1335 Multimedia Contents Multiple53.MultipleMobileRobotSystems Mo 53 | E Part b

Lynne E. Parker, Daniela Rus, Gaurav S. Sukhatme

53.4 Networked Mobile Robots ...... 1340 Within the context of multiple mobile, and net- 53.4.1 Overview...... 1341 worked robot systems, this chapter explores the 53.4.2 State of the Art and Potential .. 1342 current state of the art. After a brief introduction, 53.4.3 Research Challenges...... 1344 we first examine architectures for multirobot co- 53.4.4 Control...... 1346 operation, exploring the alternative approaches 53.4.5 Communication for Control ...... 1347 that have been developed. Next, we explore com- 53.4.6 Communication for Perception. 1347 munications issues and their impact on multirobot 53.4.7 Control for Perception ...... 1349 teams in Sect. 53.3, followed by a discussion of 53.4.8 Control for Communication ...... 1350 networked mobile robots in Sect. 53.4. Following 53.5 Swarm Robots ...... 1351 this we discuss swarm robot systems in Sect. 53.5 and modular robot systems in Sect. 53.6.While 53.6 Modular Robotics ...... 1354 swarm and modular systems typically assume large 53.6.1 Chain Systems ...... 1354 numbers of homogeneous robots, other types of 53.6.2 Lattice Systems ...... 1354 multirobot systems include heterogeneous robots. 53.6.3 Truss Systems ...... 1356 We therefore next discuss heterogeneity in coop- 53.6.4 Free-Form Systems ...... 1356 erative robot teams in Sect. 53.7. Once robot teams 53.6.5 Self-Assembling Systems ...... 1356 allow for individual heterogeneity, issues of task 53.7 Heterogeneity ...... 1357 allocation become important; Sect. 53.8 therefore 53.8 Task Allocation ...... 1359 discusses common approaches to task allocation. 53.8.1 Taxonomy for Task Allocation ... 1359 Section 53.9 discusses the challenges of multirobot 53.8.2 Representative Approaches...... 1360 learning, and some representative approaches. We outline some of the typical application domains 53.9 Learning ...... 1361 which serve as test beds for multirobot systems 53.10 Applications...... 1362 research in Sect. 53.10. Finally, we conclude in 53.10.1 Foraging and Coverage ...... 1362 Sect. 53.11 with some summary remarks and sug- 53.10.2 Flocking and Formations ...... 1362 gestions for further reading. 53.10.3 Object Transportation and Cooperative Manipulation . 1363 53.10.4 Multitarget Observation...... 1364 53.1 History...... 1336 53.10.5 Traffic Control and Multirobot 53.2 Architectures for Multirobot Systems ... 1337 Path Planning ...... 1365 53.2.1 The Nerd Herd ...... 1337 53.10.6 Soccer ...... 1365 53.2.2 The ALLIANCE Architecture 1338 ...... 53.11 Conclusions and Further Reading ...... 1366 53.2.3 The Distributed Robot Architecture ...... 1339 Video-References...... 1366 53.3 Communication...... 1339 References...... 1367 Ar- . swarm mobile , applica- . Finally modular robotics is or learning is a particular , while hardware ,and 53.5 Networked in multirobot systems describes these systems task allocation communication , 22. 39 the coordination of robots reconfigurable ltiple mobile robot systems; r capabilities. When robots Swarm robots s that interact implicitly with task allocation , architectures communication robots, in which team members may robot systems are those robots that move and in multirobot teams is of particular inter- heterogeneity , are the final major topic of discussion in this mobile Most of the work specific to multiple mobile robot The types of robots considered in the study of mul- deal with heterogeneity in thewhich robot team team members members, vary in capabilities. in In their these sensor teams, andcan be effector very different from collectivesince swarm approaches, robots are no longer interchangeable. cooperation can be categorizedics into of a study. set Thesechapter, of topics, include key which top- are the foci of this robots robots, as distinguished from otherinteraction. types For example, of a multirobot special casebile of robot multiple mo- systems arerobots that the interconnect with each other forof the purposes navigation or manipulation.these Algorithmic systems aspects are of covered in Sect. around in theaerial environment, vehicles, or such underwater vehicles. as Thiscuses chapter ground fo- specifically vehicles, on the interaction of multiple learning est in designingand can teams learn new thatin behaviors. Illustrating are each the adaptive advances of of over representative these application time domains; areas these often takes place in a set chitectures are relevant for all types ofapproaches multirobot systems, specify as how these theorganized robot and team interact. memberstype are of multirobot system,of typified homogeneous by robot largeeach numbers other. Suchheterogeneous systems arevary often significantly contrasted invary with thei in capabilities,which challenges robots arise shouldlenge in perform commonly referred which determining to as tasks – a chal- tions chapter. aspects are covered in Chap. very closely related tothe mu focus in networked robotics issensors, on embedded systems computers, of and human robots, users thatall are connected byvariant networked of communication. multirobot cooperation Another is multipletor arm manipula- cooperation. Chapter tiple in detail. ]. 3 sys- sys- Collective solutions allow systems. collective swarm tentionally cooperative mobile robot systems fall agree that multirobot sys- rative multirobot systems can intentionally cooperative solutions require robots to act in Weakly cooperative ]. The most common motivations for de- 2 , intentionally cooperative 1 systems are those in which robots execute their Researchers generally accomplish. easier than having a single powerful robot. allelism. bustness through redundancy. The issues that must be addressed in developing tems have knowledge ofin the the environment presence and of actactions, other together based or robots on capabilities the ofaccomplish state, their the teammates same insystems vary goal. order in the to In extent tocount the which actions robots or take state of into other ac- either robots, and strongly can or lead to weaklyStrongly cooperative solutions cooperative [53. concert to achievenot the trivially serializable. goal, Typically, these approaches executing re- quire some tasks type of that communication and synchronization among are the robots. robots to havesubsequent periods to of coordinating their operationalroles. Intentionally selection independence, coope of tasks or swarm own tasks with only minimal needother for knowledge robot about team members. Theseby systems the are assumption typified ofmobile a robots, large in number which of robotstrol homogeneous make laws to use generate of globally localwith coherent con- team little behaviors, explicitthe communication other among hand, robots robots. in On tems and veloping multirobot system solutions are that: 1. The task complexity is too high for a2. single robot to The task3. is inherently distributed. Building several resource-bounded robots is much 4. Multiple robots can solve problems faster using par- 5. The introduction of multiple robots increases ro- multirobot solutions arequirements dependent and upon the the sensorythe and task available effector robots. re- capabilities of into one of two broad categories: 53.1 History Since the earliest worktems in on the multiple 1980s, the mobile fieldcovers robot has a grown sys- significantly, and large bodylevel, approaches of to research. multiple At the most general tems have severaltems advantages [53. over single-robot sys- Moving in the Environment

Part E

Part E | 53.1 1336 Multiple Mobile Robot Systems 53.2 Architectures for Multirobot Systems 1337

53.2 Architectures for Multirobot Systems 53.2 | E Part

The design of the overall control architecture for the A plethora of multirobot control architectures have multirobot team has a significant impact on the robust- been developed over the years. We focus here on three ness and scalability of the system. Robot architectures early approaches that illustrate the spectrum of control for multirobot teams are composed of the same fun- architectures. The first, the Nerd Herd, is representative damental components as in single-robot systems, as of a pure swarm robotics approach using large num- described in Chap. 12. However, they also must address bers of homogeneous robots. The second, ALLIANCE, the interaction of robots and how the group behavior is representative of a behavior-based approach that will be generated from the control architectures of the enables coordination and control of possibly heteroge- individual robots in the team. Several different philoso- neous robots without explicit coordination. The third, phies for multirobot team architectures are possible; the distributed robot architecture (DIRA), is a hybrid ap- most common are centralized, hierarchical, decentral- proach that enables both robot autonomy and explicit ized,andhybrid. coordination in possibly heterogeneous robot teams. Centralized architectures that coordinate the entire team from a single point of control are theoretically 53.2.1 The Nerd Herd possible [53.4], although often practically unrealistic due to their vulnerability to a single point of failure, One of the first studies of social behaviors in multi- and due to the difficulty of communicating the entire robot teams was conducted by Matari´c [53.6], with system state back to the central location at a frequency results being demonstrated on the Nerd Herd team of suitable for real-time control. Situations in which these 20 identical robots (shown in Fig. 53.1). This work is approaches are relevant are cases in which the central- an example of swarm robotic systems, as described fur- ized controller has a clear vantage point from which to ther in Sect. 53.4. The decentralized control approach observe the robots, and can easily broadcast group mes- was based on the subsumption architecture (Chap. 12), sages for all robots to obey [53.5]. and assumed that all robots were homogeneous, but Hierarchical architectures are realistic for some ap- with relatively simple individual capabilities, such as plications. In this control approach, each robot oversees detecting obstacles and kin (i. e., other robot team the actions of a relatively small group of other robots, members). A set of basic social behaviors were de- each of which in turn oversees yet another group of fined and demonstrated, including obstacle avoidance, robots, and so forth, down to the lowest robot, which homing, aggregation, dispersion, following, and safe simply executes its part of the task. This architecture wandering. These basic behaviors were combined in scales much better than centralized approaches, and is various ways to yield more composite social behaviors, reminiscent of military command and control. A point including flocking (composed of safe wandering, aggre- of weakness for the hierarchical control architecture is gation, and dispersion), surrounding (composed of safe recovering from failures of robots high in the control wandering, following, and aggregation), herding (com- tree. posed of safe wandering, surrounding, and flocking), Decentralized control architectures are the most and foraging (composed of safe wandering, dispersion, common approach for multirobot teams, and typically following, homing, and flocking). The behaviors were require robots to take actions based only on knowledge local to their situation. This control approach can be highly robust to failure, since no robot is responsible for the control of any other robot. However, achiev- ing global coherency in these systems can be difficult, because high-level goals have to be incorporated into the local control of each robot. If the goals change, it may be difficult to revise the behavior of individual robots. Hybrid control architectures combine local control with higher-level control approaches to achieve both ro- bustness and the ability to influence the entire team’s actions through global goals, plans, or control. Many multirobot control approaches make use of hybrid ar- chitectures. Fig. 53.1 The Nerd Herd robots ] allows a robot 10 has recently begun work on Robots using the ALLIANCE architecture for ior set is needed activated this task based on how long it has been attempting the task. Effectively, the motivation continues to increase at is no longer needed. first time. In any of these four situations, the motivation re- The L-ALLIANCE extension [53. Fig. 53.3 time step, theon: motivation level is recalculated based 1. The previous motivation2. level The rate3. of impatience Whether the sensory feedback indicates the4. behav- Whether the robot has another behavior set5. already Whether another robot 6. Whether the robot is willing to give up the task, some positive ratecurs: unless one of four situations1. oc- The sensory feedback indicates that the behavior set 2. Another behavior set3. in the robot Some activates. other robot has just taken over the task4. for the The robot has decided to acquiesce the task. turns to zero. Otherwise,crosses the a motivation threshold grows value, untilset is at it activated and which the robot time canan be the said action. to behavior have When selected anwithin action that is robot selected, prevents cross-inhibition tivated within other that tasks same from robot.active in being When a robot, a ac- the behavior robot broadcasts setto its other is current activity robots at a periodic rate. to adapt the rate of change of the impatience and ac- a mock clean-up task Actuators ), developed set 2 behavior Behavior 53.2 Motivational ]. 8 set 1 behavior Behavior Alliance Motivational Cross-inhibition ] for fault-tolerant task allocation in 9 [53. ], and learning [53. 7 set 0 Layer 2 Layer 1 Layer 0 Layer behavior Behavior turn towardand aggregation go. centroid stop. Motivational In this approach, the initial motivation to perform Parker If agentdistance is outside aggregation Else The ALLIANCE architecture heterogeneous robot teams.the This subsumption approach architecture builds byand on adding motivations behavior for sets out achieving explicit negotiations action between robots. selectiongroup Behavior low-level sets behaviors with- together fora the execution particular of task. Theimpatience motivations and consist acquiescence of thata levels can robot’s of raise interest andsponding in to lower a activating task that a must be behavior accomplished. seta corre- given behavior set is set to zero. Then, at each by 53.2.2 The ALLIANCE Architecture Another early workis in the multirobot ALLIANCE team architecture (Fig. architectures This work showedgenerated that through collective the behaviors combinationsic could of be behaviors. lower-level ba- Relatedissues such work as using on bucketence brigades [53. to this reduce interfer- project studied implemented as rules, such as thegregate: following rule for ag- Aggregate: Moving in the Environment cation

Sensors communi-

Interrobot Part E Fig. 53.2

Part E | 53.2 1338 Multiple Mobile Robot Systems 53.3 Communication 1339 atE|53.3 | E Part

Planner Planner Planner

Executive Executive Executive

Behaviors Behaviors Behaviors

Robot 1 Robot 2 Robot 3

Fig. 53.4 The distributed robot architecture quiescence values depending on the quality with which Fig. 53.5 Robots using the distributed robot architecture that robot is expected to accomplish a given task. The for assembly tasks result is that robots that have demonstrated their abil- ity to better accomplish certain tasks are more likely consists of a planning layer that decides how to achieve to choose those tasks in the future. Additionally, if high-level goals; an executive layer that synchronizes problems occur during team performance, then robots agents, sequences tasks, and monitors task execution; may dynamically reallocate their tasks to compensate and a behavioral layer that interfaces to the robot’s sen- for the problems. This approach was demonstrated sors and effectors. Each of these layers interacts with on a team of three heterogeneous robots performing those above and below it. Additionally, robots can in- a mock clean-up task, two robots performing a box- teract with each other via direct connections at each of pushing task, and four robots performing a cooperative the layers. target observation problem. The approach has also been This architecture has been demonstrated in a team demonstrated in the simulation of a janitorial service of three robots – a crane, a roving eye, and a mo- task and a bounding overwatch task. Figure 53.3 shows bile manipulator – performing a construction assembly robots using ALLIANCE to perform the mock clean-up task (Fig. 53.5). This task requires the robots to work task. together to connect a beam at a given location. In these demonstrations, a foreman agent decides which 53.2.3 The Distributed Robot robot should move the beam at which times. Initially, Architecture the crane moves the beam to the vicinity of the em- placement based on encoder feedback. The foreman Simmons et al. [53.11] have developed a hybrid archi- then sets up a behavioral loop between the roving eye tecture called the distributed robot architecture (DIRA). and the crane robot to servo the beam closer to the Similar to the Nerd Herd and ALLIANCE approaches, point of emplacement. Once the beam is close enough, the DIRA approach allows autonomy in individual the foreman tasks the roving eye and the mobile ma- robots. However, unlike the previous approaches, DIRA nipulator to servo the arm to grasp the beam. After also facilitates explicit coordination among robots. This contact is made, the foreman tasks the roving eye approach is based on layered architectures that are and the mobile manipulator to coordinate to servo the popular for single-robot systems (Chap. 12). In this beam to the emplacement point, thus completing the approach (Fig. 53.4), each robot’s control architecture task.

53.3 Communication

A fundamental assumption in multirobot systems re- a number of ways; the three most common techniques search is that globally coherent and efficient solutions are: can be achieved through the interaction of robots lack- ing complete global information. However, achieving 1. The use of implicit communication through the these globally coherent solutions typically requires world (called stigmergy), in which robots sense the robots to obtain information about their teammates’ effects of teammate’s actions through their effects states or actions. This information can be obtained in on the world [53.6, 12–16] and ]. How- 25 et al.’s tax- [53. ] includes axes MacLennan . 26 Dudek Arkin pieces of information and optimal Balch ], which investigates the evolution of 24 networked robot systems [53. Several related issues of active research in commu- onomy of multirobot systems [53. related to communication,range, including communication communication topology,bandwidth. and These communication characteristics canpare and be contrast used multirobot systems. to com- nications for multirobot teamswork deal connectivity with and dynamic topologies; net- teams for must example, either be robot connectivity able to as maintain they communications move,gies or that allow employ the robot recovery teammunications strate- to recover when connectivity the com- ismay require broken. robots to These adaptthe their concerns anticipated actions effects in on response the to communicationsor network, in responseagation to behavior knowledge of of information thenetwork. These through anticipated and the prop- related issues dynamic the are context discussed of next in ever, more information doesto not improve necessarily performance, continue the as communications bandwidth it withoutapplication can providing benefit. quickly an Thetems overload challenge is in to multirobot discoverto exchange sys- the that yield these performance improvements without saturating the communications bandwidth. Cur- rently, no general approaches toinformation identifying are this available; critical thus, thecommunicate decision is of what an to application-specificanswered question to by be the system designer. communication in simulated worlds and concludesthe that communication of local robot information canin result significant performance improvements. Interestingly, for many representative applications,found researchers a have nonlinear relationshipinformation between communicated the amount and of itsformance of impact the team. on Typically, even a small theinformation amount of per- can have aas significant found impact in on the study the of team, difficult with just perceptiontion and control. enables Communica- new controlthe and system perception (e.g., capabilities access in ception to range information of outside the the robotenables per- system). Conversely, solutions control for problems that are difficult with- communications approaches to determinethat the can method reliablyperformance. achieve Researchers the generally required agree that levelnication commu- can of have system aformance strong of positive impact the on team.of the this One per- impact of was given the inBurghardt earliest the work illustrations of Passive ], in which approach is 23 networked com- is appealing because , fallible communication problem [53. ] – an area widely studied in Stigmergy 21 ]. are multiple robots operating to- is appealing because it does not de- 18– 22 explicit communication , 9 ] hidden-state 17 to accomplish a specified task. Multiple the field of networked robot systems. Each of these mechanisms for exchanging infor- sensors to directly observe the actions ofmates their [53. team- Selecting the appropriate use of communication in robots directly and intentionally communicate rele- vant information through some activeas means, radio such [53. munication robots enable new capabilitiesnetwork and enables the new communication approaches and solutions that are 53.4 Networked Mobile Robots Networked robots gether coordinating and cooperating by mation between robotsdisadvantages [53. has its own advantages and 2. Passive action recognition, in which robots use of its simplicity andplicit its communications lack channels ofever, dependence and it upon protocols. is ex- How- perception limited of by thethe the world mission extent reflects the to robotaction the recognition which team salient must a states accomplish. pend robot’s upon of a limited-bandwidth mechanism. As with implicitis cooperation, limited however, by it thefully degree to interpret which its adifficulty sensory robot of information, can analyzing the success- as actionsbers. well of Finally, robot as team the the mem- appealing because ofwhich its robots directness can and becomegoals the aware ease of of with the its actionscommunication and/or teammates. in The multirobotnize major teams actions, uses are exchangebetween of information, to robots. and explicit synchro- Explicit todealing communication negotiate with is the alimited way sensors of cannotstates distinguish of between the world different mance. that However, explicit are communication importantterms is for of limited fault task in tolerance perfor- ically and depends reliability, upon because amunications it channel noisy, typ- that limited-bandwidth may com- all not continually members connect of themake robot use of team. explicit communications Thus, must approaches alsomechanisms provide to that handle communication failuresmessages. and lost a multirobot team is a design choicetasks dependent to upon be the achieved byto the carefully multirobot consider team. the One costs needs and benefits of alternative 3. Explicit (intentional) communication, in which Moving in the Environment

Part E

Part E | 53.4 1340 Multiple Mobile Robot Systems 53.4 Networked Mobile Robots 1341

out mobility (e.g., localization). Section 53.4.1 defines tion, perception, and control to enable such new 53.4 | E Part the field, examines the benefits of networking in robot capabilities. coordination, and discusses applications. Section 53.4.2 highlights a few projects focused on networked robotics This definition of autonomous networked robots and discusses the application potential of the field. Sec- also includes a third class of distributed systems, mobile tion 53.4.3 discusses the research challenges at the sensor networks, which is a natural evolution of sen- intersection of control, communication, and perception. sor networks. Robot networks allow robots to measure Section 53.4.4 defines a model for the control of a net- spatially and temporally distributed phenomena more worked system which is used in Sects. 53.4.5–53.4.8 efficiently. The robots in turn can deploy, repair, and to examine specific research issues and opportunities maintain the sensor network to increase its longevity, facilitated by the interplay between communication, and utility. The focus of this chapter is autonomous net- control, and perception. worked robots. Embedded computers and sensors are becoming 53.4.1 Overview ubiquitous in homes and factories, and increasingly wireless ad hoc networks or plug-and-play wired net- The term networked robots refers to multiple robots works are becoming commonplace. Human users inter- operating together coordinating and cooperating by act with embedded computers and sensors to perform networked communication to accomplish a specified tasks ranging from monitoring (e.g., supervising the op- task. Communication between entities is fundamental eration of a factor and surveillance in a building) to to cooperation (and coordination), hence there is a cen- control (e.g., running an assembly line consisting of tral role for the communication network in networked sensors, actuators, and material-handling equipment). robots. Networked robots may also involve coordina- In most of these cases, the human users, embedded tion and cooperation with stationary sensors, embedded computers, and sensors are not collocated and the co- computers, and human users. The central feature of net- ordination and communication happens through a net- worked robots is the ability of the system to perform work. Networked robots extends this vision to multiple tasks that are well beyond the abilities of a single robot robots functioning in a wide range of environments per- or multiple uncoordinated robots. forming tasks that require them to coordinate with other The IEEE (Institute of Electrical and Electronics robots, cooperate with humans, and act on information Engineers) Technical Committee on Networked Robots derived from multiple sensors. has adopted the following definition of a networked Figure 53.6 shows prototype concepts derived from robot: academic laboratories and industry. In all these exam- A networked robot is a robotic device connected to a communications network such as the Internet or a) b) local-area network (LAN). The network could be wired or wireless, and based on any of of a variety of proto- cols such as the transmission control protocol (TCP), the user datagram protocol (UDP), or 802.11. Many new applications are now being developed ranging from automation to exploration. There are two subclasses of networked robots: c) d)

1. Teleoperated, where human supervisors send com- mands and receive feedback via the network. Such systems support research, education, and public awareness by making valuable resources accessible to broad audiences. 2. Autonomous, where robots and sensors exchange data via the network. In such systems, the sensor Fig.53.6a–d Small modules (after [53.27]) can automat- network extends the effective sensing range of the ically connect and communicate information to perform robots, allowing them to communicate with each locomotion tasks (a); robot arms (after [53.28]) on mo- other over long distances to coordinate their ac- bile bases can cooperate to perform household chores (b); tivity. Sensing, actuation, and computation need swarms of robots (after [53.29]) can be used to explore an no longer be collocated. A broad challenge is to unknown environment (c); and industrial robots can coop- develop a science base that couples communica- erate in welding operations (d) ]. and 32 plug ]. Robot networks are analogous 31 can be swapped in and out automatically to provide A system of robots, embedded computers, actua- Finally, networked robots have the potential to pro- Applications for networked robots abound. The US 53.4.2 State of the Art and Potential The growth in networked robot systemsacross is broad-based, many industries. Therebetween this industry is and the a industry connectedsor to strong sen- networks. connection Sensor networksgrow have been dramatically projected in to market terms of value [53. commercializationto and sensor networkshave except mobility that and they allowof allow the the sensors geographical sensors to to distribution acquired. be adapted based on the information tors, and sensors hasdefense, tremendous and potential manufacturing in applications.vides civilian, Nature the proof pro- of conceptGroup of behaviors what in is nature possiblethat can [53. are be only found microns inin to length. organisms There those are that numerousmals are examples that of execute several simple simple meters behaviors ani- with modestand sensors actuators but communicateneighbors with to and enable sense nearest complexare emergent fundamental behaviors to that navigation, foraging,structing nests, hunting, survival, con- and eventually growth. As seen to robot failures.multiple This gateways, is routers, and seena computers fault-tolerant in provide system (although the for thebust Internet Internet in is where other not ways). ro- Similarly, robotsplay that can for a robust operating environment. vide great synergywith by complementary bringing together benefitsgreater components than and the sum making of the the parts. whole military routinely deploys unmannedreprogrammed vehicles remotely based that on are intelligenceby gathered other unmanned vehicles, sometimesThe automatically. deployment of satellitesnauts in in space, a often shuttlethe by with coordination astro- the of shuttle complexthe robot space instrumentation shuttle, arm, onboard human requires operatorsthe on shuttle a arm, ground and station, aappliances human now user contain on the sensors shuttle.worked. and As Home are domestic and becoming personal net- robotscommonplace, become more it is natural towith see sensors these robots and working operating appliances with in one or therobots will more house likely human while be users.environmental used co- Networked observatories as of critical the ingredientsecological future. in monitoring the Large-scale precludes theinfrastructure, use and of monolithic istributed, envisioned networked robotic to system. be built as a dis- a snake-like ]) can be re- ) ) and manip- b ( 30 53.7 a four-legged walk- ) c ( oiting the efficiency that b)b) c)c) a wheel-like rolling system, Robotic modules (after [53. ) a ( Besides being able to perform tasks that individual Another advantage of using the network to connect The ability to network robots also enables fault tol- a) ples, independent robot orerate robotic to modules perform can tasks coop- cannot that perform. a single Robotsperform robot can (or locomotion automatically module) tasks couple (alsoulation Fig. tasks to that either aor single that robot would cannot require perform, aperform. special-purpose They larger can robot to alsoand reconnaissance coordinate tasks to expl performis search inherent in parallelism.pendent They tasks can that also need perform tothe be inde- manufacturing coordinated. industry Examples include, for in example, fixtur- ing and welding. robots cannot perform, networked robotsimproved also efficiency. result Networking in givescess each to robot information ac- outside itssuch perception range. as Tasks searchingperformed or faster mapping withrobots. A can, an speed up increase in inachieved in principle, manufacturing by operations can the deploying be be multiple numbererations robots in performing of parallel op- but in a coordinated fashion. robots is the abilityremoved to assets. connect and Mobile harnesstion robots sensed physically by can other mobile react robotsIndustrial at to a robots remote informa- can location. adaptparts their being end-effectors to manufacturedline. new upstream Human users in canlocated the use via the machines assembly network. that are remotely erance in design. If robots canthemselves dynamically using reconfigure the network, they are more tolerant Fig.53.7a–c ing system undulatory locomotion system, and configured to morphincluding into different locomotion systems Moving in the Environment

