. Th e challenges are to (1)

interface with real-world sensors and reports

to determ ine th e situation at th e disaster site,

(2) represent such information in th e sim ula-

t ion, (3) transmit information to PDAs and

other devices in the appropriate way for per-

sons on si te, and (4) produce comm ands and

instructions for robots and their operators at

an ad justable level of auton om y.

Although it is too early to discuss full-scale

deployment, m any interesting research issues

involve man-machine interaction at various

levels as well as information representation

and sensing.

Evaluat ion Session s

A series of evaluation session s can be organ ized

to measure the performance of various

approaches . Al though the theme of the

research is life and d eath , it is no t app ropriate

to organize evaluation sessions as entertain-

ment events, as with RoboCup Soccer. Never-

th eless, certain forms of com petition wo uld be

useful to evaluate overall characteristics of th e

various approaches and prom ote techn ological

development.

In the simulation project , evaluations for

rescue strategies can be m ade on th e basis of

the number of lives saved and the amount of

property damage prevented. Figures 9 and 10

show evaluation results for two teams (TRIDENT

Art icles

SPRING 2001 49

Simulation and

Decision Support System

Comm and, Control, and Com mu nication

PDARescue

RobotsData Collection

Figure 8. Integrated System .



Art icles

50 AI MAGAZINE

Figure 9. Snapshot from a RoboCup-2000 Rescue Simulator

Dem onstration — Successful Fire Fighting Rescue Operation.

Top: Onset. Middle: 160 min. Bottom: 300 min.

(Courtesy of Milind Tambe, ISI/USC)

Figure 10. Snapshot from a RoboCup-2000 Rescue Simulator

Dem onstration — Unsuccessful Fire Fighting Rescue Operation.

Top: Onset. Middle: 160 min. Bottom: 300 min.

(Courtesy of Milind Tambe, ISI/USC)



team from University of Southern California

Information Sciences Inst i tute an d th e team

from the Nagoya Institute of Technology) for

small-scale fire-fightin g simu lation using actual

data from the Kobe earthquake (Nair et al .

2000). In addition , th e accuracy of specific sim -

ulat ion submodules can be evaluated and

always upgraded t o th e best-available mod ule.

In th e robotics and infrastructure project ,evaluation for search-and-rescue systems can

be don e by locatin g num bers of victim s (faked

or actual human) under the simulated debris

and measuring the accuracy and speed of fin d-

ing these victim s under th e debris with no or

minimum visual contact. The evaluation can

be extended to measure the speed of pulling

victims from the debris and sendin g them to a

designated first-aid station. Th is time, th e level

of ph ysical damage to th e victims can be eval-

uated using sensors attached to th e sim ulated

victims. Currently, the RoboCup Rescue tech-

nical com m ittee is working with th e National

Institute of Standards and the American Asso-ciatio n for Artificial In telligen ce (AAAI) to

define the simulated debris field used in the

competitive evaluation at the RoboCup/AAAI

Rescue Robot Competi t ion to take place at

RoboCu p-2001 in Seattle, Washin gton , in con-

ju n ct io n wi th th e Seven teen th In tern at ion al

Joint Conference on Artificial Intelligence

(Murph y, Casper, and Micire 2000). This field

was already used at th e Twelfth Nation al Con-

ference on Artificial Intelligence (AAAI-2000)

rescue robot com petition (figure 11).

ConclusionIn th is article, we described RoboCu p Rescue as

the grand challenge project for RoboCup. The

basic motivation behind the project is the

sense of obligation as scientists and engineer-

ing researchers in AI and robotics to h elp miti-

gate th e suffering of p eople from disasters. It is

also consistent with RoboCup ’s initial inten-

tion to apply technologies developed through

Robo Cup Soccer to serious social problem s. It is

design ed to ensure the smoot h tran sfer of tech-

n ologies developed in RoboCu p Soccer as well

as promote inn ovation as it com plements fea-

tures missing in soccer. RoboCup Rescue has

both simu lation and real robot aspects, each of

which initially focus on different areas of the

overall activity. As was clearly illustrated,

RoboCup Rescue is a rich source of research,

and direct contribution to society is expected.

