You are on page 1of 12

Advanced Engineering Informatics 20 (2006) 59–70

www.elsevier.com/locate/aei

Multi-agent robot systems as distributed autonomous systems


Jun Ota*
Department of Precision Engineering, The University of Tokyo, 7-3-1 Hongo, Bunkyo-ku, Tokyo 1138656, Japan
Received 25 March 2005; revised 17 June 2005; accepted 24 June 2005

Abstract
In the numerous existing studies dealing with multi-agent robot systems, the systems are positioned on the crossover area of robotics and
distributed autonomous systems. Multi-agent robots perform many tasks, which are classified into six types according to the dimension of the
goal state and the number of iterations of the tasks. This paper surveys earlier studies on multi-agent robots for each type, such as multi-robot
motion-planning algorithms and exploration algorithms of multiple robots. The tasks that multi-agent robots can perform are becoming
increasingly more complex as they move from single, one-time tasks to those involving many iterations. This study is an investigation of the
current trends and the potentials for multi-agent robot systems.
q 2005 Elsevier Ltd. All rights reserved.

Keywords: Multi-agent systems; Robotics; Distributed systems

1. Introduction multi-agent systems (distributed autonomous systems), the


above-mentioned characteristics can also apply to multi-
The environments in which robots operate include agent robot systems.
factories, offices, and homes. Because various tasks However, multi-agent robot systems have the following
are required of robots, cooperation becomes important. special characteristics: (s1) since a robot has its own physical
Currently, various studies have been undertaken on the issue parts, including sensors, processors, actuators, and
of the accomplishment of tasks by multi-agent robots. communication devices, its limitations must be considered,
The characteristics of distributed autonomous systems or such as those related to calculation power, sensing errors, and
multi-agent systems were described as follows in Ref. [1]: errors in actuating and communicating devices; and (s2)
(1) incomplete information or capabilities for solving the robots move in actual environments; therefore, geometry must
problem in each agent, (2) no system global control, be considered. In other words, physical interaction, such as
(3) decentralized data, and (4) asynchronous computation. collision avoidance and the manipulation of physical objects
One more important aspect to be added to the above list is by the cooperation of robots, must be taken into account.
that, although systems are evaluated according to the Several reviews on these issues are currently available.
performance of multi-agents, group performance cannot For example, Ref. [2] presents a review of artificial systems,
be programmed directly. Designs are limited to the multi-agent robot systems, and software agents as well as
individual performance of each agent. The performance of natural systems, such as human and animal groups. In
multi-agents emerges from that of individual agents. The Ref. [3], in order to clarify the mechanism of cooperation
gap between the evaluated object (the behavior of multi- among multiple robots, five key research areas applying to
agents) and the designed object (the behavior of an all of the reviewed papers were examined. They included
individual agent) makes it difficult to design a multi-agent group architecture, resource conflict, origin of cooperation,
system. Because multi-agent robot systems are part of learning, and geometry problems. Ref. [4] focused on three
topics as he surveyed existing studies: arbitration among
agents, communication, and dynamic configurable
* Tel.: C81 35 841 6456; fax: C81 35 841 8547. architecture.
E-mail address: ota@prince.pe.u-tokyo.ac.jp In this review, the focus is on the task assigned to the
1474-0346/$ - see front matter q 2005 Elsevier Ltd. All rights reserved. robots. Multi-agent robots deal with many kinds of tasks.
doi:10.1016/j.aei.2005.06.002 The task specification determines the level and kind of
60 J. Ota / Advanced Engineering Informatics 20 (2006) 59–70

intelligence that robots need. Although many method- Table 1


ologies have been proposed for some specific tasks or Task classification of multi-agent robots
general task frameworks, an overall definition of robotic One-time Many-times
tasks is still lacking. It has not been determined what types Point-reaching Motion-planning Coming and going
of tasks multiple robots should perform or what is Cooperative handling of a between two positions
appropriate for robots. In order to overcome the current large object
deficiency, in this paper, we (a) classify tasks from two Pattern formation
Region-sweeping Sweeping Periodical cooperative
viewpoints and (b) survey previous studies on the basis of Map generation sweeping
the classification obtained in (a), describing what kind of Compound Cooperative transpor- Collecting objects/
knowledge robots need to create to realize these tasks, and tation in unknown foraging
(c) we discuss future trends for multi-agent robot systems. environments Robot soccer
It is natural to consider that behavior is essential for
robots. Here, the robot tasks are categorized according to the Compound tasks are those in which the two types of tasks
various dimensions of the tasks and the numbers of tasks referred to above are combined.
that are to be performed.
1.2. Number of iterations of tasks: one-time/many-times
1.1. Dimension of the goal state: point-reaching tasks
(zero-dimension)/region-sweeping tasks (more than Ordinary tasks should be performed once only. Some
one-dimension)/compound tasks tasks, however, must be performed many-times. Here,
‘many-times’ means that we expect the effect of adaptation,
Point-reaching tasks are those in which the target is learning, and self-organization to occur as a result of
expressed with a specific goal configuration of the robot. several trials.
The goal configuration is expressed as a specific point in the
configuration space of the robots. Here, the configuration of 1.3. Task classification
the robot is a specification of the position of every point in
the robot in a certain coordinate system. The configuration Tasks can be classified into six (Z3!2) by a
space is the space of the entire configuration of the robot [5]. combination of the two categories given above. Examples
Standard motion-planning tasks of multi-agent robots are of tasks are shown in Table 1. Most multi-agent robotic
grouped into this category. Here, the word ‘dimension’ is tasks are included in this table. Tasks in the same class are
used to compare the expressive form of the goal, that is, the similar in that the same behavior is required of the robots.
goal point (which is zero-dimensional in the configuration Former studies in this field are surveyed on the basis of
space). In the point-reaching task, the robot’s trajectory is these classes.
generated by connecting the robot’s start and goal Many studies have researched the multi-agent robotic
configurations. In general, a robot’s trajectory is expressed tasks in Table 1. For example, there are studies on
with a space curve and parameterized with an arc length cooperative manipulation and transportation, such as those
(one parameter). Hence, we can regard the dimension of a by Refs. [6–18]. An example is shown in Fig. 1, in which
robot’s trajectory as one. The point-reaching task is to four circular robots are tasked with transporting one large
generate a one-dimensional solution (the robot’s trajectory) rectangle-shaped object cooperatively. The main topic in
to accomplish the zero-dimensional goal. this field is how to achieve stable and robust transportation
Region-sweeping tasks are those in which the target is in environments that involve errors in sensing, actuating,
expressed with a specific goal region. The target is to and communication. Other important topics are the
generate the trajectory of the robot covering the required development of self-reconfigurable robots [19–24], in
region with the robot’s sensing area, i.e. to generate a one- which many small robotic units gather and connect
dimensional solution (trajectory) to complete a multi- dynamically depending on the demands, and the achieve-
dimensional goal. The term ‘multi-dimensional goal’ ment of cooperative localization [25–31], in which several
indicates that the goal is expressed as a certain region that robots maintain formation to know their location within
has more than one-dimension. Sweeping or exploring tasks
are included in the second category. For a sweeping task of a
flat region, the dimension is two.
It is noteworthy that, when the dimension of the target is
equal to one, the problem becomes a control problem rather
than the task realization problems addressed here. An
example of a control problem is a tracking task on a given
trajectory by a mobile robot. This type of task will not be
discussed in this paper. Fig. 1. An example of cooperative transportation.
J. Ota / Advanced Engineering Informatics 20 (2006) 59–70 61