Part E

Part E | 53.4 1342 Multiple Mobile Robot Systems 53.4 Networked Mobile Robots 1343

with humans in assembly and material-handling tasks. 53.4 | E Part Workcells consist of multiple robots, numerous sensors and controllers, automated guided vehicles, and one or two human operators working in a supervisory role. However, in most of these cells, the networked robots operate in a structured environment with very little vari- ation in configuration and/or operating conditions. There is a growing emphasis on networking robots in applications of field robotics, for example, in the mining industry. Like the manufacturing industry, oper- ating conditions are often unpleasant and the tasks are repetitive. However, these applications are less struc- tured and human operators play a more important role. Fig. 53.8 Ants are able to cooperatively manipulate and In the health care industry, networks allow health transport objects often in large groups, without identified care professionals to interact with their patients, other or labeled neighbors, and without centralized coordination professionals, expensive diagnostic instruments, and surgical robots. Telemedicine is expected to provide in Fig. 53.8, relatively small agents are able to manip- a major growth impetus for remote networked robotic ulate objects that are significantly larger in terms of devices that will take the place of today’s stand-alone size and payload by cooperating with fairly simple in- medical devices. dividual behaviors. The coordination between agents is There are already many commercial products, no- completely decentralized, allowing scaling up to large tably in Japan, where robots can be programmed via numbers of robots and large objects [53.33]. Individu- and communicate with cellular phones. For example, als do not recognize each other. In other words, there the MARON robot developed by Fujitsu lets a human is no labeling or identification of robots. The number user dial up their robot and instruct it to conduct sim- of agents in the team is not explicitly encoded. Agents ple tasks including sending pictures back to the user are identical, enabling robustness to failures and modu- via a cellular phone. Indeed these robots will inter- larity. There is minimal communication, and even that act with other sensors and actuators in the home – which is present is only between neighbors. Further- door openers equipped with Bluetooth cards and actua- more, the optimal mode of group coordination may be tors and computer-controlled lighting, microwaves, and scale dependent. Studies of wasps show strong evidence dishwashers. Indeed the Network Robot Forum [53.35] of centralized coordination among species with small is already setting standards for how stationary sensors colony sizes, but a distributed, decentralized coordina- and actuators can interact with other robots in domestic tion in larger colonies [53.34]. All these attributes are and commercial settings. relevant to networked robots. Environmental monitoring is a key application for Biology has shown how simple decentralized be- networked robots. By exploiting mobility and commu- haviors in unidentified individuals (e.g., insects and nication, robotic infrastructure enables observation and birds exhibiting swarming behaviors) can exhibit a wide data collection at unprecedented scales in various as- array of seemingly intelligent group behaviors. Sim- pects of ecological monitoring. This is significant for ilarly networked robots can potentially communicate environmental regulatory policies (e.g., clean air and and cooperate with each other, and even though indi- water legislation), as well as an enabler of new scientific vidual robots may not be sophisticated, it is possible for discovery. For example, it is possible to obtain maps networked robots to provide a range of intelligent be- of salinity gradients in oceans, temperature and humid- haviors that are beyond the scope of intelligent robots. ity variations in forests, and chemical composition of The significance and potential impact of networked air and water in different ecological systems [53.36]. robots is apparent from the following examples. In addition to mobile sensor networks, it is also pos- The manufacturing industry has always relied on in- sible to use robots to deploy sensors and to retrieve tegration between sensors, actuators, material-handling information from the sensors. Mobile platforms allow equipment, and robots. Today companies are finding the same sensor to collect data from multiple locations it easier to reconfigure existing infrastructure by net- while communication allows the coordinated control working new robots and sensors with existing robots and aggregation of information. Examples include sys- via wireless networks. There is also an increasing trend tems built for aquatic [53.37], terrestrial [53.38], and toward robots interacting with each other in operations subsoil monitoring [53.39]. There are many efforts to like welding and machining, and robots cooperating developed networked underwater platforms [53.40–42]. ]and 49 ]. Projects such as -wide coordinated 52 ]. A multi-university EU wherein robots map an ] are examples of swarm ) has a strong group in 50 53.10 LAAS ]) ) [53. 48 ) funded several EPFL EU A single operator commanding a network of Many research projects are addressing group be- The problem of coordinating multiple autonomous aerial and ground vehiclesvehicle from in an a urban commandnaissance environment and in for control a scouting recentPennsylvania, and demonstration Georgia recon- by Tech. the andCalifornia University (after University of [53. of Southern setting is shown inenvironment Fig. and deploy themselvesnetwork to to detect form intruders. a sensor haviors or collective intelligence by realizingbehaviors swarming observed in nature.pean For Union example, ( the Euro- 53.4.3 Research Challenges While there are manyworked successful robots embodiments of with net- applicationsdustry, to the manufacturing in- defensemestic assistance, industry, and space civiliansignificant infrastructure, exploration, challenges that there have do- are to be overcome. units and making themthe cooperate intersection of creates communication, problems control,tion. and at percep- Who should talkshould to be whom and conveyed, what and information how? How does each unit Fig. 53.9 projects on collectivegence. The intelligence I-Swarm or project in swarm Karlsruhe [53. intelli- the Swarm-Bot project at Ecole Polytechnique Fédérale de Lausanne ( intelligence. The Laboratorytecture for Analysis of and Systemsrobotics Archi- and ( artificial intelligence.a This long group history has ofrobot had basic systems. and The applied integrationvehicles research for of in applications multiple multi- such unmanned fire-fighting as is addressed terrain in mapping51 [53. and US project addressed the development of networkedhicles ve- for swarming behaviors [53. these are exploring the scalabilityto of large the numbers basic of concepts robots, sensors, and actuators. ] 43 ]has 38 ]andthe 45 10 depend- ]arebeing ,  ) has focused 39 44 ), University of RPI 53.9.Anexample ] and to obtain high- UCLA 37 47] to detect and track in- , ) is operated from a tactical ), University of California, 46 UAV USC 10 operators.  The eventual goal, however, is to enable a single In the defense industry, countries like the USA have human user toground, deploy surface, networks and of underwaterbeen unmanned vehicles. several aerial, There recent have demonstrationstems of exploring urban multirobot environments sys- [53. interiors of buildings [53. of a project with heterogeneous vehicles in an indoor truders, and transmita all remote operator. These of examples show the thatble it above is to possi- information deploy networked to 802.11b robots wireless using an networkmotely off-the-shelf and tasked have and monitored theexample by team of a be a single operator.an re- project An urban with setting heterogeneous vehicles is in shown in Fig. focused on themonitoring development the of forest robotic canopy,ing networks with data for a for modeling viewNetworked canopy to and robotic provid- undercover mini-rhizotrons growth. [53. deployed in thesoil. forest to monitor root growth ininvested the heavily in the concept ofically networked, geograph- distributed assets. Unmanned aerialthe vehicles Predators like are operatedsensors remotely. on Information from theother Predators vehicles triggers and themote weapon deployment location systems of and atcation allows a to commanders different control re- inUS and a military command engaged third all intems lo- these the initiative assets. large to The Future developto Combat deploying autonomous network-centric Sys- vehicles. The approaches network-centric tactical paradigms fornetworked robots modern for warfare defenseWhile have and networked robots homeland created are security. rent already approaches in are operation, limited cur- toing human users a command- singletakes many vehicle human or operatorsing on sensor (between the complexity 2 system. of thesystems However, system) like to it unmanned deploy complex aerialmanned vehicles. aerial A vehicle Predator ( un- control station, which may be ona an basic aircraft crew carrier, with of 3 Riverside, and University of California,networked Merced on infomechanical the system project [53. on the developmentmonitoring of a river robotic ecosystem.sity sensor Recent of work California, networks Los at Angeles for Univer- Southern ( California ( Networks of static and robotic devices haveoped been for devel- aquatic monitoring [53. resolution information on the spatialtributions of and plankton assemblages temporal and dis- concomitant en- vironmental parameters. The RiverNetat project Rensselaer [53. Polytechnic Institute ( Moving in the Environment

Part E

Part E | 53.4 1344 Multiple Mobile Robot Systems 53.4 Networked Mobile Robots 1345 atE|53.4 | E Part

Networking Sensor Robotics networking community community Communi- Control cation Networked humans, sensors and Fig. 53.10 Under the DARPA SDR program, a team from robots the University of Southern California, the University of Adaptation Perception Decision Tennessee, and Science Applications and International making Corporation (SAIC) demonstrated mapping, and intruder detection by a team of networked robots (after [53.46]) AI community move in order to accomplish the task? How should the team members acquire information? How should Fig. 53.11 The paradigm of networked robots introduces the team aggregate information? These are all basic fundamental challenges at the intersection of control, per- questions that need basic advances in control theory, ception, and communication that is of interest to the perception, and networking. In addition, because hu- robotics, sensor networks, and artificial intelligence com- mans are part of the network (as in the case of the munities Internet), we have to devise an effective way for mul- tiple humans to be embedded in the network and problem in navigation – behaviors for controlling in- command/control/monitor the network without worry- dividuals to achieve a specified aggregate motion and ing about the specificity of individual robots in the shape of the group, and the application to active per- network. Thus the underlying research challenges lie at ception and coverage. An overview of some of these the intersection of control theory, perception, and com- methods is provided in Sect. 53.4.4. munication/networks, as shown in Fig. 53.11. Problems of perception have been studied exten- It is also worth noting that robot networks are dy- sively in the robotics community. However, the percep- namic, unlike networks of sensors, computers or ma- tion problems in a system of networked, mobile sensor chines which might be networked together in a fixed platforms bring a new set of challenges; for example, topology. When a robot moves, its neighbors change and consider the problem of estimating the state of the net- its relationship to the environment changes. As a con- work. State estimation requires the estimation of the sequence, the information it acquires and the actions it state of robots and the environment based on local, executes must change. Not only is the network topology limited-range sensory information. Localization of n dynamic, but the robot’s behavior also changes as the vehicles in an m-dimensional configuration space re- topology changes. It is very difficult to predict the per- quires O..nm/k/ computations, where k is somewhere formance of such dynamic robot networks, yet it is this between 3 and 6, depending on the algorithm and analysis problem that designers of robot networks must domain-specific assumptions. The estimation problem solve before deploying the network. is further exacerbated by the fact that not all robots in This notion of a changing topology inevitably leads the network may be able to get the necessary informa- us to complicated mathematical models. Traditionally, tion in a time-critical fashion. There are deep issues of models of group behavior have been built on continu- representation and algorithmic development, which are ous models of dynamics of individuals, including local discussed in Sect. 53.4.6. interactions with neighbors, and models of control and The paradigm of active perception [53.53] links sensing with a fixed set of neighbors. While dynamics the control of sensor platforms to perception, bring- at the level of individual units may be adequately de- ing control theory and perception together in a com- scribed by differential equations, the interactions with mon framework. Extending this paradigm to networked neighbors are best described by edges on a graph. Mod- robots requires approaches of distributed control to be eling, analysis, and control of such systems will require merged with decentralized estimation. Robots can move a comprehensive theoretical framework and new rep- in order to localize themselves with respect to their resentational tools. New mathematical tools that marry neighbors, to localize their neighbors, and also to iden- dynamical system theory, switched systems, discrete tify, localize, and track features in the environment. mathematics, graph theory, and computational geom- These problems are discussed in Sect. 53.4.7. etry are needed to solve the underlying problems. We As discussed earlier, the communication network need a design methodology for solving the inverse is central to the functioning of a network of robots. j , I B are that  may / I i (53.2) (53.3) (53.1) ij r and its z . s ! j r (and sim- N . Thus the N s s  agents with . Clearly the . All agents i j i A r x N U N X and  D 2 i / i x ! ij X . Note that r . Each behavior r i and i i .  b D A X c c n 1 ij behaviors, which we  B x X N N i W . b R /; s d n W i consists of 2 j r E and may therefore not con- : ;:::; k . communications graph s 2 S n D B n ; adjacency matrix, N 1 that define the range and field ), and these estimates are derived 2 B N j j r A r  and a by sensing or communication chan- -dimensional space, and its velocity, D with time because the processors are /; i d otherwise if , with i N ij can encode T /: will consist of the position (and orienta- A B z i  i ; i ; ; u T i i x processors. Furthermore, even for a fixed has estimates of its own state and the states P 0 1 r ; A x i i ; . ( x is the estimator used by A T i h . change represents measurements of the state of agent i r ) has entries h f D in some  c ij D i ij z s / r D A D i . i j i O P x x A network of robots A Agent The state Finally, x available to sensing graph W j i P neighborhoods of of view of therespectively. communication hardware and sensors, a is defined byThe the sensing physical graph distribution (andgraph) of similarly is the the defined communications by agents. a map moving in and out of the sets relative position vector denoted by can be assigned identicalbehavior or represents different a behaviors.executed set Each computations (for of control unsynchronized,ing or locally carried estimation) out be- forprocessor some using collective in purpose, its withneighboring computations each only data fromassignment its of behaviors,typically each processor’s neighbors from information associated with edgesand communication in graph the sensing of neighbors (e.g., where the edges ofdepending the on the graph physical are proximity of formedSpecifically, pairs the dynamically of agents. nels and have dimension less than tain complete information about where is a controller, a function magnitude are important quantities thatestimated may for need biological to and be artificial agents. dynamics ilarly A methodology for modeling and analyzing suchwill systems require the merging of graphsystem theory theory and at dynamical a fundamental level. will denote by tion), r , 54 ], among 61 – specifying the i 59 , and control in- n TX R .  ! is characterized by an i i i ts map making, tracking X U A 2 -hop connectivity, or alge-  53.4.8 ], information integration, i k i x X 57 W , i f 56 ], and coverage [53. ,astate Z , with 58  R I  2 i i U 2 i u Given a group of mobile sensors, we would like In a robot network, we have multiple agents or ], localization [53. of objects and events, andusers goal-directed of navigation the for network. Finally,accomplish mobility allows tasks robots such to transportation, as and navigation, search and reconnaissance, rescue. to have distributed controlsirable capabilities global that specifications. realizebe Thus, de- able it to issition automatically necessary determine and to the orientationthe necessary distribution of po- of the group members,achieve group and the their members desired motion task. to must and/or At be a able lowercation to level, network use the and robots informationlocal from from estimates, reason their the about own communi- theneighbors spatial sensors and network to (their their derive and relationship then to use the the appropriatethe control environment), policies desired to group achieve specifications.simplest We briefly mathematical outline model the thatmulate such is problems necessary in order toof to the for- provide underlying a challenges. better sense nodes in whichcan each be agent aa robot, is sensor a a platform vehicle (possibly physical with static)cation relay or entity actuators node. even and Each that a agent sensors, communi- braic connectivity, enabling message delivery fromrobot one to another. Theposition group sensors so can as also toto cover self-organize shifts a to in desired the arealing and focus sensor adapt of position also monitoring suppor activities. Control- 53.4.4 Control The control of individual robots ismance critical and to scope the of perfor- robotordination networks. Indeed algorithms motion have co- been proposedpose of for improving communication the performance [53. pur- 55 deployment [53. However, if the network consiststransmitters of mobile and agents with receiversno with guarantee finite that power,Unlike all there a agents is static cancan sensor move talk network, toward to robots eachtion and teach other in adaptively maintain to other. a a communication facilitateSome network network. communica- basic algorithmic problems andresults several pertinent are provided in Sect. other tasks. Mobility allows thedeploy, group of and robots self-organize to self- bysupport relocating of themselves communication, in sensing,example, or task they needs; can for communication reconfigure bandwidth, to guarantee a desired identifier, puts Moving in the Environment

Part E

Part E | 53.4 1346 Multiple Mobile Robot Systems 53.4 Networked Mobile Robots 1347

The reader is directed to the many survey articles In such communication-enabled cooperative con- 53.4 | E Part on this subject for further information. An overview trol and planning (see also [53.75]), the communication of challenges for the controls community is presented network plays an important role in the creation of in [53.62]. The underlying theory for networked mo- a shared representation of information. This notion of bile systems has been explored in the context of au- a shared representation is important to the scaling of tomated highway systems [53.63], cooperative robot coordinated control algorithms to large numbers of de- reconnaissance [53.46] and manipulation [53.64], for- vices. For example, in [53.67], the information form mation flight control [53.65], and the control of groups of the Kalman filter is used to derive a framework of unmanned vehicles [53.45]. Our goal in the fol- for decentralizing estimation and fusion algorithms. lowing sections is to explore the connections between This approach was shown to be applicable to multi- communication, perception, and control. ple heterogenous ground and aerial platforms [53.56]. Such methodologies are transparent to the specificity 53.4.5 Communication for Control and identity of the cooperating vehicles. This is be- cause vehicles share a common representation, which Communication networks allow physically discon- consists of a certainty grid that contains information nected entities to exchange information. At the lowest about the probability of detection of targets, and an level, when groups of vehicles coordinate their actions, information vector–matrix pair that is used in the infor- communication allows vehicles to exchange state infor- mation form of the Kalman filter [53.45]. Observations mation [53.66–68]. At a higher level, robots can plan are propagated through the network by changing both navigation and exploration tasks based on an integrated the certainty grid and the information vector/matrix. map of the world derived from information acquired This allows each vehicle to choose the action that from different robots [53.52]. maximizes a utility function, which is the combined The use of communication for control in the mul- mutual information gain from onboard sensors to- tivehicle context has been addressed in the PATH wards the detection and localization of features in the project where formations of inline vehicles were stud- environment. ied [53.63]. Problems of the stability of the forma- Thus, in summary, at the lowest level, communica- tion [53.69], the convergence of the formation to tion enables either partial or complete state feedback of shapes [53.70], and the overall performance of the sys- the network and allows agents to exchange information tem [53.71] are of great interest. The performance of for feedforward control. At the higher levels, agents can the system is directly influenced by the interconnections share information for planning and for control. This is between agents. In addition to impacting on stabil- also discussed in Sect. 53.4.6 where the communication ity [53.63], feedback of states from different agents network is shown to enable a network-centric approach and feedforward information from the plans of differ- to perception. ent agents affects the rates at which the system of agents can respond to external stimuli [53.71]ortocommands 53.4.6 Communication for Perception from human operators [53.72]. In addition, communication can be used for high- While individual robots have sensors and the ability to level control and planning of robots. There is great build maps and models by integrating sensory informa- interest in using static sensor nodes as beacons to guide tion, networked robots can exchange information and robot navigation. In [53.73], the problem of coverage leverage sensory data, maps, and models from other and exploration of an unknown dynamic environment robots. The challenge is to exploit communication for using a mobile robot is considered. An algorithm is perception in tasks such as distributed mapping in the presented which assumes that global information is not presence of the delays, limited bandwidth, and disrup- available (neither a map, nor global positioning system tion that are typical of communication networks. (GPS) information). The algorithm deploys a network Distributed localization is the term used to de- of radio beacons that assists the robot in coverage. The scribe the merging of communication and perception network is also used by the robot for navigation. The for state estimation. Localization is an essential tool deployed network can also be used for applications for the development of low-cost robot networks for use other than coverage (such as multirobot task allocation). in location-aware applications and ubiquitous network- A similar idea was presented using potential-field-based ing [53.76]. Location information is needed to track navigation in [53.52]. In this work the notion of no- the placement of the nodes and to correlate the values go or danger areas was incorporated into the navigation measured by the node with their physical location. Dis- cost function. Recent work along these lines with exper- tributed computation and robustness in the presence of imental data from sensor nodes is reported in [53.74]. measurement noise are key ingredients for a practical , 73 , ]. A tiered ]andfield 46 88 ]. The networked ]. For example, the -funded software for ) project has focused NIMS ) program. The goal of these DARPA addresses the design of network ]. SDR 89 USC Another important set of problems arises when The information collected by the nodes in a sensor experiments was to developrobot system and capable demonstrate of carrying a outThis a multi- specific required mission. therobots ability into to an deploy unexplored ainterior, building, detect large map and number the track of the intruders, building above information and to transmit aone remote all set operator. of A of report experiments is on presented in [53. distributed robotics ( search Projects Agency ( strategy for deploying thehighly robots capable is robots described,and formed map where the a first building,used followed wave by the to a resulting second enter mapenvironment wave to for which self-deploy intruders. and Bothtensively approaches monitor on relied the ex- networking the802.11b robots wireless using technology. This commercial communication task for involved building both awell shared control representation for perception. as robot networks areand used then for trackingembedded identifying, targets stationary localizing, wireless ina sensor a virtual network sensor dynamic is spreadSuch setting. over like a a An large network geographicalrobots area. can about remote provide locations.this information Robot virtual to networks sensor allow mobile stimuli and to to move tracksible in moving to targets. response Indeed, cast itwith to this robotic is sensor external scenario pos- networks87 [53. asTenet project a at pursuit-evasion game on sensor-assisted techniquesadaptive for sampling mobile for event robot-based response [53. primitives and abstractionstectures, for with tiered robotictarget network pursuit applications. archi- evasion Algorithms aspling for strategy one guiding of of thephenomena a sam- the of robotic interest boatvironments (e.g., to are hotspots) model in discussed andinfomechanical aquatic in locate systems en- 37 [53. ( reconstruction [53. network can be processed at a centralcentralized location or fashion. in a Such de- techniques in-network make better data use of processing and network communication computation resourcesing. This than also centralized enables theand process- network up-to-date to global compute pictures accurate oflandscape the that global are perception available totem. all Methods the for robots in-network data in processing with thenodes static sys- include artificial potential-field computation, gra- dient computations, particle filters, Bayesianand inference, signal processing. Algorithms have been developed for computing maps, paths,90]. and predictors52 [53. ] ]. 84 86 , gener- ], each 85 85 ]. Other techniques ]. Mobility-assisted 81 , ] the Cramér–Rao lower 80 ]. In a recent paper [53. 77 83 , landmarks, or maps of any , 77 GPS ], maximum-likelihood estimation . In [53. 86 ] a theoretical foundation for network lo- 54 ] a distributed algorithm that uses no beacons ]. Localization based on the propagation of lo- )CRLB for network localization is derived. This 78 79 The methods for distributed localization can be clas- In [53. Two approaches for cooperative relative localiza- Akeyissueistobeabletoscalethesecomputa- MLE) and numerical optimization is used to achieve localization algorithm that will give reliablea results large-scale over network. sified into two broadanchor nodes for classes: localization and algorithms algorithms that thatbeacons. use Localization no rely may be on computed using rangeformation in- between nodes, bearing information, or both. calization in terms of graphThe rigidity problem theory is is solved provided. information when and nodes it have perfect islocalization range show if that and a only if network itsically has underlying a globally graph is unique rigid kind; instead, robots make directrelative measurements pose of of the nearbyformation robots to and the broadcast team this in- as a whole. In [53. and is guaranteedmation to under compute measurement correctrange noise location to for infor- neighbors nodes ison presented. that the This can notion algorithm ofbustly relies a robust global system quadrilaterals of to coordinatesThe among compute the computation nodes. ro- supports movingof nodes. this Extensions workin to [53. passive trackingcation have information been from discussed known referenceon nodes connectivity includes based [53. work computes thean expected ideal error algorithm, and characteristicsror compares this for in to an the algorithmthe actual er- based important conclusion on that multilateration,the the algorithm drawing error is just introducederror as by important in as theIn assessing measurement [53. end-to-end localization accuracy. bound ( localization is introduced inuse [53. 82 distributed propagation of locationing information multilateration us- [53. the problem of evaluatingwork the rigidity is of treated aof planar real-world while systems: net- decentralization, satisfying asynchronicity, and common parallelization. objectives tion of mobile robot teams are given in [53. Neither method uses robot processes this information independentlyerate to an gen- egocentric estimate for theusing pose a of Bayesian other robots formalism withmentation. a In particle [53. filter imple- ( a similar result. tions for building a sharedbers representation of to large robots num- andin sensors. experiments This under problem the was US studied Defense Advanced Re- Moving in the Environment

Part E

Part E | 53.4 1348 Multiple Mobile Robot Systems 53.4 Networked Mobile Robots 1349