For further information, please visit www.

robocup .org/rescue/ as well as the m ailing list

[email protected] .

Art icles

SPRING 2001 51

Figure 11. Rescue Robot Evaluation.

National Institute of Standards test-bed red zone: Scenes from AAAI-2000 rescue

robot comp etit ion. Photo courtesy of W illiam A dams (NRL), Alan Schultz (NRL), and

Robin Murphy (USF).



Hiroaki Kitano is a

senior researcher at Sony

Com puter Science Labo-

ratories, Inc.; director of

ERATOKitan o Sym biot ic

Systems Project, Japan

Science and Technology

Corporat ion, a govern-

ment organizat ion for

basic science; and president of The

RoboCup Federation. Kitano was a visiting

researcher at Carnegie Mellon University

from 1988 to 1993 an d received a Ph .D. in

comp uter science from Kyoto Un iversity in

1991. Kitano received The Com puters and

Though t Award from IJCAI in 1993 an d th e

Prix Ars Electronica in 2000. His e-mail

add ress is kitan o@symb io.jst.go.jp.

Satoshi Tadokoro

received a B.E. in preci-

s ion machinery engi -

neering in 1982, an M.E.

in precision m achinery

engineer ing in 1984from the University of

Tokyo, an d a D .E. in sys-

tem science in 1991

from Kobe University. From 1984 to 1993,

he was a research associate in the Depart-

ment of Inst rumentat ion Engineer ing,

Kobe Un iversity, and h as been an associate

professor in the Department of Computer

and Systems Engineering since 1993. His

research interest is in rescue robotics and

molluscan robotics. He received the

Hydraul ic and Pneumat ic Technology

Foun dation Award in 1993, the Japan Soci-

ety of Artificial Intell igence Incentive

Award in 2000, and th e Robotics Symp osia

Best Paper Award in 200 0. His e-m ail

add ress is tadoko [email protected].

Mobile Robots in RoboCup Rescue. Paper

presented at th e RoboCup Workshop 2000,

31 August, Melbourne, Australia.

Murphy, R.; Casper, J.; Micire, M.; and

Hyams, J. 2000. Assessment of the NIST

Stan dard Test Best for Urban Search an d

Rescue. Paper presented at the NIST Work-

shop on Performance Metrics for Intell i-

gent Systems, 14–16 August, Gaithersburg,

Maryland.Nair, R.; Ito, T.; Tam be, M.; an d Ma rsella, S.

2000. RoboCu p Rescue: A Proposal an d Pre-

l iminary Experiences. Paper presented at

th e Worksho p on RoboCup Rescue 2000, 8

July, Boston , Massachu setts.

Pollack, M., and Zhang, W. N. 1998. Plan

Generat ion, Plan Management , and the

Design of Computational Agents. Paper

presented at the International Conference

on Multiagent Systems 1998, 2–8 July,

Paris, Fran ce.

Russel, S., and Norvig, P. 1995. Arti fi cia l

Int elligence: A Modern Approach. New York:

Prentice Hall.

Scholl, K.; Kepp lin , V.; Bern s, K.; an d D ill-

m ann , R. 1999. An Articulated Service

Robot for Auton om ous Sewer Inspection

Tasks. Paper presented at th e Intern ational

Conference on Intell igent Robotics and

System s, 17–21 Octo ber, Kyon gju, Korea.

Shinjoh, A. 1999. The Wearable System for

the Disaster Mitigation Problem: A Mis-

sion-Critical Man-Machin e Interface of th e

RoboCup-Rescue Simulator. Paper present-

ed at the Ninth International Conference

on Artificial Reality and Tele-Existence,

16 –18 December, Waseda, Tokyo, Japan.

