cooperative mobile robotics: antecedents and …autonomous robots kl402-02-cao december 19, 1996...

P1: MVG

Autonomous Robots KL402-02-Cao December 19, 1996 14:57

Autonomous Robots 4, 7–27 (1997)c© 1997 Kluwer Academic Publishers. Manufactured in The Netherlands.

Cooperative Mobile Robotics: Antecedents and Directions∗

Y. UNY CAOComputer Science Department, University of California, Los Angeles, CA 90024-1596

[email protected]

ALEX S. FUKUNAGAJet Propulsion Laboratory, California Institute of Technology, Pasadena, CA 91109-8099

[email protected]

ANDREW B. KAHNGComputer Science Department, University of California, Los Angeles, CA 90024-1596

[email protected]

Abstract. There has been increased research interest in systems composed of multiple autonomous mobile robotsexhibiting cooperative behavior. Groups of mobile robots are constructed, with an aim to studying such issuesas group architecture, resource conflict, origin of cooperation, learning, and geometric problems. As yet, fewapplications of cooperative robotics have been reported, and supporting theory is still in its formative stages. In thispaper, we give a critical survey of existing works and discuss open problems in this field, emphasizing the varioustheoretical issues that arise in the study of cooperative robotics. We describe the intellectual heritages that haveguided early research, as well as possible additions to the set of existing motivations.

Keywords: cooperative robotics, swarm intelligence, distributed robotics, artificial intelligence, mobile robots,multiagent systems

1. Preliminaries

There has been much recent activity toward achievingsystems of multiple mobile robots engaged in collec-tive behavior. Such systems are of interest for severalreasons:

• tasks may be inherently too complex (or impossible)for a single robot to accomplish, or performance ben-efits can be gained from using multiple robots;• building and using several simple robots can be eas-

ier, cheaper, more flexible and more fault-tolerantthan having a single powerful robot for each sepa-rate task; and

∗This is an expanded version of a paper which originally appearedin the proceedings of the 1995 IEEE/RSJ IROS conference.

• the constructive, synthetic approach inherent in co-operative mobile robotics can possibly yield insightsinto fundamental problems in the social sciences(organization theory, economics, cognitive psychol-ogy), and life sciences (theoretical biology, animalethology).

The study of multiple-robot systems naturally ex-tends research on single-robot systems, but is also adiscipline unto itself: multiple-robot systems can ac-complish tasks that no single robot can accomplish,since ultimately a single robot, no matter how capa-ble, is spatially limited. Multiple-robot systems arealso different from other distributed systems becauseof their implicit “real-world” environment, which ispresumably more difficult to model and reason about

P1: MVG


8 Cao, Fukunaga and Kahng

than traditional components of distributed system en-vironments (i.e., computers, databases, networks).

The termcollective behaviorgenerically denotes anybehavior of agents in a system having more than oneagent. Cooperative behavior, which is the subject ofthe present survey, is a subclass of collective behaviorthat is characterized by cooperation. Webster’s dictio-nary (Merriam-Webster, 1963) defines “cooperate” as“to associate with another or others for mutual, ofteneconomic, benefit”. Explicit definitions of coopera-tion in the robotics literature, while surprisingly sparse,include:

1. “joint collaborative behavior that is directed towardsome goal in which there is a common interest orreward” (Barnes and Gray, 1991);

2. “a form of interaction, usually based on communi-cation” (Mataric, 1994a); and

3. “[joining] together for doing something that createsa progressive result such as increasing performanceor saving time” (Premvuti and Yuta, 1990).

These definitions show the wide range of possiblemotivating perspectives. For example, definitions suchas (1) typically lead to the study of task decomposition,task allocation, and other distributed artificial intelli-gence (DAI) issues (e.g., learning, rationality). Def-initions along the lines of (2) reflect a concern withrequirements for information or other resources, andmay be accompanied by studies of related issues suchas correctness and fault-tolerance. Finally, definition(3) reflects a concern with quantified measures of co-operation, such as speedup in time to complete a task.Thus, in these definitions we see three fundamentalseeds: thetask, the mechanismof cooperation, andsystemperformance.

We define cooperative behavior as follows:Givensome task specified by a designer, a multiple-robot system displayscooperative behaviorif, due tosome underlying mechanism(i.e., the “mechanism ofcooperation”), there is an increase in the total utilityof the system.Intuitively, cooperative behavior entailssome type of performance gain over naive collectivebehavior. The mechanism of cooperation may lie inthe imposition by the designer of a control or commu-nication structure, in aspects of the task specification,in the interaction dynamics of agent behaviors, etc.

In this paper, we survey the intellectual heritage andmajor research directions of the field of cooperativerobotics. For this survey of cooperative robotics toremain tractable, we restrict our discussion to works

involving mobile robots orsimulations of mobilerobots, where amobile robot is taken to be an au-tonomous, physically independent, mobile robot. Inparticular, we concentrated on fundamentaltheoreticalissues that impinge on cooperative robotics. Thus, thefollowing related subjects were outside the scope ofthis work:

• coordination of multiple manipulators, articulatedarms, or multi-fingered hands, etc.• human-robot cooperative systems, and user-inter-

face issues that arise with multiple-robot systems(Yokota et al., 1994; Arkin and Ali, 1994; Noreilsand Recherde, 1991; Adams et al., 1995).• the competitive subclass of collective behavior,

which includes pursuit-evasion (Reynolds, 1994;Miller and Cliff, 1994) and one-on-one competitivegames (Asada et al., 1994). Note that a coopera-tive team strategy for, e.g., work on the robot soc-cer league recently started in Japan (Kitano, 1994)would lie within our present scope.• emerging technologies such as nanotechnology

(Drexler, 1992) and Micro Electro-Mechanical Sys-tems (Mehregany et al., 1988) that are likely to bevery important to cooperative robotics are beyondthe scope of this paper.

Even with these restrictions, we find that over thepast 8 years (1987–1995) alone, well over 200 pa-pers have been published in this field of cooperative(mobile) robotics, encompassing theories from suchdiverse disciplines as artificial intelligence, game the-ory/economics, theoretical biology, distributed com-puting/control, animal ethology and artificial life.

We are aware of two previous works that have sur-veyed or taxonomized the literature. Asama (1992)is a broad, relatively succinct survey whose scopeencompassesdistributed autonomous robotic systems(i.e., not restricted to mobile robots). Dudek et al.(1993) focuses on several well-known “swarm” archi-tectures (e.g., SWARM and Mataric’s Behavior-basedarchitecture—see Section 2.1) and proposes a taxon-omy to characterize these architectures. The scope andintent of our work differs significantly from these, inthat (1) we extensively survey the field of coopera-tive mobile robotics, and (2) we provide a taxonomicalorganization of the literature based onproblemsandsolutionsthat have arisen in the field (as opposed to aselected group of architectures). In addition, we sur-vey much new material that has appeared since theseearlier works were published.

P1: MVG


Cooperative Mobile Robotics: Antecedents and Directions 9

1.1. Towards a Picture of Cooperative Robotics

In the mid-1940’s Grey Walter, along with Wiener andShannon, studied turtle-like robots equipped with lightand touch sensors; these simple robots exhibited “com-plex social behavior” in responding to each other’smovements (Dorf, 1990). Coordination and interac-tions of multiple intelligent agents have been activelystudied in the field of distributed artificial intelligence(DAI) since the early 1970’s (Bond and Gasser, 1988),but the DAI field concerned itself mainly with prob-lems involving software agents. In the late 1980’s,the robotics research community became very activein cooperative robotics, beginning with projects suchas CEBOT (Fukuda and Nakagawa, 1987), SWARM(Beni, 1988), ACTRESS (Asama et al., 1989), GOFER(Caloud et al., 1990), and the work at Brussels (Steels,1990). These early projects were done primarily insimulation, and, while the early work on CEBOT,ACTRESS and GOFER have all had physical imple-mentations (with≤3 robots), in some sense these im-plementations were presented by way of proving thesimulation results. Thus, several more recent works(cf., Kube and Zhang, 1992; Mataric, 1992b; Parker,1992) are significant for establishing an emphasison the actual physical implementation of cooperativerobotic systems. Many of the recent cooperative roboticsystems, in contrast to the earlier works, are based ona behavior-based approach (cf. Brooks, 1986). Var-ious perspective on autonomy and on the connectionbetween intelligence and environment are strongly as-sociated with the behavior-based approach (Brooks,1991), but are not intrinsic to multiple-robot systemsand thus lie beyond our present scope. Also note thata recent incarnation of CEBOT, which has been im-plemented on physical robots, is based on a behavior-based control architecture (Cai et al., 1995).

The rapid progress of cooperative robotics since thelate 1980’s has been an interplay ofsystems, theoriesandproblems: to solve a given problem, systems areenvisioned, simulated and built; theories of coopera-tion are brought from other fields; and new problemsare identified (prompting further systems and theories).Since so much of this progress is recent, it is not easy todiscern deep intellectual heritages from within the field.More apparent are the intellectual heritages from otherfields, as well as the canonical task domains which havedriven research. Three examples of the latter are:

• Traffic Control. When multiple agents move withina common environment, they typically attempt to

avoid collisions. Fundamentally, this may be viewedas a problem ofresource conflict, which may beresolved by introducing, e.g., traffic rules, priori-ties, or communication architectures. From anotherperspective, path planning must be performed tak-ing into consideration other robots and the globalenvironment; this multiple-robot path planning isan intrinsicallygeometricproblem in configurationspace-time. Note that prioritization and communi-cation protocols—as well as the internal model-ing of other robots—all reflect possible variantsof the group architectureof the robots. For ex-ample, traffic rules are commonly used to reduceplanning cost for avoiding collision and deadlockin a real-world environment, such as a networkof roads. (Interestingly, behavior-based approachesidentify collision avoidance as one of the mostbasic behaviors (Brooks, 1986), and achieving acollision-avoidance behavior is the natural solutionto collision avoidance among multiple robots. How-ever, in reported experiments that use the behavior-based approach, robots are never restricted to roadnetworks.)• Box-Pushing/Cooperative Manipulation. Many

works have addressed the box-pushing (or couch-pushing) problem, for widely varying reasons. Thefocus in Parker (1994b) is on task allocation, fault-tolerance and (reinforcement)learning. By contrast,(Donald et al., 1994) studies two box-pushing pro-tocols in terms of their intrinsic communication andhardware requirements, via the concept of informa-tion invariants. Cooperative manipulation of largeobjects is particularly interesting in that cooperationcan be achieved without the robots even knowing ofeach others’ existence (Sen et al., 1994; Tung andKleinrock, 1993). Other works in the class of box-pushing/object manipulation include (Wang et al.,1994; Stilwell and Bay, 1993; Johnson and Bay,1994; Brown and Jennings, 1995; Kube and Zhang,1992, 1993; Kube et al., 1993; Mataric et al., 1995;Rus et al., 1995; Hara et al., 1995; Sasaki et al.,1995).• Foraging. In foraging, a group of robots must pick

up objects scattered in the environment; this isevocative of toxic waste cleanup, harvesting, searchand rescue, etc. The foraging task is one of thecanonical testbeds for cooperative robotics (Brookset al., 1990; Steels, 1990; Arkin, 1992; Goss andDeneubourg, 1992; Lewis and Bekey, 1992; Drgouland Feber, 1993; Mataric, 1994a; Arkin and Hobbs,1993; Beckers et al., 1994). The task is interesting

P1: MVG



because (1) it can be performed by each robot inde-pendently (i.e., the issue is whether multiple robotsachieve a performance gain), and (2) as discussed inSection 3.2, the task is also interesting due to moti-vations related to the biological inspirations behindcooperative robot systems. There are some concep-tual overlaps with the related task of materials han-dling in a manufacturing workcell (Doty and Aken,1993). A wide variety of techniques have been ap-plied, ranging from simple stigmergy (essentiallyrandom movements that result in the fortuitous col-lection of objects (Beckers et al., 1994) to more com-plex algorithms in which robots form chains alongwhich objects are passed to the goal (Drgoul andFeber, 1993). Beckers et al. (1994) defines stig-mergy as “the production of a certain behaviour inagents as a consequence of the effects produced inthe local environment by previous behaviour”. Thisis actually a form of “cooperation without communi-cation”, which has been the stated object of severalforaging solutions since the corresponding formu-lations become nearly trivial if communication isused. On the other hand, that stigmergy may notsatisfy our definition of cooperation given above,since there is no performance improvement over the“naive algorithm”—in this particular case, the pro-posed stigmergic algorithmis the naive algorithm.Again, group architecture and learning are major re-search themes in addressing this problem.