A recent DARPA demonstration showed how com- based objective functions [53.91]. Stability results are 53.4 | E Part munication networks can be used effectively in per- derived without concerns for the optimality of the ception tasks involving heterogenous robots [53.44]. In network configuration, but local guarantees are pro- cooperative search, identification, and localization un- vided. Topology aware coordinated behavior is treated manned aerial vehicles (UAVs) can be used to cover in [53.92]. A body of results reported in [53.93] large areas, searching for targets. However, sensors on and [53.94] describes decentralized control laws for po- UAVs are typically limited in their accuracy of localiza- sitioning mobile sensor networks optimally with respect tion of targets on the ground. On the other hand, ground to a known event distribution density function. This ap- robots can be deployed to accurately locate ground tar- proach is advantageous because it guarantees that the gets but have the disadvantage of not being able to move network (locally) minimizes a cost function relevant to rapidly and see through obstacles such as buildings or the coverage problem. However, the control strategy fences. In [53.56], the synergy between these two de- requires that each agent have a complete knowledge vices is exploited by creating a seamless network of of the event distribution density, thus it is not reac- UAVs and unmanned ground vehicles (UGVs).Asdis- tive to the sensed environment. The work by [53.95, cussed in Sect. 53.2,thekeytosuchnetwork-centric 96] generalizes these results to situations in which the approaches for search and localization is a shared rep- nodes estimate rather than know ahead of time the event resentation of state information, which in this case is distribution density function. A local (decentralized) easily scalable to large numbers of UAVsandUGVs control law requires that each agent can measure the and is transparent to the specificity of individual plat- value and gradient of the distribution density function forms. However, how to do this more generally and for at its own position. This results in a sensor network more unstructured information remains an issue for fu- that is reactive to its sensed environment while main- ture research. taining or seeking a near-optimal sensing configuration. In addition, the distribution density function approxi- 53.4.7 Control for Perception mation yields a closed-form expression for the control law in terms of the vertices of an agent’s Voronoi Networked mobile robots enable the exploration of region. This eliminates the need for the numerical in- dynamic environments and the recovery of three- tegration of a function over a polynomial domain at dimensional information via distributed active percep- every time step, thereby providing a significant reduc- tion [53.53]. Since the nodes are mobile, a natural tion in computational overhead for each agent. Other question is: where should the nodes be placed in or- work in event monitoring for unknown distributions der to ensure successful integration of information from includes [53.59]. Krause et al. [53.97] have recently multiple nodes, and to maximize the quality of the es- proposed an approach for sensor placement that con- timates returned by the team? Since there is a cost siders both the sensing quality and communication cost associated with transmitting and processing data, it is of imperfect sensing and communication components. important to consider which sensor readings should They use a parametric model for link reception rate that be used in the state estimation and what information assumes no acknowledgement and no temporal correla- should be communicated to the rest of the system. The tion of lossy links. quality of the information computed by the network de- Beginning with the art gallery problem, there have pends on the locations of the sensor platforms both in been multiple efforts to determine an optimal configu- an absolute and relative sense. The quality also depends ration of sensors to cover a given region [53.98–100]. on the noise characteristics of each sensor, and the com- A variant which allows the use of mobile sensors is munication network. known as the watchmen tours problem. In these ap- A robot network goes well beyond a fixed sensor proaches the sensor model is abstract and not well network, which can only collect data at fixed positions suited to real environments and cameras. Distributed in space; for example, when an event is detected at geometric optimization methods [53.94] have also been a specific location it is possible to direct more sen- used for mobile sensor network reconfiguration. A re- sors toward the location of observation of the event for lated class of methods is the use of estimation-theoretic more information (for example, higher-resolution data optimization metrics and the application of informa- or higher sampling frequency). Reconfiguring the node tion filters to coordinate network-wide motion [53.56]. locations for adaptive resolution sampling relies on dis- There are other distributed optimization methods which tributed control strategies. use a distributed control law and show that it optimizes Various strategies have been introduced for control- a global metric of interest, such as using a potential ling mobile sensor network coverage. Mobile sensing field or other linear control law based only on local agents are controlled using gradients of information- neighbor interactions [53.101]. Research focusing on ], 107 , ]using 106 105 ]; nodes are deployed s is controlled either by 58 ]. In contrast, controlled de- 109 ]. Given an overdeployed network, 108 ]. It is also possible to reposition nodes 55 Mobile robots can be used to create desired network A distributed algorithm for the deployment of mo- Most work on network topology control has dealt a parent in the spanningregion, tree. the When the parent robot is leavestoo informed. its far When away, the the spanning robot tree is moves modified. topologies under suitable modelsnication. of If network commu- aenvironment (or robot if sensor is nodes robotically self-deploy) usedto build to a network, emplace the problemment. nodes is It referred is in to possible as an to deploy-nodes control to the guarantee motion that oftained individual a [53. specified topology is main- with the explicit aimogy of – changing the theproblem. so-called network topol- mobility-based topology control bile robot teams has been described by [53. the concept of virtualfrom pheromones: localized one messages robotto to generate another. either These adeployment messages gas model. expansion are or Similar used a algorithmsficial guided based growth potential on fields arti- are described in [53. where the latter incorporates aAn connectivity incremental constraint. deployment algorithmsor for networks mobile is sen- given in [53. one at a time intonode making an use unknown of environment, information gathered withdeployed by each previously nodes to determineThe its algorithm deployment is location. designed toage maximize while network ensuring cover- thatone nodes another. retain line of sight with with uncontrolled deployments, whereplicit control there of is the noprimary positions ex- of mechanisms individual proposed nodes. are The sleep scheduling. power These control methods involve pruning and anready al- existing, well-connected communication graph in order to save powersubgraph preserves while connectivity. Given ensuring ais that network connected the that when all resultant nodespower, are the operating aim at of maximum power controlpower level is at to each use node the forconnected which minimum the [53. network remains sleep scheduling seeksof to nodes activate todesired a maintain metrics minimal connectivity and [53. subset achieve other ployments are feasible when the positionsnodes of can individual be altered. Suchfor deployments are two interesting reasons. First,communication network topology relates with directly wireless and to hence the proximity position of relations thecreasing nodes. evidence Second, that there is a in- are large number likely of to deployments involvenodes. careful, The nonrandom positioning placement of node of the nodes themselves or by external agents. Such net- ] ]an 104 ], the 60 102 ]. In [53. 103 , 102 , robots to keep track of each 60 , we briefly discussed the benefits of 53.4.5 One simple control strategy that can affect network In Sect. using the communicationimprove network controller to design. Conversely, synthesizeof the and robots movement affects thethe network network. and This data gives transmissioncontrollers rise in for to individual robots many are known, challenges. canvide If we pro- the guarantees about communicationand can in we the develop robust information network working routing and algorithms net- in theAnother presence challenge of concerns robotgates how and motion? diffuses information in these propa- networks.under If the a robots given move controlformation model, propagate through how the does network a andwe what say piece can about of when and in- wherewill that be piece of heard? information Iftions, it we may know be the possibledesired to answers design communication to controllers network to characteristics. such realize ques- performance is to controlmessages the are robot transmitted motionThe between to movement ensure designated of robotssensors nodes. may in cause network a partitioningout network when of nodes of range. go However, robots the abilityin and of a the robots controlled to way move alsodress leads the to information an routing problem opportunitynetworks to in by ad- disconnected turning thekey robots idea into here is relaymessage to nodes. to enable The an the unavailabletrajectory robot in destination holding order to to a relay modify current abeen message. their This formulated problem as has anis optimization problem. to The minimize goal to the send trajectory a modifications message necessary have to been its proposed depending destination. on Severalis the solutions available information to that the robots.known, If path the planning robots’ techniques trajectories canpute are be who used moves totrajectories where com- are to not relay known,can a what. be distributed If spanning created tree the to robots’ enable 53.4.8 Control for Communication other. Each robot is assigned a region of movement and the control ofpabilities cameras is with dueapproach pan, to tilt, is [53. and developedcamera zoom automatically to ca- over calibrate itsbuild full very a high-resolution zoom pan–tilt–zoom panoramas. range In and [53. to cameras are constantly moved to trackusing observed a targets, factor graph. A recent algorithm due to [53. significantly improves on this by positioning camerasmake to the network bettertargets suited as to theythe detect emerge. construction and of Pan–tilt–zoom far classify more cameras flexiblestatic vision allow cameras. systems than Moving in the Environment

Part E

Part E | 53.4 1350 Multiple Mobile Robot Systems 53.5 Swarm Robots 1351

works present a different and interesting scenario for where the former range is twice as long as the latter. 53.5 | E Part topology control since it is possible to exploit control of Zhang and Hou proved that, if the communication range the motion and placement of the nodes to build efficient is at least twice the sensing range, complete coverage topologies. A local, completely decentralized technique of a convex area guarantees network communication for topology control using mobility is given in [53.110]. connectivity, and then used this theorem as a basis for An important application for networked robots is a localized density control algorithm [53.109]. This was in monitoring and surveillance, where it is impor- subsequently generalized to show that the condition that tant that the robots cover the space while remaining the communication range is twice the sensing range within communication range [53.111–113]. In a recent is sufficient and is the tight lower bound to guarantee development [53.114] the problem of how to design that complete coverage preservation implies communi- communication models and scheduling protocols for cation connectivity among nodes if the original network choosing the appropriate path planning algorithms for topology is connected [53.117]. robotic data collection is discussed. Probing environ- In summary, if the state of the communication ment and adaptive sleeping protocol (PEAS) was one of network and the desired state of the communication the first attempts to address communication connectiv- network is known to each agent, it should be possible ity and sensing coverage simultaneously using heuris- to synthesize distributed controllers to move agents to tic algorithms [53.115]. Wang et al. [53.116] proposed attain desired network characteristics. However, the as- a new coverage configuration protocol (CCP)topro- sumptions on the global state are clearly not justified. duce an approach that simultaneously optimizes cov- Also, the desired motion to optimize network charac- erage and connectivity while maximizing the number teristics will conflict with the motion that is required to of nodes that are placed into sleep mode. Furthermore, perform the desired task. However, as the brief discus- they also identified three different classes of coverage– sion above illustrates, there are many interesting studies connectivity problems with respect to the ratio of radio that point to promising directions for future work in this and sensing ranges and recognized the critical ratio very fertile research field.

53.5 Swarm Robots

Historically, some of the earliest work in multirobot achieve team-level tasks. Ideally, the entire team should systems [53.12, 13, 118–125] dealt with large numbers be able to achieve much more than individual robots of homogeneous robots, called swarms; swarm robotics working alone (i. e., it is superadditive, meaning that continues to be a very active area of research. Most the whole is bigger than the sum of the parts). These swarm approaches obtain inspiration from biological systems assume very large numbers of robots (at least societies – particularly ants, bees, and birds – to develop dozens, and often hundreds or thousands) and explicitly similar behaviors in multirobot teams. Because biolog- address issues of scalability. Swarm robotic approaches ical societies are able to accomplish impressive group achieve high levels of redundancy because robots are capabilities, such as the ability of termites to build large assumed to be identical, and thus interchangeable with complex mounds, or the ability of ants to collectively each other. carry large prey, robotics researchers aim to reproduce Many types of swarm behaviors have been studied, these capabilities in robot societies. such as foraging, flocking, chaining, search, herding, Swarm robotics systems are often called collective aggregation, and containment. The majority of these robotics, indicating that individual robots are often un- swarm behaviors deal with spatially distributed multi- aware of the actions of other robots in the system, other robot motions, requiring robots to coordinate motions than information on proximity. These approaches aim either: to achieve a desired team-level global behavior from 1. Relative to other robots the interaction dynamics of individual robots follow- 2. Relative to the environment ing relatively simple local control laws. Swarm robotic 3. Relative to external agents systems typically involve very little explicit communi- 4. Relative to robots and the environment cation between robots, and instead rely on stigmergy 5. Relative to all (i. e., other robots, external agents, (i. e., communication through the world) to achieve and the environment). emergent cooperation. Individual robots are assumed to have minimal capabilities, with little ability to solve Table 53.1 categorizes swarm robot behaviors ac- meaningful tasks on their own. However, when grouped cording to these groupings, citing representative exam- with other similar robots, they are collectively able to ples of relevant research. ). ], 132 ] VIDEO 195 129 58], coverage [53. robots are developed that 128], dispersion [53. ], forced herding/shepherding ] s-bot 138 ], 139 135] 142 ). -based SWARM-BOTS project studied new The SwarmBot robots ] 53.13, which shows the robots self-assembling VIDEO 212 ], target tracking [53. 140 121], natural herding (as in herds of cattle), schooling, ], exploration [53. Another notable effort in swarm robotics research The EU 137 134 is the US multi-universitythe University SWARMS of initiative Pennsylvania. Research led in thisaimed by project at developing a new system-theoretic frame- Fig. 53.12 as collective perceptiona tasks. hardware This and a projecting software microscale was robots challenge, that are in both fully autonomous thatperform and meaningful develop- can cooperative behaviors will require significant advances in the current state of the art. concepts in theorganizing and design self-assembling robots and ( implementation of self- In this work [53. have grippers enablinglinks the with robots other tosemblies s-bots create of or physical robots.together objects, These for thus assemblies navigation across can creating roughlectively then as- terrain, transport or work objects. to Thewith col- s-bots a are flexible cylindrical, armnect and one toothed s-bot gripperof that to object can another. transport con- An usingFigure interesting SWARM-BOTs is application showninto in four chains in order( to pull a child across the floor ], condensation, aggregation [53. 131], grazing, harvesting, deployment [53. 14 dis- 53.1 ]. ], flocking [53. 127 129 ], mapping [53. , 53.10 133 126 ], predator–prey [53. ], foraging [53. as primitives 136 14], clumping [53. 130 ) developed an exten- (as in shepherding sheep) localization [53. Swarm behaviors Formations [53. sorting [53. Search [53. Pursuit [53. Containment, orbiting, surrounding, perimeter search [53. Evasion, tactical overwatch, soccer [53. basis behaviors VIDEO 215 , and were able to locate an object 2 ]( 53.12. He created several group behav- 141 ). Some of these swarm behaviors have ], involving about 20 physical robots per- 6 Categories of swarm behaviors 53.2, the first demonstrations were by [53. [53. c VIDEO 214 Much of the current research in swarm robotics The European Union has sponsored many swarm Demonstration of physical robot swarms is both Relative to other robots andenvironment the Relative to other robots, external agents, and the environment Relative motion requirements Relative to other robots Relative to the environment Relative to external agents for structuring more complex systems. InMcLurkin later research, forming aggregation, dispersion,work defined and composable flocking. This ( is aimed atmore developing of specific the solutions swarm to behaviors one listed or in Table of interest and lead a human to its location [53. robot projects, leadingsized toward individual decreasingly robots, smaller- robot and teams. The increasingly I-SWARM project, for larger-sized instance, aimed at developing millimeter-sizedboard sensing, robots computation, withing and biologically full power inspired for on- swarming perform- behaviors, as well sive catalog of swarmstrated behavior these software, behaviors and on(called demon- the about SwarmBot robots), 100 developed byshown physical iRobot, in robots as Fig. iors, such as avoidManyRobots,disperseFromLeaves, disperseFromSource, disperseUniformly, computeAv- erageBearing, avoidManyRobots, followTheLeader, or- bitGroup, navigateGradient, clusterOnSource, and clus- terIntoGroups. A swarm of 108oped dispersion robots algorithms used in the an empty devel- area schoolhouse of of about 300 m Table 53.1 received particular attention, notably formations,ing, flock- search, coverage, and foraging. Section cusses these behaviors in morecurrent detail. work In general, in most theis development aimed of not swarm justare at behaviors similar demonstrating group to motionsstanding biological the that formal systems, control theoretic principles butpredictably that can also converge to at theand under- remain desired in stable group states. behaviors, a hardware andin a Sect. softwareMatari´ challenge. As discussed Moving in the Environment

Part E

Part E | 53.5 1352 Multiple Mobile Robot Systems 53.5 Swarm Robots 1353

a different approach to the synthesis of local decision 53.5 | E Part policies by making use of the decentralized sparse- interaction Markov decision process. This technique allows agents to recognize states when interactions with other robots might occur, thus enabling them to choose better motions based on possible future inter-robot ac- tions. Tsiotras and Castro [53.146] synthesize local controllers by generalizing the standard consensus al- gorithm, applied to geometric pattern formation. A common theme in most of the works cited above is that they first make use of formal methods to de- scribe the desired macro-level behavior, and then show how to use this macro-level goal to synthesize individ- ual robot controllers. Another use of formal methods is to show how the individual goals of robot team mem- bers can be considered collectively, with the objective of maximizing the system’s achievement of individual goals. Toward this end, game-theoretic techniques have Fig. 53.13 SWARM-BOTs self-assembling to move been shown useful in a variety of distributed robot for- a child across the floor mulations. Cheng and Dasgupta [53.147]makeuseof the game-theoretic technique called Weighted Voting work for swarming, developing models of swarms and Games to address the problem of multi-robot team for- swarming behavior, analyzing swarm formation, stabil- mation control amidst obstacles. Taheri et al. [53.148] ity, and robustness, synthesizing emergent behaviors for also make use of game-theoretic principles, building active perception and coverage, and developing algo- upon the Local Interaction Game diffusion model to in- rithms for distributed localization. vestigate how a small number of agents can influence Besides the hardware challenges of dealing with the global society’s behavior through local interactions. large numbers of small robots, there are many import- An interesting question in the design of distributed ant software challenges that remain to be solved. From robot coordination mechanisms is the extent to which a practical perspective, a common approach to creat- identical controllers can lead to diversity, specializa- ing homogeneous multirobot swarms is to hypothesize tion, or changes in robot behavior. Hsieh et al. [53.149] a possible local control law (or laws), and then study study the emergence of specialization in robot swarms the resulting group behavior, iterating until the desired by making use of a distributed adaptation algorithm. global behavior is obtained. However, the longer-term They present a top-down analytical approach that de- objective is to be able to both predict group perfor- fines the system equilibrium using waiting time param- mance based on known local control laws, and to eters, and then present adaptive optimization strategies generate local control laws based upon a desired global that converge to the optimal configurations that achieve group behavior. system equilibrium. Temporal changes in system-level More recent research has focused on the devel- swarm behavior are addressed by Hoff et al. [53.150] opment of analytical techniques that can synthesize who show how a swarm can change and improve its for- distributed controllers that achieve the desired macro- aging behavior by switching between algorithms based level system behaviors. Mather and Hsieh [53.143] on the environment in which the swarm finds itself. address this challenge by proposing a technique that Once the distributed controller is synthesized, most first identifies robot–robot interactions at the macro- of the works mentioned above presume that individual scopic level; they then use this analysis to improve local robots execute their controller successfully. The typi- robot control policies by filtering out spurious robot– cal presumption is that large swarms of interchangeable robot interactions. Another top-down design approach robots automatically result in robust and scalable swarm is presented by Chen et al. [53.144] who show how behavior. However, this presumption is challenged by to automatically synthesize control and communication Winfield and Nembrini [53.151], who illustrate that strategies for a robot team based on global specifica- overall swarm reliability quickly falls in the presence of tions of the desired system-level behavior, stated using worst-case, partially failed robots. They conclude that regular expressions. The resulting control strategies are future large scale swarm systems must develop new ap- formally proven to correctly achieve the desired global proaches for achieving high levels of fault tolerance. behavior. The work of Melo and Veloso [53.145] takes One example approach is shown in VIDEO 194 . , ], ]. 168 161 163 et al. added improved upon et al. is another ] was developed Murata 166 , ]. ]. The Molecube sys- Ijspeert Lipson ] which uses a connected ] also builds upon on the ] which has undergone 165 170 172 and 167 163 (one-dimensional) lattice sys- can self-replicate. – ) configurations. Lipson et al. 1-D 160 3-D SMA) springs when heated beyond Marbach et al. developed the M-TRAN modular ], developed by 171 et al. designed another unique chain-type sys- C. Interestingly, the RATChET modules possess no Murata The ATRON system [53. The PolyBot is chain-type modular robot [53. ] with a single rotational degree of freedom. Poly- Yim ı intelligence. Instead, they relynal on actuator an which intelligentdangling rotates exter- chain. to One unique control propertysystem of one is the its end RATChET relatively strength. of the 53.6.2 Lattice Systems Lattice-type modular robot systemsinterconnected are robotic collections modules in of situated which the at units the are dimensional grid. intersection (A points of a two or three tem is simply a chain-type robot.) The main characteris- robotic system [53. multiple revisions and improvements. In [53. Kamimura et al. employof phase a oscillators set toTRAN achieve of walking system. interconnected, gaits in out the M- the ability of systemsin like real-time M-TRAN by applying tomodular function generate system, optimization YaMoR gaits 164 [53. to their cameras to theTRAN M-TRAN modules system could sotasks, separate, that and perform then a rejoin independent into set a larger of structure [53. M- to improve upon thekeep M-TRAN. M-TRAN’s Lund ability et al.taking to advantage wanted of form to the two dense orthogonaldom, degrees lattices of (pitch free- while and yaw),The Superbot found system in [53. themechanical design CONRO of M-TRAN system. by adding andegree additional of rotational freedom betweenrotation the axes. two existing 169 Bot evolved into CKBotability which to reassemble has itself demonstratedintentionally after destroyed the being [53. accidentallytem or [53. tem named RATChET [53. chain of inter-latching right anglestructures. tetrahedrons to Neighboring form RATChET modulesgether latch when the to- angleical between them value, passes and some theymemory crit- unlatch alloy through the ( 70 use of shape example of a chain-type modular systemdegree with of only freedom one butthree-dimensional still ( able toshowed achieve that interesting a short chainwith of some Molecube free modules, modules, along ] ]in 156 Nak- , 159 – and 155 o accomplish [53. Fukuda Yim can combine t mposed of two orthogonal ]. 154 , 153 ] describing the abstract concept of a re- ). The polypod system was composed of 152 [53. VIDEO 196 One of the first chain-type modular robotic systems Modular robots are collections of physically con- 53.6.1 Chain Systems The defining characteristic of chain-type modularsystems robot is the factto that their the modules, neighbors,chains when may connected are be arrangedbut one-dimensional, three-dimensional in or chains two-dimensional, are afact not that chain. a as chain-type modular common. robot These is The or two-dimensional, even one-dimensional, does notoperate in mean three that dimensions. In it fact, cannot robots snake-like modular composed of segments with orthogonal jointsquite are common. was the polypod system developed by ( The modular roboticssented field at began International withAutomation in Conference a the Spring on paper of 1988 pre- Robotics by and 53.6 Modular Robotics two types of modules: segments and nodes. It coulda form variety of shapes including rolling loops andand hexapods, it went ontems. to One inspire many was other the chain-based sys- CONRO system157 [53. a variety of tasks.lar Over robotics the research past developeddesign; twenty many planning facets: years, and hardware modu- controlbetween hardware algorithms; and the algorithmic complexity; efficient trade-off simulation; and system integration. nected, electromechanically activea whole, modules form that, roboticgreater systems than as that those exhibit of the capabilities modular individual robots modules. can Typically change their shapein or order configuration to adapt toample, a a variety collection of of different modulesa tasks. could For closed reconfigure ex- from chain thata rolls legged robot quickly that over moreModular open easily robots traverses ground are rough typically to terrain. toutedity, for their their adaptabil- fault-tolerance, andthe the unit relative modules. simplicity Modularscribed robotic of and systems classified can on be severalproperties. de- axes using In a what variety follows, of route of we classifying choose modular robotic the systemsometry traditional by of the the ge- system: chain,For lattice, truss, a or free more form. field, detailed consult [53. history of the modular robotics agawa configurable robotic system thatshapes and can envisioned a assume robot different systemferent composed types of of dif- modules that which each module wasservo motors co controlling each module’s pitch and yaw. Moving in the Environment