their environment without the use of landmarks. The last communication is made when entering/leaving the area,
two papers mentioned above deal with the multi-robot where two robots intersect.
Simultaneous Localization and Mapping (SLAM) problem. The coordination diagram method consists of two
The standard SLAM problem asks if it is possible for a steps: (1) each robot generates its own trajectory, and (2)
mobile robot to start in an unknown location in an unknown each robot determines a velocity profile to avoid collisions
environment and then incrementally build a map of this among the robots. Ferrari et al. applied the above
environment while simultaneously using this map to methodology [41]. The calculation cost could be smaller
compute absolute vehicle location [32]. The structure of by utilizing the anytime algorithm, whose quality of
multi-agent robot systems is similar to that used in results improves gradually as the computation time
Distributed Artificial Intelligence (for example, [33,34]) increases [42].
because it deals with the intelligence of many agents Lavalle et al. presented a method that can compute the
interacting in the same domain. solutions that simultaneously optimize an independent
performance index for each robot [43]. This is a kind of
multi-objective optimization problem, and the method is
2. ‘One-time’ and ‘point-reaching’ tasks based on dynamic programming. Siméon et al. offered the
planning algorithm for multiple mobile robots with fixed
There are many studies of one-time and point-reaching paths based on a bounding-box representation of the
tasks. One of the most fundamental topics is ‘motion- obstacles [44]. The method can solve the problem of more
planning of multi-agent robots’, in which individual robots than 100 robots in practical cases.
reach goal configurations from initial configurations Yoshioka et al. introduced a sensor-based path-planning
while avoiding each other. Some examples are shown in methodology for multi-agent robots [45]. Sensor-based
Fig. 2. In Fig. 2, four robots try to move to their destinations path-planning is a method in which no robot uses a map of
as quickly as possible while avoiding other robots and static the working environments but, rather, makes use of the
obstacles. knowledge (rules) about the relationship between sensor
Several approaches exist to solve the motion-planning information and action. For example, robots ‘go straight to
problem. Erdmann et al. introduced the priority method [35]. the goal’ when nothing is in front and ‘turn to the right’
Here, priority is added to each robot in advance by utilizing when they encounter something in front of them.
the initial and goal configurations of the robots. The plan is Arai et al. presented a method using the concept of the
obtained from the robot with the highest priority. The distributed artificial potential field [46], in which robots
problem can be solved by iterating a single robot motion- avoid each other by receiving a virtual repulsive force from
planning N times, where N is the number of robots. Planning other robots and obstacles while obtaining a virtual
is executed with a visibility graph [36] in configuration attractive force from each goal position. This method is
space-time [37]. Ref. [38] adopted the same kind of approach similar to the motion-generation algorithm of flocks/herds in
when navigating with multiple all-terrain vehicles. the field of computer graphics called Boids [47]. Ota et al.
Azarm et al. proposed a dynamic priority method [39]. extended algorithm in Ref. [46] and actualized a feasible
With this method, as several robots are about to collide, a motion generation by introducing a real-time search
motion plan is prepared for these robots while considering methodology [48]. Ref. [49] offers a detailed survey of
all possible combinations of priorities among the robots. multi-agent motion-planning.
The plan with the least cost is then selected. Finally, each From the application side, the problem is similar to the
robot follows this plan. motion-planning problem of Automated Guided Vehicles
Ref. [40] proposed the distributed motion-planning (AGVs) in factory environments (Fig. 3). In Fig. 3, many
method with the concept of ‘plan merging’. In order to mobile robots (AGVs) in factory environments transport
assure synchronization among nearby robots, mutual objects from one place to another. In ordinal cases, magnetic

Fig. 2. Motion-planning problem of multi-agent robots.


62 J. Ota / Advanced Engineering Informatics 20 (2006) 59–70

Fig. 3. AGV transport system in factory environments.

tapes are attached to the factory floor for the AGVs to There are other tasks included in this category.
follow. Yamashita et al. introduced a motion-planning method of
AGV systems are an example of the application of multi- multiple mobile robots for the cooperative transportation of
agent robot systems in factory environments. AGV control a large object [54]. In Fig. 4, the motion of the robots and the
problems are divided into two types: scheduling and routing objects should be adequately considered. Here, the
[50]. The scheduling problem is to determine what AGV orientation of the object should be mapped depending on
should be assigned to a given task (for example, transport the environmental conditions (for example, the object
Object A from point B to point C). The routing problem is to should be placed vertically as it is transported in a narrow
determine which route should be selected for a certain AGV corridor). Other important considerations are that the
to achieve the assigned task. The latter problem is similar to sensing and actuating errors of mobile robots should be
this class. Several routing algorithms have been proposed taken into account in the planning phase. In order to tackle
(for example, [51–53]). The timing of the holding and the high dimensionality of the problem, the motion planner
release of the resource (working space) in the vicinity of is divided into a global path planner and a local
AGV travel is the focus of this study. manipulation planner. In the global planner, the motion of
Following are the differences between the multi-agent the object is generated by using A* search to assure the
motion-planning problem and the AGV motion-planning optimality of the solution (for an explanation of A* search,
problem: (a) robots in the former move freely within the please see textbooks on artificial intelligence, such as
working area. On the other hand, robots in the latter are Ref. [55]). In the local planner, a position-based motion plan
restricted to the network established by the magnetic tapes. (b) of robot manipulators is made to consider the position errors
In the former, the problem is settled when all the robots reach of the robots.
their goal configurations. In the latter, there are many transport Ref. [56] solved the problem of the ‘coordinated target
tasks. Here, AGVs need to move within the path network. assignment problem’, in which several unmanned air

Fig. 4. Cooperative transportation of large objects.