Other interesting task domains that have receivedattention in the literature include multi-robot securitysystems (Everett et al., 1993), landmine detection andclearance (Franklin et al., 1995), robotic structural sup-port systems (i.e., keeping structures stable in case of,say, an earthquake) (Ma et al., 1994), map making(Singh and Fujimura, 1993), and assembly of objectsusing multiple robots (Wang et al., 1994).

Organization of Paper

With respect to our above definition of cooperative be-havior, we find that the great majority of the cooperativerobotics literature centers on themechanismof cooper-ation (i.e., few works study a task without also claimingsome novel approach to achieving cooperation). Thus,our study has led to the synthesis of five “ResearchAxes” which we believe comprise the major themes ofinvestigation to date into the underlying mechanism ofcooperation.

Section 2 of this paper describes these axes, whichare: 2.1 Group Architecture, 2.2 Resource Conflict,2.3 Origin of Cooperation, 2.4 Learning, and 2.5 Ge-ometric Problems. In Section 3, we present more syn-thetic reviews of cooperative robotics: Section 3.1discusses constraints arising from technological lim-itations; and Section 3.2 discusses possible lacunae inexisting work (e.g., formalisms for measuring perfor-mance of a cooperative robot system), then reviewsthree fields which we believe must strongly influencefuture work. We conclude in Section 4 with a list ofkey research challenges facing the field.

2. Research Axes

Seeking amechanismof cooperation may be rephrasedas the “cooperative behavior design problem”:Givena group of robots, an environment, and a task, howshould cooperative behavior arise? In some sense, ev-ery work in cooperative robotics has addressed facetsof this problem, and the major research axes of thefield follow from elements of this problem. (Note thatcertain basic robot interactions are not task-performinginteractions per se, but are rather basic primitives uponwhich task-performing interactions can be built, e.g.,following ((Connell, 1987; Donald et al., 1994) andmany others) or flocking (Reynolds, 1987; Mataric,1994a). It might be argued that these interactions entail“control and coordination” tasks rather than “cooper-ation” tasks, but our treatment does not make such adistinction).

First, the realization of cooperative behavior mustrely on some infrastructure, thegroup architecture.This encompasses such concepts as robot heterogene-ity/homogeneity, the ability of a given robot to recog-nize and model other robots, and communication struc-ture. Second, for multiple robots to inhabit a sharedenvironment, manipulate objects in the environment,and possibly communicate with each other, a mech-anism is needed to resolveresource conflicts. Thethird research axis,origins of cooperation, refers tohow cooperative behavior is actually motivated andachieved. Here, we do not discuss instances wherecooperation has been “explicitly engineered” into therobots’ behavior since this is the default approach. In-stead, we are more interested in biological parallels(e.g., to social insect behavior), game-theoretic justi-fications for cooperation, and concepts of emergence.Because adaptability and flexibility are essential traitsin a task-solving group of robots, we viewlearning as a

P1: MVG



fourth key to achieving cooperative behavior. One im-portant mechanism in generating cooperation, namely,task decomposition and allocation, isnotconsidered aresearch axis since (i) very few works in cooperativerobotics have centered on task decomposition and allo-cation (with the notable exceptions of Noreils (1993),Lueth and Laengle (1994), Parker (1994b)), (ii) co-operative robot tasks (foraging, box-pushing) in theliterature are simple enough that decomposition andallocation are not required in the solution, and (iii) theuse of decomposition and allocation depends almostentirely on the group architectures (e.g., whether it iscentralized or decentralized). Note that there is also arelated, geometric problem of optimizing the allocationof tasks spatially. This has been recently studied in thecontext of the division of the search of a work area bymultiple robots (Kurabayashi et al., 1995). Whereasthe first four axes are related to thegenerationof co-operative behavior, our fifth and final axis—geometricproblems—covers research issues that are tied to theembedding of robot tasks in a two- or three-dimensionalworld. These issues include multi-agent path planning,moving to formation, and pattern generation.

2.1. Group Architecture

Thearchitectureof a computing system has been de-fined as “the part of the system that remains unchangedunless an external agent changes it” (VanLehn, 1991).The group architectureof a cooperative robotic sys-tem provides the infrastructure upon which collec-tive behaviors are implemented, and determines thecapabilities and limitations of the system. We nowbriefly discuss some of the key architectural featuresof a group architecture for mobile robots: central-ization/decentralization, differentiation, communica-tions, and the ability to model other agents. We thendescribe several representative systems that have ad-dressed these specific problems.

Centralization/Decentralization. The most fundam-ental decision that is made when defining a grouparchitecture is whether the system is centralized ordecentralized, and if it is decentralized, whether thesystem is hierarchical or distributed.Centralizedarchi-tectures are characterized by a single control agent.De-centralizedarchitectures lack such an agent. There aretwo types of decentralized architectures:distributedarchitectures in which all agents are equal with respectto control, andhierarchical architectures which are

locally centralized. Currently, the dominant paradigmis the decentralized approach.

The behavior of decentralized systems is often de-scribed using such terms as “emergence” and “self-organization”. It is widely claimed that decentralizedarchitectures (e.g., Beckers et al., 1994; Arkin, 1992;Steels, 1994; Mataric, 1994a) have several inherent ad-vantages over centralized architectures, including faulttolerance, natural exploitation of parallelism, reliabil-ity, and scalability. However, we are not aware of anypublished empirical or theoretical comparison that sup-ports these claims directly. Such a comparison wouldbe interesting, particularly in scenarios where the teamof robots is relatively small (e.g., two robots pushing abox), and it is not clear whether the scaling propertiesof decentralization offset the coordinative advantage ofcentralized systems.

In practice, many systems do not conform to astrict centralized/decentralized dichotomy, e.g., manylargely decentralized architectures utilize “leader”agents. We are not aware of any instances of sys-tems that are completely centralized, although thereare some hybrid centralized/decentralized architectureswherein there is a central planner that exerts high-levelcontrol over mostly autonomous agents (Noreils, 1993;Lueth and Laengle, 1994; Aguilar et al., 1995; Causseand Pampagnin, 1995).

Differentiation. We define a group of robots to behomogeneousif the capabilities of the individual robotsare identical, andheterogeneousotherwise. In general,heterogeneity introduces complexity since task alloca-tion becomes more difficult, and agents have a greaterneed to model other individuals in the group. Parker(1994b) has introduced the concept oftask coverage,which measures the ability of a given team member toachieve a given task. This parameter is an index of thedemand for cooperation: when task coverage is high,tasks can be accomplished without much cooperation,but otherwise, cooperation is necessary. Task coverageis maximal in homogeneous groups, and decreases asgroups become more heterogeneous (i.e., in the limitonly one agent in the group can perform any giventask).

The literature is currently dominated by works thatassume homogeneous groups of robots. However,some notable architectures can handle heterogene-ity, e.g., ACTRESS and ALLIANCE (see Section 2.1below). In heterogeneous groups, task allocationmay be determined by individual capabilities, but in

P1: MVG



homogeneous systems, agents may need todifferenti-ate into distinct roles that are either known at design-time, or arise dynamically at run-time.

Communication Structures. The communicationstructure of a group determines the possible modes ofinter-agent interaction. We characterize three majortypes of interactions that can be supported. (Dudeket al. (1993) proposes a more detailed taxonomy ofcommunication structures.)

Interaction via Environment. The simplest, most lim-ited type of interaction occurs when the environmentitself is the communication medium (in effect, a sharedmemory), and there is no explicit communication or in-teraction between agents. This modality has also beencalled “cooperation without communication” by someresearchers. Systems that depend on this form of inter-action include Goss and Deneubourgh (1992), Beckerset al. (1994), Arkin (1992), Steels (1990), Tung andKleinrock (1993), Tung (1994), Sen et al. (1994).

Interaction via Sensing. Corresponding to arms-length relationships in organization theory (Hewitt,1993), interaction via sensing refers to local interac-tions that occur between agents as a result of agentssensing one another, but without explicit communica-tion. This type of interaction requires the ability ofagents to distinguish between other agents in the groupand other objects in the environment, which is called“kin recognition” in some literatures (Mataric, 1994a).Interaction via sensing is indispensable for modelingof other agents (see Section 2.1.4 below). Because ofhardware limitations, interaction via sensing has oftenbeen emulated using radio or infrared communications.However, several recent works attempt to implementtrue interaction via sensing, based on vision (Kuniyoshiet al., 1994a, b; Sugie et al., 1995). Collective behav-iors that can use this kind of interaction include flock-ing and pattern formation (keeping in formation withnearest neighbors).

Interaction via Communications.The third form ofinteraction involves explicit communication with otheragents, by either directed or broadcast intentional mes-sages (i.e., the recipient(s) of the message may be eitherknown or unknown). Because architectures that enablethis form of communication are similar to communica-tion networks, many standard issues from the field ofnetworks arise, including the design of network topolo-gies and communications protocols. For example, in

Wang (1994) a media access protocol (similar to thatof Ethernet) is used for inter-robot communication. InIchikawa et al. (1993), robots with limited communica-tion range communicate to each other using the “hello-call” protocol, by which they establish “chains” in orderto extend their effective communication ranges. Gage(1993) describes methods for communicating to many(“zillions”) robots, including a variety of schemes rang-ing from broadcast channels (where a message is sentto all other robots in the system) to modulated retrore-flection (where a master sends out a laser signal toslaves and interprets the response by the nature of thereflection). Wang et al. (1995) describes and simulatesa wireless CSMA/CD (Carrier Sense Multiple Accesswith Collision Detection) protocol for the distributedrobotic systems.

There are also communication mechanisms designedspecially for multiple-robot systems. For example,Wang and Beni (1990) proposes the “sign-board” asa communication mechanism for distributed roboticsystems. Arai et al. (1993) gives a communicationprotocol modeled after diffusion, wherein local com-munication similar to chemical communication mech-anisms in animals is used. The communication isengineered to decay away at a preset rate. Similarcommunications mechanisms are studied in Lewis andBekey (1992), Drgoul and Feber (1993), Goss andDeneubourg (1992). Additional work on communi-cation can be found in Yoshida et al. (1994), whichanalyzes optimal group sizes for local communicationsand communication delays. In a related vein, Yoshidaet al. (1995a, b) analyzes optimal local communicationranges in broadcast communication.

Modeling of Other Agents. Modeling the intentions,beliefs, actions, capabilities, and states of other agentscan lead to more effective cooperation between robots.Communications requirements can also be lowered ifeach agent has the capability to model other agents.Note that the modeling of other agents entails morethan implicit communication via the environment orperception: modeling requires that the modeler hassome representation of another agent, and that this rep-resentation can be used to make inferences about theactions of the other agent.

In cooperative robotics, agent modeling has beenexplored most extensively in the context of manipulat-ing a large object. Many solutions have exploited thefact that the object can serve as a common medium bywhich the agents can model each other.

P1: MVG



The second of two box-pushing protocols in (Donaldet al., 1994) can achieve “cooperation without com-munication” since the object being manipulated alsofunctions as a “communication channel” that is sharedby the robot agents; other works capitalize on the sameconcept to derive distributed control laws which relyonly on local measures of force, torque, orientation,or distance, i.e., no explicit communication is neces-sary (cf., Stilwell and Bay, 1993; Hashimoto and Oba,1993). In a two-robot bar carrying task, Fukuda andSekiyama’s agents (Fukuda and Sekiyama, 1994) eachuses a probabilistic model of the other agent. Whena risk threshold is exceeded, an agent communicateswith its partner to maintain coordination. In Donald(1993a, b), the theory of information invariants is usedto show that extra hardware capabilities can be added inorder to infer the actions of the other agent, thus reduc-ing communication requirements. This is in contrast toSen et al. (1994), where the robots achieve box pushingbut are not aware of each other at all. For a more com-plex task involving the placement of five desks in Sugieet al. (1995), a homogeneous group of four robots sharea ceiling camera to get positional information, but donot communicate with each other. Each robot relies onmodeling of other agents to detect conflicts of paths andplacements of desks, and to change plans accordingly.

Representative Architectures.All systems imple-ment some group architecture. We now describe severalparticularly well-defined representative architectures,along with works done within each of their frameworks.It is interesting to note that these architectures encom-pass the entire spectrum from traditional AI to highlydecentralized approaches.