Part E

Part E | 53.6 1354 Multiple Mobile Robot Systems 53.6 Modular Robotics 1355

tic separating a lattice system from a densely configured robot developed by Koseki [53.190] is able to actu- 53.6 | E Part chain-type robot is the density of the interconnections ally perform the sliding motion assumed by Chiang and between the modules. In a lattice-type system, each Chirikjian in [53.189]. More recently, An developed the module is typically connected to all of its neighbors. In EM-Cube system [53.191] which is also capable of slid- a dense chain-type system, two modules may be neigh- ing motion. bors, but they will not be physically connected. Another unique lattice is the I-Cube developed by Additionally, lattice-type systems tend to be built Khosla et al. [53.192, 193]. The 3-D I-Cube system con- with modules that contain no rotational degrees of free- sists of passive cubes which are connected by active links dom. While the modules in a lattice system typically with three rotational degrees of freedom that are able to have mechanisms which enable the modules to move grab, reposition, and release the cubes. The 3-D I-Cube relative to, and bond with, their neighbors, they gener- system was an improvement of the two-dimensional ally cannot bend themselves. In comparison, chain-type (2-D) system [53.194] developed by Hosokawa et al. for systems are often built from modules that contain one or rearranging cubic modules in a vertical plane. more rotational degrees of free so that the modules can Goldstein et al. initiated the Claytronics project flex like links in a chain. There is some overlap between by publishing several papers [53.195, 196] propos- between the two types of system. ing lattice-based claytronic atoms or catoms. These Chirikjian et al. developed one of the first vertically-oriented cylindrical robots, which were inca- lattice-based modular robotic systems [53.173–175] pable of independent motion, used 24 electromagnets ( VIDEO 198 ) in which the modules are deformable around their perimeters to achieve rolling locomotion hexagons capable of bonding with their neighbors. Oth- about their neighbors. Goldstein et al. envisioned a sys- ers, such as Walter et al. [53.176] further analyzed these tem in which millions of smaller catoms could form hexagonal type systems to create distributed motion arbitrary shapes using a randomized algorithm that planners capable of reconfiguring the system from one avoided conveying a complete description of the shape state to another. to each module in the system. Murata et al. were also early contributors to the The catoms continue to evolve. One of the newest development of lattice-based modular robotic systems instantiations [53.197] employs hollow cylinders rolled with their development of a roughly hexagonal module from SiO2 rectangles patterned with aluminum elec- capable of rolling around its neighbors in two dimen- trodes. The authors hope that two of these cylinders, sions [53.177, 178]. Kurokawa et al. presented a three when placed in close proximity with their axes aligned, dimensional adaptation [53.179] composed of cubes will be able to rotate with respect to one another using with six protruding arms capable of rotation. Yoshida electrostatic forces. Specifically, the electrodes, (which et al. improved on this system with a new design that reside on the inside of each cylinder and are electrically used shape-memory alloy actuators to rotate one robot isolated by the SiO2), will be charged so that they attract module around the perimeter of a neighbor [53.180]. and repel mirror charges on the neighboring cylinder in One of the simplest lattice systems is the the Digital a way that causes rotation. Currently, the system ap- Clay project [53.181]. The system was a set of com- pears to be constrained to form 2-D structures. The pletely passive modules that relied on the user to make authors claim the completed system will have a yield changes to it topology. The 2:5 cm rhombic dodecahe- strength similar to that of plastic and that the modules drons were able to sense and communicate with their will be able to transfer power and communication sig- neighbors in order to create a virtual model of the phys- nals capacitively from neighbor to neighbor. ical arrangement of modules. The Claytronics project has proposed, but not yet Rus et al. also explored the idea of 3-D mod- demonstrated with hardware, the use of sub-millimeter ules capable reconfiguration through a series of latch- intelligent particles as sensing and replication de- ings, rotations, and unlatchings with the Molecule vices [53.198]. In particular, Pillai et al. present a the- system [53.182–185]. In [53.186, 187], Vona and Rus oretical 3-D fax machine in which the object to be describe a different type of deformable lattice system. faxed is immersed in a container of intelligent particles The Crystal system is composed of square modules able that sense and encode the object’s dimensions. At the to expand and contract by a factor of two in the x–y receiving end, these same Claytronic particles decode plane. Suh et al. expanded on the Crystal concept with the shape description sent by the transmitter and bond the Telecubes [53.188] that could move in three dimen- together to replicate the original object. Unlike our ap- sions by expanding all six faces. proach, Pillai’s approach is completely centralized and Chiang and Chirikjian analyzed how to perform relies on an external computer for computation. motion planning in a lattice of rigid cubic modules White et al. developed hardware and algorithms for able to slide past each other [53.189]. The CHOBIE several 2-D stochastically-driven self-assembling sys- . ]. ]. 215 217 , et al. in- 216 Whitesides tches oscillate radially et al. developed a localization ]. The system consists of identi- ]. ]. Once the shape is formed, mod- 211 ] that is capable of localizing tens- , 220 – 214 Funiak et al. performed a more theoretical , 212 210 218 et al. developed a 1000 modules hardware 213 Researchers are also developing algorithms for free- Miyashita Researchers have also explored the use of folding algorithm [53. 53.6.5 Self-Assembling Systems Self-assembling modular robotiction systems of are collec- modulescoalescing and that bonding are witha their greater capable neighbors structure. of toneed The form autonomously not result is be.assembling often is Whether robotic, independent a of buta system whether chain, it a it is lattice, is oraforementioned capable a free-form, modular truss-based robot of system. systems Almost self- rely allintervention of on to assemble. human In anprocess attempt of to creating automate the intricateresearchers modular have attempted robotic to mimic systems, andnatural improve upon self-assembling systems. vestigated a wide variety of engineeredsystems [53. self-assembling 53.6.4 Free-Form Systems Free-form systems are able toleast aggregate semi-arbitrary modules positions. in at OneSlimebot [53. such system iscal vertical the cylindrical modules that move ontal a plane. horizon- The perimetersix of gender-less hook each and loop module patches is usedneighboring to covered bond modules. with by These pa in and outling from the frequency the and phase center ofneighbors, the of the oscillations between system the can achieve body.a aggregate given By motion direction. in control- form systems. of-thousands of irregularlyRubenstein and packed Shen developed modules a numbermation in of shape algorithms 3-D for- for collectionsmodules. These of algorithms allow two-dimensional an arbitrary-sized col- lection of modules to formshapes arbitrary [53. scale-independent analysis of self-assembly using pie-shaped pieces ules can be added tothe or system removed from will thenew system, reconfigure modules. and The itself resulting shape tobut will its incorporate grow or the basic shrink, Rubenstein form will remainplatform on unchanged. which to deploy Recently, these algorithms [53. to create reconfigurable foldable systems [53. These systems useory flexible alloy wiring actuatorsprogrammatically embedded and in create shape composite origami-inspired mem- sheetscontrolling shapes. to which By actuatorscan are form energized, multiple different the shapes. system ] ] 209 208 et al. ]. , struc- ]. The 203 207 ]byusing 206 3-D Lipson 202 ) are based, at – 200 et al. [53. VIDEO 211 ] consisting of 45 mm cu- Lyder ]. A group of modules may 3-D [53. itchable magnets, each of 204 205 et al. is similar to Odin. It also system to actuation capabilities from the modules ’sLipson group has worked to move the Nagpal ]. To form specific shapes, each module is 199 One of the newest lattice-type modular robotics The Miche system [53. least in principle, on the Miche modules. is an aerial systemsingle-rotor modules composed [53. of identical,connect hexagonal, to form arangement flying of platform multiple with rotors. anto In arbitrary fly, addition each ar- to module the containsmay ability self-reconfigure wheels on so the ground that for theat the hand. system specific task 53.6.3 Truss Systems Truss systems, asrobotic systems their in name whichedges implies, the in modules a are arenectors truss modular nodes structure. may and Both belattice-based the systems, active truss-based trusses systems in and doto con- not such operate need on systems. any regulartems Unlike lattice. under Most the truss-based development sys- employcontract struts that to expand achieve or structuralfirst deformation. such One system of to the do so was Tetrobot [53. extended their 2-D tems [53. provided with a representation of thedecides, desired based shape on and itsto location allow in other modules the to structure, bond to whether its faces. tures. The Robot Pebbles ( to the tank in which the modules circulate [53. cubic modules suspended inself-assembly turbulent and fluid reconfiguration. As to the achieve freecirculate modules in the fluid,of they assembled pass modules. by When a theythey growing come are structure close accreted enough, onto the structure.or The modules repel attract each othersure. with Early fluid versions suction of orinterval the positive values system pres- that used couldMore modules redirect recently, with these suction forces. intelligence and Odin system, conceived by bic modules capable of mating withing their mechanically neighbors switchable us- permanent magnets.module Each contains threewhich mated with sw a steel faceBecause on a the neighboring module. connectors wereof gendered, modules had any to be collection assembled bynectors hand so were that the always con- orientedwas correctly, capable but the of system self-disassembling to form consists of three physically differentactive types strut of modules: modules capablepassive of strut changing modules their length; ofules. The fixed biologically inspired length; Morpho system and [53. joint mod- uses active links, passive links, and connector cubes. developed by Moving in the Environment

Part E

Part E | 53.6 1356 Multiple Mobile Robot Systems 53.7 Heterogeneity 1357

to form complete circles [53.221] from pie-shaped Werfel [53.231] also applied the idea of a transition rule 53.7 | E Part pieces. In the process, they followed Hosokawa et al.’s set when studying the use of swarms to assemble com- lead [53.222] and modeled the system as a chemical plex structures from passive materials. reaction. Shimizu and Suzuki developed a system of Other groups have attempted to make self- passive modules capable of self-repair when placed on assembly more deterministic. The MEMS (micro- a vibrating table [53.223]. electromechanical system) robots developed by Don- Computer scientists have also investigated theoreti- ald et al. [53.232, 233] consists of thin (720 m), cal aspects of self-assembly in the context of 2-D tiles rectangular ( 260 m  60 m), scratch-drive devices which selectively bond with their neighbors to form capable of moving on an insulating substrate embed- simple well-defined shapes like squares [53.224–226]. ded with electrodes. The authors used four of these Each side of every tile in the system has an associated robots to build larger composite structures. The Sitti bonding strength. When two tiles collide, they remain group has developed a similar system of micro-meter attached only if their cumulative bond strength exceeds sized robots [53.234]. Instead of using a scratch drive a globally defined system entropy. To form a specific for locomotion, the robots are manipulated by exter- shape, one must design a set of tiles with the appropri- nal magnetic fields. The authors can electrostatically ate bonding strengths. clamp any number of robots to the stage on which they Klavins et al. worked to develop intelligent self- move. With all but one robot immobilized, the remain- assembling systems that employ triangular modules ing robot may be moved independently. The system driven by oscillating fans on an air table to self- naturally self-assembles because the robots contain per- assemble different shapes [53.227]. The authors employ manent magnets that attract their neighbors. knowledge of the module’s local topology and in- The majority of existing self-assembly systems aim ternal module state so that each module decides, in to form structures in one of two ways. Some systems a distributed fashion, when to maintain or break a con- such as [53.221, 223–226] use a collection of applica- nection with its immediate neighbors. Griffith et al. also tion specific differentiated modules, that are only capa- worked with intelligent modules capable of selective ble of assembling in a particular fashion to form a spe- bonding to show that self-assembling systems may self- cific shape. In contrast, other systems such as [53.199– replicate [53.228]. 201, 227, 229–231, 235] use completely generic mod- Rus et al. [53.185] present the first generic rule- ules with more computation and communication ability based approach to self-assembly, shape formation, and embedded in each module. Both types of systems aim to locomotion by reconfiguration. The rules can be used form complex shapes in a direct manner: as these struc- on any modular robot system that can implement the tures grow from a single module, new modules are only sliding cube model of relative motion. The result is an allowed to attach to the structure in specific locations. abstract set of rules for each of these tasks, that can be An alternative approach eliminates many of the compiled down to module motions, taking into account complexities of shape formation by active assembly how the physical module implementes translation and by using dissasemly for shape formation. The Smart convex and concave transitions. Pebble system [53.154, 204, 236, 237] employs a set of Jones and Matari´c [53.229] presented rule-based distributed algorithms to perform two discrete steps: approach to self-assembly termed transition rule sets. 1) rely on stochastic forces to self-assemble a close- In particular, they present a method that, given a goal packed crystalline lattice of modules and 2) use the structure, produces a set of rules shared among all mod- process of self-disassembly to remove the extra material ules that govern when and where new modules are from this block leaving behind the goal structure. By allowed to attach to the growing structure. Kelly and approaching shape formation in this manner, the entire Zhang [53.230] expanded on this work by optimizing shape-formation process is sped-up and the robustness the size of the rule sets used to form a specific shape. of the system is increased.

53.7 Heterogeneity

Robot heterogeneity can be defined in terms of vari- through the use of homogeneous robots, which are ety in robot behavior, morphology, performance quality, completely interchangeable (i. e., the swarm approach, size, and cognition. In most large-scale multirobot sys- as described in Sect. 53.4). However, certain complex tems work, the benefits of parallelism, redundancy, and applications of large-scale robot teams may require solutions distributed in space and time are obtained the simultaneous use of multiple types of sensors and ] ). 240 53.14. Parker ] of 100 [53. VIDEO 206 (automated 244 e that could ]. This work ] demonstrated ] demonstrated 9 11 ]( Zelinsky 245 [53. et al. [53. 241 ] demonstrated a he- and ASyMTRe 242 has studied heterogeneity econnaissanc 53.2, one of the earliest et al. [53. Parker et al. [53. Jung ] developed modular milli- Howard 134 ] demonstrated assistive naviga- Murphy et al. [53. Simmons ] developed 243 246 Chaimowicz ). robot assists smaller robots in applications et al. [53. [53. Heterogeneous team of an air and two ground Sukhatme et al. [53. VIDEO 200 Tang As discussed in Sect. mothership demonstrated the abilityheterogeneity of robots in to teamtion compensate and for members execution. during task alloca- Fig. 53.14 vehicles that can performsurveillance cooperative reconnaissance and Grabowski in the context ofa marsupial robot deployment,such where as search and rescue [53. research demonstrations ofcal robot heterogeneity teams inLIANCE was in physi- architecture the by development of the AL- be composed ofcomponents, interchangeable thus sensor creating anderogeneous a teams. effector variety ofthe different use het- ofsembly heterogeneous and robots construction for taskscations. relevant autonomous to as- space appli- licopter robot cooperating withtasks involving two marsupial-inspired payload deployment ground robotsand in recovery, cooperative localization, andsance reconnais- and surveillance tasks,Parker as shown intion Fig. for sensorintelligent leader robot network for guiding deploymentlenged navigationally simple chal- sensor using robots to goal aa locations, larger as demonstration more by part of bots for surveillance and r robots performing exploration, mapping,and detection. deployment, a team ofsurveillance applications aerial in urban and environments. and ground robots cooperating for symbols in their language;derstanding developing of a communicated common symbolswith un- among different robots physicalchallenge, capabilities addressed is by a fundamental ( construc- ]. Techniques as 238 d: heterogeneity may be ssumes that robots have A particular challenge to ]. Most research in hetero- dual robot design, can typically deal with het- 239 [53. 53.6 Balch The motivation for developing heterogeneity in There are a variety of research challenges in hetero- A second compelling reason to study heterogeneity achieving efficient autonomous control is whenin overlap team member capabilities occurs,allocation thus or affecting role task assignments [53. robots, all of whichtype cannot of be robot. designed Somesmaller into robots sizes, a may which need single willtain to required limit sensors be may their be scaled too payloads,across to expensive to all or duplicate robots cer- onto the be large team. to Other carrysors, application-specific robots payload may or or sen- need toThese navigate applications, therefore, long require distances theof collaboration in large numbers a of heterogeneous limited robots. time. multirobot teams is thus twofol tion, and experience will inevitablysystem cause to a multirobot driftrecognized by to experienced roboticists, heterogeneitythat who have over several seen copies time.vary of widely This in the capabilities is same duetuning, to model calibration, differences etc. of in sensor differences Over robot among time, can robots even willrobot minor grow drift initial due and to wearto individual and employ tear. robot teams The effectively, we implicationdiversity, must predict is understand that, how itenable will robots to impact adapt performance, to thepeers. and diverse In capabilities of fact, their it isexplicitly often into advantageous the to design build of diversity a robot team. geneous multirobot systems. a design feature beneficial toheterogeneity particular may applications, be or a necessity.heterogeneity As can a offer economic design benefits, feature, be since easier it to can distributetiple varying team capabilities members across mul- of rather monolithic than robots. to Heterogeneity buildgineering can benefits, many also as copies offer it en- design may individual robots simply that incorporate be all ofing, too the computational, and difficult sens- effector requirements to of a given application. Heterogeneity in behavior may also arisean in emergent manner in physically homogeneousas teams, a result of behavior specialization. is that itpossible may in practice be to build a ateam. necessity, truly The realities homogeneous in robot of indivi that it is nearly im- a common language and a common understanding of erogeneous robots for theAnother purposes important topic of in task heterogeneity is allocation. ognize how and to quantify rec- heterogeneity inSome multirobot types teams. of heterogeneity cantively, be using evaluated metrics quantita- suchdeveloped as by the social entropygeneous metric multirobot systems a described in Sect. Moving in the Environment

Part E

Part E | 53.7 1358 Multiple Mobile Robot Systems 53.8 Task Allocation 1359

synthesis of multirobot task solutions through software application dictates certain constraints on the physi- 53.8 | E Part reconfiguration), which enables heterogeneous robots cal design of the robot team members. However, it is to share sensory resources to enable the team to ac- also clear that multiple choices may be made in de- complish tasks that would be impossible without tightly signing a solution to a given application, based upon coupled sensor sharing. cost, robot availability, ease of software design, flexi- Many open research issues remain to be solved in bility in robot use, and so forth. Designing an optimal heterogeneous multirobot teams; for example, the issue robot team for a given application requires significant of optimal team design is a very challenging problem. analysis and consideration of the tradeoffs in alternative Clearly, the required behavioral performance in a given strategies.

53.8 Task Allocation

In many multirobot applications, the mission of the team tion or coordination between subtasks, meaning that is defined as a set of tasks that must be completed. Each each task must be aware of the current state of the co- task can usually be worked on by a variety of differ- ordinated subtasks within a small time delay. As this ent robots; conversely, each robot can usually work on time delay becomes progressively larger, coordinated a variety of different tasks. In many applications, a task subtasks become more loosely coupled, representing is decomposed into independent subtasks [53.9], hier- weakly cooperative solutions. archical task trees [53.247], or roles [53.11, 245, 248, Robots can also be categorized as either single-task 249] either by a general autonomous planner or by the robots (ST), which work on only one task at a time or human designer. Independent subtasks or roles can be multitask robots (MT), which are able to make progress achieved concurrently, while subtasks in task trees are on more than one task at a time. Most commonly, task al- achieved according to their interdependence. Once the location problems assume robots are single-task robots, set of tasks or subtasks have been identified, the chal- since more capable robots that perform multiple tasks in lenge is to determine the preferred mapping of robots to parallel are still beyond the current state of the art. tasks (or subtasks). This is the task allocation problem. Tasks can either be assigned to optimize the in- The details of the task allocation problem can vary stantaneous allocation of tasks (IA), or to optimize the in many dimensions, such as the number of robots re- assignments into the future (TA, for time-extended as- quired per task, the number of tasks a robot can work signment). In the case of instantaneous assignment, on at a time, the coordination dependencies among no consideration is made for the effect of the cur- tasks, and the time frame for which task assignments rent assignment on future assignments. Time-extended are determined. Gerkey and Matari´c [53.250]defined assignments attempt to assign tasks so that the perfor- a taxonomy for task allocation that provides a way of mance of the team is optimized for the entire set of tasks distinguishing task allocation problems along these di- that may be required, not just the current set of tasks that mensions, which is referred to as the multirobot task need to be achieved at the current time step. allocation (MRTA) taxonomy. Using the MRTA taxonomy, triples of these abbre- viations are used to categorize various task allocation 53.8.1 Taxonomy for Task Allocation approaches, such as SR-ST-IA, which refers to an assignment problem in which single-robot tasks are as- Generally, tasks are considered to be of two principal signed once to single-task robots. Different variations types: single-robot tasks (SR, according to the MRTA of the task allocation problem have different computa- taxonomy) are those that require only one robot at tional complexities. The easiest variant is the ST-SR- a time, while multirobot tasks (MR) are those that re- IA problem, which can be solved in polynomial time quire more than one robot working on the same task since it is an instance of the optimal assignment prob- at the same time. Commonly, single-robot tasks that lem [53.251]. Other variants are much more difficult, have minimal task interdependencies are referred to as and do not have known polynomial time solutions. For loosely coupled tasks, representing a weakly coopera- example, the ST-MR-IA variant can be shown to be an tive solution. On the other hand, multirobot tasks are instance of the set partitioning problem [53.252], which often considered to be sets of subtasks that have strong is strongly NP-hard. The ST-MR-TA, MT-SR-IA,and interdependencies. These tasks are therefore often re- MT-SR-TA variants have also all been shown to be ferred to as tightly coupled tasks that require a strongly NP-hard problems. Because these problems are compu- cooperative solution. The subtasks of a loosely coupled tationally complex, most approaches to task allocation multirobot task require a high level of synchroniza- in multirobot teams generate approximate solutions. , and 247 prob- ]was ]) deal- d others ST 256 ] applies - an 260 Botelho ]. In the M+ – 263 ), MR ]. An example , ) [53. IA 255 258 problem variant, , 269 247 – CNP 11 SR - 264 , ], comparing alternative ], TraderBots [53. ]. The MURDOCH ap- nation in dynamic environ- problem variant is found 246 257 258 265 ]) addressing time-extended as- MT - utilities for performing particular 263 MR – ] employs a resource-centric, publish– ). More recent methods are beginning to 261 , 258 TA ]. 135 ’s contract net protocol ( with their M+ architecture [53. 270 Some representative market-based techniques in- ], and Hoplites [53. The TraderBots approach [53. Smith Since these early developments, many alternative Most of the current approaches in market-based (e.g., [53. signments ( address the coaltionallocation formation of problem, multirobot which tasks is (i. e., the the lem variant), including [53. approach to the in [53. clude MURDOCH [53. 263 proach [53. subscribe communication model to carrywhich has out the advantage auctions, of anonymous communication. In this approach aresources, such task as the is environmental sensors. The represented meth- ods for by how to the use such required results a sensor is to preprogrammed generate into satisfactory the robot. based on their capabilitiestiation and process availability. The isthe nego- team based seeks on toupon optimize market individual an robot objective theory, function in based tasks. which The approachestasks typically to greedily the assignhighest robot utility. sub- that can perform the task with the market economy techniquesand robust for multirobot coordi generatingments. In efficient a market economy, robots actinterests. A based robot on receives selfish revenue and incurstrying cost when to accomplish a task. The goal is for robots to the first to addressgotiate the problem to of how collectivelyof agents solve a can market-based ne- a approachtask specifically set for of allocation multirobot tasks. wasAlami The first use developedapproach, robots by plan their own individualtask plans for they the have been assigned.other They teammates then to negotiate incrementally adapt with theirsuit the actions to team as athat whole, through facilitate the the use merging of of social plans. rules approaches to market-based taskdeveloped. allocation A have thorough been surveythe on art the in current market-basedallocation state techniques is of for given multirobot task in [53. approaches in termsdynamic events of and environments, solutionteams. and heterogeneous quality, scalability, task allocation addresswith the some approaches ST (e.g., [53. ing with instantaneous assignment ( ] 9 vari- ]. AL- variant TA 10 - IA - ] for a com- SR - SR - 250 ST ’s M+ architec- ST ] that allows robots and Alami 254 IA - and ts without explicitly dis- SR - tion in multirobot teams can uses a subsumption style be- ST uses an assignment algorithm that BLE Botelho BLE ], which addresses the ]. , ALLIANCE achieves adaptive action se- 253 255 53.2.2 ) [53. Behavior-Based Task Allocation One of the earliest architectures for multirobot task Another behavior-based approach to multirobot Market-Based Task Allocation BLE Approaches to task alloca 53.8.2 Representative Approaches be roughly divided into behavior-based approachesmarket-based and (sometimes calledauction-based) approaches. negotiation-style The following or describe subsections some representative architecturesthese for general approaches. each Refer of parative to analysis [53. of some ofof these computation approaches, and in communications terms requirementssolution and quality. Behavior-based approaches typically enabledetermine robots to taskcussing assignmen individual tasks. In these approaches, robots use knowledge of the currentsion, state robot of team member the capabilities, robot andto robot team decide, actions mis- in aperform distributed which fashion, task. which robot should allocation that waswas demonstrated the behavior-based ALLIANCE on architecture [53. physicaland the robots related L-ALLIANCE architecture [53. LIANCE addresses the ants of the task allocation problem without explicitmunication com- among robots about tasks.Sect. As described in lection through the use of motivational behaviors, which are levels of impatiencerobot and that determine acquiescence its within own and each fitness its teammates’ for relative performing certainare tasks. calculated based These upon motivations the mission requirements,activities the and capabilities of teammates, andinternal the states. robots’ These motivationsutility effectively measures for calculate each robot–task pair. task allocation( is broadcast of local eligibility havior control architecture [53. of task allocation. to efficiently execute tasks bying locally continuously computed broadcast- eligibilities androbot only selecting with the the bestthis eligibility case, to task performior allocation inhibition. the is task. achieved In through behav- is very similar to Market-based (or negotiation-based) approachescally typi- involve explicit communicationsabout between the robots required tasks, in which robots bid for tasks ture [53. Moving in the Environment

Part E

Part E | 53.8 1360 Multiple Mobile Robot Systems 53.9 Learning 1361

trade tasks through auctions/negotiations such that the change. Strategies are predefined for a robot to accom- 53.9 | E Part team profit (revenue minus cost) is optimized. plish a selected plan. The Hoplites approach [53.265] focuses on the se- Some alternative approaches formulate the objects lection of an appropriate joint plan for the team to to be assigned as roles, which typically package a set execute by incorporating joint revenue and cost into of tasks and/or behaviors that a robot should undertake the bid. This approach couples planning with passive when acting in a particular role. Roles can then be dy- and active coordination strategies, enabling robots to namically assigned to robots in a similar manner as in change coordination strategies as the needs of the task the auction-based approaches [53.11, 248].

