You are on page 1of 7



How to control and coordinate teams of cooperating robots
embedded in wireless networks connected to the Internet? Use
communication to reduce uncertainty in robotics, and robotics for
network communication through physical mobility.

ith the explosive growth of work connectivity—have contributed to the phe-
embedded computing hard- nomenal increase in the number of computers in
ware, it is possible to conceive our physical environment. Researchers now increas-
many new networked robotic ingly accept the notion that future appliances (in
applications for diverse offices, transportation, homes, and schools) will be
domains ranging from urban based on a multitude of small embedded computers
industrial and environmental disaster search and with limited (but growing) functionality and net-
rescue to house cleaning. Designing reliable work connections. On their own, they are not phys-
software for such systems is a challeng- ically mobile—an often-ignored
ing problem. But Internet communi- characteristic of these appliances.
cation can facilitate such robotics by Instead, they will depend on
reducing uncertainty while human users for their placement
providing direct user and transport. Our focus here is
input and assistance; the class of embedded systems
robotics facilitate commu- with built-in capacity for
nication by providing physi- autonomous mobility, better
cal mobility at a distance. Here, we known as robots. Introducing these devices into
explore methods for controlling and coordinating environments built primarily for people raises inter-
embedded mobile systems, or robots, interacting esting and challenging questions:
with other computers over wireless networks in
human environments. • What is the best way to control and coordinate
Ubiquitous embedded computing is here to stay ubiquitous robots?
[12]. Information appliances, laptops, palmtops, • How should they be used? What services can
and wearable computers are examples of the first they provide?
wave of this emerging information environment. • How does their software differ from software for
Two factors—Moore’s law [12] and improved net- other embedded systems?

Gaurav S. Sukhatme and Maja J. Matarić

COMMUNICATIONS OF THE ACM May 2000/Vol. 43, No. 5 67

Figure 1. Example scenario: These robots are on a wireless network connected to the Internet,
receiving information and instructions and providing in situ data.

• What related safety and human factors issues do ure includes several robots exploring the interior of
we have to account for? a building. They are on a local wireless network con-
nected to the Internet backbone, allowing them to
We use the term “embedded” to reflect the fact that receive information and instructions and send back
these robots communicate over a wireless network. data. For example, if the building being explored is
Although complex, this communication represents partially collapsed due to an earthquake, a human
the ability to provide richer interaction among specialist may want to direct the robots to find peo-
robots, as well as between robots and other network ple trapped in the rubble and send back their vital
resources. This robust but specialized interaction has signs. Other remote users might exploit audio and
strong implications for the sharing of tasks among video information the robots can provide and com-
robots, as well as for human-robot interaction and municate with people in the building. Moreover,
for on-the-fly reprogramming and adaptation of the connectivity to the Internet allows the robots to
robots on the network (see Figure 1). access Web-based information repositories, includ-
At the University of Southern California’s Robotics ing building maps, to aid their exploration.
Research Laboratories, we are working on a National The traditional approach to robotics (largely
Science Foundation-funded research study called before wireless networks) was an off-line program-
Scalable Coordination of Wireless Robots in collab- ming process in which robot controllers were devel-
oration with the USC Computer Networks and Dis- oped on a desktop computer and downloaded to a
tributed Systems Research Laboratory to address robot’s microcontroller, usually through a serial link.
some of these issues (see The new paradigm, enabled by wireless communi-
Our focus is scalable algorithms for the distributed cation, is to develop controllers on the robot itself,
control and coordination of wireless nodes that may since its computer is accessible remotely over the
be robotic, or autonomously mobile. Our goals are network at all times. By using wireless communica-
to address networking and robotics issues in the tion, robot controllers can now take advantage of a
problem domain while recognizing that the wireless variety of networked resources that may be physi-
network also connects conventional computers, cally attached to another robot, an immobile com-
wearables, portables, and immobile sensors, along puter, or an online database.
with robots. Our efforts addressing the issues involved in
Here, we address the issues and ideas related to the embedding robots in the Internet seek to achieve a
control and coordination of robots as entities dual synergy: communication facilitating robotics
embedded in wireless networks connected to the by reducing uncertainty, and robotics facilitating
Internet. The depiction of such a scenario in the fig- communication by providing physical mobility. We