CEBOT. CEBOT (CEllular roBOTics System) is adecentralized, hierarchical architecture inspired by thecellular organization of biological entities (cf., Fukudaand Nakagawa, 1987; Fukuda and Kawauchi, 1993;Ueyama and Fukuda, 1993a, b; Fukuda and Iritani,1995). The system is dynamically reconfigurable inthat basic autonomous “cells” (robots), which can bephysically coupled to other cells, dynamically recon-figure their structure to an “optimal” configuration inresponse to changing environments. In the CEBOThierarchy there are “master cells” that coordinate sub-tasks and communicate with other master cells. A so-lution to the problem of electing these master cellswas discussed in Ueyama et al. (1993b). Formationof structured cellular modules from a population of

initially separated cells was studied in Ueyama andFukuda (1993b). Communications requirements havebeen studied extensively with respect to the CEBOTarchitecture, and various methods have been proposedthat seek to reduce communication requirements bymaking individual cells more intelligent (e.g., enablingthem to model the behavior of other cells). Fukudaand Sekiyama (1994) studies the problem of model-ing the behavior of other cells, while Kawauchi et al.(1993a, b) present a control method that calculates thegoal of a cell based on its previous goal and on its mas-ter’s goal. (Fukuda et al., 1990) gives a means of esti-mating the amount of information exchanged betweencells, and (Ueyama et al., 1993a) gives a heuristic forfinding master cells for a binary communication tree.A new behavior selection mechanism is introduced in(Cai et al., 1995), based on two matrices, the prioritymatrix and the interest relation matrix, with a learn-ing algorithm used to adjust the priority matrix. Re-cently, a Micro Autonomous Robotic System (MARS)has been built consisting of robots of 20 cubic mm andequipped with infrared communications (Mitsumotoet al., 1995).

ACTRESS. The ACTRESS (ACTor-based Robot andEquipments Synthetic System) project (Asama et al.,1989; Ishida et al., 1991; Asama et al., 1992) is inspiredby the Universal Modular ACTOR Formalism (Hewittet al., 1973). In the ACTRESS system, “robotors”,including 3 robots and 3 workstations (one as inter-face to human operator, one as image processor andone as global environment manager), form a hetero-geneous group trying to perform tasks such as objectpushing (Asama et al., 1991a) that cannot be accom-plished by any of the individual robotors alone (Ishidaet al., 1994; Suzuki et al., 1995). Communication pro-tocols at different abstraction levels (Matsumoto et al.,1990) provide a means upon which “group cast” andnegotiation mechanisms based on Contract Net (Smith,1980) and multistage negotiation protocols are built(Asama et al., 1994). Various issues are studied, suchas efficient communications between robots and en-vironment managers (Asama et al., 1991b), collisionavoidance (Asama et al., 1991c).

SWARM. A SWARMis a distributed system witha large number of autonomous robots (Jin et al.,1994). (Note that the work on SWARM systemsbegan as work on Cellular Robotic Systems (Beni,1988), where many simple agents occupied one- or

P1: MVG



two-dimensional environments and were able to per-form tasks such as pattern generation and self-organization). SWARM intelligenceis “a propertyof systems of non-intelligent robots exhibiting col-lectively intelligent behavior” (Hackwood and Beni,1991). Self-organizationin a SWARM is the abil-ity to distribute itself “optimally” for a given task,e.g., via geometric pattern formation or structural or-ganization. SWARM exhibits a distributed architec-ture, usually with no differentiation among members(an exception is Hackwood and Beni (1992), wheretwo different types of robots were used). Interactiontakes place by each cell reacting to the state of itsnearest neighbors. Mechanisms for self-organizationin SWARM are studied in Hackwood and Beni (1992),Beni and Wang (1991), Beni and Hackwood (1992),Hackwood and Beni (1991), Liang and Beni (1995),Jin et al. (1994). Examples for possible applicationsinclude large-scale displays and distributed sensing(Hackwood and Wang, 1988). Communication prim-itives have been an important part of research inSWARM (Wang and Beni, 1988; Wang and Premvuti,1995) (see Section 3.2 below for more details).

GOFER. The GOFER architecture (Caloud et al.,1990; LePape, 1990) was used to study distributedproblem solving by multiple mobile robots in an in-door environment using traditional AI techniques. InGOFER, a central task planning and scheduling sys-tem (CTPS) communicates with all robots and has aglobal view of both the tasks to be performed and theavailability of robots to perform the tasks. The CTPSgenerates a plan structure (template for an instance ofa plan) and informs all available robots of the pendinggoals and plan structures. Robots use a task alloca-tion algorithm like the Contract Net Protocol (Smith,1980) to determine their roles. Given the goals as-signed during the task allocation process, they attemptto achieve their goals using fairly standard AI planningtechniques. The GOFER architecture was successfullyused with three physical robots for tasks such as fol-lowing, box-pushing, and wall tracking in a corridor.

Other architectures that make significant use of con-cepts studied within the classical distributed paradigmare described in Lueth and Laengle (1994), Noreils(1990, 1992, 1993), Noreils and de Nozay (1992),Alani et al. (1995), Albus (1993).

ALLIANCE/L-ALLIANCE. The ALLIANCE archi-tecture was developed by Parker (1994a, b) inorder to study cooperation in a heterogeneous,

small-to-medium-sized team of largely independent,loosely coupled robots. Robots are assumed able to,with some probability, sense the effects of their ownactions and the actions of other agents through per-ception and explicit broadcast communications. Indi-vidual robots are based on a behavior-based controllerwith an extension for activating “behavior sets” thataccomplish certain tasks. These sets are activated bymotivational behaviors whose activations are in turndetermined by the robots’ awareness of their team-mates. L-ALLIANCE (Parker, 1994b) is an extensionto ALLIANCE that uses reinforcement learning to ad-just the parameters controlling behavior set activation.The ALLIANCE/L-ALLIANCE architecture has beenimplemented both on real robots and in simulation, andhas been successfully demonstrated for tasks includingbox-pushing, puck-gathering, marching in formation,and simulations of hazardous waste cleanup and jani-torial service.

Behavior-Based Cooperative Behavior. Mataric(1992a, c, 1993, 1994a) proposes a behavior-based ar-chitecture for the synthesis of collective behaviors suchas flocking, foraging, and docking based on the directand temporal composition of primitive basic behaviors(safe-wandering, following, aggregation, dispersion,homing). A method for automatically constructingcomposite behaviors based on reinforcement learningis also proposed. The architecture has been imple-mented both on groups of up to 20 real robots (thelargest group reported in the works we surveyed) andin simulation.

Similar behavior-based architectures include thework by Kube et al., which is based on an AdaptiveLogic Network, a neural network (Kube and Zhang,1992, 1993, 1994; Kube et al., 1993), the Tropism-Based Cognitive Architecture (Agah and Bekey, 1994),and an architecture based on “instinctive behaviors”(Dario et al., 1991).

2.2. Resource Conflict

When a single indivisible resource is requested by mul-tiple robots,resource conflictarises. This issue hasbeen studied in many guises, notably the mutual exclu-sion problem in distributed algorithms and the multi-access problem in computer networks. With multiplerobots, resource conflict occurs when there is a needto share space, manipulable objects or communica-tion media. Few works have dealt specifically with

P1: MVG



object sharing or sharing of communication media (i.e.,sharing of communication media is usually achievedusing very basic techniques—wireless LAN, straight-forward time-division multiplexing, or broadcast overan RF channel; recently Wang and Premvuti (1994),Yoshida et al. (1995a, b) have considered some prob-lems in sharing communications channels). We there-fore center on the space sharing problem, which hasbeen studied primarily via multiple-robot path plan-ning (the “traffic control” formulation from above) andthe collision and deadlock avoidance problems.

In a multi-robot system, each robot can conceiv-ably plan a path that accounts for other robots andthe global environment via configuration space-time,explicit models of other agents, or other techniquesFor example, (Fukuda and Sekiyama, 1994) proposesa “hierarchical prediction model” which essentiallyuses simulation to achieve collision avoidance. (Rude,1994) considers the problem of crossing an intersec-tion: event transforms into the local space-time co-ordinate frame of a robot are applied, and each robot(i) iteratively updates the local frame and its objects,(ii) evaluates collision risk, and (iii) generates a modi-fied path depending on the collision risk. (See also Sec-tion 2.5). However, researchers considering real-worldmulti-robot systems typically conclude that planningpaths in advance is impossible. Thus, robots are oftenrestricted to prescribed paths or roads, with rules (muchlike traffic laws in the human world) and communi-cations used to avoid collision and deadlock (Caloudet al., 1990; Asama et al., 1991a).

Grossman (1988) classifies instances of the trafficcontrol problem into three types: (i) restricted roads,(ii) multiple possible roads with robots selecting au-tonomously between them, and (iii) multiple possibleroads with centralized traffic control. When individualrobots possess unique roads from one point to another,no conflict is possible; when there is global knowl-edge and centralized control, it is easy to prevent con-flict. Thus, the interesting case is (ii), where robotsare allowed to autonomously select roads. Analysis inGrossman (1988) shows that restricted roads are highlysuboptimal, and that the autonomous road choice cou-pled with a greedy policy for escaping blocked situ-ations is far more effective (cf. “modest cooperation”(Premvuti and Yuta, 1990), where robots are assumedto be benevolent for the common good of the system).

Solutions to the traffic control problem range fromrule-based solutions to approaches with antecedents indistributed processing. In Kato et al. (1992), robots

follow pre-planned paths and use rules for collisionavoidance. Example rules include “keep-right”, “stopat intersection”, and “keep sufficient space to the robotin front of you”. Asama et al. (1991c) solves colli-sion avoidance using two simple rules and a commu-nication protocol that resolves conflict by transmittingindividual priorities based on the task requirement, theenvironment, and the robot performance. In Yuta andPremvuti (1992), the robots stop at an intersection andindicate both the number of robots at the intersectionand the directions in which they are traveling. If dead-lock is possible, each robot performs “shunting” (tryingto obtain high priority) and proceeds according to theagreed-upon priorities. Wang (1991) takes a distributedcomputing approach to traffic control, where the par-ticular problem solved is to keep the number of robotstraveling on any given path below a threshold value.Robots use a mutual exclusion protocol to compete forthe right to travel on each path. Wang and Beni (1990)adapt two distributed algorithms to solve two problemsin their CRS/SWARM architecture. The “n-way inter-section problem” is solved using an algorithm similarto mutual exclusion, and the “knot detection problem”is solved using an algorithm similar to distributed dead-lock detection.

2.3. The Origin of Cooperation

In almost all of the work in collective robotics sofar, it has been assumed that cooperation is explicitlydesigned into the system. An interesting research prob-lem is to study how cooperation can arise without ex-plicit human motivation among possibly selfish agents.

McFarland (1994) distinguishes between two signif-icantly different types of group behaviors that are foundin nature: eusocial behaviorandcooperative behav-ior. Eusocial behavior is found in many insect species(e.g., colonies of ants or bees), and is the result of ge-netically determined individual behavior. In eusocialsocieties, individual agents are not very capable, butseemingly intelligent behavior arises out of their inter-actions. This “cooperative” behavior is necessary forthe survival of the individuals in the colonies. Wernerand Dyer (1992) studies the evolution of herding be-havior in “prey” agents in a simulated ecology, wherethere is noa priori drive for cooperation. Recently,McFarland (1994), Steels (1994) have laid the initialgroundwork to address the problem of emergent co-operation in an ecological system inhabited by actual

P1: MVG



mobile robots. In their ecosystem, individual robotsare selfish, utility-driven agents that must cooperate inorder to survive (i.e., maintain some minimal energylevel).

On the other hand, McFarland (1994) defines cooper-ative behavior as the social behavior observed in higheranimals (vertebrates); i.e., cooperation is the result ofinteractions between selfish agents. Unlike eusocialbehavior, cooperative behavior is not motivated by in-nate behavior, but by an intentional desire to cooperatein order to maximize individual utility.

Inspired by economics and game-theoretic ap-proaches (Bond and Gasser, 1988; Geneserethet al., 1986; Rosenschein and Genesereth, 1985;Rosenschein and Zlotkin, 1994) and others havestudied the emergence of cooperation in selfish ratio-nal agents in the field of distributed artificial intelli-gence (DAI). A recent work in the robotics literaturethat adopts this game-theoretic approach is Numaoka(1993).