53.9 Learning

Multirobot learning is the problem of learning new not explicitly share their intentions. Two different vari- cooperative behaviors, or learning in the presence of ations of the credit assignment problem are common in other robots. The other robots in the environment, how- multirobot learning. The first is when robots are learn- ever, have their own goals and may be learning in ing individual behaviors in the presence of other robots parallel [53.271]. The challenge is that having other that can affect their performance. The second is when robots in the environment violates the Markov prop- robots are attempting to learn a task with a shared fit- erty that is a fundamental assumption of single-robot ness function. It can be difficult to determine how to learning approaches [53.271]. The multirobot learning decompose the fitness function to appropriately reward problem is particularly challenging because it combines or penalize the contributions of individual robots. the difficulties of single-robot learning with multiagent While learning has been explored extensively in learning. Particular difficulties that must be consid- the area of single-robot systems (see, for example, ered in multirobot learning include continuous state the discussion of learning in behavior-based systems and action spaces, exponential state spaces, distributed in Chap. 13, and a discussion of fundamental learn- credit assignment, limited training time and insufficient ing techniques in Chap. 15) and in multiagent sys- training data, uncertainty in sensing and shared infor- tems [53.280], much less work has been done in the area mation, nondeterministic actions, difficulty in defining of multirobot learning, although the topic is gaining in- appropriate abstractions for learned information, and creased interest. Much of the work to date has focused difficulty of merging information learned from differ- on reinforcement learning approaches. Some examples ent robot experiences. of this multirobot learning research include the work by The types of applications that have been studied Asada et al. [53.281], who propose a method for learn- for multirobot learning include multitarget observa- ing new behaviors by coordinating previously learned tion [53.272, 273], air fleet control [53.274], predator– behaviors using Q-learning, and apply it to soccer- prey [53.137, 275, 276], box pushing [53.277], forag- playing robots. Matari´c [53.8] introduces a method ing [53.23], and multirobot soccer [53.140, 278]. Partic- for combining basic behaviors into higher-level be- ularly challenging domains for multirobot learning are haviors through the use of unsupervised reinforcement those tasks that are inherently cooperative. Inherently learning, heterogeneous reward functions, and progress cooperative tasks are those that cannot be decomposed estimators. This mechanism was applied to a team into independent subtasks to be solved by individual of robots learning to perform a foraging task. Kubo robots. Instead, the utility of the action of one robot is and Kakazu [53.282] proposed another reinforcement dependent upon the current actions of the other team learning mechanism that uses a progress value for members. This type of task is a particular challenge in determining reinforcement, and applied it to simu- multirobot learning, due to the difficulty of assigning lated ant colonies competing for food. Fernandez and credit for the individual actions of the robot team mem- Parker [53.272] apply a reinforcement learning algo- bers. rithm that combines supervised function approximation The credit assignment problem is a particular chal- with generalization methods based on state-space dis- lenge, since it is difficult for a robot to determine cretization, and apply it to robots learning the multi- whether the fitness (either good or bad) is due to its own object tracking problem. Bowling and Veloso [53.271] actions, or due to the actions of another robot. As dis- developed a general-purpose, scalable learning algo- cussed by Pugh and Martinoli in [53.279], this problem rithm called GraWoLF (gradient-based win or learn can be especially difficult in situations where robots do fast), which combines gradient-based policy learning ). ], ]. 120 131 VIDEO 293 ]. 294 , – 289 , 132 VIDEO 217 d demining. Additionally, , ], forming chains [53. ] apply particle swarm op- 23 14 279 multirobot systems because of domain has similar issues to the [53. coverage Martinoli and In foraging and coverage applications, a fundamen- 53.10.2 Flocking and Formations Coordinating the motionsother of has robots beenrobot relative a systems since topic to the inception of of each thelar, field. interest much In attention particu- has in been paid multiplemation to control the mobile problems flocking ( and for- The flocking problemof could the be formation viewed controlmove as problem, together a requiring along robots subcase somewith to only path minimal requirements in for paths thecific taken robots. by aggregate, spe- Formations but are stricter, requiring robots to foraging application. In coverage,to robots visit are all required areasing of for their objects environment, (such perhapsaction as search- in landmines) or all executingcleaning). parts some The of coverage application the alsorelevance has environment to real-world tasks (e.g., such as for demining, lawnmental care, floor environ- mapping, and agriculture. tal question is how toenvironments enable quickly the without duplicating robots actions to or explore in- terfering their with each other.clude Alternative basic strategies stigmergy [53. can in- timization techniques to distributed unsupervised robot learning in groups,avoidance. for the task of learning obstacle or simulated food pellets arenar distributed terrain, across the and pla- objects robots and are delivering tasked themlocations, with to such collecting as one a the or homethe more base. Foraging study gathering lends of itself weakly to the cooperative actions robot of systems, individual robots in do not that synchronized have to with be each tightly other.ally This been task of has interestits tradition- in close analogy to the biologicalswarm systems robotics that research. motivate However, it alsoto has relevance several real-world applications, suchcleanup, search as and rescue, toxic an waste since foraging usually requires robots to completelyplore ex- their terrain ininterest, the order to discover the objects of and making use of heterogeneous robots [53. statistical experience data, to learnent the fitness heterogeneous of robots differ- inPugh performing a set of tasks. Other research demonstrated in the foraging and/or cov- erage domain includes [53. ], ]for 284 ], secu- 288 9 ] and [53. 2 mentions some of the ’s L-ALLIANCE archi- approaches not based on le robot systems and the 46 ile robot systems is often ). Multiple robot systems le mobile robot systems. ], mineral mining, transporta- Parker 241 VIDEO 210 ], agriculture, and warehouse manage- ], extraplanetary exploration [53. ], which uses parameter tuning, based on ]( 286 283 , ], hazardous waste cleanup [53. 10 287 11 285 Research in multiple mob Other multirobot learning are also usedping, and in exploration; Chap. thework in domain multirobot systems of appliedParts to F localization, these and problems. map- cation G areas of that this aresystems, handbook relevant but not outline only also manyTo to appli- to date, single-robot relatively multip fewthese real-world multirobot systems implementations have of occurred,to primarily the due complexities ofrelative multip newness of theertheless, supporting technologies. many Nev- proof-of-principlephysical demonstrations multirobot of systems havethe expectation been is achieved, that these and into systems practical will find implementations their as way tinues the to technology mature. con- Foraging is a popular testingsystems, application particularly for for multirobot those approachesswarm that robotics, involving very address large numbers of mobile robots. In the foraging domain, objects such as pucks 53.10.1 Foraging and Coverage 53.10 Applications Many real-world applicationsfit can from potentially theExample bene- use applications of includein multiple container ports [53. management mobilesearch robot and rescue systems. [53. techniques with a variablestrated learning the results rate, in andapplication. the demon- adversarial multirobot soccer a discussion of these domainsing and of related a research. more detailed list- explored in the context ofmains. common While application not test yet do- elevated totasks, the level these of benchmark commonties domains for do provide researcherstive opportuni- to strategies to compare multirobot control. andthough Additionally, even contrast these alterna- commonlaboratory test experiments, they do domains have relevance areworld to applications. real- usually This section outlines just theseapplication common domains; see also [53. tion, industrial and householdtion maintenance, [53. construc- rity [53. reinforcement include tecture [53. ment [53. Moving in the Environment

Part E

Part E | 53.10 1362 Multiple Mobile Robot Systems 53.10 Applications 1363

maintain certain relative positions as they move through push alone. Sometimes there are several boxes to be 53.10 | E Part the environment. In these problems, robots are assumed moved, with ordering dependencies constraining the to have only minimal sensing, computation, effector, sequence of motions. Cooperative manipulation is sim- and communications capabilities. A key question in ilar, except it requires robots to lift and carry objects both flocking and formation control research is deter- to a destination. This test bed domain lends itself to mining the design of local control laws for each robot the study of strongly cooperative multirobot strategies, that generate the desired emergent collective behavior. since robots often have to synchronize their actions to Other issues include how robots cooperatively localize successfully execute these tasks. The domain of box themselves to achieve formation control [53.133, 295], pushing and cooperative manipulation is also popular and how paths can be planned for permutation-invariant because it has relevance to several real-world appli- multirobot formations [53.296]. cations [53.288], including warehouse stocking, truck Early solutions to the flocking problem in artifi- loading and unloading, transporting large objects in cial agents were generated by Reynolds [53.297]using industrial environments, and assembly of large-scale a rule-based approach. Similar behavior- or rule-based structures. approaches have been used physical robot demonstra- Researchers usually emphasize different aspects of tions and studies, such as in [53.121, 298]. These earlier their cooperative control approach in the box push- solutions were based on human-generated local con- ing and cooperative manipulation domain. For example, trol rules that were demonstrated to work in practice. Kube and Zhang [53.13]( VIDEO 199 ) demonstrate More recent work is based on control theoretic princi- how swarm-type cooperative control techniques could ples, with a focus on proving stability and convergence achieve box pushing (Fig. 53.15), Parker [53.10, 316] properties in multirobot team behaviors. Examples of illustrates aspects of adaptive task allocation and learn- this work include [53.128, 299–307]. Refer to [53.308, ing, Donald et al. [53.317]( VIDEO 208 ) illustrates 309] for surveys of relevant control theoretic work. concepts of information invariance and the interchange- ability of sensing, communication, and control, and 53.10.3 Object Transportation Simmons et al. [53.11] demonstrate the feasibility of co- and Cooperative Manipulation operative control for building planetary habitats. In general, the manipulation techniques used for Some of the earliest work in swarm robotics was aimed collective object transportation can be grouped into at the object transportation task [53.13, 123, 310–313], three primary methods [53.318]: pushing, grasping, and which requires a team of robots to move an object caging. The pushing approach [53.10, 11, 13, 316, 317] from its current position in the environment to some requires contact between each robot and the object, in goal destination ( VIDEO 193 ). The primary benefit order to impart force in the goal direction; however, of using collective robots for this task is that the indi- the robots are not physically connected with the ob- vidual robots can combine forces to move objects that ject. In the grasping approach [53.123, 142, 310–312, are too heavy for individual robots working alone or in 319–322], each robot in the team physically attaches to small teams. However, the task is not without its chal- the object being transported. See for example Fig. 53.16 lenges; it is non-trivial to design decentralized robot control algorithms that can effectively coordinate robot team members during object transportation. A further complication is that the interaction dynamics of the robots with the object can be sensitive to certain object geometries [53.314, 315] and object rotations during transportation [53.315], thus exacerbating the control problem. Object transportation and cooperative manipula- tion are popular domains for demonstrating multirobot cooperation, because they offer a clear domain where close coordination and cooperation is required. A com- mon type of object transportation – box pushing – requires robot teams to move boxes from their start- ing positions to defined goal configurations, sometimes along specified paths. Typically, box pushing operates in the plane, and the assumption is made that the boxes are too heavy or too long to enable single robots to Fig. 53.15 Collective pushing of lighted box ]. In Parker ]. 331 ], which 344 336 [53. – 338 ] proposes a dis- , Murata 334 323 , , ], in which robots use et al. in [53. and ]. Early ideas of this con- 333 310 332 , 335 [53. 284 Brooks , Terada 258 et al. [53. , Russell 123 , ], further develop this idea by proposing 6 and , 3 337 Wawerla Hardware challenges of collective robot construc- Another type of construction is called blind bull- A significant body of additional research has been Stewart 53.10.4 Multitarget Observation The domain oftiple multitarget robots observation to requiresgets mul- monitor moving and/or throughis the observe environment. to multiple The tar- maximizehood, objective that the the amountmember targets throughout of remain task in time,especially execution. view or challenging The by task if the somerobots. This can there team likeli- application domain be are can be useful moreing for study- targets strongly than cooperativehave task to solutions, coordinate sincetargets their robots to motions follow or inIn the order the switching to of context maximizethe of their planar objective. multiple version mobile of robot this test applications, bed was first introduced structure. The goal of their worksome is number of to robots be and able freetion to blocks zone, deploy along into with a a construc- singlefor block the that structure, serves and as then aceed seed have the autonomously according construction to to pro- theof provided blueprint the desired structure. tion are addressed by this work, a hardwarepassive design building blocks, is along proposed withthat an that constructs assembler defines structures robot with thework robots. on Other the related topic of collective constructionwork includes the of a behavior-based approach toblocks build equipped a with either linear positive or wall negativedistinguished Velcro, using by blockadding color. Their 1 results bit of showcolor state that of information the to lastimprovement communicate attached in the block the provides collectiveby a performance. significant The work tributed approach to building a loosea wall robot structure swarm. with dozing, which iscertain inspired ant by colonies. a Rathermulating than behavior materials, constructing this observed by approach in achieves accu- constructionby removing materials.cation This in task site hasplanetary practical clearing, exploration [53. appli- such ascept were would discussed be by needed for argues for large numbersered of to small the robotset to lunar al. be surface [53. deliv- for site preparation. robots using force sensorsmaterial to to clear the an edges area of by the work pushing site. illustrated in this domain;clude [53. representative examples in- ] )in- 329 – 2-D or 327 ]. Finally, VIDEO 292 320 2-D [53. approach, the sys- ]( 3-D 326 – have extensively explored 323 ), developing distributed algo- Nagpal ]. ]. In their 323 and 330 VIDEO 216 for an example of collective transport via Collective transport via caging Werfel 3-D [53. 53.17 . This task is distinguished from self-reconfigurable A closely related task is that of collective construc- Cooperative stick pulling volves robots encircling themoves in object the so desired thatstant direction, the even contact object without of theFig. con- all thecaging, robots from [53. with the object. See tion and wall building.construction The and objective of wallbuild the building collective structures task of is3-D a for specified form, robotsrobots, in to whose bodies either themselves servestructure. as the dynamic this topic ( rithms that enable simplified robotsbased on to provided blueprints, build both structures in for cooperative stick pulling work of [53. Fig. 53.17 the caging approach [53. tem consists ofthe idealized construction, and mobile smart blocks robots thatsive that serve structure. as The perform the robots’ pas- job is towhile provide the the mobility, blocks’ role ising to structure at identify which places an inthat additional the block is grow- can be on placed the path toward obtaining the desired final and in Moving in the Environment

Part E Fig. 53.16

Part E | 53.10 1364 Multiple Mobile Robot Systems 53.10 Applications 1365

in [53.345] as cooperative multirobot observation of 53.10 | E Part multiple moving targets (CMOMMT). Similar prob- lems have been studied by several researchers, and extended to more complex problems such as environ- ments with complex topography or three-dimensional versions for multiple aerial vehicle applications. This domain is also related to problems in other areas, such as art gallery algorithms, pursuit evasion, and sen- sor coverage. This domain has practical application in many security, surveillance, and reconnaissance prob- lems. Research applied to the multitarget observation problem in multirobot systems includes [53.138, 253, 346–353]. Fig. 53.18 Legged robot teams competing in robot soccer 53.10.5 Traffic Control and Multirobot Path Planning for the problem at hand. In these cases, approxima- tion approaches may be sufficient, such as centralized When multiple robots are operating in a shared envi- techniques that limit the search space through roadmap- ronment, they must coordinate their actions to prevent ping [53.361, 362], and decoupled approaches that use interference. These problems typically arise when the either prioritized planning [53.363–365] (i. e., generat- space in which robots operate contains bottlenecks, ing robot paths one by one) or path coordination (i. e., such as networks of roadways, or when the robots take first planning individual paths for robots, then handling up a relatively large portion of the navigable space. collision avoidance). In these problems, the open space can be viewed as a resource that robots must share as efficiently as possi- 53.10.6 Soccer ble, avoiding collisions and deadlocks. In this domain, robots usually have their own individual goals, and must Since the inception of the RoboCup multirobot soc- work with other robots to ensure that they receive use of cer domain as a proposed challenge problem for the shared space to the extent needed to achieve their studying coordination and control in multirobot sys- goals. In some variants, the entire paths of multiple tems [53.366], research in this domain has grown robots need to be coordinated with each other; in other tremendously. This domain incorporates many chal- variants, robots must simply avoid interfering with each lenging aspects of multirobot control, including col- other. laboration, robot control architectures, strategy acqui- A variety of techniques have been introduced to ad- sition, real-time reasoning and action, sensor fusion, dress this problem, including traffic rules, subdividing dealing with adversarial environments, cognitive mod- the environment into single-ownership sections, and ge- eling, and learning. Annual competitions show the ometric path planning ([53.354] for an overview). Many ever-improving team capabilities of the robots in a var- of the earliest research approaches to this problem were iety of settings, as shown in Fig. 53.18. A key aspect based on heuristic approaches, such as predefining mo- of this domain that is not present in the other mul- tion control (or traffic) rules that were shown to prevent tirobot test domains is that robots must operate in deadlock [53.355–358], or using techniques similar to adversarial environments. This domain is also pop- mutual exclusion in distributed computing [53.359, ular because of its educational benefits, as it brings 360]. These approaches have the benefit of minimiz- together students and researchers from across the ing the planning cost for obtaining a solution. Other, world in competitions to win the RoboCup chal- more formal, techniques view the application as a ge- lenges. The RoboCup competitions have added an ometric multirobot path planning problem that can be additional search-and-rescue category to the competi- solved precisely in configuration space–time. Chapter 7 tion [53.367], which has also become a significant area includes a discussion of motion planning for multiple of research (Chap. 66 for more details on this field). robots relevant to this domain. While geometric motion Annual proceedings of the RoboCup competitions planning approaches provide the most general solu- document much of the research that is incorporated tions, they can often be too computationally intensive into the multirobot soccer teams. Some representative for practical application, impractical due to the dy- research works include [53.368–372]( VIDEO 202 , namic nature of the environment, or simply unnecessary VIDEO 209 ). , ]. ]. ]. 55 gen- , 378 388 391 , – – 52 , 288 379 389 46 , , ]. Addition- 26 35 , 374 , open questions , 31 ]. Some taxonomies 373 , 377 – 288 , 2 375 , heterogeneity 1 (distributed autonomous robotic DARS include enabling the robot team to reason ]. 104 , For further reading on the topic of multiple mobile 90 requires basic researchperception, at and the communication. intersection Openeralization of issues control, in about context and increasingso the that versatility of they systems cancations. operate In in a dealing variety with include of determining different theoretical appli- approachesing to system performance predict- when alland robots determining are how not to equal, designfor a a robot given team application. optimally have Advances provided over the human lastact users decade with with hundreds theInternet. or ability thousands It to of is inter- centric computers necessary approaches on to to interfacing, the bothfor develop for monitoring. similar control Finally, and network- asystems major that are challenge proactive is andcommands anticipate to our rather create needs than and reactingcommands. (with delays) to human robot systems, the readerin is the referred to field, survey includingally, articles [53. several special journalappeared, including [53. issues on this topic have of multirobot systemsA are variety given of in symposiaon and [53. a workshops regular have basis beenin on held particular the the topicsystems) of multirobot series systems, ofthese symposia. workshops and Recent symposia proceedings include [53. of Additional texts on this topic include [53. 75, For some excellent furtherrobotics background we on direct networked the reader to [53. ,an , multi- software forpractical multi-robot applications integration robustness is still a challeng- ing architectures, com- http://handbookofrobotics.org/view-chapter/53/videodetails/199 http://handbookofrobotics.org/view-chapter/53/videodetails/192 http://handbookofrobotics.org/view-chapter/53/videodetails/193 http://handbookofrobotics.org/view-chapter/53/videodetails/194 http://handbookofrobotics.org/view-chapter/53/videodetails/195 http://handbookofrobotics.org/view-chapter/53/videodetails/196 http://handbookofrobotics.org/view-chapter/53/videodetails/197 http://handbookofrobotics.org/view-chapter/53/videodetails/198 Scalability in multirobot teams is still in its Multi-robot box pushing available from available from Self-assembly and morphology control inavailable a from swarm-bot CKBOTS reconfigurable robots available from Biologically inspired multi vehiclesavailable control from algorithm Metamorphic robotic system available from Agents at play: Off-the-shelf Handling of a single objectavailable by from multiple mobile robots basedSynchronization on and caster-like fault dynamics detection inavailable autonomous from robots learning VIDEO 199 VIDEO 195 VIDEO 196 VIDEO 197 VIDEO 198 VIDEO 192 VIDEO 193 VIDEO 194 For example, in the area of system munications issues, swarmneous teams, robot task allocation, learning, systems, and applications. Clearly, heteroge- significant advancesfield have in been the last madeof decade. research, in The however, field the since is manystill still open remain an research to active issues be area remain solved. in Key the open broad areas researchness, of questions system learning, integration, scalability, robust- generalization,with heterogeneity. and dealing This chapter hasin surveyed the multirobot current systems, examin state of the art 53.11 Conclusions and Further Reading ing problem, in termsas of well more as complex environments yet ever-larger numbers have of a robots.robot methodology We networks for do that creating not bering), are with self-organizing robust completely decentralized toestimators, controllers and labeling with and provable (or emergent num- response. This robot teams stilldegrade need gracefully, improvements in toto the reason achieve ability for complexity to without fault escalatingThe tolerance, failure area rates. and of infancy, with open questions includingcontinual how learning to in multirobot achieve teams, howthe to facilitate use ofable complex humans to representations, influence and and/orof understand how the the results to team en- learning. open question is howto to combine effectively a allow spectrum robot ofcomplete teams approaches systems toward that achieving can performset more than of a limited tasks.bust We robot are networksin a that the can long perform real way physical world. from tasks In creating the ro- area of Video-References Moving in the Environment

Part E

Part E | 53.11 1366 Multiple Mobile Robot Systems References 1367

VIDEO 200

Elements of cooperative behavior in autonomous mobile robots 53 | E Part available from http://handbookofrobotics.org/view-chapter/53/videodetails/200 VIDEO 201 Coordination of multiple mobile platforms for manipulation and transportation available from http://handbookofrobotics.org/view-chapter/53/videodetails/201 VIDEO 202 Robots in games and competition available from http://handbookofrobotics.org/view-chapter/53/videodetails/202 VIDEO 203 A robotic reconnaissance and surveillance team available from http://handbookofrobotics.org/view-chapter/53/videodetails/203 VIDEO 204 MARS (multiple autonomous robots) available from http://handbookofrobotics.org/view-chapter/53/videodetails/204 VIDEO 205 Amethodfortransportingateamofminiaturerobots available from http://handbookofrobotics.org/view-chapter/53/videodetails/205 VIDEO 206 Reconfigurable multi-agents with distributed sensing for robust mobile robots available from http://handbookofrobotics.org/view-chapter/53/videodetails/206 VIDEO 207 Miniature air vehicle cooperative timing missions available from http://handbookofrobotics.org/view-chapter/53/videodetails/207 VIDEO 208 Distributed manipulation with mobile robots available from http://handbookofrobotics.org/view-chapter/53/videodetails/208 VIDEO 209 Autonomous robot soccer – Through the wormhole with Morgan Freeman available from http://handbookofrobotics.org/view-chapter/53/videodetails/209 VIDEO 210 A day in the life of a Kiva robot available from http://handbookofrobotics.org/view-chapter/53/videodetails/210 VIDEO 211 Robot Pebbles – MIT developing self-sculpting smart sand robots available from http://handbookofrobotics.org/view-chapter/53/videodetails/211 VIDEO 212 Transport of a child by swarm-bots available from http://handbookofrobotics.org/view-chapter/53/videodetails/212 VIDEO 213 Towards a swarm of nano quadrotors available from http://handbookofrobotics.org/view-chapter/53/videodetails/213 VIDEO 214 Swarm robotics at CU-Boulder available from http://handbookofrobotics.org/view-chapter/53/videodetails/214 VIDEO 215 Swarm robot system available from http://handbookofrobotics.org/view-chapter/53/videodetails/215 VIDEO 216 Swarm construction robots available from http://handbookofrobotics.org/view-chapter/53/videodetails/216 VIDEO 217 Multi robot formation control – Khepera team available from http://handbookofrobotics.org/view-chapter/53/videodetails/217 VIDEO 292 Experiments of escorting a target available from http://handbookofrobotics.org/view-chapter/53/videodetails/292 VIDEO 293 Formation control via a distributed controller-observer available from http://handbookofrobotics.org/view-chapter/53/videodetails/293