68 May 2000/Vol. 43, No. 5 COMMUNICATIONS OF THE ACM

also deal with key research problems associated with from a basis set for executing complex tasks. In this
uncertainty and mobility. Here, we propose some context, we are developing strategies for resource
ideas toward their solution and touch on related transport using a large group of robots [11].
work in these areas. For convenience in experimentation, as well as to
offer a teaching and research tool, we use the multi-
Behavior-based Robotics robot simulator Arena (available at
Behavior-based robotics is the most active and pop- Arena simulates
ular approach to mobile robot control in the multi- the movement and sensing of many small mobile
robot domain [1, 7]. It is based on the notion of robots in a 2D world. All sensing and actuation is
“behavior,” a unifying representation for control, modeled at low fidelity to achieve high update rates.
reasoning, and learning. Behaviors are real-time Simple noise models for the robots’ sensors and
processes taking inputs from sensors and other effectors are also provided. While our goal is to
behaviors and sending outputs to the robot’s actua- always validate our methods on real robots, a well-
tors, as well as to other behaviors. The controller is a designed simple simulator enables fast incremental
network of such communication, executing behav- construction of controllers that run well in the pres-
iors concurrently. The behavior metaphor has excel- ence of large perturbations injected into the simu-
lent real-time and scaling properties. For example, lated world. Arena also uses a TCP/IP socket server
applied in a multi-robot scenario, it can eliminate that provides an identical controller-robot interface
the distinction between a collection of processes on to that of our Pioneer robots, which incorporate
one robot and a collection of processes on multiple small differential drives for mobility. This arrange-

> Behaviors interact not only within a robot system but

also throughout the environment, allowing designers to exploit
emergent properties.
robots across a wireless network. In either case— ment facilitates the rapid transfer of controllers
with or without the abstraction barrier—the entire developed in simulations to the physical robots.
system is a collection of communicating behaviors.
Much of our work involves the application of this Internet Robots
metaphor. The major goals of our efforts are to develop, test,
The problem of behavior coordination within a and characterize algorithms for scalable, application-
robot controller (as well as among multiple robot driven, wireless network services using a heteroge-
controllers) is an active area of robotics research. Var- neous collection of communicating mobile nodes.
ious approaches—ranging from fuzzy control to Some of these nodes will be autonomous, that is,
decision theory to neural network learning—have robots, in that their movements will not be con-
been applied effectively. Behaviors interact not only trolled by humans. The others will be portable,
within a robot system but throughout the environ- dependent on humans for their transportation.
ment, allowing designers to exploit emergent prop- While the focus of our work is the mobile nodes, we
erties. Various methods for principled behavior include immobile computers on the network as well.
design and coordination have been proposed [1]. For We emphasize that most, though not all, mobile
example, in 1992, we introduced the concept of nodes have modest sensing, computational, and
“basis behaviors,” a small set of necessary and suffi- communication resources.
cient behaviors that could be composed (through As a concrete example, consider an earthquake sce-
sequencing or fusion) as a means of handling con- nario in which people are trapped inside a partially
troller complexity and simplifying design. We have collapsed building. The rescue team has to quickly
demonstrated this principle on large groups of identify the areas where people are likely to be, so
robots performing spatial exploration, as well as on heavy machinery can be brought in to assist them.
other behaviors, including coordinated movement in One solution is a group of small autonomous robots
the form of aggregation, dispersion, flocking, and introduced into the building at various entry points.
more. We are expanding the principle to the more These robots communicate with one another and
general problem of producing reusable robotics soft- with the outside world through multihop wireless
ware that can be composed effectively at runtime radio while exploring the building trying to detect