2.4. Learning

Finding the correct values for control parameters thatlead to a desired cooperative behavior can be a difficult,time-consuming task for a human designer. Therefore,it is highly desirable for multiple-robot systems to beable tolearncontrol parameter values in order to opti-mize their task performance, and to adapt to changes inthe environment. Reinforcement learning (Barto et al.,1983; Kaelbling, 1993) has often been used in cooper-ative robotics.

Mataric (1994a, b) proposes a reformulation of thereinforcement learning paradigm using higher levelsof abstraction (conditions, behaviors, and heteroge-neous reward functions and progress estimators in-stead of states, actions, and reinforcement) to enablerobots to learn a composite foraging behavior. Parker(1994b) uses standard reinforcement algorithms to im-prove the performance of cooperating agents in the L-ALLIANCE architecture by having the agents learnhow to better estimate the performance of other agents.Sen et al. (1994) uses reinforcement learning in a two-robot box-pushing system, and Yanco and Stein (1992)applies reinforcement learning to learn a simple, artifi-cial robot language. Other relevant works in multiagentreinforcement learning (done in simulation, in contrastto the above works which were implemented on actualrobots) include Whitehead (1991), Tan (1993), Littman(1994).

In addition, techniques inspired by biological evo-lution have also been used in cooperative robotics.Werner and Dyer (1992) uses a genetic algorithm(Goldberg, 1989) to evolve neural network controllersfor simulated “prey” creatures that learn a herding be-havior to help avoid predators. Reynolds (1992) usesgenetic programming (Koza, 1990) to evolve flockingbehavior in simulated “boids”.

2.5. Geometric Problems

Because mobile robots can move about in the physicalworld and must interact with each other physically,geometric problems are inherent to multiple-robotsystems. This is a fundamental property that distin-guishes multiple-robot systems from traditional dis-tributed computer systems in which individual nodesare stationary. Geometric problems that have beenstudied in the cooperative robotics literature includemultiple-robot path planning, moving to (and main-taining) formation, and pattern generation.

(Multiple-Robot) Path Planning.Recall that multiple-robot path planning requires agents to plan routes thatdo not intersect. This is a case of resource conflict,since the agents and their goals are embedded in a finiteamount of space. However, we note path planningseparately because of its intrinsic geometric flavor aswell as its historical importance in the literature.

Detailed reviews of path planning are found inFujimura (1991), Latombe (1991), Arai and Ota(1992). Fujimura (1991) views path planning as ei-ther centralized (with a universal path-planner makingdecisions) or distributed (with individual agents plan-ning and adjusting their paths). Arai and Ota (1992)make a similar distinction in the nature of the planner,and also allow hybrid systems that can be combina-tions of on-line, off-line, centralized, or decentralized.Latombe (1991) gives a somewhat different taxonomy:his “centralized planning” is planning that takes intoaccount all robots, while “decoupled” planning en-tails planning the path of each robot independently. Forcentralized planning, several methods originally usedfor single-robot systems can be applied. For decou-pled planning, two approaches are given: (i)prioritizedplanningconsiders one robot at a time according to aglobal priority, while (ii) thepath coordination methodessentially plans paths by scheduling the configura-tion space-time resource. The work of Erdmann andLozano-Perez (1986) is a typical decoupled approach

P1: MVG



where every robot is prioritized and robots plan globalpaths with respect to only higher-priority robots (e.g.,the highest-priority robot plans only around the ob-stacles in the environment). Note that this is still acentralized method according to the terminology of(Fujimura, 1991; Arai and Ota, 1992). On the otherhand, Yeung and Bekey (1987) presents a distributedapproach (per Fujimura’s taxonomy) where each robotinitially attempts a straight-line path to the goal; if aninterfering obstacle is seen, then the robot will scan thevisible vertices of the obstacle and move toward theclosest one. In general, this continues until the goalis reached. Dynamically varying priorities are givento each robot (based on current need) to resolve pathintersection conflicts, and conflicting robots can eithernegotiate among themselves or allow a global black-board manager to perform this function.

Some recent works have addressed some nontradi-tional motion planning problems. For example, Hertand Lumelsky (1995) proposes an algorithm for pathplanning in tethered robots, and (Ota et al., 1995)consider the problem of moving while grasping largeobjects.

The Formation and Marching Problems.The Forma-tion and Marching problems respectively require mul-tiple robots to form up and move in a specified pattern.Solving these problems is quite interesting in termsof distributed algorithms (Sugihara and Suzuki, 1990),balancing between global and local knowledge (Parker,1994b), and intrinsic information requirements for agiven task. Solutions to Formation and Marching arealso useful primitives for larger tasks, e.g., moving alarge object by a group of robots (Stilwell and Bay,1993; Chen and Luh, 1994a, b) or distributed sensing(Wang and Beni, 1988).

The Formation problem seems very difficult, e.g.,no published work has yet given a distributed “circle-forming” algorithm that guarantees the robots will ac-tually end up in a circle. For this problem, the bestknown solution is the distributed algorithm of Sugiharaand Suzuki (1990), which guarantees only that therobots will end up in a shape of constant diameter(e.g., a Reuleaux triangle can be the result). It is as-sumed that thei th mobile robot knows the distancesDi and di to its farthest and nearest neighbors, re-spectively; the algorithm attempts to match the ratiosDi /di to a prescribed constant. No method of de-tecting termination was given. Chen and Luh (1994a,b) extend the method of Sugihara and Suzuki (1990)

to incorporate collision avoidance when the robotsare moving. Yamaguchi and Arai (1994) approachesthe shape-generation problem using systems of linearequations; starting at some initial location, each robotchanges its(x, y) position according to a linear func-tion of its neighbors’ positions and some fixed constant.Simulations of the method show that a group of ini-tially collinear robots will converge into the shape of anarc.

We observe that the circle-forming problem, whilequite simple to state, reveals several pitfalls in formu-lating distributed geometric tasks. For example, theability of an individual agent to sense attributes ofthe formation must be carefully considered: too muchinformation makes the problem trivial, but too littleinformation (e.g., returns from localized sensors) mayprevent a solution (e.g., robots may never find eachother). Information lower bounds, e.g., for robots to beable to realize that they have achieved the prescribedformation, are also largely unexplored in the literature.Interestingly, we note that the algorithm of (Sugiharaand Suzuki, 1990) can be slightly modified: rather thaneach robot seeking to achieve a prescribed ratioD/d,each robot could seek to achieve a prescribed angle(close to 90 degrees) subtended by its farthest neigh-bor and its closest neighbor to the right. This uses verysimilar sensing capabilities but guarantees the desiredcircular shape.

For Marching, Chen and Luh (1994a) employs po-sitional constraint conditions in a group of robots thatmakes turns while maintaining an array pattern. InChen and Luh (1994b) a leader-follower approach isused to solve a similar task. Parker (1993) studies theproblem of keeping four marching robots in a side-by-side formation; this increases in difficulty whenthe leader has to perform obstacle avoidance or othermaneuvers. Parker also defines the concepts of globalgoals and global/local knowledge. To study the effectsof different distributions of global goals and globalknowledge, four strategies are compared both in sim-ulation and on mobile robots. Simplified instances ofthe Marching problem require robots to reliably fol-low each other and to move in a group (without tightconstraints on their relative positions). Some worksthat address this problem (sometimes referred to as theherding/flocking problem) include Reynolds (1987),Mataric (1994a), Hodgins and Brogan (1994), Broganand Hodgins (1995) Milios and Wilkes (1995). A some-what related problem is the problem of cooperativepositioning (determining the locations of the robots

P1: MVG



in a group using limited information) (Kurazume andNagata, 1994).

Related to the Formation problem is the pattern gen-eration problem in Cellular Robotic Systems, multiple-robot systems which can “encode information aspatterns of its own structural units” (Beni, 1988). Typ-ically, one- or two-dimensional grids constitute theworkspace, and sensing of neighboring cells is the onlyinput. Within these constraints, a set of rules is de-vised and applied to all agents; a standard result is toshow in simulation that convergence to some spatialpattern is guaranteed. The meaningful aspect of thiswork lies in providing a system with the capability ofspatial self-organization: a CRS will reconfigure itselfwithout intervention in certain situations or under cer-tain conditions.

In (Wang and Beni, 1988), (Liang and Beni, 1995),a CRS is characterized as an arbitrary number of robotsin a one- or two-dimensional grid. The robots are ableto sense neighboring cells and communicate with otherrobots via a signboard mechanism. Protocols are pre-sented for creating different patterns, e.g., alternatingrobots and spaces in a one-dimensional grid; coveringthe top row of a two-dimensional grid by robots; orcovering the boundary of a two-dimensional grid byrobots. Egecioglu and Zimmermann (1988) pose the“Random Pairing” problem, and seek a set of rules bywhich for any given number, a CRS will converge toa pattern such that there is a group of two robots withthat number of vacant spaces between them (see also(Beni and Hackwood, 1992)). An analogous cellularapproach is adopted by Genovese et al. (1992) who de-scribe the simulation of a system of pollutant-seekingmobile robots. The simulation uses a potential fieldmechanism to attract robots to the pollutant and to re-pulse robots from each other. The combined effect ofthese two forces yields a gradient pattern that “points”toward the source of the pollutant.

3. Perspectives

As an integrative engineering discipline, robotics hasalways had to confront technological constraints thatlimit the domains that can be studied. Cooperativerobotics has been subject to these same constraints,but the constraints tend to be more severe because ofthe need to cope with multiple robots. At the sametime, cooperative robotics is a highly interdisciplinaryfield that offers the opportunity to draw influences frommany other domains. In this section, we first outline

some of the technological constraints that face the field.We then mention some directions in whichcooperativeroboticsmight progress, and describe related fields thathave provided and will continue to provide influences.

3.1. Technological Constraints

It is clear that technological constraints have limited thescope of implementations and task domains attemptedin multiple-robot research systems.

One obvious problem that arises is the general prob-lem of researchers having to solve various instances ofthe vision problem before being able to make progresson “higher-level” problems. Often, difficulties arisingfrom having to solve difficult perceptual problems canlimit the range of tasks that can be implemented on amultiple-robot platform. For example, in cooperativerobotics systems where modeling of other agents (seeSection 2.1) is used, the lack of an effective sensor ar-ray can render the system unimplementable in practice.In addition, robot hardware is also notoriously unreli-able; as a result, it is extremely difficult to maintain afleet of robots in working condition. Again, collectiverobotics must deal with all of the hardware problems ofsingle-robotic systems, exacerbated by the multiplicityof agents.

Due to the difficulties (such as those outlined above)encountered when working with real robots, much ofcollective robotics has been studied exclusively in sim-ulation. Some researchers have argued (cf. Brooks,1991) that by ignoring most of the difficulties associ-ated with perception and actuation, simulations ignorethe most difficult problems of robotics. By makingoverly simplistic assumptions, it is possible to gener-ate “successful” systems in simulation that would beinfeasible in the real world. (Conversely, mobile re-search robots can also come to “look like the simula-tor”, i.e., circular footprint, sonar ring, synchro-driveis a common configuration.) Nevertheless, simulationmust inevitably play a role in multi-agent robotics atsome level. Although it is currently possible for re-searchers to study groups of 10–20 robots, it is unlikelythat truly large-scale collective behavior involving hun-dreds or thousands of real robots will be feasible atany time in the near future. Thus, cooperative mobilerobot researchers have used a variety of techniques tosimulate perception, while using physical robots. Forinstance, the use of a global positioning system can inpart compensate for the lack of vision, but can placesevere environmental constraints under which robots

P1: MVG



can operate (because many objects and acoustic fea-tures of the environment can interfere with the GPS).For the basic problem of differentiating between otheragents and all other objects in the environment. Someresearchers (Parker, 1994b) use radio communicationto solve this problem. In other works (Donald, 1993a;Parker, 1994b) interaction via sensing is done by ex-plicit radio communication. There are recent attemptsto perform recognition via vision (Kuniyoshi et al.,1994a, b).

An approach taken by some researchers is to usesimulations as prototypes for larger-scale studies, andsmall numbers of real robots as a proof-of-conceptdemonstration (Mataric, 1994a; Parker, 1994b). On theother hand, some researchers, citing the necessity ofworking in the real world domain, have chosen to es-chew simulations altogether and implement their the-ories directly on actual robots (Beckers et al., 1994;McFarland, 1994; Steels, 1994). In studies of loco-motion in large herds of (upto 100) one-legged robotsand simulated human cyclists, (Hodgins and Brogan,1994; Brogan and Hodgins, 1995) take an alternate ap-proach of design a very physically realistic simulation.While this approach brings realism to actuation, theissue of perception is still simulated away; it is still un-clear whether it will be feasible to realistically modelsophisticated agents in more complex environments, orwhether the effort will outweigh the benefits.