References

53.1 T. Arai, E. Pagello, L.E. Parker: Editorial: Advances 53.6 M.J. Matarić: Issues and approaches in the design in multi-robot systems, IEEE Trans. Robotics Au- of collective autonomous agents, Robotics Auton. tom. 18(5), 655–661 (2002) Syst. 16,321–331(1995) 53.2 Y. Cao, A. Fukunaga, A. Kahng: Cooperative mo- 53.7 E. Ostergaard, G.S. Sukhatme, M.J. Matarić: Emer- bile robotics: Antecedents and directions, Auton. gent bucket brigading, 5th Int. Conf. Auton. Robots 4, 1–23 (1997) Agents (2001) 53.3 R.G. Brown, J.S. Jennings: A pusher/steerer model 53.8 M. Matarić: Reinforcement learning in the multi- for strongly cooperative mobile robot manipula- robot domain, Auton. Robots 4, 73–83 (1997) tion, IEEE/RSJ Int. Conf. Intell. Robots Syst. (1995) 53.9 L.E. Parker: ALLIANCE: An architecture for fault- pp. 562–568 tolerant multi-robot cooperation, IEEE Trans. 53.4 D. Milutinović, P. Lima: Modeling and optimal Robotics Autom. 14(2), 220–240 (1998) centralized control of a large-size robotic pop- 53.10 L.E. Parker: Lifelong adaptation in heterogeneous ulation, IEEE Trans. Robotics 22(6), 1280–1285 teams: Response to continual variation in in- (2006) dividual robot performance, Auton. Robots 8(3), 53.5 B. Khoshnevis, G.A. Bekey: Centralized sensing 239–267 (2000) and control of multiple mobile robots, Comput. 53.11 R. Simmons, S. Singh, D. Hershberger, J. Ramos, Ind. Eng. 35(3/4), 503–506 (1998) T. Smith: First results in the coordination of het- Am- multiple (National , 1567–1584 Information 357 , ed. by C. Detrain, rithms for ourg (Birkhauser, Basel idberg, S.G. Chappell: (Center for Embedded Net- , ed. by W. Weber, J.M. Rabaey, ,202 296–305 (2002) Embedded, Everywhere A global array of 3000 Free-Drifting (22), 30–34 (2002) 39 Networked Minirhizotron Planning and (1), 48–74 (2006) (2), 205–215 (2006) ocessing in Social Insects Academies Press, Washington 2001) fish schools: examinationAn erties, Biol. of emergent Bull. prop- D. Sumpter: Informationand collective flow, intelligence opinion incial house-hunting insects, polling so- Philos. Trans. R.(2002) Soc. B formation flow in socialPr wasps. In: by A. Schultz,Dordrecht L.E. 2003) Parker, F. Schneider (Kluwer, Spectrum Initial Deployment workedsensing, Los Angeles 2007) H. Henriksen, F.F. Hover, D.ham, Yoerger, J.G. B.A. Moran: Belling- An integrated approachtiple to mul- AUV communications, navigation anding, dock- MTS/IEEE Conf. Proc. OCEANS (1996) pp. 59–64 D. Rus: Data collection, storage, andan retrieval with underwater sensor network,Syst. Proc. (2005) ACM Sens. D.M. Fratantoni, R. Davis: Collectivesor motion, sen- networks95 and ocean sampling, Proc. IEEE S.S. Mupparapu, D.R.Adaptive Bl samplingautonomous algo underwaterAUV2004: A Workshop vehicles,Veh. Oper. Multiple (2004) Auton. IEEE/OES Underw. M.A. Hsieh,Deploying J.F. Keller, air-groundurban V. multi-robot Kumar, environments,Workshop C.J. teams (2005) 3rd Taylor: in Multi-Robot Syst. Profiling Floats(Argo for Information Center, Ramonville 2007) Environmental Monitoring B. Stauffer, D.A.opment Caron: of The a wireless designmicrobial robotic observing networked devel- and aquatic system,24 Environ. Eng. Sci. bient Intelligence E. Aarts (Springer, Berlin, Heidelberg 2005) J.M. Pasteels, J.L. Deneub 1999) robot system, IEEE Int.Comm. Symp. (2006) Intell. Signal Proc. J. Villasenor, D. Estrin: Networked infomechanical systems (NIMS) for ambient intelligence. In: 53.32 J. Parrish, S. Viscido, D. Grünbaum: Self-organized 53.33 N. Franks, S. Pratt, E.53.34 Mallon, N. R.L. Britton, Jeanne: Group size, productivity, and in- 53.30 M. Yim,53.31 Y. Zhang, D. Duff: Modular D. robots, IEEE Estrin: 53.40 H. Singh, J. Catipovic, R. Eastwood, L. Freitag, 53.41 I. Vasilescu, M. Dunbabin, P. Corke,53.42 K. Kotay, N. Leonard, D. Paley, F. Lekien,53.43 R. Sepulchre, D.O. Popa, A.S. Sanderson, R.J. Komerska, 53.44 L. Chaimowicz, A. Cowley, B. Grocholsky, 53.37 G.S. Sukhatme, A. Dhariwal, B. Zhang, C. Oberg, 53.39 Amarss: 53.35 T. Akimoto, N. Hagita: Introduction to53.36 a network Argo Floats: 53.38 W. Kaiser, G. Pottie, M. Srivastava, G.S. Sukhatme, , 3 (2), 1(1), 4 ,ed. (2), 189–219 (2), 195–240 2 75 (3), 39–45 (2002) 45 (5), 796–812 (2002) (2/3), 323–336 (1997) 18 19 110(2), 293–320 (1999) Multi-Robot Systems Vol- , 161–188 (1993) 2 (11), 755–764 (1996) 13 7, 83–124 (1997) , 155–167 (1997) robot awareness inIEEE/RSJ cooperative Int. roboticpp. teams, Conf. 212–219 Intell. Robots Syst.from navigation, (1995) learning, andJ. Exp. group Theor. behavior, Artif. Intell. ogy and the evolutioncation, of Adapt. cooperative Behav. communi- erogeneous robots for large-scale assembly, Proc. ISER 7th Int. Symp. Exp. Robotics (2000) P. Dario: Self-organizing collection andof transport objects in unpredictableUSA environments, Symp. Flex. Jpn.- Autom. (1990) pp. 1093–1098 cial insects to robots, Adapt.(1993) Behav. actions to global tasks:robotics, Stigmergy Proc. and 14th collective Living Int. Syst. Workshop (1994) Synth. pp. Simul. 181–189 laws for multi-agent95 mobilization, Artif. Intell. A minimal behavior-basedrobot teams, approach Artif. Intell. to multi- 122–127 (1992) solving ining industrial joint multi-agent intentions,(1995) systems Artif. Intell. us- tell. Res. rić: LOST: Localization-space trails for robot teams, IEEE Trans. Robotics Autom. multiagent robotic systems,27–52 Auton. (1995) Robots onomy for multi-agent robotics, Auton. Robots to action duringtion, Proc. observation-based 1st Int. coordina- pp. Conf. 163–170 Multi-Agent Syst. (1995) I. Endo: Developmenttem of using tasktonomous assignment communication robots, sys- for J. Robotics multiple Mechatron. au- ume II: From Swarms to Intelligent Automata 100+ mobile robots. In: berg, A. Casal:cooperation Coordination of and multipleJ. decentralized Robotic Syst. mobile manipulators, 375–397 (1996) figuring robots, Commun. ACM 53.22 L.E. Parker: The Effect of action recognition and 53.23 M. Matarić: Behavior-based control:53.24 Examples B. MacLennan, G.M. Burghardt: Synthetic ethol- 53.12 J. Deneubourg, S. Goss, G. Sandini,53.13 F. Ferrari, C.R. Kube, H. Zhang: Collective robotics: From so- 53.14 R. Beckers, O. Holland, J. Deneubourg: From local 53.15 S. Onn, M. Tennenholtz: Determination53.16 of social B.B. Werger: Cooperation without deliberation: 53.19 N. Jennings: Controlling cooperative problem53.20 M. Tambe: Towards flexible53.21 teamwork, J. Artif. In- R.T. Vaughan, K. Stoy, G.S. Sukhatme, M.J. Mata- 53.25 T. Balch, R.C. Arkin:53.26 Communiation in reactive G. Dudek, M. Jenkin, E. Milios, D. Wilkes: A tax- 53.17 M.J. Huber, E. Durfee: Deciding when to commit 53.18 H. Asama, K. Ozaki, A. Matsumoto, Y. Ishida, 53.29 L.E. Parker: The effect of heterogeneity teamsin of 53.28 O. Khatib, K. Yokoi, K. Chang, D. Ruspini, R. Holm- 53.27 Z. Butler, K. Kotay, D.L. Rus, M. Vona: Self-recon- Moving in the Environment

Part E

Part E | 53 1368 Multiple Mobile Robot Systems References 1369

53.45 B. Grocholsky, R. Swaminathan, J. Keller, V. Ku- 53.60 M. Chu, J. Reich, F. Zhao: Distributed attention 53 | E Part mar, G. Pappas: Information driven coordinated for large video sensor networks, Intell. Distrib. air-ground proactive sensing, Proc. IEEE Int. Conf. Surveill. Syst. Semin (2004) Robotics Autom. (ICRA) (2005) 53.61 B. Jung, G.S. Sukhatme: Tracking targets using 53.46 A. Howard, L.E. Parker, G.S. Sukhatme: Experi- multiple robots: The effect of environment occlu- ments with a large heterogeneous mobile robot sion, Auton. Robots 13(3), 191–205 (2002) team: Exploration, mapping, deployment and 53.62 R.M. Murray, K.J. Åström, S.P. Boyd, R.W. Brockett, detection, Int. J. Robotics Res. 25(5/6), 431–447 G. Stein: Future directions in control in an infor- (2006) mation-rich world, IEEE Control Syst. Mag. (2003) 53.47 D. Fox, J. Ko, K. Konolige, B. Limketkai, D. Schulz, 53.63 A. Pant, P. Seiler, K. Hedrick: Mesh stability of B. Stewart: Distributed multirobot exploration look-ahead interconnected systems, IEEE Trans. and mapping, Proc. IEEE 94(7), 1325–1339 (2006) Autom. Control 47, 403–407 (2002) 53.48 M.A. Hsieh, A. Cowley, J.F. Keller, L. Chaimow- 53.64 T. Sugar, J. Desai, V. Kumar, J.P. Ostrowski: Coor- icz, B. Grocholsky, V. Kumar, C.J. Talyor, Y. Endo, dination of multiple mobile manipulators, Proc. R. Arkin, B. Jung, D. Wolf, G. Sukhatme, IEEE Int. Conf. Robotics Autom. (ICRA) (2001) D.C. MacKenzie: Adaptive teams of autonomous pp. 3022–3027 aerial and ground robots for situational aware- 53.65 R.W. Beard, J. Lawton, F.Y. Hadaegh: A coordi- ness, J. Field Robotics 24(11/12), 991–1014 (2007) nation architecture for spacecraft formation con- 53.49 J. Seyfied, M. Szymanski, N. Bender, R. Estana, trol, IEEE Trans. Control Syst. Technol. 9, 777–790 M. Theil, H. Worn: The I-swarm project: Intelli- (2001) gent small world autonomous robots for micro- 53.66 A. Das, J. Spletzer, V. Kumar, C. Taylor: Ad hoc net- manipulation, SAB 2004 Int. Workshop (2004) works for localization and control, Proc. IEEE Conf. 53.50 F. Mondada, G.C. Pettinaro, A. Guignard, Decis. Control (2002) pp. 2978–2983 I.W. Kwee, D. Floreano, J.-L. Deneubourg, 53.67 J. Manyika, H. Durrant-Whyte: Data Fusion and S. Nofli, L.M. Gambardella, M. Dorigo: Swarm- Sensor Management: An Information-Theoretic Bot: A new distributed robotic concept, Auton. Approach (Prentice Hall, Upper Saddle River 1994) Robots 17, 193–221 (2004) 53.68 H.G. Tanner, A. Jadbabaie, G.J. Pappas: Stable 53.51 A. Ollero, S. Lacroix, L. Merino, J. Gancet, J. Wik- flocking of mobile agents, Part I: Fixed topology, lund, V. Remuss, I. Veiga, L.G. Gutierrez, D.X. Vie- Proc. IEEE Conf. Decis. Control (2003) pp. 2010–2015 gas, M.A. Gonzalez, A. Mallet, R. Alami, R. Chatila, 53.69 J.M. Fowler, R. D’Andrea: Distributed control of G. Hommel, F.J. Colmenero, B. Arrue, J. Ferruz, close formation flight, Proc. IEEE Conf. Decis. Con- R. Martinez de Dios, F. Caballero: Architecture trol (2002) pp. 2972–2977 and perception issues in the COMETS multi-UAV 53.70 J.P. Desai, J.P. Ostrowski, V. Kumar: Modeling and project, IEEE Robotics Autom. Mag. 12(2), 46–57 control of formations of nonholonomic mobile (2005) robots, IEEE Trans. Robotics Autom. 17(6), 905– 53.52 Q. Li, D. Rus: Navigation protocols in sensor net- 908 (2001) works, ACM Trans. Sens. Netw. 1(1), 3–35 (2005) 53.71 H.G. Tanner, V. Kumar, G.J. Pappas: Leader-to- 53.53 R. Bajcsy: Active perception, Proc. IEEE 76,996– formation stability, IEEE Trans. Robotics Autom. 1005 (1988) 20(3), 443–455 (2004) 53.54 T. Eren, D. Goldenberg, W. Whitley, Y.R. Yang, 53.72 S. Loizou, V. Kumar: Relaxed input to state sta- S. Morse, B.D.O. Anderson, P.N. Belhumeur: Rigid- bility properties for navigation function based ity, computation, and randomization of network systems, Proc. IEEE Conf. Decis. Control (2006) localization, Proc. IEEE INFOCOM (2004) 53.73 M. Batalin, G.S. Sukhatme: Coverage, exploration 53.55 A. Hsieh, A. Cowley, V. Kumar, C.J. Taylor: Towards and deployment by a mobile robot and com- the deployment of a mobile robot network with munication network, Telecommun. Syst. J. 26(2), end-to-end performance guarantees, Proc. IEEE 181–196 (2004) Int. Conf. Robotics Autom. (ICRA) (2006) 53.74 K.J. O’hara, V. Bigio, S. Whitt, D. Walker, T.R. Balch: 53.56 B. Grocholsky, S. Bayraktar, V. Kumar, C.J. Taylor, Evaluation of a large scale pervasive embedded G. Pappas: Synergies in feature localization by air- network for robot path planning, Proc. IEEE Int. ground teams, Proc. 9th Int. Symp. Exp. Robotics Conf. Robotics Autom. (ICRA) (2006) pp. 2072– (2004) 2077 53.57 B. Grocholsky, E. Stump, V. Kumar: An extensive 53.75 V. Kumar, N. Leonard, A.S. Morse (Eds.): Cooper- representation for range-only SLAM, Int. Symp. ative control, Lecture Notes in Control and Infor- Exp. Robotics (2006) mation Sciences, Vol. 309 (Springer, Berlin, Hei- 53.58 A. Howard, M.J. Matarić, G.S. Sukhatme: An in- delberg 2004) cremental self-deployment algorithm for mo- 53.76 J. Chen, S. Teller, H. Balakrishnan: Pervasive pose- bile sensor networks, Auton. Robots 13(2), 113–126 aware applications and infrastructure, IEEE Com- (2002) put. Graph. Appl. 23(4), 14–18 (2003) 53.59 Z. Butler, D. Rus: Controlling mobile networks 53.77 A. Savvides, C.-C. Han, M. Srivastava: Dynamic for monitoring events with coverage constraints, fine-grained localization in ad-hoc networks of Proc. IEEE Int. Conf. Robotics Autom. (ICRA) sensors, Proc. 7th Annu. Int. Conf. Mobile Com- (2003) put. Netw. (MOBICOM-01) (2001) pp. 166–179 11, ,39–41(1975) , 374 (1978) 18 24 1(6), 272–300 (2005) (2), 218–230 (2009) 8 (10), 1456–1477 (2001) Art Gallery Theorems and Algorithms 89 11,691–719(2005) (1), 147–159 (2005) 13 (6), 988–1001 (2003) le sensor network deployment using potential (4), 1–28 (2007) fields: A distributed, scalable solutioncoverage problem, Proc. to Int. area the Symp. Distrib. Auton. Robotic Syst. (2002) pp. 299–308 age for mobileRobotics sensor Autom. (2004) networks, pp. IEEE 165–172 Int. Conf. based distributed topology-control algorithmwireless for multi-hopNetw. networks, IEEE/ACM Trans. of alphalifetime for largeTrans. Sens. sensor Netw. networks, ACM G.S. Sukhatme: Using localtopology geometry for control tunable in sensorMobile Comput. networks, IEEE Trans. multiple robots, Proc. IEEE Int. Conf.tom. Robotics (ICRA) Au- (2006) D. Rus, G.S. Sukhatme: Persistent ocean mon- tributed coverage optimizationlimited-range interactions, and ESAIM control Control withCalc. Optim. Var. coverage control with sensoryworked feedback robots, for net- (2006) Proc. Robotics Sci. Syst.ized (RSS) adaptive control for coveragerobots, for Proc. networked Int. Conf. Robotics Autom. (2007) Near-optimal sensor placements: Maximizing in- formation while minimizing communication cost, 5th Int. Conf. Inf. Process Sens. Netw. (IPSN) (2006) ometry, J. Comb. Theory Ser. (Oxford Univ. Press, New York 1987) orem, J. Comb. Theory Ser. groups of mobile autonomous agents using near- est neighbor rules,48 IEEE Trans. Autom. Control from dynamic silhouettes, Proc.put. IEEE Vis. Pattern Conf. Recognit. (2004) Com- rithms for cooperativeProc. multisensor IEEE surveillance, G.S. Sukhatme: Reconfiguration methods for mo- bile sensor networks, ACM3 Trans. Sens. Networks C. Lee: Pheromone319–324 (2001) robotics, Auton. Robots bi 53.107 S. Poduri, G.S. Sukhatme:53.108 Constrained cover- R. Wattenhofer, L. Li, P. Bahl, Y.M. Wang: A cone- 53.109 H. Zhang, J.C. Hou: On deriving the upper bound 53.110 S. Poduri, S. Pattem,53.111 B. M.A. Hsieh, Krishnamachari, V. Kumar: Pattern53.112 generation with R.N. Smith, M. Schwager, S.L. Smith, B.H. Jones, 53.94 J. Cortes, S. Martinez, F. Bullo: Spatially-dis- 53.95 M. Schwager, J. McLurkin, D. Rus:53.96 Distributed M. Schwager, J.-J. Slotine,53.97 D. Rus: Decentral- A. Krause, C. Guestrin, A. Gupta,53.98 J. Kleinberg: V. Chvatal: A53.99 combinatorial theorem in plane ge- J. O’Rourke: 53.100 S. Fisk: A short proof of Chvatal’s watchmen the- 53.101 A. Jadbabaie, J. Lin, A.S. Morse: Coordination of 53.102 S. Sinha, M. Pollefeys: Camera network calibration 53.103 R. Collins, A. Lipton, H. Fujiyoshi, T. Kanade: Algo- 53.104 A. Kansal, W. Kaiser, G. Pottie, M. Srivastava, 53.105 D. Payton, M. Daily, R.53.106 Estkowski, M. Howard, A. Howard, M.J. Matarić, G.S. Sukhatme: Mo- (4), (4), 29 30 (Morgan Kauf- (5), 662–669 (2002) (9), 116 (2005) 4 18 EEE/RSJ Int. Conf. Intell. Wireless Sensor Networks: An (2), 243–255 (2004) ) (2001) pp. 489–498 http://www.eecs.berkeley.edu/ 20 , 333–348 (2003) 2634 Pubs/TechRpts/2002/4010.html sisted localization and navigationrobots, for Int. networked J. Robotics Res. a global coordinate system from localon information an ad hoc sensor network,Sci. Lect. Notes Comput. G.S. Sukhatme: Evaluating network rigidityalistic in re- systems: Decentralization,and asynchronicity, parallelization, IEEE Trans. Robotics distributed network localization with noisymeasurements, range Proc. 2nd Int. Conf. Embed.Sens. Netw. Syst. (SenSys) (2004) pp. 50–61 mobile robot localizationcon field, withinProc. Int. Workshop aRobotics Algorithmic (2006) Found. fixed bea- con placement, Proc. 21st Int.put. Conf. Syst. Distrib. Com- (ICDCS-01 Wireless ad hocM02/26 Networks, (2001), Tech. Rep. UCB/ERL Information Processing Approach mann, New York 2004) for localization of robotRobotics networks, Autom. IEEE (2004) Int. Conf. teraction and coordinationmultiagent in proximity-limited systems, IEEE Trans. Robotics G.S. Sukhatme, W. Kaiser,M. M. Srivastava, D. Estrin: Hansen, Call G. andiments response: Pottie, Exper- in sampling theInt. Conf. environment, Embed. Netw. Proc. Sens. 2nd Syst.pp. (SenSys) (2004) 25–38 Adaptive sampling for environmentalmation field using esti- robotic sensors,Intell. IEEE/RSJ Robots Int. Syst. Conf. (2005) pp. 747–753 950–965 (2014) the ‘i’ incooperative ‘team’: localization, AnRobotics Proc. Autom. ego-centric (ICRA) (2003) IEEE pp. approach 868–892 Int. to Conf. ization for mobile robotlikelihood teams estimation, using I maximum Robots Syst. (2002) pp. 434–459 try: Probabilistic pursuit-evasion games:implementation Theory, andIEEE experimental Trans. Robotics evaluation, Autom. 930–944 (2013) control for mobile sensingRobotics Autom. networks, IEEE Trans. 53.82 P. Corke, R. Peterson, D. Rus: Communication-as- 53.83 R. Nagpal, H.E. Shrobe, J. Bachrach: Organizing 53.84 R.K. Williams, A. Gasparri, A. Priolo, 53.78 D. Moore, J. Leonard, D. Rus, S.J. Teller:53.79 Robust C. Detweiler, J. Leonard, D. Rus, S. Teller: Passive 53.80 N. Bulusu, J. Heidemann, D. Estrin: Adaptive bea- 53.81 S.N. Simic, S. Sastry: Distributed Localization in 53.91 F. Zhang, B. Grocholsky,53.92 V. Kumar: Formations R.K. Williams, G.S. Sukhatme: Constrained in- 53.88 M. Batalin, M.H. Rahimi, Y. Yu, D. Liu, A. Kansal, 53.89 M.H. Rahimi, W. Kaiser, G.S. Sukhatme, D. Estrin: 53.90 F. Zhao, L. Guibas: 53.85 A. Howard, M.J. Matarić, G.S. Sukhatme: Putting 53.86 A. Howard, M.J. Matarić, G.S. Sukhatme: Local- 53.87 R. Vidal, O. Shakernia, H.J. Kim, D.H. Shim, S. Sas- 53.93 J. Cortes, S. Martinez, T. Karatas, F. Bullo: Coverage Moving in the Environment