J. Ota / Advanced Engineering Informatics 20 (2006) 59–70 63

3. ‘Many-time’ and ‘point-reaching’ tasks

When a certain task has to be executed many-times, it


becomes important to generate a group formation of robots
based on the learning capacity of individual robots. Several
studies have been conducted on this topic. Hand-to-hand
delivery motion was introduced to avoid deadlocks [69].
This motion is compatible with object transportation and
mutual avoidance of two robots. Yoshimura et al. settled an
iterative transportation task, as shown in Fig. 6, in which the
target of the task is to transport as many objects as possible
from the start to the goal positions in an unknown
environment [70]. All robots should explore the environ-
ments efficiently and transport objects while avoiding
Fig. 5. An example of formation control.
recognized obstacles and other robots. The learning capacity
vehicles (UAVs) are assigned to go to different target of robots for generating adequate group formation is shown
locations. A hierarchical approach has been offered to solve depending on both the characteristics of the objects (cost of
the following sub-problems: task assignment, coordinated handling over the object between two robots) and the
UAV intercept, path-planning, trajectory generation, and disposition of obstacles. The final group formation of robots
asymptotic trajectory following. is determined by taking into account the variation of the cost
Many studies have focused on the control of robot of handing over the object. When the cost is small, it
formation, in which several robots are tasked to go to their converges to a ‘relay type’, in which the objects are
destination while maintaining a specific physical arrange- transported by the hand-to-hand delivery motions of
ment. In this problem, each robot is to reach the goal individual robots. When the cost is large, the formation
configuration from the initial configuration, as shown in converges into a ‘loop’, in which each robot transports an
Fig. 5. The difficulty of satisfying the two demands entails object from the start to the goal position within a loop.
reaching the goal as quickly as possible and maintaining the Inoue et al. extended the above methodology to
desired formation. The following studies have been dynamically changing environments [71]. A three-phase
published on this topic [57–67]. The main issues discussed methodology was proposed. It includes an environmental
in this field were: (a) completeness of the algorithm, exploration phase, a path-generation phase, and a strategic
(b) stabilizing ability against many kinds of errors, and phase. They considered a trade-off between ‘searching for
(c) sensing and communication locality in control. unknown environments’ and ‘transport efficiency in
Shimoyama et al. presented a mathematical model of recognized environments’. Kawabata et al. addressed the
collective motion of multi-agents [68]. They modeled the emergence of communication in an iterative multi-robot
behavior with a differential equation and showed that the navigation problem [72]. They discussed methods for
equation can produce various kinds of dynamic patterns robots to learn appropriate actions, including types of
by introducing a set of dimensionless parameters: communication that are not directed but decided by
(a) marching, (b) oscillatory (wavy), (c) oscillatory the robots to adapt to an environment. They demonstrated
(circling), (d) oscillatory (spiral), (e) wandering, and that, using this approach, a proper communication
(f) swarming. procedure results.

Fig. 6. Iterative transportation task.


64 J. Ota / Advanced Engineering Informatics 20 (2006) 59–70

Fig. 7. Region-sweeping problem by multi-agent robots.

4. ‘One-time’ and ‘region-sweeping’ tasks [78]. In this problem, each robot is required to localize itself
in the environment. It must then determine where it is and
The region-sweeping (covering) task finishes when the reach an adequate rendezvous point in an unknown
region produced by integrating the sensing area of the robots environment. This is a combination of the cooperative
along the robots’ trajectories becomes identical to that of exploration, SLAM, and landmark selection problems.
the target region (Fig. 7). Here, the robots are assumed to The method was tested with real robots with practical
have their own sensing areas. The tasks are realized by communication limits.
covering a multi-dimensional area with one-dimensional Schneider-Fontán et al. proposed the methodology for
trajectories of robots, as described in Section 1. This is doing a cleanup and collection task [79]. They took a
fundamentally different from point-reaching tasks, in which territorial approach in which the robots were assigned
trajectories are made to reach zero-dimensional targets. The individual territories. In order to adapt to robot malfunc-
problems with these tasks are related to the path-generation tions, the territories can be dynamically resized.
problems of painting robots or those designed for cleaning Rybski et al. designed and built a set of miniature robots
or maintenance. The main issues are (a) dealing with for a surveillance task [80]. They discussed how to achieve
complicated shapes within a region and (b) assigning tasks the task with robots using a limited communication
to each robot. bandwidth and showed the relationship between the
Kurabayashi et al. [73] proposed an off-line sweep number of robots with the same bandwidth and the
planning algorithm for known environments. They intro- performance of a task.
duced both edges of the configuration space and Voronoi In Ref. [81], a distributed algorithm was introduced for
diagram so as to obtain the trajectories of a working area. multi-robot observation tasks of multiple moving targets.
The paths were then adequately divided among the robots These researchers presented a distributed-potential field
based on an evaluation of the cost. approach for achieving real-time behavior, in which each
There are several studies that deal with local communi- robot is attracted to nearby targets and repelled by other
cation networks of the multiple robots. Ichikawa et al. [74] nearby robots.
presented a multi-robot system that creates a communication Jung et al. settled the same kind of problem as that in the
network starting from the initial location area and expanding above-mentioned study [82]. They developed a region-
over a working space. They showed the effectiveness through based approach, which attempts to solve the sensor
manufacturing 10 hardware robots with simple sensors. distribution problem without explicit negotiation among
Yoshida et al. discussed a design methodology of robots’ the robots. Their approach includes two layers: the first one
group behavior and optimal local communication area from distributes robots across regions based on a topological map,
the viewpoint of task realization efficiency [75,76]. and the second tries to maximize the number of tracked
In Ref. [77], the following cycle for exploring the targets in a region.
behavior of robots in an unknown environment is presented: Pheromone-based covering task realization was also
(a) a working environment is divided into cells of the same achieved. Wagner et al. introduced three methods for
size; (b) each robot recognizes the environments by using a covering unknown environments by using pheromones [83].
range sensor and classifies the cells into open cells, The algorithm is more adaptive than the existing search
unknown cells, and occupied cells; (c) any open cell method to dynamic changes in the working environment and
adjacent to an unknown cell is labeled a frontier edge cell. the sensor noises of the robots.
Adjacent edge cells are grouped into frontier regions, and Vidal et al. dealt with the problem that both UAVs and
robots go to the center of each region. unmanned ground vehicles (UGVs) pursue a team of
Roy et al. dealt with a rendezvous problem of two robots, evaders and generate a map [84]. They modeled the
in which the two meet at a certain adequate position in problem as a probabilistic game theoretical framework
unknown environments and unknown starting positions and advanced two pursuit policies that maximize
J. Ota / Advanced Engineering Informatics 20 (2006) 59–70 65