COMMUNICATIONS OF THE ACM May 2000/Vol. 43, No. 5 69

the presence of people. When one or more robots can handle inevitable uncertainties in sensing, action,
detects a person, the location and perhaps an image communication, and control. Robot sensors provide
of that person are sent back. incomplete noisy information about the environ-
In more sophisticated versions, robots might also ment; actuators are rarely precise. Most of this uncer-
be used for bidirectional audio communication tainty is difficult to characterize, even analytically.
between people trapped in the building and the res- Moreover, because the behavior of other robots or
cuers outside or for delivering medicine and supplies humans in the environment is far from predictable,
to the victims; there’s no end to the possibilities. It is the central challenge in robotics is to perform
also easy to imagine ubiquitous robots in everyday robustly in the face of uncertainty. Embedding robots
life for applications ranging from mail delivery in into the Internet can potentially simplify some of
buildings to cleaning and security. these fundamental robotics problems.
Such small ubiquitous robots need a number of
basic abilities that make them autonomous and gen- Robot Localization Algorithms
erally useful. We emphasize three of them: There are several approaches to robot localization:

• Localization, referring to robots’ ability to use Inertial sensing and filtering. Techniques for accu-
their sensors and wireless communication to rately estimating the position and orientation of a
compute their positions over time; robot lend themselves to a natural partitioning. One
• Exploration, allowing robots to search and cover class of techniques relies on using onboard inertial
an area, perhaps with some guidelines from a sensing and odometry to keep track of changes in
user; and position. Integrating small changes over time leads to
• Mapping, supporting robots’ ability to collec- an updated position estimate [2]. The second class of
tively create a representation of the environment techniques uses some global sensing method (perhaps
or augment a representation provided by a user. a map or, if outdoors, the global positioning system,
GPS) to update position estimates. The former is
The interplay of these abilities yields robots capa- prone to drifting and depends on knowledge of the
ble of functioning autonomously in relatively initial position; the latter depends on global informa-
unstructured environments. Our concurrent devel- tion not always easy to obtain, as when there is no
opment of them focuses on several key principles: GPS signal, if indoors. The two approaches have been
combined with varying degrees of success. Each has
Multi-robot solutions. In order to be robust, we been studied extensively in robotics, but given the
investigate multi-robot solutions wherever possible, uncertain nature of sensor measurements, the prob-
especially on the key problems of collective mapping, lem of accurate position estimation remains a chal-
exploration, and localization. Our intuition is that, lenge.
with careful design, multiple robots provide redun- Radio signal strength and range. We use two
dancy and hence fault tolerance. A well-designed approaches to investigating radio as a basis for local-
multi-robot solution also reduces global uncertainty, ization. One is coarse-grained in which multiple
even if each individual robot is a relatively noisy transmitters are placed in the environment. Each
source of data. robot is equipped with a receiver that can distinguish
Distributed, bottom-up strategies. We investigate the signature of the transmitters. In any location
distributed, bottom-up strategies, emphasizing those within the environment, the receiver can detect some
scaling to large numbers of robots, thus favoring local, subset of transmitters. The set of locations where the
decentralized ones over global, centralized alterna- same transmitters can be detected form an equiva-
tives. The goals are to endow individual robots with lence class. An environment with N transmitters is
independent abilities and minimal communication thus divided into at most 2N equivalence classes.
needs (each one needs to communicate only with its Although straightforward, this approach is coarse and
nearby neighbors) and provide globally coherent and may not yield the desired granularity in realistic envi-
efficient behavior. ronments. An even more serious problem involves the
The wireless world. We treat the wireless network possibility that a given equivalence class may not be
as a key resource in distributed robotics, seeking ways spatially connected, resulting in the “holes” often
to exploit it without adopting too many restrictive or encountered in wireless communication in cluttered
simplifying assumptions. environments.
Effective autonomous mobility and interaction The other approach is a fine-grain version of the
with the physical world require robustness, so a robot first. Again, each robot in the environment is