Part E

Part E | 53 1370 Multiple Mobile Robot Systems References 1371

itoring with underwater gliders: Adapting spa- 53.128 V. Gazi: Swarm aggregations using artificial po- 53 | E Part tiotemporal sampling resolution, J. Field Robotics tentials and sliding-mode control, IEEE Trans. 28(5), 714–741 (2011) Robotics 21(6), 1208–1214 (2005) 53.113 J. Das, F. Py, T. Maughan, T. O’Reilly, M. Mes- 53.129 J. McLurkin, J. Smith: Distributed algorithms for sié, J. Ryan, G.S. Sukhatme, K. Rajan: Coordinated dispersion in indoor environments using a swarm sampling of dynamic oceanographic features with of autonomous mobile robots, Symp. Distrib. Au- AUVs and drifters, Int. J. Robotics Res. 31(5), 626– ton. Robots Syst. (2004) 646 (2012) 53.130 D. Gage: Randomized search strategies with im- 53.114 G.A. Hollinger, S. Choudhary, P. Qarabaqi, C. Mur- perfect sensors, Proc. SPIE Mobile Robots VIII phy, U. Mitra, G.S. Sukhatme, M. Stojanovic, (1993) pp. 270–279 H. Singh, F. Hover: Underwater data collection 53.131 T. Balch: The impact of diversity on performance using robotic sensor networks, IEEE J. Sel. Areas in robot foraging, Proc. 3rd Ann. Conf. Auton. Commun. 30(5), 899–911 (2012) Agents (1999) pp. 92–99 53.115 F.Ye,G.Zhong,J.Cheng,L.Zhang,S.Lu:Peas: 53.132 Z.J. Butler, A.A. Rizzi, R.L. Hollis: Cooperative cov- A robust energy conserving protocol for long- erage of rectilinear environments, Proc. IEEE Int. lived sensor networks, Int. Conf. Distrib. Comput. Conf. Robotics Autom. (ICRA) (2000) Syst. (2003) 53.133 A.I. Mourikis, S.I. Roumeliotis: Performance anal- 53.116 X. Wang, G. Xing, Y. Zhang, C. Lu, R. Pless, C. Gill: ysis of multirobot cooperative localization, IEEE Integrated coverage and connectivity configura- Trans. Robotics 22(4), 666–681 (2006) tion in wireless sensor networks, 1st ACM Conf. 53.134 R. Grabowski, L.E. Navarro-Serment, C.J. Pare- Embed. Netw. Sens. Syst. (SenSys) (2003) dis, P.K. Khosla: Heterogeneous teams of mod- 53.117 D. Tian, N.D. Georganas: Connectivity mainte- ular robots for mapping and exploration, Auton. nance and coverage preservation in wireless sen- Robots 8(3), 271–298 (2000) sor networks, Ad Hoc Networks 3(6), 744–761 53.135 M. Berhault, H. Huang, P. Keskinocak, S. Koenig, (2005) W. Elmaghraby, P. Griffin, A. Kleywegt: Robot 53.118 G. Theraulaz, S. Goss, J. Gervet, J.-L. Deneubourg: exploration with combinatorial auctions, Proc. Task differentiation in Polistes wasp colonies: IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS) (2003) A model for self-organizing groups of robots, pp. 1957–1962 Proc.1stInt.Conf.Simul.Adapt.Behav.(1990) 53.136 J. Kim, J.M. Esposito, V. Kumar: An RRT-based al- pp. 346–355 gorithm for testing and validating mulit-robot 53.119 L. Steels: Cooperation Between Distributed Agents controllers, Proc. Robotics Sci. Syst. I (2005) Through Self-Organization, Proc. IEEE Int. Work- 53.137 Z. Cao, M. Tin, L. Li, N. Gu, S. Wang: Cooperative shop Intell. Robots Syst. (IROS) (1990) hunting by distributed mobile robots based on 53.120 A. Drogoul, J. Ferber: From tom thumb to the local interaction, IEEE Trans. Robotics 22(2), 403– dockers: Some experiments with foraging robots, 407 (2006) Proc. 2nd Int. Conf. Simul. Adapt. Behav. (1992) 53.138 R.W. Beard, T.W. McLain, M. Goodrich: Coordi- pp. 451–459 nated target assignment and intercept for un- 53.121 M.J. Matarić: Designing emergent behaviors: From manned air vehicles, Proc. IEEE Int. Conf. Robotics local interactions to collective intelligence, Proc. Autom. (ICRA) (2002) pp. 2581–2586 2nd Int. Conf. Simul. Adapt. Behav. (1992) pp. 432– 53.139 J. Clark, R. Fierro: Cooperative hybrid control of 441 robotic sensors for perimeter detection and track- 53.122 G. Beni, J. Wang: Swarm intelligence in cellu- ing, Proc. Am. Control Conf. (2005) pp. 3500–3505 lar robotics systems, Proc. NATO Adv. Workshop 53.140 P. Stone, M. Veloso: A layered approach to learn- Robots Biol. Syst. (1989) ing client behaviors in the RoboCup soccer server, 53.123 D. Stilwell, J. Bay: Toward the development of Appl. Artif. Intell. 12, 165–188 (1998) a material transport system using swarms of ant- 53.141 J. McLurkin: Stupid Robot Tricks: Behavior-Based like robots, Proc. IEEE Int. Conf. Robotics Autom. Distributed Algorithm Library for Programming (1993) pp. 766–771 Swarms of Robots, Ph.D. Thesis (Massachusetts 53.124 T. Fukuda, S. Nakagawa, Y. Kawauchi, M. Buss: Self Institute of Technology, Cambridge 2004) organizing robots based on cell structures – CE- 53.142 F. Mondada, L.M. Gambardella, D. Floreano, BOT, Proc. IEEE Int. Workshop Intell. Robots Syst. S. Nolfi, J.L. Deneuborg, M. Dorigo: The coopera- (1988) pp. 145–150 tion of swarm-bots: Physical interactions in col- 53.125 J.H. Reif, H.Y. Wang: Social Potential fields: A dis- lective robotics, IEEE Robotics Autom. Mag. 12(2), tributed behavior control for autonomous robots, 21–28 (2005) Robotics Auton. Syst. 27(3), 171–194 (1999) 53.143 T.W. Mather, M.A. Hsieh: Macroscopic modeling of 53.126 L.E. Parker: Designing control laws for cooperative stochastic deployment policies with time delays agent teams, Proc. IEEE Int. Cont. Robotics Autom. for robot ensembles, Int. J. Robotics Res. 30(5), (1993) pp. 582–587 590–600 (2011) 53.127 K. Sugihara, I. Suzuki: Distributed algorithms for 53.144 X.C. Ding, M. Kloetzer, Y. Chen, C. Belta: Automatic formation of goemetric patterns with many mo- deployment of robot teams, IEEE Robotics Autom. bile robots, J. Robotic Syst. 13(3), 127–139 (1996) Mag. 18(3), 75–86 (2022) (4), 21 (5), 317–388 (2), 308–319 13 23 (4), 531–545 (1997) 13 7(4), 442–451 (2003) 996) (1 ration control of modular robotConf. M-TRAN, Mechatron. IEEE Autom. Int. (2005) pp. 254–259 periments of a modular robot byProc. visual feedback, IEEE/RSJ Int.(IROS) (2006) Conf. pp. 625–630 Intell. Robots Systems modular robot locomotion, IEEE Int. Conf. Mecha- tron. Autom. (2005) pp. 248–253 modular robotic units,Intell. IEEE Robotics Autom. Int. (2003) pp. Symp. 886–892 Comput. ular ATRON: Modulesrobot, Proc. IEEE/RSJ for Int. Conf. a(IROS) Intell. (2004) Robots pp. self-reconfigurable Syst. 2068–2073 ployable, multi-functional, andreconfigurable robotic modular system,Proc. self- Conf. IEEE/RSJ Intell. Int. Robots Syst. (IROS) (2006) pp. 3636– reconfigurable robot,Proc. IEEE Int.Conf. Autom. Robotics (ICRA) (2000) pp. 514–520 shaw: Connecting andreconfigurationwith disconnecting PolyBot, forMechatron. IEEE/ASME self– Trans. M. Dugan, C.J.assembly after Taylor: explosion, IEEE/RSJ Int. Towards Conf. Intell. Robots robotic Syst. (2007) self-re- pp. 2767–2772 Evolved and designedlar self-reproducing robotics, modu- IEEE(2007) Trans. Robotics ysis of miniature foldedchain programmable right matter, angle Proc. IEEE tetrahedron Robotics Int. Autom. Conf. (ICRA) (2010) pp. 2785–2790 robotic system, Proc. IEEE Int.tom. Conf. (ICRA) Robotics (1994) Au- pp. 449–455 ating efficiency of self-reconfigurationof modular robots, in a class J. Robotic Syst. ful metrics forIEEE modular Trans. robot Robotics Autom. motion planning, for fast concurrent reconfigurationmetamorphic of robots, hexagonal IEEE621–631 Trans. (2005) Robotics bling machine, Proc. IEEEtom. Int. (ICRA) Conf. (1994) Robotics pp. Au- 441–448 S. Kokaji:a modular Distributed mechanical system,Conf. Proc. formation IEEE/RSJ Intell. Int. Robots Syst. control (IROS) (1997) for pp. 1090– 3641 1097 53.163 S. Murata, K. Kakomura, H. Kurokawa: Docking ex- 53.164 D. Marbach, A.J. Ijspeert: Online optimization of 53.165 E.H. Østergaard, H.H. Lund: Evolving control for 53.166 M.W. Jørgensen, E.H. Østergaard, H.H. Lund: Mod- 53.167 B. Salemi, M. Moll, W.-M. Shen: SUPERBOT: A de- 53.168 M. Yim, D.G. Duff, K.D. Roufas: PolyBot: A modular 53.169 M. Yim, Y. Zhang, K. Roufas, D. Duff, C.53.170 Elder- M. Yim, B. Shirmohammadi, J. Sastra, M. Park, 53.171 V. Zykov, E. Mytilinaios, M. Desnoyer, H. Lipson: 53.172 P.J. White, M.L. Posner, M. Yim: Strength anal- 53.173 G.S. Chirikjian: Kinematics of53.174 a metamorphic G. Chirikjian, A. Pamecha, I. Ebert-Uphoff: Evalu- 53.175 A. Pamecha, I. Ebert-Uphoff, G.S. Chirikjian: Use- 53.176 J.E. Walter, E.M. Tsai, N.M. Amato: Algorithms 53.177 S. Murata, H. Kurokawa, S.53.178 Kokaji: Self-assem- E. Yoshida, S. Murata, K. Tomita, H. Kurokawa, 10(3), 314– (11), 1757–1789 175 ssembly, IEEE Robotics self-disa 1(1), 30–37 (2006) 17(3), 38–53 (2010) (1), 43–52 (2007) 7(4), 431–441 (2002) 14 (2011) consensus protocols formetric the patterns generation in of the(2011) geo- pp. plane, 3850–3855 Am. Controls Conf. based multi-robot team formation for distributed area coverage, Proc. 3rd Int.Agents Robots Symp. (2010) Pract. pp. Cognit. 9–15 maximization fortive informed behavior, Distrib. Auton. agents Roboticpp. 389–402 Syst. in (2013) collec- A. Martinoli: Specialization as an optimalunder strategy varying external conditions, Proc.Conf. IEEE Robotics Int. Autom. (2009) pp. 1941–1946 level algorithm switching foring, robot swarm Distrib. forag- Auton.430 Robotic Syst. (2013) pp. 417– Fault-tolerance in robotIdentif. systems, Control Int. J. Model. urable robotic system,Autom. IEEE (1988) pp. Int. 1581–1586 Conf. Robotics H. Lipson, E. Klavins, G.S. Chirikjian: Modularreconfigurable robot self- systems: Challengesand op- portunities for theMag. future, IEEE Robotics Autom. self-assembly to Autom. Mag. many modes of locomotion,Mechatron. JSME (1993) Int. pp. Conf. 283–288 Adv. Robotics Autom. (ICRA) (1994) pp. 2508–2514 for reconfigurable robots, Proc. IEEE/RSJIntell. Int. Conf. Robots Syst. (IROS) (2000) pp. 2203–2209 urable robots,Robots Proc. Syst. (IROS) IEEE/RSJ (2001) pp. Int. 1049–1054 Conf.for Intell. reconfigurable robots, IEEE Trans.7(4), 403–409 Mechatron. (2002) K. Tomita, S. Kokaji: M-TRAN:modular robotic Self-reconfigurable system, IEEE/ASME Trans. Mecha- tron. rata, K. Tomita, S.design Kokaji: and Automatic locomotion experimentssystem, for IEEE/ASME a Trans. Mechatron. modular robotic S. Kokahji, S. Murata: Distributed self-reconfigu- sparse interactions, Artif. Intell. 325 (2005) 53.146 P. Tsiotras, L.I.R. Castro: Extended multi-agent 53.147 K. Cheng, P. Dasgupta: Weighted voting game 53.148 A.A. Taheri, M. Afshar, M. Asadpour: Influence 53.149 M.A. Hsieh, A. Halász, E.D. Cubuk, S. Schoenholz, 53.150 N. Hoff, Wood,R. Nagpal: R. Distributed colony- 53.151 A. Winfield, J. Nembrini:53.152 Safety T. in Fukuda,numbers: S. Nakagawa: Dynamically reconfig- 53.153 M. Yim, W.-M. Shen, B. Salemi, D. Rus, M. Moll, 53.154 K. Gilpin, D. Rus: Modular robot53.155 systems: From M. Yim: A reconfigurable modular53.156 robot with M. Yim: New locomotion gaits, Proc.53.157 IEEE Int. Conf. A. Castano, P. Will: Mechanical design of a module 53.158 W.-M. Shen, P. Will: Docking in53.159 self-reconfig- A. Castano, A. Behar, P. Will: The conro53.160 modules S. Murata, E. Yoshida, A. Kamimura, H. Kurokawa, 53.161 A. Kamimura, H. Kurokawa, E. Yoshida, S. Mu- 53.162 H. Kurokawa, K. Tomita, A. Kamimura, E. Yoshida, 53.145 F.S. Melo, M. Veloso: Decentralized MDPs with Moving in the Environment

Part E

Part E | 53 1372 Multiple Mobile Robot Systems References 1373

53.179 H. Kurokawa, S. Murata, E. Yoshida, K. Tomita, 53.195 S. Goldstein, J. Campbell, T. Mowry: Pro- 53 | E Part S. Kokaji: A 3-D self-reconfigurable structure grammable Matter, IEEE Computer 38(6), 99–101 and experiments, Proc. IEEE/RSJ Int. Conf. Intell. (2005) Robots Syst. (IROS) (1998) pp. 860–865 53.196 S.C. Goldstein, J.D. Campbell, T.C. Mowry: Pro- 53.180 E. Yoshida, S. Murata, S. Kokaji, A. Kamimura, grammable matter, Computer 38(6), 99–101 (2005) K. Tomita, H. Kurokawa: Get back in shape! 53.197 M.E. Karagozler, S.C. Goldstein, J.R. Reid: Stress- A hardware prototype self-reconfigurable modu- driven MEMS assembly + electrostatic forces = lar microrobot that uses shape memory alloy, IEEE 1 mm diameter robot, Proc. IEEE/RSJ Int. Conf. In- Robotics Autom. Mag. 9(4), 54–60 (2002) tell. Robots Syst. (IROS) (2009) pp. 2763–2769 53.181 J. Gargus, B. Kim, I. Llamas, J. Rossignac, C. Shaw: 53.198 P. Pillai, J. Campbell, G. Kedia, S. Moudgal, Finger Sculpting with Digital Clay, Tech. Rep. GIT- K. Sheth: A 3-D fax machine based on claytronics, GVU-02-22 (2002) Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS) 53.182 K. Kotay, D. Rus, M. Vona, C. McGray: The self-re- (2006) pp. 4728–4735 configuring robotic molecule, Proc. IEEE Int. Conf. 53.199 P. White, K. Kopanski, H. Lipson: Stochastic Robotics Autom. (ICRA) (1998) pp. 424–431 self-reconfigurable cellular robotics, IEEE Conf. 53.183 K. Kotay, D. Rus: Motion synthesis for the self- Robotics Autom. (2004) pp. 2888–2893 reconfiguring robotic molecule, IEEE Int. Conf. In- 53.200 P. White, V. Zykov, J. Bongard, H. Lipson: Three di- tell. Robots Syst. (1998) pp. 843–851 mensional stochastic reconfiguration of modular 53.184 K. Kotay, D. Rus: Algorithms for self-reconfigur- robots, Robotics Sci. Syst. (2005) ing molecule motion planning, Proc. IEEE/RSJ Int. 53.201 M. Tolley, J. Hiller, H. Lipson: Evolutionary design Conf. Intell. Robots Syst. (IROS) (2000) and assembly planning for stochastic modular 53.185 Z.J. Butler, K. Kotay, D. Rus, K. Tomita: Generic robots, Proc. IEEE/RSJ. Int. Conf. Intell. Robotics decentralized control for lattice-based self-re- Syst.(IROS)(2009)pp.73–78 configurable robots, Int. J. Robotics Res. 23(9), 53.202 M. Tolley, H. Lipson: Fluidic manipulation for scal- 919–937 (2004) able stochastic 3-D assembly of modular robots, 53.186 D. Rus, M. Vona: A basis for self-reconfiguring Proc. IEEE Int. Conf. Robotics Autom. (ICRA) (2010) robots using crystal modules, Proc. IEEE/RSJ Int. pp. 2473–2478 Conf. Intell. Robots Syst. (IROS) (2000) pp. 2194– 53.203 M.T. Tolley, H. Lipson: Programmable 3-D stochas- 2202 tic fluidic assembly of cm-scale modules, Proc. 53.187 D. Rus, M. Vona: Crystalline robots: Self-recon- IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS) (2011) figuration with compressible unit modules, Int. pp. 4366–4371 J. Robotics Res. 22(9), 699–715 (2003) 53.204 K. Gilpin, K. Kotay, D. Rus, I. Vasilescu: Miche: 53.188 J.W. Suh, S.B. Homans, M. Yim: Telecubes: Me- Modular shape formation by self-disassembly, chanical design of a module for self-reconfig- Int. J. Robotics Res. 27,345–372(2008) urable robotics, Proc. IEEE Int. Conf. Robotics Au- 53.205 R. Oung, F. Bourgault, M. Donovan, R. D’Andrea: tom. (ICRA) (2002) pp. 4095–4101 The distributed flight array, Proc. IEEE Int. Conf. 53.189 C.-J. Chiang, G.S. Chirikjian: Modular robot mo- Robotics Autom. (ICRA) (2010) tion planning using similarity metrics, Auton. 53.206 G.J. Hamlin, A.C. Sanderson: Tetrobot: A modu- Robots 10, 91–106 (2001) lar system for hyper-redundant parallel robotics, 53.190 M. Koseki, K. Minami, N. Inou: Cellular robots Proc. IEEE Int. Conf. Robotics Autom. (ICRA) (1995) forming a mechanical structure (evaluation of pp. 154–159 structural formation and hardware design of CHO- 53.207 A. Lyder, R.F.M. Garcia, K. Stoy: Mechanical de- BIE II), Proc. 7th Int. Symp. Distrib. Auton. Robotic sign of Odin, an extendable heterogeneous de- Syst. (DARS) (2004) pp. 131–140 formable modular robot, Proc. IEEE/RSJ Int. Conf. 53.191 B.K. An: Em-cube: Cube-shaped, self-recon- Intell. Robots Syst. (IROS) (2008) pp. 883–888 figurable robots sliding on structure surfaces, 53.208 A. Lyder, H.G. Peterson, K. Stoy: Representation Proc. IEEE Int. Conf. Robotics Autom. (ICRA) (2008) and shape estimation of Odin, a parallel un- pp. 3149–3155 der-actuated modular robot, Proc. IEEE/RSJ Int. 53.192 C. Ünsal, P.K. Khosla: Mechatronic design of Conf. Intell. Robots Syst. (IROS) (2009) pp. 5275– a modular self-reconfiguring robotic system, 5280 Proc. IEEE Int. Conf. Robotics Autom. (ICRA) (2000) 53.209 C.-H. Yu, K. Haller, D. Ingber, R. Nagpal: Morpho: A pp. 1742–1747 self-deformable modular robot inspired by cellar 53.193 K.C. Prevas, C. Ünsal, M.Ö. Efe, P.K. Khosla: A hi- structure, Proc. IEEE/RSJ Int. Conf. Intell. Robots erarchical motion planning strategy for a uni- Syst. (IROS) (2008) pp. 3571–3578 form self-reconfigurable modular robotic system, 53.210 M. Shimizu, A. Ishiguro, T. Kawakatsu: A modu- Proc. IEEE Int. Conf. Robotics Autom. (ICRA) (2002) lar robot that exploits a spontaneous connectivity pp. 787–792 control mechanism, Proc. IEEE/RSJ Int. Conf. Intell. 53.194 K. Hosokawa, T. Tsujimori, T. Fujii, H. Kaetsu, Robots Syst. (IROS) (2005) pp. 1899–1904 H. Asama, Y. Kuroda, I. Endo: Self-organizing col- 53.211 M. Shimizu, T. Mori, A. Ishiguro: A Development lective robots with morphogenesis in a vertical of a modular robot that enables adaptive recon- plane, Proc. IEEE Int. Conf. Robotics Autom. (ICRA) figuration, Proc. IEEE/RSJ Int. Conf. Intell. Robots (1998) pp. 2858–2863 Syst. (IROS) (2006) pp. 174–179 , 437 17(4), 789–808 , 431–447 (2006) 25 (2), 14–19 (2000) 15 Multi-Robot Systems ssembly, IEEE Int. Conf. (3), 269–292 (2000) 8 Robot Teams: From Diversity to Poly- , ed. by T. Balch, L.E. Parker (A K Peters, (3), 209–238 (2000) 8 (1), 1–15 (2006) 15 rdrecht 2003) theoretic measure of robot teamRobots diversity, Auton. munication between heterogeneous cooperating robots, Auton. Robots and rescue, IEEE Intell. Syst. Experiments withrobots. In: cooperativemorphism aerial-ground Natick 2002) coupled navigation assistance inmulti-robot heterogeneous teams, Proc.Robots Syst. IEEE (2004) Int. Conf. Intell. ments with a largeteam: heterogeneous Exploration, mobile mapping, robot detection, Int. deployment, J. and Robotics Res. Self-replication from random parts,636 Nature (2005) havior in intelligent self-assembly, IEEE Int.Robotics Conf. Autom. (ICRA) (2003) pp. 721–726 of sensing for rule-basedsembly, planar IEEE distributed Int. Conf. as- Intell.pp. Robots 3728–3734 Syst. (2006) struction with Artificial Swarms, Ph.D.Cambridge Thesis(MIT, 2006) D. Rus: An untethered, electrostatic,trollable MEMS globally micro-robot, J. con- Microelectromech. Syst. croassembly by parallel actuator ofrobots, MEMS J. micro- Microelectromech. Syst. (2008) disassembly of magnetictowards 2-D mobile reconfigurablemicro-systems, micro-robots Symp. Robotics Int. Res. (2009) namics of programmedConf. self-assembly, IEEE Robotics Int. Autom. (ICRA) (2006) pp. 1469– shape duplication with smart pebble robots,Int. IEEE Conf. Robotics Autom. (ICRA) (2012) centimeter robotic modulesmatter through for self-disa programmable Robotics Autom. (ICRA) (2010) of 100+ mobileVolume robots. II: In: From Swarms toed. Intelligent by, Automata A. Schultz, L.E. Parker, F. SchneiderDo (Kluwer, 1476 53.239 T. Balch: Hierarchic social entropy: An information 53.240 D. Jung, A. Zelinsky: Grounded53.241 symbolic com- R.R. Murphy: Marsupial robots for urban search 53.242 G. Sukhatme, J.F. Montgomery, R.T. Vaughan: 53.243 L.E. Parker, B. Kannan, F. Tang, M. Bailey: Tightly- 53.244 A. Howard, L.E. Parker, G.S. Sukhatme: Experi- 53.228 S. Griffith, D. Goldwater, J.M. Jacobson: Robotics: 53.229 C. Jones, M.J. Matarić: From local to53.230 global be- J. Kelly, H. Zhang: Combinatorial optimization 53.231 J. Werfel: Anthills Built to Order: Automating Con- 53.232 B. Donald, C.G. Levey, C.D. McGray, I. Paprotny, 53.233 B.R. Donald, C.G. Levey, I. Paprotny: Planar mi- 53.234 C. Pawashe, S. Floyd, M. Sitti: Assembly53.235 and N. Napp, S. Burden, E. Klavins: The statistical dy- 53.236 K. Gilpin, D. Rus: A distributed algorithm for53.237 2-D K. Gilpin, A. Knaian, D. Rus: Robot pebbles: One 53.238 L.E. Parker: The effect of heterogeneity teams in , 32rd Annu. Conf. Robotics Autom. 1(4), 413–427 (1994) wski: Self-assembly at all mbled squares mbly of mesoscope and , 2418–2421 (2002) (8), 946–961 (2009) 28 295 99(8), 4769–4774 (2002) (28),107 12441–12445 (2010) xity of self-asse (5482), 1170–1172 (2000) J.D. Campbell,calization S.C. of Goldstein:J. Distributed Robotics modular Res. lo- robot ensembles, Int. folding sheets, IEEE(ICRA) (2012) Int. scales, Sci. USA molecules: Self-asse macroscopic components,Sciences Proc. Natl.G.M. Acad. Whitesides:in Forming three electrical dimensions289 by networks self-assembly, Science phology affectsmodular self-assembly robot, in IEEE(2008) Int. pp. 3533–3538 Conf. a stochastic Robotics Autom. of self-assembling systems: Analogy withical chem- kinematics, Artif. Life bly and self-repair in aIEEE/RSJ collective Int. of Conf. robots, Intell. Proc. Robots Syst.pp. (IROS) 1484–1489 (2009) size selectionrobotic for collective, Proc. the IEEE/RSJRobots Int. shape Syst. Conf. (IROS) of Intell. (2010) pp. a 508–513 distributed low cost scalable robothaviors, system for IEEE collective be- Int.(2012) Conf. Robotics Autom. (ICRA) S. Kim, E.D.grammable Demaine, matter D. by Rus, folding,Sci. USA Proc. R.J. Natl. Wood: Acad. Pro- for modules andtion its mechanisms, control IEEE Int. by Conf.(ICRA) vibrating Robotics (2009) Autom. actua- pp. 4281–4286 comple ACM Symp. Theory Comput. (2000) pp. 459–468 ning time andsquares, program size 33rd Annu. for(2001) self-assembled ACM pp. Symp. 740–748 Theory Comput. R.T. Schweller:models of Complexities self-assembly,Symp. 15th for Annu. Discrete ACM-SIAM generalized 889 Algorithms (2004) pp.W. Malone, 880– N.parts: Napp, A T. demonstration Nguyen: ofproach Programmable the to grammatical self-organization, ap- Conf. Intell. Proc. Robots IEEE/RSJ Syst. (IROS) Int. (2005) pp. 3684– 3691 53.212 S. Funiak, P. Pillai, M.P. Ashley-Rollman, 53.217 B. An, D. Rus: Programming and controlling self- 53.218 G. Whitesides, B. Grzybo 53.219 G.M. Whitesides, M.53.220 Boncheva: D.H. Beyond Garcias, J. Tien,53.221 T.L. Breen, S. Miyashita, M. Kessler, C. M. Lungarella: How mor- Hsu, 53.222 K. Hosokawa, I. Shimoyama, H. Miura: Dynamics 53.213 M. Rubenstein, W.-M. Shen: Scalable self-assem- 53.214 M. Rubenstein, W.-M. Shen: Automatic scalable 53.215 M. Rubenstein, C. Ahler, R. Nagpal: Kilobot: A 53.216 E. Hawkes, B. An, N.M. Benbernou, H. Tanaka, 53.223 M. Shimizu, K. Suzuki: A Self-repairing structure 53.224 P.W.K. Rothemund, E. Winfree: The program-size 53.225 L. Adleman, Q. Cheng, A. Goel, M.-D. Huang: Run- 53.226 G. Aggarwal, M.H. Goldwasser, M.-Y. Kao, 53.227 J. Bishop, S. Burden, E. Klavins, R. Kreisberg, Moving in the Environment