the probability of capturing the evaders with various time in which a Predicate/Transition net is introduced for the
scales: local-max and global-max methods. coordinated protocol.
Payton et al. proposed the concept of a virtual pheromone Parker proposed a fault-tolerant architecture called
by using simple transceivers mounted atop a mobile robot ALLIANCE for heterogeneous robots [92]. This hierarch-
[85,86]. A gas expansion methodology is implemented ical behavior-based architecture, which utilizes mathemat-
among the robots to solve the problem of providing ically modeled motivational behaviors, such as impatience
surveillance for an unknown environment. This can be and acquiescence, constantly monitors the sensory feedback
modeled as a kind of dynamic formation problem. of the tasks. These behaviors are effective for achieving an
adaptive action selection of robots for the environmental
changes and failures of robots.
5. ‘Many-time’ and ‘region-sweeping’ tasks Gerkey et al. presented an auction-based dynamic task
allocation methodology based on the Contract Net Protocol
An example in this class is the patrolling of a [93]. Experimental results for loosely coupled and box-
dynamically changing environment. This problem can be pushing tasks demonstrated that the system is robust for
modeled as a Multiple Traveling Salesmen Problem changes in the environment. Ref. [94] extended the study to
(MTSP) [87] by considering observation points as nodes include a more general architecture. These researchers
in graph theory. We can calculate the observation points by focused on the effects of two key aspects of distributed
utilizing the method of Ref. [77] or other methodologies. control, ‘commitment’ and ‘coordination’. They compared
There are very few studies that have examined this method. four task-allocation strategies to solve an emergency
Trevai et al. advanced a method to generate an handling problem and showed that there is no single
exploration path in a restricted working environment [88]. strategy that produces the best performance in all cases.
The observation points are distributed in the environment, Murphy et al. proposed a hierarchical architecture
and they are calculated by utilizing a reaction-diffusion utilizing a concept of emotions [95]. They applied it to the
equation on a graph proposed in Ref. [89]. The positions of task of serving finger food at a reception, in which waiter
these points can be moved in real-time with respect to the robots serve items to an audience and refiller robots serve
recognized environmental information. By obtaining the refills upon request. The experimental results showed the
result of the MTSP, robots can share their exploration effectiveness of the emotions used from the viewpoint of
regions efficiently within an environment. Because the emergent cooperation.
derived solution is a set of closed-loop trajectories, Miyata et al. developed an architecture for cooperative
the patrolling task is continuous. transport by multiple mobile robots in an unknown
environment [96], as shown in Fig. 8. They consider the
unit task as a job that should be done by one robot, such as,
6. ‘One-time’ and ‘compound’ tasks for example, searching, handling an object, maintaining a
safe distance from humans, or cooperatively transporting a
When the task is to (a) find a specific object whose initial large object to its destination. The unit tasks are assigned to
location is unknown to the robots and (b) transport it to a
desired position, it is categorized as a compound task. This
is because it includes a region-sweeping task (searching the
object in the working environment) and a point-reaching
task (transporting an object to a target position). More
generally, compound tasks consist of several subtasks, such
as a searching subtask or a transporting subtask. Multi-agent
robots share the role of realizing each subtask in real-time.
This can be modeled as a task assignment problem.
In Ref. [90], a review of multi-agent robot systems is
presented with a focus on the problem of a multi-robot task
allocation. These researchers utilized a domain-independent
taxonomy to solve the problem. They dealt with the
relationship between the problem and the relevant theory
in the field of operations research and combinatorial
optimization.
Noreils offered an architecture composed of three levels:
function, control, and planner [91]. Cooperation among
robots is achieved in two phases: (1) collaboration, in which
a task is divided into subtasks; and (2) coordination, Fig. 8. Transport of a large object in an unknown environment.
66 J. Ota / Advanced Engineering Informatics 20 (2006) 59–70

each robot in real-time with a priority-based assignment a finite interaction between robots [107]. Here, instead of
planner by using a linear programming method. utilizing the crumb, the robot that obtains a sample radiates
In Ref. [97], a task-programming framework for light for a short period. After that, it goes home. Other
collective robots is presented. They extended the finite- nearby robots are attracted to the light. In Ref. [108], with a
state automata theory with the concept of the Q-machine mathematical formulation similar to that in Ref. [107], it is
and perceptual cues. The validity of the framework was shown that there is an optimal group size that maximizes
shown by solving a multi-robot box-pushing problem. group performance. Ijspeert et al. proposed a test-bed
Jung et al. proposed a multi-agent planning methodology experiment in which the task of the robots is to pull sticks
using an action selection scheme similar to behavior out of the ground [109]. Here, the cooperation of two robots
networks [98] so that spatial and topological navigation, is necessary in order to pull them out. They applied the
cooperation, and communication could be achieved [99]. probabilistic model to the quantitative evaluation of robot
They used two autonomous robots to solve a task involving behavior. The same kind of task was also dealt with in
cooperative cleaning. They discussed how symbolic Ref. [110].
systems could be used in a multi-agent system [100]. As the architecture for general tasks, in Ref. [111], the
Fierro et al. proposed a software framework for multi- L-ALLIANCE was developed as an extended version of the
agent robot systems [101], including a formal architecture ALLIANCE architecture, including the learning behavior of
and high-level language for programming robot behavior. robots. Here, robots monitor the performance of other robots
They utilized a high-level language named CHARON to working on common tasks and evaluate it on the basis of the
describe the complex behaviors of multi-agent robots. time of task completion. The robots then update their
Brumitt et al. developed a multiple robot system that is control parameters.
capable of achieving multiple goals in natural terrain Several types of architecture have been analyzed for
[102,103]. In their study, they advanced a mission planner RoboCup competitions; some examples are found in the
to assign multiple goals (destinations) to robots dynamically work of Refs. [112–115]. It is natural for the RoboCup task
by solving the Traveling Salesman Problem (TSP) and to be included in this class. However, in the present
Multiple TSP, including a dynamic route planner based on situation, it seems that the architecture presented in the
the D* planner [104]. The architecture was adopted to a real study is unsuccessful in its attempt to introduce learning
robot system. ability in robots. The following factors are discussed in the
multi-agent architecture for RoboCup: real-time decision-
making, restriction of communication range, and sensing/
7. ‘Many-time’ and ‘compound’ tasks actuating error and restriction of calculation power.
In Ref. [116], a reinforcement learning approach was
As previously reported, learning or adaptive ability used for the design of multiple space robots. For space
becomes important in this task class. applications, a fail-safe system is necessary, in particular, to
The task of collecting samples scattered in the working prevent communication mishaps. They introduced a new
environments by many robots in unknown environments has learning method and tested it on a truss construction task.
been thoroughly studied. The task is popular in the field of Human-robot interaction is also a very important topic.
swarm intelligence, where there is no explicit communi- Some studies on the following issues have been conducted:
cation among robots and very simple behavior is expected transportation of unknown large objects by human-
of the robots. In order to complete the task efficiently, an supervised multi-agent robot systems [117–120]
adequate formation of robots is necessary. Such a task and cooperation between men and machines for rescue
could then be called one of the simplest ‘many-time’ and missions [121].
‘compound’ tasks.
In Ref. [105], a behavioral approach was applied to the
above-mentioned problem. The effectiveness of phero-
mones was compared with a purely random-movement 8. Conclusion
method and with a gradient-field method. The pheromone
effect can be realized by embedding the following simple In this paper, we classified tasks performed by multiple
rules into each robot: (a) a robot drops two crumbs when robots based on the dimension of the goal state and the
each carries a sample, and (b) it picks up one crumb when number of iterations they perform. This classification is
neither one carries a sample. This advantage of the chain- made by considering how a motion-planning problem,
making behavior of robots, which is a hand-to-hand delivery which is one of the most fundamental tasks, can be
motion among robots in the same vicinity, was developed in extended. The dimension of the goal state is an extension
Ref. [106]. The performance was shown to be superior. of the geometric dimension. The number of iterations is an
Sugawara et al. showed that the workability of the group extension of the time dimension.
increased drastically when the density of robots exceeded The results of the review, according to the classifications
the critical value under phase-transition behavior with listed above, clarify the following points:
J. Ota / Advanced Engineering Informatics 20 (2006) 59–70 67