Tadokoro, S.; Kitano, H.; Takahashi, T.;

Noda, I.; Matsubara, H.; Shijoh, A.; Koto, T.;Tekeuch i, I.; Takah ash i, H.; Matsun o, F.;

Hatayam a, M.; Nobe, J.; and Shim ada, S.

2000. Th e RoboCup Rescue Project: A Mul-

ti-Agent Approach to the Disaster Mitiga-

tion Problem. Paper presented at the IEEE

International Conference on Robotics and

Autom ation (ICRA-00), 24–28 April, San

Francisco, Californ ia.

Tam be, M., and Zhan g, W. N. 1998. Toward

Flexible Teamwork in Persistent Teams.

Paper presented at th e International Con-

ference o n Multiagent Systems, 2–8 July,

Paris, Fran ce.

Takah ash i, H.; Akakura , Y.; Nakam oto , T.;

an d Yoshim ura, H. 1998. Simu lation of Street Blockages Caused by Earthquake

through the Analysis of the Hyougoken-

Nanbu Earthquake. Paper presented at the

Tenth Japan Earthq uake Engineering Sym-

posium, 25–27 Novem ber, Yokoh am a,

Japan.

Acknowledgments

This article is based on discussions of

the RoboCup Rescue Technical Com-

mittee, which includes Satoshi

Tadokoro, Koichi Osuka, Tomoichi

Takahashi, Atsushi Shinjou, Ikuo

Takeuchi, Kazuhiko Noguchi, Jun

Nobe, Mitunori Hatayama, Takeshi

Matsui, Fum itoshi Matsun o, Sh u Ishig-uro, Yuuich i Oh tan i, Toshiyuki Kan e-

da, Tetsuhiko Koto, Hironao Taka-

hashi, Haruki Nishi, Hajime Asama,

tomohiko Sakao, Hitoshi Matsubara,

Itsuki Noda, Najit Nair, Milind Tambe,

and Stacy Marsella. The authors want

to th ank Robin Murph y (University of

South Florida), Manuela Veloso

(Carnegie Mellon University [CMU]),

Silvia Co radesh i (Orebro University,

Sweden), John Blitch (Defense

Advanced Research Projects Agency

and National Inst i tute for Urban

Search an d Rescue), Adam Jacoff (National In stitute of Stan dards), Alan

Schultz (Naval Research Laboratory),

Gal Kaminka (University of Southern

Californ ia Inform ation Scien ces In sti-

tute, CMU), Pedro Lima (Lisbon Tech-

n ical University), Klaus Fisher (GMD),

and the execut ive members of

RoboCup for useful comments and

discussions.

References

Kambhampati, S.; Mali, A.; and Srivastava,

B. 1998. Hybrid Plan nin g for Partially Hier-

archical Domains. In Proceedings of theTenth National Conference on Artificial

Intelligence, 882–888. Men lo Park, Calif.:

American Association for Artificial Intelli-

gence.

Kitan o, H. 2000. RoboCu p Rescue: A Gran d

Challenge for Multi-Agent Systems. Paper

presented at the International Conference

on Mult i-Agen t System s (ICMAS-2000),

10 –12 July, Boston , Massachu setts.

Kitan o, H.; Tadokoro, S.; Noda, I.; Matsub-

ara, H.; Takah ashi, T.; Sh injoh , A.; and Shi-

mada, S. 1999. RoboCup Rescue: Search

an d Rescue in Large-Scale Disasters as a

Dom ain for Auton om ous Agents Research.

Paper presented at the IEEE InternationalConference on System, Man, and Cyber-

netics (SMC-99), 12–15 October, Tokyo,

Japan.

Kitan o, H.; Asada , M.; Kun iyosh i, Y.; Noda,

I.; Osawa, E.; and Matsubara, H. 1997.

RoboCup: A Challenge Problem for AI. AI

Magaz ine 18(1): 73–86 .

Murp h y, R.; Casper, J.; and Micire, M. 2000.

Potential Tasks and Research Issues for

Art icles

52 AI MAGAZINE

Documents

RoboCup Rescue A Grand Challenge for Multiagent and Intelligent Systems