3.2. Towards a Science of Cooperative Robotics

The field of cooperative mobile robotics offers an in-credibly rich application domain, integrating a hugenumber of distinct fields from the social sciences, lifesciences, and engineering. That so many theories havebeen brought to bear on “cooperative robotics” clearlyshows the energy and the allure of the field. Yet, co-operative robotics is still an emerging field, and manyopen directions remain. In this subsection, we pointout some promising directions that have yet to be fullyexplored by the research community. By way of a pref-ace, we also point out three “cultural” changes whichmay come as the field matures: (1) Because of theyouth of the field, cooperative robotics research hasbeen necessarily rather informal and “concept” ori-ented. However, the development of rigorous for-malisms is desirable to clarify various assumptionsabout the systems being discussed, and to obtain a moreprecise language for discussion of elusive concepts

such as cooperation (there are some exceptions, suchas (Parker, 1994b), which presents a formalization ofmotivational behavior in the ALLIANCE architecture).(2) Formal metrics for cooperation and system perfor-mance, as well as for grades of cooperation, are notice-ably missing from the literature. While the notion ofcooperation is difficult to formalize, such metrics willbe very useful in characterizing various systems, andwould improve our understanding of the nature of agentinteractions. Although Mataric (1994a) has suggestedparameters such as agent density for estimating inter-ference in a multi-robot system, much more work inthis area is necessary. (3) Experimental studies mightbecome more rigorous and thorough, e.g., via standardbenchmark problems and algorithms. This is challeng-ing in mobile robotics, given the noisy, system-specificnature of the field. Nevertheless, it is necessary forclaims about “robustness” and “near-optimality” to beappropriately quantified, and for dependencies on vari-ous control parameters to be better understood. For ex-ample, we have noted that despite a number of claimsthat various decentralized approaches are superior tocentralized approaches, we have not seen any thorough,published experimental comparisons between the ma-jor competing paradigms on a particular task. How-ever, we note that recently, researchers have begunto empirically study quantitative measures of coopera-tion, trying to identify conditions under which mecha-nisms for cooperation are beneficial (Arkin et al., 1993;Arkin and Hobbs, 1993; Parker, 1995).

Finally, several basic analogies remain incomplete,and must be revisited and resynthesized as the fieldmatures. For instance, many multi-robot problems are“canonical” for distributed computation and are inter-esting primarily when viewed in this light. A typicalexample is moving to formation, which has been solvedoptimally in the computational geometry literature (itis the “geometric matching under isometry” problem(Heffernan and Schirra, 1992)), but which is difficultin the distributed context due to issues like synchro-nization, fault-tolerance, leader election, etc. However,the distributed context can be selectively ignored, e.g.,Sugihara and Suzuki (1990) use “human intervention”to perform what is essentially leader election (breakingsymmetry in a circle of robots to choose vertices of thedesired polygonal formation). The introduction of suchdevices runs counter to the implicit assumption that itis thedistributedproblem that holds research interest.

More generally, it is likely that more structural andless superficial analogies with other disciplines will

P1: MVG



be needed in order to obtain “principled” theories ofcooperation among (mobile) robots; integration of for-malisms and methodologies developed in these moremature disciplines is likely to be an important step in thedevelopment of cooperative robotics. Disciplines mostcritical to the growth of cooperative robotics are: dis-tributed artificial intelligence, biology, and distributedsystems.

Distributed Artificial Intelligence. The field of dis-tributed artificial intelligence (DAI) concerns itselfwith the study of distributed systems of intelligentagents. As such, this field is highly relevant to co-operative robotics. Bond and Gasser (1988) defineDAI as “the subfield of artificial intelligence (AI) con-cerned with concurrency in AI computations, at manylevels”. Grounded in traditional symbolic AI and thesocial sciences, DAI is composed of two major areas ofstudy: Distributed Problem Solving (DPS) and Multi-agent Systems (MAS).

Research in DPS is concerned with the issue of solv-ing a single problem using many agents. Agents cancooperate by independently solving subproblems (task-sharing), and by periodically communicating partialsolutions to each other (result-sharing). DPS involvesthree possibly overlapping phases: (i) problem decom-position (task allocation), (ii) subproblem solution, and(iii) solution synthesis. Of these, problem decompo-sition has attracted the greatest interest among DAIresearchers. The critical issue in task sharing is find-ing the appropriate agent to assign to a subproblem.This is nontrivial, since if the most appropriate agentfor a subtask is not obvious, then the system must try todetermine which of the many eligible agents should beassigned the task, and often there are too many eligibleagents to attempt an exhaustive search. Perhaps the bestknown scheme for task allocation is the Contract NetProtocol (Smith, 1980), which has been used in the AC-TRESS (Ishida et al., 1994; Asama et al., 1994; Ozakiet al., 1993) and GOFER (Caloud et al., 1990) projects.

One important assumption in DPS is that the agentsare predisposed to cooperate. Research in DPS is thusconcerned with developing frameworks for coopera-tive behavior between willing agents, rather than de-veloping frameworks to enforce cooperation betweenpotentially incompatible agents, as is the case with mul-tiagent systems and distributed processing.

Multiagent Systems (MAS) research is the study ofthe collective behavior of a group of possibly hetero-geneous agents with potentially conflicting goals. In

other words, researchers in MAS discard the “benev-olent agent” assumption of DPS (Genesereth et al.,1986). Genesereth et al. (1986) states the centralproblem of MAS research as follows: “in a worldin which we get to design only our own intelligentagent, how should it interact with other intelligentagents?” Therefore, areas of interest in MAS researchinclude game-theoretic analysis of multi-agent inter-actions (cf. Genesereth et al., 1986; Rosenschein andGenesereth, 1985; Rosenschein and Zlotkin, 1994),reasoning about other agents’ goals, beliefs, and ac-tions (cf. Rosenschein, 1982; Georgeff, 1983, 1984),and analysis of the complexity of social interactions(Shoham and Tennenholtz, 1992).

Work in MAS has tended to be theoretical and in veryabstract domains. A common underlying assumption isthat although the agents may be selfish, they are rationaland highly deliberative. This is in stark contrast withresearch in swarm intelligence (see Section 2.5), inwhich individual agents are assumed to be relativelyunintelligent.

However, the influence of DAI on cooperativerobotics has been limited. This is in part because re-searchers in DAI have mostly concentrated on domainswhere uncertainty is not as much of an issue as it is in thephysical world. Work in MAS has tended to be theoret-ical and in very abstract domains where perfect sensingis usually assumed; typical DPS domains are in disem-bodied, knowledge-based systems. Another assump-tion of DAI that has prevented its application in cooper-ative robotics is the assumption is that although agentsmay be selfish, they are rational and highly deliberative.However, achieving strict criteria of rationality and de-liberativeness can often be prohibitively expensive incurrent robotic systems. Thus, it has been argued thatDAI, while suited for unsituated, knowledge-based sys-tems, will not succeed in the domain of cooperativerobotics (Parker, 1994b; Mataric, 1994a). However,we observe that direct comparisons of DAI and alterna-tive paradigms are notably missing from the literature;such comparisons are needed to evaluate the true util-ity of DAI techniques in cooperative robotics. Also,as lower-level processes (perception and actuation) arebetter understood and implemented, and as computa-tional power increases, the high-level results of DAIresearch may become increasingly applicable to col-lective mobile robotics.

Distributed Systems.A multiple-robot system is infact a special case of a distributed system. Thus, the

P1: MVG



field of distributed systems is a natural source of ideasand solutions. Beni (1988) describes cellular roboticsas belonging to the general field of distributed com-puting. It is noted, however, that distributed comput-ing can only contribute general theoretical foundationsand that further progress needs to be made concerningthe application of such methods to collective robotics.Wang and Beni (1990) states, “a distributed comput-ing system contains a collection of computing deviceswhich may reside in geographically separated locationscalled sites”. By noting the similarities with distributedcomputing, theories pertaining to deadlock (Wang andBeni, 1988, 1990; Lin and Hsu, 1995), message passing(Wang and Beni, 1990) and resource allocation (Wang,1991), and the combination of the above as primitives(Wang, 1995; Wang and Premvuti, 1995), have beenapplied to collective robotics in a number of works.

In work done on multiple AGV systems and Dis-tributed Robotic Systems, deadlock detection and re-source allocation methods are applied to allow manyrobots to share the limited resource of path space (Wangand Beni, 1990; Wang, 1991). Pattern generation in aCRS may also rely on distributed computing to resolveconflicts (Wang, 1991; Wang and Beni, 1990). Fi-nally, (Wang, 1993, 1994) describe a task allocationalgorithm where the robots vie for the right to partici-pate in a task. See also the discussion in Sections 2.1and 2.5.

Broadcast communication, which is widely assumedin cooperative robotics, exhibits poor scaling proper-ties. As robots become more numerous and widelydistributed, techniques and issues from the field ofcomputer networks become relevant. A rich body ofresearch on algorithms, protocols, performance model-ing and analysis in computer networks can be applied tocooperative robotics. There is currently a great amountof effort being put into studying networking issuesrelated to mobile/nomadic/ubiquitous computing (cf.,Awerbuch and Peleg, 1991; Weiser, 1993; Kleinrock,1995; Badrinath et al., 1994). Results from this fieldcould be applied in a straightforward way to multi-robotsystems.

Finally, distributed control is a promising frameworkfor the coordination of multiple robots. Due to diffi-culty of sensing and communication, a parsimoniousformulation which can coordinate robots having min-imal sensing and communication capabilities is desir-able. In an ideal scenario, maximal fault tolerance ispossible, modeling of other agents is unnecessary, andeach agent is controlled by a very simple mechanism. A

distributed control scheme (known as the GUR game)developed originally by Tsetlin (1964) and recentlystudied in Tung and Kleinrock (1993), Tung (1994)provides a framework in which groups of agents withminimal sensing capability and no communication arecontrolled by simple finite state automata and convergeto optimal behaviors. (Tung, 1994) describes possiblecooperative robotics applications in moving platformcontrol and perimeter guarding.

Biology. Biological analogies and influences aboundin the field of cooperative robotics. The majority of ex-isting work in the field has cited biological systems asinspiration or justification. Well-known collective be-haviors of ants, bees, and other eusocial insects Wilson(1971) provide striking existence proofs that systemscomposed of simple agents can accomplish sophisti-cated tasks in the real world. It is widely held that thecognitive capabilities of these insects are very limited,and that complex behaviorsemergeout of interactionsbetween the agents, which are individually obeyingsimple rules. Thus, rather than following the AI tradi-tion of modeling robots as rational, deliberative agents,some researchers in cooperative robotics have chosento take a more “bottom-up” approach in which individ-ual agents are more like ants—they follow simple rules,and are highly reactive (this is the approach taken inthe field of Artificial Life). Works based on this insect-colony analogy include (Mataric, 1994a; Beckers et al.,1994; Stilwell and Bay, 1993; Doty and Aken, 1993;Johnson and Bay, 1994; Deneubourg et al., 1991a, b).The pattern generation of CRS’s can also be consid-ered as bottom-up (see Section 2.5), since each robotis designed as a very simple agent which follows a setof prespecified rules.

A more general, biological metaphor that is oftenused in cooperative robotics is the concept of aself-organizing system(Nicolis and Prigogine, 1977; Yates,1987). (Note that researchers from many fields havestudied self-organization; it is by no means an exclu-sively biological concept. However, in the field of co-operative robotics, references to self-organization haveoften been made in a biological context.) The behaviorof insect colonies described above can be character-ized more generally as that of self-organizing systems.Representative work that is based on this concept in-cludes (Wang and Beni, 1988; Steels, 1990; Hackwoodand Beni, 1991, 1992, Beni and Hackwood, 1992).Self-organization in multi-cellular biological systemshas been an inspiration for (Hackwood and Wang,

P1: MVG



1988; Beni, 1988; Egecioglu and Zimmermann, 1988;Genovese et al., 1992). Hierarchical organization ofbiological multi-cellular organisms (i.e., from cellularto tissue to organism level) has been used as a guidingmetaphor for cellular robotics in the CEBOT project(Fukuda and Nakagawa, 1987).