† Trends in realized tasks by multi-agent robots are moved This approach has the possibility to solve the problem
from single-task realization (for example, multi-agent because a rough, robust off-line plan-making of each robot
motion-planning and region-sweeping motion gener- ensures efficiency with the advantage of plan generation,
ation) to multiple-task realization and from one-time while on-line re-planning or reactive behavioral control
task realization to multiple-task realization. In other with sensor information can contribute to the two
words, in the earlier stages of this research, specific characteristics. The concept of ‘mobiligence’ [123], which
heuristics for specific tasks were embedded into the is a combination of mobility and intelligence, is also a
robots, which are not enough for general types of tasks. noteworthy element of the design of multi-agent robot
Recently, some studies have proposed task assignment systems. The concept of ‘mobiligence’ was derived from
methodologies for multi-agent robots seeking a general studies by biologists and engineers; it has the following
task description. three characteristics: embodied plasticity, co-embodiment
† Learning, evolution, and real-time performance are with the environment, and internal generation of abductive
becoming important in the field of multiple-task information.
realization. The second theme is a design problem of multi-agent
† Mathematical analyses have been conducted for robot systems. Although more and more complicated tasks
relatively simple tasks, such as foraging. These analyses are being dealt with, there are still many design problems,
can be extended to more general tasks. In order for such as the number of robots and their working environ-
multi-agent robots to accomplish general tasks, robots ments. Such large-scale design problems remain unsolved.
need to have and use knowledge about tasks and working If the tasks and the behavior of the robots are restricted, their
environments. A concept of emergence that will promote behavior can be modeled by utilizing queuing network
cooperation among robots is necessary. It seems that no theory [124]. However, in more general cases, because it is
methodologies have been developed that promote difficult to construct mathematical models for multi-robot
enough emergent behavior among robots; however, behavior, a simulation-based optimization approach [125]
some researchers are investigating the subject. would be preferable. In order to decrease the calculation
cost, the concept of co-evolution, such as that proposed by
Based on the above discussion, the following three Refs. [126,127], will be important.
themes are important as future topics of the multi-agent The last theme is the evaluation of multi-agent robot
robot systems. systems. As shown in this paper, many algorithms have been
The first theme is the performance of multi-agent robot proposed. The tasks deal with algorithms that are slightly
systems. Multi-agent robot systems have been considered different from each other, making it very difficult to
as a distributed autonomous system with two main quantitatively evaluate the algorithms. Benchmark pro-
characteristics: flexibility and tolerance for errors. These blems [49] are one way to address this issue, and the task
characteristics have been the focus of many studies on classification presented in this paper would be essential for
multi-agent robot systems. In addition to flexibility and creating the problems. It is important for multi-agent robots
tolerance for errors, future multi-agent systems will need to to extract subtasks from a given mission. A rearrangement
be efficient, particularly from the viewpoint of engineering. task by multi-agent robots may be one in which each robot
Dwarf intelligence [122] is the intelligence that is produced rearranges all the movable objects from the initial
from a combined architecture of the rough, robust off-line configuration to the goal configuration by grasping and
plan-making of each robot and on-line re-planning or transporting them. An example task is shown in Fig. 9. Here,
reactive behavioral control with sensor information. two mobile robots need to transport eight movable objects to

Fig. 9. An example of a multi-agent rearrangement problem. (a) Initial configuration (b) goal configuration.
68 J. Ota / Advanced Engineering Informatics 20 (2006) 59–70