70 May 2000/Vol. 43, No. 5 COMMUNICATIONS OF THE ACM

equipped with a receiver, and there are N transmit- unexplored spaces, while staying localized constrains
ters in the environment. We endow each robot with their movements to small feature-rich areas. The
the ability to distinguish not only transmission sig- problem is even more difficult—though the explo-
natures but to detect signal strength as well. The goal ration could be made faster—when multiple robots
is to acquire the mapping from signal strength to are involved. Coordinating a team of robots is one of
range for each transmitter and then triangulate posi- the areas in which embedding robots into a commu-
tion based on the various range estimates available at nication network would boost task performance.
each instant. The triangulation procedure uses a Group exploration for maintaining connectivity.
noise model of the range output to provide an esti- While exploration strategies for individual robots have
mate of position, along with an estimate of uncer- been studied extensively, less is known about the
tainty. This approach, like the first, can be used in problem in a multi-robot scenario. How should mul-
conjunction with inertial sensing and odometry to tiple robots coordinate themselves to explore a given
provide a better position estimate. One advantage of environment so they provide complete and efficient

> Because the behavior of other robots or humans in

the environment is far from predictable, the central challenge in
robotics is to perform robustly in the face of uncertainty.
both approaches is that they are not tailored specifi- space coverage? When posed as a global, top-down
cally for robotics applications; therefore, each one optimization problem, this question is extremely dif-
can be used just as well to localize, say, a person car- ficult to answer in all but the most stylized, simpli-
rying a computer. fied domains. We explore this and other problems
bottom-up, in a decentralized fashion. Our earlier
Robot Exploration Algorithms work demonstrated effective distributed exploration
Robot exploration approaches have been studied in for groups of up to 13 robots [8] and has applied a
different contexts. A common abstraction is the so- variety of behavior-based controllers, including
called “art-gallery” analogy in which the robot’s goal homogeneous, heterogeneous, dominance-hierarchy,
is to move from one position to another to maximize and territorial solutions [5]. We have developed
visual coverage of its surroundings, as a human methods for online interference estimation and min-
might try to do in a gallery. A complementary set of imization [5], as well as several approaches to adap-
approaches addresses the pursuit-evasion problem tive multi-robot coordination using simple
[6] in which a robot tries to move so as to evade communication of sensory input and feedback [9] to
observation or capture by a group of mobile trackers. improve and optimize group performance over time.
Several approaches to exploration address the We are currently designing behavior-based robot
related goals of searching for a specific location or controllers for a variety of scenarios. Two are partic-
object, space coverage, and maximizing some mea- ularly relevant to communication-based exploration.
sure of novelty. Task-specific heuristics can be The objective of the first is for a group of robots to
applied to simplify the exploration, as well as to “fan out” from a common starting location and
make it more robust. For example, in our recent explore an area in search of some goal. When the
work, we employed an exploration strategy for goal is detected, an image is sent back to the starting
indoor environment mapping that forces a robot to location over the network. We are using a greedy
explore corridors all the way to their end (depth exploration strategy in which each robot tries to
first) [4]. Door openings on the way were recorded maximize the amount of space it explores, as long as
but not explored; the goal was to quickly generate a it is within communication range of the other
map of the overall structure of the building and fill robots. When a robot goes out of range, it stops,
in the details later. This strategy is heuristic, and it is stores its location, and backtracks until it reestab-
easy to imagine topologies in which it is less than lishes communication with at least one of the others.
desirable. It then shares its stored location with this robot and
Exploring space efficiently is a challenging prob- stores the second robot’s location. These two loca-
lem, due to the multiple objectives involved. Detec- tions, together, define a rendezvous pair, which is
tion of new and interesting features leads robots into later used by both robots if they need to establish