Biological analogies have also influenced the choiceof task domains studied in cooperative robotics. Whileforaging is a natural abstraction of some practical ap-plications such as waste retrieval and search and rescue,one major reason that it has become identified as thecanonical cooperative robotic task is that it is a natu-ral task, given the group architectures resulting fromanalogies to insect colonies. Another example of thisphenomenon is the flocking/herding task. It seems noaccident that biological inspirations led to “natural”models of group motion, as opposed to more struc-tured models of coordinated motion (such as movingin some arbitrary formation).

Finally, as we noted in Section 2.4, there have beensome biological influences on the learning and opti-mization algorithms used to tune control parameters inmultiple-robot systems.

4. Conclusions

We have synthesized a view of the theoretical basesfor research in cooperative mobile robotics. Key re-search axes in the field were identified, particularlywith respect to achieving a “mechanism of coopera-tion”, and existing works were surveyed in this frame-work. We then discussed technological constraints andinterdisciplinary influences that have shaped the field,and offered some general precepts for future growthof the field. Finally, we identified distributed artificialintelligence, biology, and distributed systems as disci-plines that are most relevant to cooperative robotics,and which are most likely to continue to provide valu-able influences. Based on our synthesis, a number ofopen research areas become apparent. We believe thatthe following are among the major, yet tractable, chal-lenges for the near future:

1. robust definitions and metrics for various forms ofcooperation,

2. achieving a more complete theory of informationrequirements for task-solving in spatial domains,perhaps for the canonical tasks of pattern forma-tion or distributed sensing (e.g., measures of patterncomplexity, information lower bounds for pattern

recognition and maintenance, abstraction of sen-sor models from the solution approach). The worksof (Donald et al., 1994; Rus et al., 1995; Brownand Jennings, 1995) have begun to address this is-sue, in the context of object manipulation tasks; in-terestingly, (Brown and Jennings, 1995) observesthat given a robot system, some tasks arestronglycooperative—the robots must act in concert toachieve the goal, and the strategy for the task isnot trivially serializable.

3. principled transfer of the concepts of fault-toleranceand reliability from the field of distributed and fault-tolerant computing,

4. incorporation of recent ideas in distributed control toachieve oblivious cooperation, or cooperation with-out communication (e.g., when robots have minimalsensing and communication capabilities),

5. achieving cooperation within competitive situations(e.g., for robot soccer, or pursuit-evasion with multi-ple pursuers and evaders). An interesting open prob-lem is how well solutions that have been developedin discretized abstractions of these domains (cf.,Levy and Rosenchein, 1991; Korf, 1992) translateto the physical world.

Acknowledgments

Partial support for this work was provided by NSFYoung Investigator Award MIP-9257982; the UCLACommotion Laboratory is supported by NSF CDA-9303148. Portions of this work were performed bythe Jet Propulsion Laboratory, California Institute ofTechnology, under contract with the National Aero-nautics and Space Administration. The authors wouldlike to thank B. Donald, T. Fukuda, M. Anthony Lewis,M. Mataric, J. Wang, the anonymous reviewers, andmembers of the UCLA Commotion Lab for helpfulcomments, suggestions, and discussions. Frank Mengassisted in the preparation of a previous version of thispaper.

References

Adams, J.A., Basjcsy, R., Kosecka, J., Kumar, V., Mandelbaum,R., Mintz, M., Paul, R., Wang, C., Yamamoto, Y., and Yun, X.1995. Cooperative material handling by human and robotic agents:Module development and system synthesis. InIEEE/RSJ IROS,pp. 200–205.

Agah, A. and Bekey, G.A. 1994. Autonomous mobile robot teams. InConf. on Intelligent Robotics in Filed, Factory, Service and Space(CIRFFSS94).

P1: MVG



Aguilar, L., Alami, R., Fleury, S., Herrb, M., Ingrand, F., andRobert, F. 1995. Ten autonomous mobile robots (and even more).In IEEE/RSJ IROS, pp. 260–267.

Alani, R., Robert, F., Ingrand, F., and Suzuki, S. 1995. Multi-robotcooperation through incremental plan-merging. InIEEE ICRA,pp. 2573–2579.

Albus, J.S. 1993.A Control Architecture for Cooperative IntelligentRobots, pp. 713–743.

Arai, T. and Ota, J. 1992. Motion planning of multiple robots. InIEEE/RSJ IROS, pp. 1761–1768.

Arai, T., Yoshida, E., and Ota, J. 1993. Information diffusion by localcommunication of multiple mobile robots. InIEEE Conference onSystems, Man and Cybernetics, pp. 535–540.

Arkin, R.C. 1992. Cooperation without communication: Multi-agent schema-based robot navigation.Journal of Robotic Systems,9(3):351–364.

Arkin, R. and Hobbs, J. 1993. Dimensions of communication andsocial organization in multi-agent robotic systems. InProc. Sim-ulation of Adaptive Behavior.

Arkin, R. and Ali, K. 1994. Integration of reactive and teleroboticcontrol in multi-agent robotic systems. InProc. Simulation ofAdaptive Behavior.

Arkin, R.C., Balch, T., and Nitz, E. 1993. Communication of be-havioral state in multi-agent retrieval tasks. InIEEE ICRA, vol. 3,pp. 588–594.

Asada, M., Uchibe, E., Noda, S., Tawaratsumida, S., and Hosoda, K.1994. Coordination of multiple behaviors acquired by a vision-based reinforcement learning. InIEEE/RSJ IROS.

Asama, H. 1992. Distributed autonomous robotic system configu-rated with multiple agents and its cooperative behaviors.Journalof Robotics and Mechatronics, 4(3).

Asama, H., Matsumoto, A., and Ishida, Y. 1989. Design of an au-tonomous and distributed robot system: ACTRESS. InIEEE/RSJIROS, pp. 283–290.

Asama, H., Habib, M.K., Endo, I., Ozaki, K., Matsumoto, A., andIshida, Y. 1991a. Functional distribution among multiple mobilerobots in an autonomous and decentralized robot system. InIEEEICRA, pp. 1921–1926.

Asama, H., Ozaki, K., Ishida, Y., Habib, M.K., Matsumoto, A., andEndo, I. 1991b. Negotiation between multiple mobile robots andan environment manager. InIEEE ICRA, pp. 533–538.

Asama, H., Ozaki, K., Itakura, H., Matsumoto, A., Ishida, Y., andEndo, I. 1991c. Collision avoidance among multiple mobile robotsbased on rules and communication. InIEEE/RSJ IROS, pp. 1215–1220.

Asama, H., Ishida, Y., Ozaki, K., Habib, M.K., Matsumoto, A.,Kaetsu, H., and Endo, I. 1992. A communication system betweenmultiple robotic agents. InProc. the Japan U.S.A. Symposium onFlexible Automation, M. Leu (Ed.), pp. 647–654.

Asama, H., Ozaki, K., Ishida, Y., Yokita, K., Matsumoto, A., Kaetsu,H., and Endo, I. 1994. Collaborative team organization usingcommunication in a decentralized robotic system. InIEEE/RSJIROS.

Awerbuch, B. and Peleg, D. 1991. Concurrent online trackingof mobile users.Computer Communication Review, 21(4):221–233.

Badrinath, B.R., Acharya, A., and Imielinski, T. 1994. Structur-ing distributed algorithms for mobile hosts. InProceedings of the14th International Conference on Distributed Computing Systems,pp. 21–24.

Barnes, D. and Gray, J. 1991. Behaviour synthesis for co-operantmobile robot control. InInternational Conference on Control,pp. 1135–1140.

Barto, A.G., Sutton, R.S., and Watkins, C.J.C.H. 1983. Learningand sequential decision making. InLearning and ComputationalNeuroscience: Foundations of Adaptive Networks, M. Gabriel andJ. Moore (Ed.), MIT Press, pp. 539–603.

Beckers, R., Holland, O.E., and Deneubourg, J.L. 1994. From localactions to global tasks: Stigmergy and collective robotics. InProc.A-Life IV, MIT Press.

Beni, G. 1988. The concept of cellular robotic system. InIEEE In-ternational Symposium on Intelligent Control, pp. 57–62.

Beni, G. and Wang, J. 1991. Theoretical problems for the realizationof distributed robotic systems. InIEEE ICRA, pp. 1914–1920.

Beni, G. and Hackwood, S. 1992. Stationary waves in cyclic swarms.In IEEE International Symposium on Intelligent Control, pp. 234–242.

Bond, A.H. and Gasser, L. 1988.Readings in Distributed ArtificialIntelligence. Morgan Kaufmann Publishers.

Brogan, D.C. and Hodgins, J.C. 1995. Group behaviors for systemswith significant dynamics. InIEEE/RSJ IROS, pp. 528–534.

Brooks, R.A. 1986. A robust layered control system for a mobilerobot. IEEE Journal of Robotics and Automation, RA-2(1):14–23.

Brooks, R.A. 1991. Intelligence without reason. InProc. Intl. JointConf. Artificial Intelligence, pp. 569–595.

Brooks, R.A., Maes, P., Mataric, M.J., and More, G., 1990. Lunarbase construction robots. InIEEE/RSJ IROS. IEEE.

Brown, R.G. and Jennings, J.S. 1995. A pusher/steerer model forstrongly cooperative mobile robot cooperation. InIEEE/RSJ IROS,pp. 562–568.

Cai, A.-H., Fukuda, T., Arai, F., Ueyama, T., and Sakai, A. 1995.Hierarchical control architecture for cellular robotic system. InIEEE ICRA, pp. 1191–1196.

Caloud, P., Choi, W., Latombe, J.-C., Pape, Le C., and Yin, M. 1990.Indoor automation with many mobile robots. InIEEE/RSJ IROS,pp. 67–72.

Causse, O. and Pampagnin, L.H. 1995. Management of a multi-robot system in a public environment. InIEEE/RSJ IROS, pp. 246–252.

Chen, Q. and Luh, J.Y.S. 1994a. Coordination and control of a groupof small mobile robots. InIEEE ICRA, pp. 2315–2320.

Chen, Q. and Luh, J.Y.S. 1994b. Distributed motion coordination ofmultiple robots. InIEEE/RSJ IROS, pp. 1493–1500.

Connell, J. 1987. Creature design with the subsumption architecture.In Proc. AAAI, pp. 1124–1126.

Dario, P., Ribechini, F., Genovese, V., and Sandini, G. 1991. Instinc-tive behaviors and personalities in societies of cellular robots. InIEEE ICRA, pp. 1927–1932.

Deneubourg, J., Goss, S., Franks, N., Sendova-Franks, A., De-train, C., and Chretien, L. 1991a. The dynamics of collective sort-ing: Robot-like ants and ant-like robots. InProc. Simulation ofAdaptive Behaviour.

Deneubourg, J.-L., Theraulaz, G., and Beckers, R. 1991b. Swarm-made architectures. InProc. European Conference on ArtificialLife., Verela and Bourgine (Eds.), MIT Press, pp. 123–133.

Donald, B.R. 1993a. Information invariants in robotics: I. State,communication, and side-effects. InIEEE ICRA, pp. 276–283.

Donald, B.R. 1993b. Information invariants in robotics: II. Sensorsand computation. InIEEE ICRA, vol. 3, pp. 284–90.

P1: MVG



Donald, B.R., Jennings, J., and Rus, D. 1994. Analyzing teams ofcooperating mobile robots. InIEEE ICRA, pp. 1896–1903.

Dorf, R. 1990.Concise International Encyclopedia of Robotics: Ap-plications and Automation. Wiley-Interscience.

Doty, K.L. and Aken, R.E. Van 1993. Swarm robot materials handlingparadigm for a manufacturing workcell. InIEEE ICRA, vol. 1,pp. 778–782.

Drexler, K.E. 1992.Nanosystems: Molecular Machinery, Manufac-turing, and Computation. John Wiley and Sons, Inc.

Drgoul, A. and Ferber, J. 1993. From tom thumb to the dockers:Some experiments with foraging robots. InProc. Simulation ofAdaptive Behavior.

Dudek, G., Jenkin, M., Milios, E., and Wilkes, D. 1993. A taxonomyfor swarm robots. InIEEE/RSJ IROS, pp. 441–447.

Egecioglu, O. and Zimmermann, B. 1988. The one dimensional ran-dom pairing problem in a cellular robotic system. InIEEE Inter-national Symposium on Intelligent Control, pp. 76–80.