their destinations. This problem is different from a foraging [18] Chaimowics L, Kumar V, Campos MFM. A paradigm for dynamic
task because the density of the movable objects is so high coordination of multiple robots. Autonomous Robots 2004;17(1):
7–21.
that it is impossible to transport a certain object directly to [19] Fukuda T, Kawauchi Y. Cellular robotic system (CEBOT) as one of
its goal. Instead, they need to set aside other objects before the realization of self-organizing intelligent universal manipulators.
completing the transportation (a single robot case is solved Proc IEEE Int Conf Robotics Automat 1990;662–7.
in Ref. [128]). In other words, task constraints should be [20] Tomita K, Murata S, Kurokawa H, Yoshida E, Kokaji S. Self-
considered. This means that robots should find (1) good assembly and self-repair method for a distributed mechanical
system. IEEE Trans Robotics Automation 1999;15(6):1035–45.
intermediate configurations of objects and (2) the sequence [21] Murata S, Yoshida E, Kurokawa H, Tomita K, Kokaji S. Self-
of each object’s movement in order to complete the task. In repairing mechanical systems. Autonomous Robots 2001;10(1):
such cases, it is desirable for robots to be able to adapt to 7–21.
changes in task specifications, such as the initial configur- [22] Rus D, Vona M. Crystalline robots: self-reconfiguration with
ation of the objects, the goal configuration of the objects, the compressible unit modules. Autonomous Robots 2001;10(1):
107–24.
shape of the working environments, and the number of [23] Yim M, Zhang Y, Duff D. Modular robots. IEEE Spectrum 2002;
robots. 39(2):30–4.
[24] Shen WM, Salemi B, Will P. Hormone-inspired adaptive communi-
cation and distributed control for CONRO self-reconfigurable
robots,. IEEE Trans Robotics Automat 2002;18(5):700–12.
[25] Kurazume R, Nagata S, Hirose S. Cooperative positioning with
References
multiple robots. Proc IEEE Int Conf Robotics Automat 1994;
1250–7.
[1] Sycara KP. Multi-agent systems. AI Mag 1998;19(2):79–92. [26] Fox D, Burgard W, Kruppa H, Thrun S. A probabilistic approach to
[2] Agah A. Robot teams, human workgroups, and animal sociobiology: collaborative multi-robot localization. Autonomous Robots 2000;
a review of research on natural and artificial multi-agent autonomous 8(3):325–44.
systems. Adv Robotics 1996;10(6):523–46. [27] Grabowski R, Navarro-Serment LE, Paredis CJJ, Khosla PK.
[3] Cao YU, Fukunaga AS, Kahng AB. Cooperative mobile robotics: Heterogeneous teams of modular robots for mapping and explora-
antecedents and directions. Autonomous Robots 1997;4(1):7–27. tion. Autonomous Robots 2000;8(1):293–308.
[4] Kurabayashi D. Toward realization of collective intelligence and [28] Thrun S. A probabilistic online mapping algorithm for teams of
emergent robotics (survey). Proc IEEE Int Conf Syst Man Cybern mobile robots. Int J Robotics Res 2001;20(5):335–63.
1999;IV:748–53. [29] Schmitt T, Hanek R, Beetz M, Buck S, Radig B. Cooperative
[5] Latombe JC. Robot Motion Planning. Kluwer: Boston; 1991. probabilistic state estimation for vision-based autonomous mobile
[6] Hashimoto M, Oba F, Eguchi T. Dynamic control approach for robots. IEEE Trans Robotics Automat 2002;18(5):670–84.
motion coordination of multiple wheeled mobile robots transporting [30] Roumeliotis SI, Bekey GA. Distributed multi-robot localization.
a single object. Proc IEEE/RSJ Int Conf Intell Robots Syst 1993; IEEE Trans Robotics Automat 2002;18(5):781–95.
1944–51. [31] Di Marco M, Garulli A, Gianitrapani A, Vicino A. Simultaneous
[7] Stilwell DJ, Bay JS. Towards the development of a material transport localization and map building for a team of cooperating robots: a set
system using swarms of ant-like robots. Proc IEEE Int Conf Robotics membership approach. IEEE Trans Robotics Automat 2003;19(2):
Automat 1993;766–71. 238–49.
[8] Kube CR, Zhang H. Collective robotics: from social insects to [32] Dissanayake MWMG, Neuman P, Clark S, Durrant-Whyte HF,
robots. Adaptive Behav 1994;2(2):189–219. Csorba M. A solution to the simultaneous localization and map
[9] Sasaki J, Ota J, Yoshida E, Kurabayashi D, Arai T. Cooperating building (SLAM) problem. IEEE Trans Robotics Automat 2001;
grasping of a large object by multiple mobile robots. Proc IEEE Int 17(3):229–41.
Conf Robotics Automat 1995;1205–10. [33] Lesser VR. Multi-agent systems: an emerging subdiscipline of AI.
[10] Donald BR, Jennings J, Rus D. Information invariants for distributed ACM Computing Surveys 1995;27(3):340–2.
manipulation. J Robotics Res 1997;16(5):1–30. [34] Stone P, Veloso M. Multiagent systems: a survey from a machine
[11] Ota J, Arai T. Transfer control of a large object by a group of mobile learning perspective. Autonomous Robots 2000;8(3):345–83.
robots. Robotics Autonomous Syst 1999;28(4):271–80. [35] Erdmann M, Lozano-Pérez T. On multiple moving objects.
[12] Khatib O. Mobile Manipulation: the robotic assistant. Robotics Algorithmica 1987;2(4):477–521.
Autonomous Syst 1999;26(2–3):175–83. [36] Nilsson NJ. A mobile automaton: an application of artificial
[13] Aiyama Y, Hara M, Yabuki T, Ota J, Arai T. Cooperative intelligence techniques. Proc 1st Int Joint Conf Artif Intel 1969;
transportation by two four-legged robots with implicit communi- 509–20.
cation. Robotics Autonomous Syst 1999;29(1):13–19. [37] Fujimura K. Motion planning in dynamic environments. Tokyo:
[14] Hirata Y, Kosuge K, Asama H, Kaetsu H, Kawabata K. Coordinated Springer; 1991.
transportation of a single object by multiple mobile robots without [38] Saptharishi M, Oliver CS, Diehl CP, Bhat KS, Dolan JM, Trebi-
position information of each robot. Proc IEEE/RSJ Int Conf Intell Ollennu A, Khosla PK. Distributed surveillance and reconnaissance
Robots Syst 2000;2024–9. using multiple autonomous ATVs: CyberScout. IEEE Trans
[15] Ahmadabadi MN, Nakano E. A. “constrain and move” approach to Robotics Automat 2002;18(5):826–36.
distributed object manipulation. IEEE Trans Robotics Automat [39] Azarm K, Schmidt G. A decentralized approach for the conflict-free
2001;17(2):157–72. motion of multiple mobile robots. Adv Robotics 1997;11(4):323–40.
[16] Sugar TG, Kumar V. Control of cooperating mobile manipulators. [40] Alami R, Fleury S, Herrb M, Ingrand F, Robert F. Multi-robot
IEEE Trans Robotics Automat 2002;18(1):94–103. cooperation in the MARTHA project. IEEE Robotics Automat Mag
[17] Tanner HG, Loizou SG, Kyriakopoulos KJ. Nonholonomic 1998;5(1):36–47.
navigation and control of cooperating mobile manipulators. IEEE [41] Ferrari C, Pagello E, Ota J, Arai T. Multi-robot motion coordination
Trans Robotics Automat 2003;19(1):53–64. in space and time. Robotics Autonomous Syst 1998;25(3-4):219–29.
J. Ota / Advanced Engineering Informatics 20 (2006) 59–70 69