Part E

Part E | 53 1374 Multiple Mobile Robot Systems References 1375

53.245 L. Chaimowicz, B. Grocholsky, J.F. Keller, V. Kumar, 53.262 G. Rabideau, T. Estlin, S. Schien, A. Barrett: A com- 53 | E Part C.J. Taylor: Experiments in multirobot air-ground parison of coordinated planning methods for co- coordination, Proc. IEEE Int. Conf. Robotics Autom. operating rovers, Proc. AIAA Space Technol. Conf. (ICRA) (2004) (1999) 53.246 L.E. Parker, F. Tang: Building multi-robot coali- 53.263 R. Zlot, A. Stentz, M.B. Dias, S. Thayer: Multi- tions through automated task solution synthesis, robot exploration controlled by a market econ- Proc. IEEE 94(7), 1289–1305 (2006) omy, Proc. IEEE Int. Conf. Robotics Autom. (ICRA) 53.247 R. Zlot, A. Stentz: Market-based multirobot coor- (2002) pp. 3016–3023 dination for complex tasks, Int. J. Robotics Res. 53.264 J. Guerrero, G. Oliver: Multi-robot task alloca- 25(1), 73–101 (2006) tion strategies using auction-like mechanisms, 53.248 J. Jennings, C. Kirkwood-Watts: Distributed mo- Proc. 6th Congr. Catalan Assoc. Artif. Intell. (2003) bile robotics by the method of dynamic teams. In: pp. 111–122 Distributed Autonomous Robotic Systems 3,ed.by 53.265 N. Kalra, D. Ferguson, A. Stentz: Hoplites: A mar- T. Lueth, R. Dillmann, P. Dario, H. Wörn (Springer, ket-based framework for planned tight coordi- Berlin, Heidelberg 1998) nation in multirobot teams, Proc. IEEE Int. Conf. 53.249 E. Pagello, A. D’Angelo, E. Menegatti: Cooperation Robotics Autom. (ICRA) (2005) issues and distributed sensing for multirobot sys- 53.266 L. Lin, Z. Zheng: Combinatorial bids based multi- tems, Proc. IEEE 94, 1370–1383 (2006) robot task allocation method, Proc. IEEE Int. Conf. 53.250 B. Gerkey, M.J. Matarić: A formal analysis and tax- Robotics. Autom. (ICRA) (2005) pp. 1145–1150 onomy of task allocation in multi-robot systems, 53.267 C.-H. Fua, S.S. Ge: COBOS: Cooperative backoff Int. J. Robotics Res. 23(9), 939–954 (2004) adaptive scheme for multirobot task allocation, 53.251 D. Gale: The Theory of Linear Economic Models IEEE Trans. Robotics 21(6), 1168–1178 (2005) (McGraw-Hill, New York 1960) 53.268 E.G. Jones, B. Browning, M.B. Dias, B. Argall, 53.252 E. Balas, M.W. Padberg: On the set-covering prob- M. Veloso, A. Stentz: Dynamically formed hetero- lem, Oper. Res. 20(6), 1152–1161 (1972) geneous robot teams performing tightly-coupled 53.253 B.B. Werger, M.J. Matarić: Broadcast of local el- tasks, Proc. IEEE Int. Conf. Robotics Autom. (ICRA) igibility for multi-target observation. In: Dis- (2006) pp. 570–575 tributed Autonomous Robotic Systems 4,ed.by 53.269 Y. Zhang, L.E. Parker: IQ-ASyMTRe: Forming exe- L.E. Parker, G. Bekey, J. Barhen (Springer, Tokyo cutable coalitions for tightly coupled multirobot 2000) tasks, IEEE Trans. Robotics 29(2), 400–416 (2012) 53.254 R.A. Brooks: A robust layered control system for 53.270 L. Vig, J.A. Adams: Multi-robot coalition forma- a mobile robot, IEEE J. Robotics Autom. RA-2(1), tion, IEEE Trans. Robotics 22(4), 637–649 (2006) 14–23 (1986) 53.271 M. Bowling, M. Veloso: Simultaneous adversarial 53.255 S. Botelho, R. Alami: M+: A scheme for multi- multi-robot learning, Proc. Int. Joint Conf. Artif. robot cooperation through negotiated task al- Intell. (2003) location and achievement, Proc. IEEE Int. Conf. 53.272 F. Fernandez, L.E. Parker: A reinforcement learn- Robotics Autom. (1999) pp. 1234–1239 ing algorithm in cooperative multi-robot do- 53.256 R.G. Smith: The contract net protocol: High- mains, J. Intell. Robots Syst. 43, 161–174 (2005) level communication and control in a distributed 53.273 C.F. Touzet: Robot awareness in cooperative mo- problem solver, IEEE Trans. Comput. C-29(12), bile robot learning, Auton. Robots 2, 1–13 (2000) 1104–1113 (1980) 53.274 R. Steeb, S. Cammarata, F. Hayes-Roth, P. Thorn- 53.257 B. Dias, R. Zlot, N. Kalra, A. Stentz: Market-based dyke, R. Wesson: Distributed Intelligence for Air multirobot coordination: A survey and analysis, Fleet Control, Rand Corp. Tech. Rep. R-2728-AFPA Proc. IEEE 94(7), 1257–1270 (2006) (1981) 53.258 B.P. Gerkey, M.J. Matarić: Sold! Auction methods 53.275 M. Benda, V. Jagannathan, R. Dodhiawalla: On for multi-robot coordination, IEEE Trans. Robotics Optimal Cooperation of Knowledge Sources, Boe- Autom. 18(5), 758–768 (2002) ing AI Center Tech. Rep. BCS-G2010-28 (1985) 53.259 H. Kose, U. Tatlidede, C. Mericli, K. Kaplan, 53.276 T. Haynes, S. Sen: Evolving behavioral strate- H.L. Akin: Q-learning based market-driven gies in predators and prey. In: Adaptation and multi-agent collaboration in robot soccer, Proc. Learning in Multi-Agent Systems, ed. by G. Weiss, Turk. Symp. Artif. Intell. Neural Netw. (2004) S. Sen (Springer, Berlin, Heidelberg 1986) pp. 113– pp. 219–228 126 53.260 D. Vail, M. Veloso: Multi-robot dynamic role as- 53.277 S. Mahadevan, J. Connell: Automatic program- signment and coordination through shared po- ming of behavior-based robots using reinforce- tential fields, multi-robot systems: From swarms ment learning, Proc. AAAI (1991) pp. 8–14 to intelligent automata, Proc. Int. Workshop 53.278 S. Marsella, J. Adibi, Y. Al-Onaizan, G. Kaminka, Multi-Robot Syst. (2003) pp. 87–98 I.Muslea,M.Tambe:Onbeingateammate:Ex- 53.261 M. Lagoudakis, E. Markakis, D. Kempe, periences acquired in the design of RoboCup P. Keshinocak, A. Kleywegt, S. Koenig, C. Tovey, teams, Proc. 3rd Annu. Conf. Auton. Agents (1999) A. Meyerson, S. Jain: Auction-based multi-robot pp. 221–227 routing, Robotics: Science and Systems I (MIT 53.279 J. Pugh, A. Martinoli: Multi-robot learning with Press, Cambridge 2005) particle swarm optimization, Proc. 5th Int. Jt. (1), 427– (5), 865– 9 (1), 43–63 2 20 (1), 152–174 (2004) 65 (9), 1465–1476 (2004) (11), 1963–1974 (2004) 49 49 22(4), 650–665 (2006) (4), 646–656 (2005) (5), 571–583 (2007) , 25–34 (1987) (6), 926–939 (1998) 21 22(6), 1285–1292 (2006) 21 14 129 (6), 988–1001 (2002) (5), 837–846 (2002) (3), 387–392 (2005) swarm behaviorsrobotic systems, of Proc. IEEE Int. Conf. distributedtom. Robotics Au- (2001) autonomous pp. 3993–3998 scheduling for resource-constrained localization of mobile robot formations,22(5), IEEE 917–931 (2006) Trans. Robotics mutation-invariant multirobot formations,Trans. Robotics IEEE tributed behavioral model, ACMput. SIGGRAPH Gr. Com- trol for multi-robotAutom. teams, IEEE Trans. Robotics groups of mobile autonomous agents using near- est neighbor rules,48 IEEE Trans. Autom. Control groups of robots, IEEE Trans. Robotics two-dimensional kinematic model forgroups, biological SIAM J. Appl. Math. erative control of vehicleAutom. formations, Control IEEE Trans. mations of vehicles inAutom. Control cyclic pursuit, IEEE Trans. tial trenches for multirobot formations, IEEE Trans. Robotics ity of multi-agent formations, IEEE Trans. Robotics 21 platoons ofRobotics autonomous vehicles, IEEE Trans. robot formations usingmal local communication, sensing IEEE and Trans. Robotics18 mini- Autom. recent progress in the studyagent of coordination, distributed IEEE multi- Trans. Ind.438 Inf. (2012) trol of multivehicleControl systems, J. Dyn. Syst. Meas. ulated autonomous mobilepayload transportation, robot Auton. Robots collectives in (1995) operative object manipulation using multiple be- haviour-based robots, Proc.Intell. Robots IEEE/RSJ Syst. (IROS) Int. (1996) pp. Conf. 310–317 875 (2004) P. Tabuada, G. Pappas, P. Lima: Motion feasibil- 53.294 S. Sun, D. Lee, K. Sim: Artificial immune-based 53.295 A.I. Mourikis, S.I. Roumeliotis: Optimal sensor 53.296 S. Kloder, S. Hutchinson: Path planning for53.297 per- C.W. Reynolds: Flocks, herds and schools: A dis- 53.298 T. Balch, R. Arkin: Behavior-based formation con- 53.299 A. Jadbabaie, J. Lin, A.S. Morse: Coordination of 53.300 C. Belta, V. Kumar: Abstraction and control for 53.301 C.M. Topaz, A.L. Bertozzi: Swarming patterns53.302 in J.A. Fax, R.M. Murray: Information flow and coop- 53.303 J.A. Marshall, M.E. Broucke, B.R.53.304 Francis: For- S.S. Ge, C.-H. Fua: Queues and artificial53.305 poten- 53.306 G. Antonelli, S. Chiaverini: Kinematic control53.307 of J. Fredslund, M.J. Matarić: A general algorithm for Y.Cao,W.Yu,W.Ren,G.Chen:Anoverviewof 53.308 53.309 R.M. Murray: Recent research in cooperative con- 53.310 P.J. Johnson, J.S. Bay: Distributed control of sim- 53.311 Z. Wang, E. Nakano, T. Matsukawa: Realizing co- Multiple 34(5), 2015–2028 (2), 113–123 (2006) 20 (1), 36–47 (1998) 5 Distributed Autonomous , ed. by H. Asama, T. Arai, 22(3), 52–67 (2002) (3), 345–383 (2000) (5), 815–822 (1998) 8 (1), 12–31 (2000) 15 19 G.A. Gilbreath,derieden, T. T.A. Tran, K.J. Grant, D.M. Heath-Pastore, Jaffee: R.S. In- multi-robot task division, IEEE Trans. Robotics Au- tom. resource host architecturebile (MRHA) for detection(MDARS), the SPAWAR mo- assessment3026, Systems Revision A response ( Technical 2000) Documen system orative robots for infrastructure securitytions, applica- Proc. Int. Symp. Collab. Technol.pp. Syst. 235–240 (2004) Soup: A testbed for studying resource allocationmulti-vehicle in systems, Proc. AAAI Workshoption Auc- Mech. Robot Coord. (2006) pp. 23–30 tems: A classificationIEEE focused Trans. Syst. on Man(2004) coordination, Cybern. B distributed optimization and control,Syst. IEEE Mag. Control vs. PC – Determinismplementary approaches and to randomness robotic as explorationcontinuous of com- unknown domains,Res. Int. J. Robotics T. Fukuda, T. Hasegawa (Springer,berg Berlin, 2002) Heidel- ative multi-robot approach toexploration of the Mars, mapping Proc. and AAAI/IAAI (1997) K. Hosoda: Coordinationacquired by of a vision-based multiple reinforcementing, learn- Proc. behaviors IEEE/RSJ/GI Int. Conf. Intell.(1994) Robots pp. Syst. 917–924 in a competition forProc.3rdInt.Conf.Simul.Adapt.Behav.(1994) food between ant colonies, pp. 487–492 Multi-robot cooperation inIEEE the Robotics MARTHA Autom. Mag. project, H. Aghazarian, E. Baumgartner:operative Sustainable robotic co- technologiesrobotic for outpost human infrastructuremaintenance, construction and Auton. Robots and interacting simpleranged robots foraging in field.Robotic In: a Systems clockface ar- Conf. Auton.pp. 441–448 Agents Multiagent Syst.vey from (2006) a machineRobots learning perspective, Auton. 53.285 H.R. Everett, R.T. Laird, D.M. Carroll, 53.290 M. Schneider-Fontan, M. Matarić: Territorial 53.286 Y. Guo, L.E. Parker, R. Madhavan: Towards collab- 53.287 C. Hazard, P.R. Wurman, R. D’Andrea: Alphabet 53.288 D. Nardi, A. Farinelli, L. Iocchi: Multirobot sys- 53.289 K. Passino: Biomimicry of bacterial foraging for 53.291 I. Wagner, M. Lindenbaum, A.M. Bruckstein: Mac 53.293 P. Rybski, S. Stoeter, C. Wyman, M. Gini: A cooper- 53.281 M. Asada, E. Uchibe, S. Noda, S. Tawaratsumida, 53.282 M. Kubo, Y. Kakazu: Learning coordinated motions 53.283 R. Alami, S. Fleury, M. Herrb, F. Ingrand, F. Robert: 53.284 A. Stroupe, A. Okon, M. Robinson, T. Huntsberger, 53.292 K. Sugawara, M. Sano: Cooperative behavior of 53.280 P. Stone, M. Veloso: Multiagent systems: A sur- Moving in the Environment

Part E

Part E | 53 1376 Multiple Mobile Robot Systems References 1377

53.312 K. Kosuge, T. Oosumi: Decentralized control of 53.328 J. Werfel, Y. Bar-Yam, D. Rus, R. Nagpal: Dis- 53 | E Part multiple robots handling an object, Proc. IEEE/RSJ tributed construction by mobile robots with en- Int. Conf. Intell. Robots Syst. (IROS) (1996) pp. 318– hanced building blocks, IEEE Int. Conf. Robotics 323 Autom. (2006) pp. 2787–2794 53.313 N. Miyata, J. Ota, Y. Aiyama, J. Sasaki, T. Arai: Co- 53.329 J. Werfel, R. Nagpal: Extended stigmergy in col- operative transport system with regrasping car- lective construction, IEEE Intell. Syst. 21(2), 20–28 like mobile robots, Proc. IEEE/RSJ Int. Conf. Intell. (2006) Robots Syst. (IROS) (1997) pp. 1754–1761 53.330 J. Werfel, R. Nagpal: Three-dimensional con- 53.314 C.R. Kube, E. Bonabeau: Cooperative transport by struction with mobile robots and modular blocks, ants and robots, Robotics Auton. Syst. 30(1), 85– Int. J. Robotics Res. 27(3/4), 463–479 (2008) 101 (2000) 53.331 Y. Terada, S. Murata: Automatic modular assembly 53.315 R. Groß, M. Dorigo: Towards group transport by system and its distributed control, Int. J. Robotics swarms of robots, Int. J. Bio-Inspired Comput. Res. 27, 445–462 (2008) 1(1), 1–13 (2009) 53.332 J. Wawerla, G.S. Sukhatme, M.J. Mataric: Col- 53.316 L.E. Parker: ALLIANCE: An architecture for fault lective construction with multiple robots, Proc. tolerant, cooperative control of heterogeneous IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS) (2002) mobile robots, Proc. IEEE/RSJ/GI Int. Conf. Intell. pp. 2696–2701 Robots Syst. (1994) pp. 776–783 53.333 R.L. Stewart, R.A. Russell: Building a loose wall 53.317 B. Donald, J. Jennings, D. Rus: Analyzing teams structure with a robotic swarm using a spa- of cooperating mobile robots, Proc. IEEE Int. Conf. tio-temporal varying template, Proc. IEEE/RSJ Int. Robotics Autom. (1994) pp. 1896–1903 Conf. Intell. Robots Syst. (IROS) (2004) pp. 712– 53.318 Y. Mohan, S.G. Ponnambalam: An extensive re- 716 view of research in swarm robotics, World Congr. 53.334 R.L. Stewart, R.A. Russell: A distributed feedback Nat. Biol. Insp. Comput. (2009) pp. 140–145 mechanism to regulate wall construction by a 53.319 A.J. Ijspeert, A. Martinoli, A. Billard, L.M. Gam- robotic swarm, Adapt. Behav. 14(1), 21–51 (2006) bardella: Collaboration through the exploitation 53.335 T. Huntsberger, G. Rodriguez, P. Schenker: of local interactions in autonomous collective Robotics challenges for robotic and human robotics: The stick pulling experiment, Auton. Mars exploration, Proc. Robotics 2000 (2000) Robots 11(2), 149–171 (2001) pp. 340–346 53.320 A. Martinoli, K. Easton, W. Agassounon: Modeling 53.336 R.A. Brooks, P. Maes, M.J. Mataric, G. More: Lunar swarm robotic systems: A case study in collabo- based construction robots, Proc. IEEE Int. Work- rative distributed manipulation, Int. J. Robotics shop Intell. Robots Syst. (IROS) (1990) pp. 389–392 Res. 23(4-5), 415–436 (2004) 53.337 C.A.C. Parker, H. Zhang, C.R. Kube: Blind bull- 53.321 S. Berman, Q. Lindsey, M.S. Sakar, V. Kumar, dozing: Multiple robot nest construction, Proc. S.C. Pratt: Experimental study and modeling of IEEE/RSJ Int. Conf. Intell. Robots Syst. (IROS) (2003) group retrieval in ants as an approach to collec- pp. 2010–2015 tive transport in swarm robotic systems, Proc. IEEE 53.338 S. Sen, M. Sekaran, J. Hale: Learning to coor- 99(9), 1470–1481 (2011) dinate without sharing information, Proc. AAAI 53.322 J.M. Esposito: Distributed grasp synthesis for (1994) pp. 426–431 swarm manipulation with applications to au- 53.339 B. Tung, L. Kleinrock: Distributed control meth- tonomous tugboats, IEEE Int. Conf. Robotics Au- ods, Proc. 2nd Int. Symp. High Perform. Distrib. tom. (2008) pp. 1489–1494 Comput. (1993) pp. 206–215 53.323 Z. Wang, V. Kumar: Object closure and manip- 53.340 Z.-D. Wang, E. Nakano, T. Matsukawa: Cooper- ulation by multiple cooperating mobile robots, ating multiple behavior-based robots for object Proc. IEEE Int. Conf. Robotics Autom. (ICRA) (2002) manipulation, Proc. IEEE/RSJ/GI Int. Conf. Intell. pp. 394–399 Robots Syst. (IROS) (1994) pp. 1524–1531 53.324 Z. Wang, Y. Hirata, K. Kosuge: Control a rigid 53.341 D. Rus, B. Donald, J. Jennings: Moving furniture caging formation for cooperative object trans- with teams of autonomous robots, Proc. IEEE/RSJ portation by multiple mobile robots, Proc. of IEEE Int. Conf. Intell. Robots Syst. (1995) pp. 235–242 Int. Conf. Robotics Autom. (ICRA) (2004) pp. 1580– 53.342 F. Hara, Y. Yasui, T. Aritake: A kinematic analysis 1585 of locomotive cooperation for two mobile robots 53.325 J. Fink, N. Michael, V. Kumar: Composition of along a general wavy road, Proc. IEEE Int. Conf. vector fields for multi-robot manipulation via Robotics Autom. (1995) pp. 1197–1204 caging, Robotics Sci. Syst. Conf. (2007) 53.343 J. Sasaki, J. Ota, E. Yoshida, D. Kurabayashi, T. Arai: 53.326 F. Arrichiello, H.K. Heidarsson, S. Chiaverini, Cooperating grasping of a large object by multi- G.S. Sukhatme: Cooperative caging and transport ple mobile robots, Proc. IEEE Int. Conf. Robotics using autonomous aquatic surface vehicles, In- Autom. (1995) pp. 1205–1210 tell. Serv. Robotics 5(1), 73–87 (2012) 53.344 C. Jones, M.J. Matarić: Automatic synthesis of 53.327 J. Werfel: Building patterned structures with communication-based coordinated multi-robot robot swarms, Proc. 19th Int. Jt. Conf. Artif. Intell. systems, Proc. IEEE/RSJ Int. Conf. Intell. Robots (IJCAI) (2005) pp. 1495–1502 Syst. (IROS) (2004) pp. 381–387 (2), 283– , 217–226 9 24 (2), 89–99 Distributed 41 (2/3), 133–143 17(2/3), 111–113 29 (Springer, Berlin, ,477–521(1987) 2 (1), 73–86 (1997) (2), 67–86 (2001) 18 , 3–12 (2000) , 1253–1256 (2006) 36 4 94 , 219–229 (1998) 22(1), 39–52 (2001) 25 Robot Teams: From Diversity to Poly- , ed. by T. Balch, L.E. Parker (A K Peters, (3), 207–208 (2000) , 33–52 (2005) 8 110(2), 241–273 (1999) 219 (18), 685–699 (2002) 1999) scalable strategy for coordinating multiplewithin robots roadmaps, IEEE Trans. Robotics ing objects, Algorithmica motion coordination in spaceAuton. and Syst. time, Robotics optimizing solvable priority schemespled path for planning decou- techniquesbiel for teams robots, of mo- Robotics Auton. Syst. (2002) awa, H. Matasubara: RoboCup:lem A of challenge AI, prob- AI Magazine challenge for multiagent and intelligent systems, AI Magazine Skills, tactics andin plays adversarial for environments, IEEE multi-robotEng. J. control Control Syst. robotic soccer team:control, Perception Robotics and( Auton. multiagent Syst. B. Nebel: CS Freiburg: Coordinating robotscessful for soccer suc- playing, IEEE Trans. Robotics Autom. 5 role assignemnt, andnicaiton low-bandwidth for commu- real-timeIntell. strategic teamwork, Artif. gio: Coordination in multi-agent RobocupRobotics teams, Auton. Syst. tributed autonomous mobileAuton. Robotic Syst. robotics, Distrib. munication methodologies fortems, multi-robot sys- Intell.(2003) Autom. Soft Comput. on heterogeneous multi-robotRobots systems, Auton. on swarm robotics, Auton.(2004) Robots systems, Proc. IEEE reward. In: morphism Natick 2002) Autonomous Robotic Systems 4 Heidelberg 2000) 292 (2008) 53.362 M. Peasgood, C. Clark, J. McPhee: A complete and 53.363 M. Erdmann, T. Lozano-Perez: On53.364 multiple mov- C. Ferrari, E. Pagello, J. Ota,53.365 T. Arai: Multirobot M. Bennewitz, W. Burgard, S. Thrun: Finding and 53.366 H. Kitano, M. Asada, Y. Kuniyoshi, I.53.367 Noda, E. Os- H. Kitano, S. Tadokoro: RoboCup rescue: A grand 53.368 B. Browning, J. Bruce, M. Bowling, M. Veloso: STP: 53.369 M. Veloso, P. Stone, K. Han: The CMUnited-97 53.370 T. Weigel, J.-S. Gutmann, M. Dietl, A.53.371 Kleiner, P. Stone, M. Veloso: Task decomposition, dynamic 53.372 C. Candea, H.S. Hu, L. Iocchi, D.53.373 Nardi, M. Piag- L.E. Parker: Current state53.374 of the K.R. Baghaei, art A. Agah: in Task allocation dis- and com- 53.375 T. Balch, L.E. Parker: Guest editorial, special issue 53.376 M. Dorigo, E. Sahin: Guest editorial, special issue 53.377 M. Veloso, D. Nardi: Special issue on multirobot 53.378 T. Balch: Taxonomies of multi-robot task and 53.379 L.E. Parker, G. Bekey, J. Barhen (Eds.): , 23 (1), 5– 5 ,ed.by Encyclo- , 491–497 (1988) 4 (4), 247–263 (2009) 2 (3), 191–205 (2002) ystem Science 13 (5), 898–908 (2005) : Coordinating autonomous 21 mplexity and S observation, Intell. Autom. Soft Comput. 125–152 (1998) sensor fusion formulti-robot systems, object Proc. IEEE Int. position Conf. Robotics Autom. estimation (ICRA) by (2001) pp. 1092–1098 decentralized algorithms forobservation, Proc. cooperative 4th Int. target Jt.Multiagent Conf. Syst. Auton. (2005) Agents pp. 911–917 C. Latombe: Motionvisibility of strategies a for movingRobotics target, Autom. maintaining Proc. (1997) pp. IEEE 731–736 Int. Conf. tralized motion coordination strategy for dynamic target tracking, Proc. IEEE Int.tom. Conf. (ICRA) Robotics (2006) Au- pp. 2416–2422 surveillance of multiple movingbehavioral targets approach, Proc. – IEEE Int. A Conf. Robotics new Autom. (ICRA) (2006) pp. 1311–1316 tiple mobile robots:occlusion, The Auton. Robots effect of environment titarget surveillance withIEEE mobile Trans. sensor Robotics agents, and practicalrobots, pursuit-evasion Intell. Serv. with Robotics networked tion in multiple mobile robot teams. In: pedia of Co R.A. Meyers (Springer, Berlin, Heidelberg 2009) hicles, IEEE J. Robotics Autom. and centralized decision makingoperative to behaviors achieve co- betweenrobots, Proc. IEEE/RSJ Int. multiple Conf. Intell.(1992) Robots mobile Syst. pp. 1566–1574 for many-AGV systems, IEEE Int.Robots Workshop Syst. Intell. (1991) pp. 1199–1204 in cellular robotic systems,shop Intell. Proc. Robots IEEE Syst. (1990) Int. pp. Work- 819–826 ning for multiple robots, Robotics Auton. Syst. M. Yim: Indoorrobots, automation IEEE with Int. many(IROS) (1990) Workshop mobile pp. 67–72 Intell. Robots Syst. Y. Ishida, I.multiple mobile Endo: robots based on Collisionmunication, rules Proc. IEEE/RSJ and avoidance Int. com- Conf. Intell. among Robots Syst. (IROS) (1991) pp. 1215–1220 19 (1999) 53.345 L.E. Parker: Cooperative robotics for multi-target 53.346 A.W. Stroupe, M.C. Martin, T. Balch: Distributed 53.347 S. Luke, K. Sullivan, L. Panait, G. Balan: Tunably 53.348 S.M. LaValle, H.H. Gonzalez-Banos, C. Becker, J.- 53.349 T.H. Chung, J.W. Burdick, R.M. Murray: A decen- 53.350 A. Kolling, S. Carpin: Multirobot cooperation for 53.351 B. Jung, G. Sukhatme: Tracking targets using mul- 53.352 Z. Tang, U. Ozguner: Motion planning53.353 for mul- M.A. Vieira, R. Govindan, G.S. Sukhatme: Scalable 53.354 L.E. Parker: Path planning and motion coordina- 53.355 D. Grossman: Trafficcontrol multiple of ve- robot 53.359 J. Wang: Fully distributed traffic controlstrategies 53.360 J. Wang, G. Beni: Distributed computing problems 53.361 P. Svestka, M. Overmars: Coordinated path plan- 53.356 P. Caloud, W. Choi, J.-C. Latombe, C. Le53.357 Pape, H. Asama, K. Ozaki, H. Itakura, A. Matsumoto, 53.358 S. Yuta, S. Premvuti Moving in the Environment

Part E

Part E | 53 1378 Multiple Mobile Robot Systems References 1379

53.380 H. Asama, T. Arai, T. Fukuda, T. Hasegawa 53.386 A. Schultz, L.E. Parker, F. Schneider (Eds.): 53 | E Part (Eds.): Distributed Autonomous Robotic Systems 5 Multi-Robot Systems Volume II: From Swarms (Springer, Berlin, Heidelberg 2002) to Intelligent Automata (Kluwer, Dordrecht 53.381 R. Alami, R. Chatila, H. Asama (Eds.): Distributed 2003) Autonomous Robotic Systems 6 (Springer, Berlin, 53.387 E. Sahin, W.M. Spears (Eds.): Swarm Robotics: SAB Heidelberg 2006) 2004 Int. Workshop (Springer, Berlin, Heidelberg 53.382 M. Gini, R. Voyles (Eds.): Distributed Autonomous 2004) Robotic Systems 7 (Springer, Berlin, Heidelberg 53.388 L.E. Parker, F. Schneider, A. Schultz (Eds.): Multi- 2006) Robot Systems Volume III: From Swarms to Intel- 53.383 H. Asama, H. Kurokawa, J. Ota, K. Sekiyama: Dis- ligent Automata (Kluwer, Dordrecht 2005) tributed Autonomous Robotic Systems 8 (Springer, 53.389 T. Balch, L.E. Parker (Eds.): Robot Teams: From Berlin, Heidelberg 2009) Polymorphism to Diversity (A K Peters, Natick 53.384 A. Martinoli, F. Mondada, N. Correll, G. Mermoud, 2002) M. Egerstedt, M.A. Hsieh, L.E. Parker, K. Stoy 53.390 F. Bullo, J. Cortés, S. Martinez: Distributed Control (Eds.): Distributed Autonomous Robotic Systems, of Robotic Networks: A Mathematical Approach to Springer Tracts in Advanced Robotics, Vol. 83 Motion Coordination Algorithms (Princeton Uni- (Springer, Berlin, Heidelberg 2013) versity Press, Princeton 2009) 53.385 A. Schultz, L.E. Parker (Eds.): Multi-Robot Systems: 53.391 S. Kernbach (Ed.): Handbook of Collective From Swarms to Intelligent Automata (Kluwer, Robotics: Fundamentals and Challenges (CRC, Dordrecht 2002) Boca Raton 2013)