Erdmann, M. and Lozano-Perez, T. 1986. On multiple moving ob-jects. InIEEE ICRA, pp. 1419–1424.

Everett, H., Gilbreath, G.A., Heath-Pastore, T.A., and Laird, R.T.1993. Coordinated control of multiple security robots.MobileRobots VIII, 2058:292–305.

Franklin, D.E., Kahng, A.B., and Lewis, M.A. 1995. Distributedsensing and probing with multiple search agents: Toward system-level landmine detection solutions. InDetection Technologies forMines and Minelike Targets, Proceedings of SPIE, vol. 2496,pp. 698–709.

Fujimura, K. 1991.Motion Planning in Dynamic Environments.Springer-Verlag, New York, NY.

Fukuda, T. and Nakagawa, S. 1987. A dynamically reconfigurablerobotic system (concept of a system and optimal configurations).In International Conference on Industrial Electronics, Control,and Instrumentation, pp. 588–595.

Fukuda, T., Kawauchi, Y., and Asama, H. 1990. Analysis and eval-uation of cellular robotics (CEBOT) as a distributed intelligentsystem by communication amount. InIEEE/RSJ IROS, pp. 827–834.

Fukuda, T. and Kawauchi, Y. 1993.Cellular Robotics, Springer-Verlag, pp. 745–782.

Fukuda, T. and Sekiyama, K. 1994. Communication reduction withrisk estimate for multiple robotic system. InIEEE ICRA, pp. 2864–2869.

Fukuda, T. and Iritani, G. 1995. Construction mechanism of groupbehavior with cooperation. InIEEE/RSJ IROS, pp. 535–542.

Gage, D. 1993. How to communicate to zillions of robots. InMobileRobots VIII, SPIE, pp. 250–257.

Genesereth, M.R., Ginsberg, M.L., and Rosenschein, J.S. 1986. Co-operation without communication. InProc. AAAI, pp. 51–57.

Genovese, V., Dario, P., Magni, R., and Odetti, L. 1992. Self orga-nizing behavior and swarm intelligence in a pack of mobile minia-ture robots in search of pollutants. InIEEE/RSJ IROS, pp. 1575–1582.

Georgeff, M. 1983. Communication and interaction in multi-agentplanning. InProc. AAAI, pp. 125–129.

Georgeff, M. 1984. A theory of action for multi-agent planning. InProc. AAAI, pp. 121–125.

Goldberg, D. 1989.Genetic Algorithms in Search, Optimization, andMachine Learning, Addison Wesley.

Goss, S. and Deneubourg, J. 1992. Harvesting by a group of robots.In Proc. European Conference on Artificial Life.

Grossman, D. 1988. Traffic control of multiple robot vehicles.IEEEJournal of Robotics and Automation, 4:491–497.

Hackwood, S. and Wang, J. 1988. The engineering of cellular roboticsystems. InIEEE International Symposium on Intelligent Control,pp. 70–75.

Hackwood, S. and Beni, G. 1991. Self-organizing sensors by deter-ministic annealing. InIEEE/RSJ IROS, pp. 1177–1183.

Hackwood, S. and Beni, G. 1992. Self-organization of sensors forswarm intelligence. InIEEE ICRA, pp. 819–829.

Hara, F., Yasui, Y., and Aritake, T. 1995. A kinematic analysis oflocomotive cooperation for two mobile robots along a generealwavy road. InIEEE ICRA, pp. 1197–1204.

Hashimoto, M. and Oba, F. 1993. Dynamic control approach for mo-tion coordination of multiple wheeled mobile robots transportinga single object. InIEEE/RSJ IROS, pp. 1944–1951.

Heffernan, P.J., and Schirra, S. 1992. Approximate decision algo-rithms for point set congruence. In8th Annual CompuationalGeometry, pp. 93–101.

Hert, S. and Lumelsky, V. 1995. Moving multiple tethered robotsbetween arbitrary configurations. InIEEE/RSJ IROS, pp. 280–285.

Hewitt, C. 1993. Toward an open systems architecture. InInforma-tion Processing 89. Proceedings of the IFIP 11th World ComputerCongress, pp. 389–392.

Hewitt, C., Bishop, P., Greif, I., Smith, B., Matson, T., and Steiger, R.1973. A universal modular actor formalism for artificial intelli-gence. InProc. Intl. Joint Conf. Artificial Intelligence, pp. 235–245.

Hodgins, J. and Brogan, D. 1994. Robot herds: Group behaviors forsystems with significant dynamics. InProc. A—Life IV.

Ichikawa, S., Hara, F., and Hosokai, H. 1993. Cooperative route-searching behavior of multi-robot system using hello-call com-muniction. InIEEE/RSJ IROS, pp. 1149–1156.

Ishida, Y., Endo, I., and Matsumoto, A. 1991. Communication andcooperation in an autonomous and decentralized robot system.In IFAC Int. Symp. on Distributed Intelligent Systems, pp. 299–304.

Ishida, Y., Asama, H., Tomita, S., Ozaki, K., Matsumoto, A., andEndo, I. 1994. Functional complement by cooperation of multipleautonomous robots. InIEEE ICRA, pp. 2476–2481.

Jin, K., Liang, P., and Beni, G. 1994. Stability of synchronizeddistributed control of discrete swarm structures. InIEEE ICRA,pp. 1033–1038.

Johnson, P.J. and Bay, J.S. 1994. Distributed control of autonomousmobile robot collectives in payload transportation. Technical re-port, Virginia Polytechnic Institute and State University, BradleyDept. of Elec. Engr.

Kaelbling, L.P. 1993.Learning in Embedded Systems. MIT Press.Kato, S., Nishiyama, S., and Takeno, J. 1992. Coordinating mobile

robots by applying traffic rules. InIEEE/RSJ IROS, pp. 1535–1541.

Kawauchi, Y., Inaba, M., and Fukuda, T. 1993a. A principle of dis-tributed decision making of cellular robotic system (CEBOT). InIEEE ICRA, vol. 3, pp. 833–838.

Kawauchi, Y., Inaba, M., and Fukuda, T. 1993b. A relation betweenresource amount and system performance of the cellular roboticsystem. InIEEE/RSJ IROS, pp. 454–459.

Kitano, H. 1994. Personal communication.Kleinrock, L., 1995. Nomadic computing—An opportunity.Com-

puter Communications Review.

P1: MVG



Korf, R. 1992. A simple solution to pursuit games. InProc. 11thInternational Workshop on Distributed Artificial Intelligence.

Koza, J. 1990.Genetic Programming: On the Programming of Com-puters By the Means of Natural Selection. MIT Press.

Kube, C.R. and Zhang, H. 1992. Collective robotic intelligence. InProc. Simulation of Adaptive Behavior, pp. 460–468.

Kube, C.R. and Zhang, H. 1993. Collective robotics: From socialinsects to robots.Adaptive Behavior, 2(2):189–219.

Kube, C.R., Zhang, H., and Wang, X. 1993. Controlling collectivetasks with an ALN. InIEEE/RSJ IROS, pp. 289–293.

Kube, C.R. and Zhang, H. 1994. Stagnation recovery behaviours forcollective robotics. InIEEE/RSJ IROS, pp. 1883–1890.

Kuniyoshi, Y., Kita, N., Rougeaux, S., Sakane, S., Ishii, M., andKakikura, M. 1994a. Cooperation by observation—the frameworkand basic task patterns. InIEEE ICRA, pp. 767–774.

Kuniyoshi, Y., Riekki, J., Ishii, M., Rougeaux, S., Kita, N., Sakane,S., and Kakikura, M. 1994b. Vision-based behaviors for multi-robot cooperation. InIEEE/RSJ IROS, pp. 925–931.

Kurabayashi, D., Ota, J., Arai, T., and Yoshida, E. 1995. An algorithmof dividing a work area to multiple mobile robots. InIEEE/RSJIROS, pp. 286–291.

Kurazume, R. and Nagata, S. 1994. Cooperative positioning withmultiple robots. InIEEE ICRA, pp. 1250–1257.

Latombe, J. 1991.Robot Motion Planning. Kluwer Academic,Boston, MA.

LePape, C. 1990. A combination of centralized and distributed meth-ods for multi-agent planning and scheduling. InIEEE ICRA,pp. 488–493.

Levy, R. and Rosenschein, J.S. 1991. A game theoretic approachto distributed artificial intelligence and the pursuit problem. InEuropean Workshop on Modelling Autonomous Agents in a Multi-Agent World, pp. 129–146.

Lewis, M.A. and Bekey, G.A. 1992. The behavioral self-organizationof nanorobots using local rules. InIEEE/RSJ IROS, pp. 1333–1338.

Liang, P. and Beni, G. 1995. Robotic morphogenesis. InIEEE ICRA,pp. 2175–2180.

Lin, F.-C. and Hsu, J.Y.-J. 1995. Cooperation and deadlock-handlingfor an object-sorting task in a multi-agent robotic system. InIEEEICRA, pp. 2580–2585.

Littman, M. 1994. Markov games as a framework for multi-agent re-inforcement learning. InProceedings of the International MachineLearning Conference, pp. 157–163.

Lueth, T.C. and Laengle, T. 1994. Task description, decompositionand allocation in a distributed autonomous multi-agent robot sys-tem. InIEEE/RSJ IROS, pp. 1516–1523.

Ma, S., Hackwood, S., and Beni, G. 1994. Multi-agent supportingsystems (MASS): Control with centralized estimator of distur-bance. InIEEE/RSJ IROS, pp. 679–686.

Mataric, M.J. 1992a. Designing emergent behaviors: From localinteractions to collective intelligence. InFrom Animals to Animats2, Second International Conference on Simulation of AdaptiveBehavior(SAB-92), edited by J.-A. Meyer, H. Roitblat, and S.Wilson, MIT Press, pp. 432–441.

Mataric, M.J. 1992b. Distributed approaches to behavior control. InSPIE—Sensor Fusion V, vol. 1828, pp. 373–382.

Mataric, M.J. 1992c. Minimizing complexity in controlling a mobilerobot population. InIEEE ICRA, pp. 830–835.

Mataric, M.J. 1993. Kin recognition, similarity, and group behav-ior. In Fifteenth Annual Cognitive Science Society Conference,Lawrence Erlbaum Associates, pp. 705–710.

Mataric, M. 1994a. Interaction and intelligent behavior. Ph.D. thesis,MIT, EECS.

Mataric, M. 1994b. Reward functions for accelerated learning. InProceedings of the International Machine Learning Conference,pp. 181–189.

Mataric, M.J., Nilsson, M., and Simsarian, K.T. 1995. Cooperativemulti-robot box-pushing. InIEEE/RSJ IROS, pp. 556–561.

Matsumoto, A., Asama, H., Ishida, Y., Ozaki, K., and Endo, I. 1990.Communication in the autonomous and decentralized robot sys-tem: ACTRESS. InIEEE/RSJ IROS, pp. 835–840.

McFarland, D. 1994. Towards robot cooperation. InProc. Simulationof Adaptive Behavior.

Mehregany, M., Gabriel, K.J., and Trimmer, W.S. 1988. Integratedfabrication of polysilicon mechanisms.IEEE Trans. Electron De-vices, 35(6):719–723.

Merriam-Webster 1963.Webster’s 7th Collegiate Dictionary.Merriam-Webster, Inc.

Milios, G., Dudek, M., Jenkin, E., and Wilkes, D. 1995. Experimentsin sensing and communication for robot convoy navigation. InIEEE/RSJ IROS, pp. 268–273.

Miller, G.F. and Cliff, D. 1994. Protean behavior in dynamic games:Arguments for the co-evolution of pursuit-evasion tactics. InProc.Simulation of Adaptive Behavior, D. Cliff, P. Husbands, J.-A.Meyer, and S.W. Wilson (Ed.).

Mitsumoto, N., Fukuda, T., Shimojina, K., and Ogawa, A. 1995.Micro autonomous robotic system and biologically inspired im-mune swarm strategy as a multi agent robotic system. InIEEEICRA, pp. 2187–2192.

Nicolis, G. and Prigogine, I. 1977.Self-Organization in Nonequilib-rium Systems. Wiley-Interscience.

Noreils, F. 1990. Integrating multirobot coordination in a mobilerobot control system. InIEEE/RSJ IROS, pp. 43–49.

Noreils, F.R. 1992. Multi-robot coordination for battlefield strategies.In IEEE/RSJ IROS, pp. 1777–1784.