[42] Zilberstein S, Russel SJ. Anytime sensing, planning and action: [66] Feddema JT, Lewis C, Schoenwald DA. Decentralized control of
a practical model for robot control. Proc 13th Int Joint Conf Artif cooperative robotic vehicles: theory and application. IEEE Trans
Intel 1993;1402–7. Robotics Automat 2002;18(5):852–64.
[43] Lavalle SM, Hutchinson SA. Optimal motion planning for multiple [67] Lowton JR, Beard RW, Young BJ. A decentralized approach to
robots having independent goals. IEEE Trans Robotics Automat formation maneuvers. IEEE Trans Robotics Automat 2003;19(6):
1998;14(6):912–25. 933–41.
[44] Siméon T, Leroy S, Laumond JP. Path coordination for multiple [68] Shimoyama N, Sugawara K, Mizuguchi T, Hayakawa Y, Sano M.
mobile robots: a resolution-complete algorithm. IEEE Trans Collective motion in a system of motile elements. Phys Rev Let
Robotics Automat 2002;18(1):42–9. 1996;76(20):3870–3.
[45] Yoshioka T, Noborio H. Sensor-based traffic rules for multiple [69] Fukuda T, Ishihara H, Hiraoka N. Deadlock resolution in distributed
automata based on a geometric deadlock-free characteristic. autonomous robotic system with hand-to-hand motion. Proc IEEE
J Robotics Mechatronics 1996;8(1):48–56. Int Conf Robotics Automat 1996;2049–54.
[46] Arai T, Ota J. Motion planning of multiple mobile robots using [70] Yoshimura Y, Ota J, Inoue K, Kurabayashi D, Arai T. Iterative
virtual impedance. J Robotics Mechatronics 1996;8(1):67–74. transportation planning of multiple objects by cooperative mobile
[47] Reynolds CW. Flocks, herds, and schools: a distributed behavioral robots. In: Asama H, Fukuda T, Arai T, Endo I, editors. Distributed
model. Comput Graphics 1987;21(4):25–34. Autonomous Robotic Systems 2. Tokyo: Springer; 1996. p. 171–82.
[48] Ota J, Arai T, Yoshimura Y, Miyata N, Yoshida E, Kurabayashi D, [71] Inoue K, Ota J, Hirano T, Kurabayashi D, Arai T. In: Lueth T,
Sasaki J, et al. Motion planning of multiple mobile robots by Dillmann R, Dario P, Wörn H, editors. Distributed Autonomous
combination of learned visibility graph and virtual impedance. Adv Robotic Systems 3. Berlin: Springer; 1998. p. 3–12.
Robotics 1996;10(6):605–20. [72] Kawabata K, Asama H, Tanaka M. A study of communication
[49] Arai T, Ota J. Motion planning of multiple mobile robots. Proc emergence among mobile robots: simulation of intention trans-
IEEE/RSJ Int Conf Intel Robots Syst 1992;1992:1761–8. mission. In: Asama H, Fukuda T, Arai T, Hasegawa T, editors.
[50] Qiu L, Hsu WJ, Huang SY, Wang H. Scheduling and routing Distributed Autonomous Robotic Systems 5. Tokyo: Springer; 2002.
algorithm for AGVs: a survey. Int J Prod Res 2002;40(3):745–60. p. 71–80.
[51] Broardbent AJ, Besant CB, Premi SK, Walker SP. Free ranging AGV [73] Kurabayashi D, Ota J, Arai T, Yoshida E. Cooperative sweeping by
systems: promises, problems, and pathways. Proc 2nd Int Conf multiple mobile robots. Proc IEEE Int Conf Robotics Automat 1996;
Automat Mater Handling 1985;2:221–37. 1744–9.
[52] Kim CW, Tanchoco JMA. Conflict-free shortest-time bi-directional [74] Ichikawa S, Hara F. An experimental realization of cooperative
AGV routing. Int J Prod Res 1991;29(12):2377–91. behavior of multi-robot system. In: Asama H, Fukuda T, Arai T,
[53] Taghaboni-Dutta F, Tanchoco JMA. Comparison of dynamic routing Endo I, editors. Distributed Autonomous Robotic Systems. Tokyo:
techniques for automated guided vehicle systems. Int J Prod Res Springer; 1994. p. 224–34.
1995;33(10):2653–69. [75] Yoshida E, Arai T, Yamamoto M, Ota J. Local communication of
[54] Yamashita A, Arai T, Ota J, Asama H. Motion planning of multiple multiple mobile robots: design of optimal communication area for
mobile robots for cooperative manipulation and transportation. IEEE cooperative tasks. J Robotic Syst 1998;7(15):408–19.
Trans Robotics Automat 2003;19(2):223–37. [76] Yoshida E, Arai T, Ota J. Local communication of multiple mobile
[55] Nilsson NJ. Principles of artificial intelligence. Tioga: Palo Alto; robots: design of group behavior for efficient communication. Adv
1980. Robotics 1998;11(8):759–79.
[56] Beard RW, McLain TW, Goodrich MA, Anderson EP. Coordinated [77] Yamauchi B. Decentralized coordination for multi-robot explora-
target assignment and intercept for unmanned air vehicles. IEEE tion. Robotics Autonomous Syst 1999;29(2–3):111–8.
Trans Robotics Automat 2002;18(6):911–22. [78] Roy N, Dudek G. Collaborative robot exploration and rendezvous:
[57] Sugihara K, Suzuki I. Distributed algorithms for formation of algorithms, performance bounds and observations. Autonomous
geometric patterns with many mobile robots. J Robotic Syst 1996; Robots 2001;11(2):117–36.
13(3):127–39. [79] Schuneider-Fontán M, Matarić MJ. Territorial multi-robot task
[58] Balch T. Behavior-based formation control for multi-robot teams. division. IEEE Trans Robotics Automat 1998;14(5):815–22.
IEEE Trans Robotics Automat 1998;14(6):926–39. [80] Rybski PE, Stoeter SA, Gini M, Hougen DF, Papanikolopoulos NP.
[59] Ando H, Oasa Y, Suzuki I, Yamashita M. A distributed memoryless Performance of a distributed robotic system using shared communi-
point convergence algorithm for mobile robots with limited cations channels. IEEE Trans Robotics Automat 2002;18(5):713–26.
visibility. IEEE Trans Robotics Automat 1999;15(5):818–28. [81] Parker LE. Distributed algorithms for multi-robot observation of
[60] Yamaguchi H, Arai T, Beni G. A distributed control scheme for multiple moving targets. Autonomous Robots 2002;12(3):231–55.
multiple robotic vehicles to make group formations. Robotics [82] Jung B, Sukhatme GS. Tracking targets using multiple robots: the
Autonomous Syst 2001;36(4):125–47. effect of environmental occlusion. Autonomous Robots 2002;13(3):
[61] Egerstedt M, Hu X. Formation constrained multi-agent control. 191–205.
IEEE Trans Robotics Automat 2001;17(6):947–51. [83] Wagner IA, Lindernbaum M, Bruckstein AM. Distributed covering
[62] Fredslund J, Matarić MJ. A general algorithm for robot formations by ant-robots using evaporating traces. IEEE Trans Robotics
using local sensing and minimal communication. IEEE Trans Automat 1999;15(5):918–33.
Robotics Automat 2002;18(5):837–46. [84] Vidal R, Shakernia O, Kim HJ, Sim DH, Sastry S. Probabilistic
[63] Vaughan RT, Støy K, Sukhatme GS, Matarić MJ. LOST: pursuit-evasion games: theory, implementation, and experimental
localization-space trails for robot teams. IEEE Trans Robotics evaluation. IEEE Trans Robotics Automat 2002;18(5):662–9.
Automat 2002;18(5):796–812. [85] Payton D, Daily M, Estowski R, Howard M, Lee C. Pheromone
[64] Ögren P, Egerstedt M, Hu X. A control lyapunov function approach robotics. Autonomous Robots 2001;11(3):319–24.
to multi-agent coordination. IEEE Trans Robotics Automat 2002; [86] Payton D, Estkowski R, Howard M. Compound behaviors in
18(5):847–51. pheromone robotics. Robotics Autonomous Syst 2003;44(3-4):
[65] Das AK, Fierro R, Kumar V, Ostrowski JP, Spletzer J, Taylor CJ. A 229–40.
vision-based formation control framework. IEEE Trans Robotics [87] The traveling salesman problem. In: Lawler EL, Lenstra JK,
Automat 2002;18(5):813–25. Rinnoy KAHG, Shmoys DB, editors.. New York: Wiley; 1985.
70 J. Ota / Advanced Engineering Informatics 20 (2006) 59–70

[88] Trevai C, Fukazawa Y, Yuasa H, Ota J, Arai T, Asama H. [109] Ijspeert AJ, Martinoli A, Billard A, Gambardella LM. Collaboration
Exploration path generation for multiple mobile robots using a through the exploitation of local interactions in autonomous
reaction-diffusion equation on a graph. Integr Comput-aided Eng collective robotics: the stick pulling experiment. Autonomous
2004;11(3):195–212. Robots 2001;11(2):149–71.
[89] Yuasa H, Ito M. Self-organizing system theory by use of reaction- [110] Martinoli A, Easton K, Agassounon W. Modeling swarm robotic
diffusion equation on a graph with boundary. Proc IEEE Int Conf systems: a case study in collaborative distributed manipulation. Int
Syst Man Cybern 1999;I:211–6. J Robotics Res 2004;23(4–5):415–36.
[90] Gerkey BP, Matarić MJ. A formal analysis and taxonomy of task [111] Parker LE. Lifelong adaptation in heterogeneous multi-robot teams:
allocation in multi-robot systems. Int J Robotics Res 2004;23(9): response to continual variation in individual robot performance.
939–54. Autonomous Robots 2000;8(3):239–67.
[91] Noreils FR. Toward a robot architecture integrating cooperation [112] Candea C, Hu H, Iocchi L, Nardi D. Piaggio M. Coordination in
between mobile robots: application to indoor environment. Int multi-agent RoboCup teams. Robotics Autonomous Syst 2001;
J Robotics Res 1993;12(1):79–98. 36(2–3):67–86.
[92] Parker LE. ALLIANCE: an architecture for fault-tolerant multi-
[113] Weigel T, Gutmann JS, Dietl M, Keiner A, Nebel B. CS Freiburg:
robot cooperation. IEEE Trans Robotics Automat 1998;14(2):
coordinating robots for successful soccer playing. IEEE Trans
220–40.
Robotics Automat 2002;18(5):685–99.
[93] Gerkey BP, Matarić MJ. Sold!: auction methods for multi-robot
[114] Iocchi L, Nardi D, Piaggio M, Sgorbissa A. Distributed coordination
coordination. IEEE Trans Robotics Automat 2002;18(5):758–68.
in heterogeneous multi-robot systems. Autonomous Robots 2003;
[94] Matarić MJ, Sukhatme GS, Østergaard EH. Multi-robot task
15(2):155–68.
allocation in uncertain environments. Autonomous Robots 2003;
14(2–3):253–63. [115] Pagello E, D’Angelo A, Ferrari C, Polesel R, Rosati R, Speranzon A.
[95] Murphy RR, Lisetti CL, Tardif R, Irish L, Gage A. Emotion-based Emergent behaviors of a robot team performing cooperative tasks.
control of cooperating heterogeneous mobile robots. IEEE Trans Adv Robotics 2003;17(1):3–19.
Robotics Automat 2002;18(5):744–57. [116] Takadama K, Matsumoto S, Nakasuka S, Shimohara K. A
[96] Miyata N, Ota J, Arai T, Asama H. Cooperative transport by multiple reinforcement learning approach to fail-safe design for multiple
mobile robots in unknown static environments associated with real- space robots - cooperation mechanism without communication and
time task-assignment. IEEE Trans Robotics Automat 2002;18(5): negotiation schemes. Adv Robotics 2003;17(1):21–39.
769–80. [117] Nakamura A, Ota J, Arai T. Human-supervised multiple
[97] Kube CR, Zhang H. Task modeling in collective robotics. mobile robot system. IEEE Trans Robotics Automat 2002;18(5):
Autonomous Robots 1997;4(1):53–72. 728–43.
[98] Maes P. Situated agents can have goals. Robotics Autonomous Syst [118] Beltran-Escavy J, Nakamura A, Kakita S, Arai T, Ota J. Human-
1990;6(1):49–70. robot GUI, system with group control for multiple mobile robot
[99] Jung D, Zelinsky A. An architecture for distributed cooperative systems—graphic interface and comparison with previous joystick-
planning in a behavior-based multi-robot system. Robotics Auton- based system. Int J Jpn Soc Prec Eng 1998;32(4):291–6.
omous Systems 1999;26(2–3):149–74. [119] Beltran-Escavy J, Nakamura A, Kakita S, Arai T, Ota J. Human-
[100] Jung D. Zelinsky A.Ground Symbolic Communication between robot GUI, system with group control for multiple mobile robot
Heterogeneous Cooperating Robots. Autonomous Robots 2000;8(3): systems—system monitoring. Int J Jpn Soc Prec Eng 1999;33(2):
269–92. 149–55.
[101] Fierro R, Das A, Spletzer J, Esposito J, Kumar V, Ostrowski JP, [120] Beltran-Escavy J, Nakamura A, Kakita S, Arai T, Ota J. Human-
Pappas G, Taylor CJ, Hur Y, Alur R, Lee I, Grudic G, Southall B, robot GUI, system with group control for multiple mobile robot
et al. A framework and architecture for multi-robot coordination. Int systems—introduction to virtual constraints. Int J Jpn Soc Prec Eng
J Robotics Res 2002;21(10–11):977–95. 1999;33(3):265–70.
[102] Brumitt B, Stentz A, Hebert M. The C.M.U.U.G.V., Group. [121] Murphy RR. Human-robot interaction in rescue robotics. IEEE Trans
Autonomous driving with concurrent goals and multiple vehicles: Syst Man Cybern Part C 2004;34(2):138–53.
mission planning and architecture. Autonomous Robots 2001;11(2): [122] Arai T, Ota J. Dwarf intelligence—a large object carried by seven
103–15.
dwarves. Robotics Autonomous Syst 1996;18(1–2):149–55.
[103] Brumitt B, Stentz A, Hebert M. The C.M.U.U.G.V., Group.
[123] Asama H, Yano M, Tsuchiya K, Ito K, Yuasa H, Ota J, et al.
Autonomous driving with concurrent goals and multiple vehicles:
System principle on emergence of mobiligence and its engineering
experiments and mobility components. Autonomous Robots 2002;
realization. Proc IEEE/RSJ Int Conf Intell Robots Syst 2003;
12(2):135–56.
1715–20.
[104] Stentz A. The focused D* algorithm for real-time replanning. Proc
[124] Hoshino S, Ota J, Shinozaki A, Hashimoto H. Comparison of an
14th Int Joint Conf Artif Intel 1995;14:1652–9.
[105] Steels L. Cooperation among distributed agents through self- AGV transportation system by using the queuing network theory.
organization. In: Demazeau Y, Müller JP, editors. Decentralized Proc IEEE/RSJ Int Conf Intell Robots Syst 2004;3785–90.
AI. North Holland: Elsevier; 1990. p. 175–96. [125] Gosavi A. Simulation-based optimization. Kluwer: Boston; 2003.
[106] Drogoul A, Ferber J. From Tom Thumb to the dockers: some [126] Chiba R, Ota J, Arai T. Integrated design for routing and network in
experiments with foraging robots. Proc 2nd Int Conf Simul of AGV systems using co-evolution. Proc IEEE Int Conf Robotics, Intel
Adaptive Behav 1992;2:451–9. Syst Signal Process 2003;318–23.
[107] Sugawara K, Sano M. Cooperative acceleration of task performance: [127] Chiba R, Ota J, Arai T. Design of robust flow-path network for AGV
foraging behavior of interacting multi-robot system. Physica D 1997; systems using competitive co-evolution with packaging. Proc
100(3–4):343–54. IEEE/RSJ Int Conf Intell Robots Systems in press.
[108] Lerman K, Galstyan A. Mathematica model of foraging in a group of [128] Ota J. Rearrangement of multiple movable objects— integration of
robots: effect of interference. Autonomous Robots 2002;13(2): global and local planning methodology. Proc IEEE Int Conf
127–41. Robotics Automat 2004;1962–7.

You might also like