Noreils, F.R. 1993. Toward a robot architecture integrating coopera-tion between mobile robots: Application to indoor environment.The International Journal of Robotics Research, 12(1).

Noreils, F. and Recherche, A. 1991. Adding a man/machine interfaceto an architecture for mobile robots. InIEEE/RSJ IROS.

Noreils, F.R. and Nozay, R.de 1992. Coordinated protocols: Anapproach to formalize coordination between mobile robots. InIEEE/RSJ IROS, pp. 717–724.

Numaoka, C. 1993. Collective alteration of strategic types withdelayed global information. InIEEE/RSJ IROS, pp. 1077–1084.

Ota, J., Miyata, N., Arai, T., Yoshida, E., Kurabayashi, D., andSasaki, J. 1995. Transferring and regrasping a large object by co-operation of multiple mobile robots. InIEEE/RSJ IROS, pp. 543–548.

Ozaki, K., Asama, H., Ishida, Y., Matsumoto, A., Yokota, K.,Kaetsu, H., and Endo, I. 1993. Synchronized motion by multiplemobile robots using communication. InIEEE/RSJ IROS, pp. 1164–1169.

Parker, L.E. 1992. Adaptive action selection for cooperative agentteams. InSecond Annual International Conference on Simulationof Adaptive Behavior, MIT Press, pp. 442–450.

Parker, L.E. 1993. Designing control laws for cooperative agentteams. InIEEE ICRA, vol. 3, pp. 582–587.

Parker, L.E. 1994a. ALLIANCE: an architecture for fault tolerant,cooperative control of heterogeneous mobile robots. InIEEE/RSJIROS, pp. 776–783.

P1: MVG



Parker, L.E. 1994b. Heterogeneous multi-robot cooperation. Ph.D.Dissertation, MIT EECS Dept.

Parker, L.E. 1995. The effect of action recognition and robot aware-ness in cooperative robotic teams. InIEEE/RSJ IROS, pp. 212–219.

Premvuti, S. and Yuta, S. 1990. Consideration on the cooperation ofmultiple autonomous mobile robots. InIEEE/RSJ IROS, pp. 59–63.

Reynolds, C.W. 1987. Flocks, herds and schools: A distributed be-havioural model.Computer Graphics, 21(4):71–87.

Reynolds, C. 1992. An evolved, vision-based behavioral model of co-ordinated group motion. InProc. Simulation of Adaptive Behavior.

Reynolds, C. 1994. Competition, coevolution and the game of tag.In Proc. A-Life IV.

Rosenschein, J.S. 1982. Synchronization of multi-agent plans. InProc. AAAI, pp. 115–119.

Rosenschein, J.S. and Genesereth, M.R. 1985. Deals among rationalagents. InProc. Intl. Joint Conf. Artificial Intelligence, pp. 91–99.

Rosenschein, J.S. and Zlotkin, G. 1994.Rules of Encounter: Design-ing Conventions for Automated Negotiation among Computers.MIT Press.

Rude, M. 1994. Cooperation of mobile robots by event transformsinto local space-time. InIEEE/RSJ IROS, pp. 1501–1507.

Rus, D., Donald, B., and Jennings, J. 1995. Moving furniturewith teams of autonomous robots. InIEEE/RSJ IROS, pp. 235–242.

Sasaki, J., Ota, J., Yoshida, E., Kurabayashi, D., and Arai, T. 1995.Cooperating grasping of a large object by multiple mobile robots.In IEEE ICRA, pp. 1205–1210.

Sen, S., Sekaran, M., and Hale, J. 1994. Learning to coordinatewithout sharing information. InProc. AAAI, pp. 426–431.

Shoham, Y. and Tennenholtz, M. 1992. On the synthesis of usefulsocial laws for artificial agent societies (preliminary report). InProc. AAAI, pp. 276–281.

Singh, K. and Fujimura, K. 1993. Map making by cooperating mobilerobots. InIEEE ICRA, vol. 2, pp. 254–259.

Smith, R. 1980. The contract net protocol: High-level communi-cation and control in a distributed problem solver.IEEE Trans.Computers, pp. 1104–1113.

Steels, L. 1990. Cooperation between distributed agents through self-organization. InEuropean Workshop on Modelling AutonomousAgents in a Multi-Agent World, pp. 175–195.

Steels, L. 1994. A case study in the behavior-oriented design ofautonomous agents. InProc. Simulation of Adaptive Behavior.

Stilwell, D.J. and Bay, J.S. 1993. Toward the development of a ma-terial transport system using swarms of ant-like robots. InIEEEICRA, vol. 6, pp. 766–771.

Sugie, H., Inagaki, Y., Ono, S., Aisu, H., and Unemi, T. 1995. Placingobjects with multiple mobile robots—Mutual help using intentioninference. InIEEE ICRA, pp. 2181–2186.

Sugihara, K. and Suzuki, I. 1990. Distributed motion coordination ofmultiple mobile robots. InProc. IEEE International Symposiumon Intelligent Control.

Suzuki, T., Asama, H., Uegaki, A., Kotosaka, S., Fujita, T., Mat-sumoto, A., Kaetsu, H., and Endo, I. 1995. An infra-red sensorysystem with local communication for cooperative multiple mobilerobots. InIEEE/RSJ IROS, vol. 1, pp. 220–225.

Tan, M. 1993. Multi-agent reinforcement learning: independent vs.cooperative agents. InProceedings of the International MachineLearning Conference.

Tsetlin, M.L. 1964. Finite automata and modeling the simplestforms of behavior. Ph.D. Dissertation, V.A. Steklov Mathemat-ical Institute.

Tung, B. and Kleinrock, L. 1993. Distributed control methods. InProceedings of the 2nd International Symposium on High Perfor-mance Distributed Computing, pp. 206–215.

Tung, Y.-C. 1994. Distributed control using finite state automata.Ph.D. thesis, UCLA Computer Science Department.

Ueyama, T. and Fukuda, T. 1993a. Knowledge acquisition and dis-tributed decision making. InIEEE ICRA, vol. 3, pp. 167–172.

Ueyama, T. and Fukuda, T. 1993b. Self-organization of cellularrobots using random walk with simple rules. InIEEE ICRA, vol. 3,pp. 595–600.

Ueyama, T., Fukuda, T., Arai, F., Kawauchi, Y., Katou, Y.,Matsumura, S., and Uesugi, T. 1993a. Communication architec-ture for cellular robotic system.JSME International Journal, Se-ries C, 36:353–360.

Ueyama, T., Fukuda, T., Arai, F., Sugiura, T., Sakai, A., and Uesugi,T. 1993b. Distributed sensing, control and planning—cellularrobotics approach. InIMACS, Elsevier Science Publ. (North-Holland), pp. 433–438.

VanLehn, K. (Ed.) 1991.Architectures for Intelligence: The 22ndCarnegie Mellon Symposium on Cognition. Lawrence ErlbaumAssociates.

Wang, J. 1991. Fully distributed traffic control strategies for many-AGV systems. InIEEE/RSJ IROS, pp. 1199–1204.

Wang, J. 1993. DRS operating primitives based on distributed mutualexclusion. InIEEE/RSJ IROS, pp. 1085–1090.

Wang, J. 1994. On sign-board based inter-robot communication indistributed robotic systems. InIEEE ICRA, pp. 1045–1050.

Wang, J. 1995. Operating primitives supporting traffic regulationand control of mobile robots under distributed robotic systems. InIEEE ICRA, pp. 1613–1618.

Wang, J. and Beni, G. 1988. Pattern generation in cellular roboticsystems. InIEEE International Symposium on Intelligent Control,pp. 63–69.

Wang, J. and Beni, G. 1990. Distributed computing problems incellular robotic systems. InIEEE/RSJ IROS, pp. 819–826.

Wang, J. and Premvuti, S. 1994. Resource sharing in distributedrobotic systems based on a wireless medium access protocol(CSMA/CD-W). In IEEE/RSJ IROS, pp. 784–791.

Wang, J. and Premvuti, S. 1995. Distributed traffic regulation andcontrol for multiple autonomous mobile robots operating in dis-crete space. InIEEE ICRA, pp. 1619–1624.

Wang, J., Premvuti, S., and Tabbara, A. 1995. A wireless mediumaccess protocol (CSMA/CD-W) for mobile robot based distributedrobotic system. InIEEE ICRA, pp. 2561–2566.

Wang, Z.-D., Nakano, E., and Matsukawa, T. 1994. Cooperating mul-tiple behavior-based robots for object manipulation. InIEEE/RSJIROS, pp. 1524–1531.

Weiser, M. 1993. Some computer science issues in ubiquitous com-puting.Communications of the ACM, 36(7):74–84.

Werner, G. and Dyer, M. 1992. Evolution of herding behavior inartificial animals. InProc. Simulation of Adaptive Behavior.

Whitehead, S. 1991. A complexity analysis of cooperative mecha-nisms in reinforcement learning. InProc. AAAI, pp. 607–613.

Wilson, E.O. 1971.The Insect Societies. Harvard University Press.Yamaguchi, H. and Arai, T. 1994. Distributed and autonomous con-

trol method for generating shape of multiple mobile robot group.In IEEE/RSJ IROS, pp. 800–807.

P1: MVG



Yanco, H. and Stein, L. 1992. An adaptive communication protocolfor cooperating mobile robots. InProc. Simulation of AdaptiveBehavior, pp. 478–485.

Yates, F.E. (Ed.) 1987.Self-Organizing Systems: The Emergence ofOrder. Plenum Press.

Yeung, D. and Bekey, G. 1987. A decentralized approach to themotion planning problem for multiple mobile robots. InIEEEICRA, pp. 1779–1784.

Yokota, K., Suzuki, T., Asama, H., Masumoto, A., and Endo, I. 1994.A human interface system for the multi-agent robotic system. InIEEE ICRA, pp. 1039–1044.

Yoshida, E., Arai, T., Ota, J., and Miki, T. 1994. Effect of group-ing in local communication system of multiple mobile robots. InIEEE/RSJ IROS, pp. 808–815.

Yoshida, E., Yamamota, M., Arai, T., Ota, J., and Kurabayashi, D.1995a. A design method of local communication area in multiplemobile robot system. InIEEE ICRA, pp. 2567–2572.

Yoshida, E., Yamamoto, M., Arai, T., Ota, J., and Kurabayashi, D.1995b. A design method of local communication range in multiplemobile robot system. InIEEE/RSJ IROS, pp. 274–279.

Yuta, S. and Premvuti, S. 1992. Coordinating autonomous and cen-tralized decision making to achieve cooperative behaviors betweenmultiple mobile robots. InProc. of the 1992 IEEE/RSJ Interna-tional Conference on Intelligent Robots and Systems, Raleigh, NC,pp. 1566–1574.

Y. Uny Cao is a Ph.D. student in the Computer Science Departmentat the University of California at Los Angeles. He received a B.E. in

Electrical Engineering from Zhejiang University, Hangzhou, China,and a M.S. in Computer Science in 1993 from the University ofLouisville, Kentucky. At UCLA, he has been working on severalresearch projects related to the Internet and robotics.

Alex S. Fukunagais a Member of the Technical Staff in the Arti-ficial Intelligence Group, Information and Computing TechnologiesResearch Section at the Jet Propulsion Laboratory, California Insti-tute of Technology. He holds an A.B. in Computer Science fromHarvard University, and a M.S. in Computer Science from the Uni-versity of California at Los Angeles, where he is currently a Ph.D.student. His research interests include optimization, search, machinelearning, and automated planning/scheduling.

Andrew B. Kahng is an Associate Professor in the Computer Sci-ence Department at the University of California at Los Angeles.He received the A.B. degree in Applied Mathematics/Physics fromHarvard College in 1983. His M.S.(1986) and Ph.D.(1989) degreesin Computer Science are from the University of California at SanDiego. Dr. Kahng has received the NSF Research Initiation Awardand an NSF Young Investigator Award. Currently, he is co-Directorof the UCLA VLSI CAD Laboratory. He is also Director of theUCLA Commotion Laboratory, studying cooperation and distributedtask-solving using multiple mobile robots. His research areas includediscrete algorithms for VLSI layout synthesis, computational geome-try, and search/recognition tasks, as well as the theory of large-scaleglobal optimization.

cooperative mobile robotics: antecedents and …autonomous robots kl402-02-cao december 19, 1996...

Documents