COMMUNICATIONS OF THE ACM May 2000/Vol. 43, No. 5 71

The other class of approaches produces topological
Figure 2. A subset of the robot testbeds maps in which the significant or salient features in
used in our research. the environment (such as doors, windows, and cor-
ners), or so-called landmarks, correspond to the
nodes of a graph. Whenever such a feature is
detected, the robot decides whether it has seen it
before (in which case it may improve its position
estimate) or the feature is new (in which case it can
add it to its map with the appropriate links to the
other features mapped already). Our early approach
[7] to topological mapping using a graph represen-
tation introduced the notion of representation into
behavior-based systems by developing an integrated
communication again. If the robots fall out of com- system that did not distinguish between the control
munication range again, they backtrack to the last program and the map, embedding both into con-
stored rendezvous location and wait for contact. current, communicating behaviors. An example of a
The second scenario we are experimenting with recent approach to learning a topological map of an
assumes that a wave of robots has arranged itself in office building in the presence of odometric uncer-
some pattern throughout the environment. These tainty is explored in [10].
robots are most likely separated into disjoint groups Our current research, focusing on multi-robot
that cannot communicate directly. Our “communi- topological mapping, involves a group of robots, as in
cation-hole-filling” algorithm then explores the area Figure 2, building individual topological maps of
by using a second wave of robots to detect commu- the environment concurrently with no a priori
nication voids and fill them by either dropping radio information about one another’s respective loca-
tags or stationing robots at “bridge” locations. These tions. Each robot tracks its own position in a private
exploration and space-coverage techniques support reference frame; this information is communicated
basic robot abilities in a variety of applications, to other robots, and a graph-matching algorithm
including mapping. combines individual maps. The matching algorithm
seeks to find the transformation among the maps to
Robot Mapping Algorithms maximize “feature overlap” among the individual
Robot mapping approaches fall into two main cate- maps. To keep the number of candidate transforma-
gories: those producing metric maps of the environ- tions manageable and thus keep the algorithm scal-
ment and those producing topological maps. able, preprocessing heuristics are employed. The
Significant research has been done in each category, match produces a final transformation among the
though the former has seen more. The best example maps that is used to correct each robot’s position
to date of metric mapping—using laser rangefinders estimates. The resulting map combines the features
to produce precise floorplans—involved robotic from each of the individual maps in a probabilistic
tour guides in 1998 in the Smithsonian National way, since the individual mapping algorithms keep
Museum of American History in Washington, D.C., track of each robot’s belief in a feature once it is
and in 1997 the Deutsches Museum in Bonn, Ger- detected.
many; see [3] for a description of this deployment Mapping is a basic ability that can be used by a
and Web-based control of the robotic tour guides. robot to build or augment a representation of an
The map was built by composing successive laser environment, thereby helping it stay localized. Stay-
scans into a grid-based representation. This ing localized in turn provides a basis for navigation
approach decides which cells of the grid are occu- and purposeful movement—the basic abilities
pied and which are empty and incrementally underlying the application of ubiquitous Internet-
improves the confidence in each grid state with suc- embedded robots.
cessive scans. However, such mapping is made sig-
nificantly easier if localization is perfect, since in Conclusion
truly dynamic environments, accurate position esti- We’ve sought to spell out some of the key challenges
mates are needed to match successive scans. Simi- in embedding robots into the Internet and the
larly, localization is significantly easier if a map is approaches we use to address them. We’ve empha-
available. Simultaneous localization and mapping sized the benefits of bottom-up, distributed control
remain a difficult problem for autonomous robots. in this domain and of behavior-based robotics.

72 May 2000/Vol. 43, No. 5 COMMUNICATIONS OF THE ACM

Three specific robotic abilities—localization, explo-
ration, and mapping—are especially important for
mobile, robotic nodes on the Internet.
While many interesting and difficult problems still
need to be solved to realize the goal of ubiquitous
robots in human environments, the combination of
robotics technology and wireless communication,
along with the interaction of various types of com-
municating nodes, is already a rich and promising
area of research. c

1. Arkin, R. Behavior-Based Robotics. MIT Press, Cambridge, Mass., 1998.
2. Barshan, B. and Durrant-Whyte, H. Inertial navigation systems for
mobile robots. IEEE Trans. Robot. Automat. 11, 3 (June 1995),
3. Burgard, W., Cremers, A., Fox, D., Haehnel, D., Lakemeyer, G.,
Schulz, D., Steiner, W., and Thrun, S. Experiences with an interactive
museum tour-guide robot. Artif. Intel. 114, 1-2 (Oct. 1999), 3–55.
4. Dedeoglu, G., Mataric, M., and Sukhatme, G. Incremental, online
topological map building with a mobile robot. In Proceedings of Mobile
Robots XIV SPIE’99 (Boston, Sept. 19–22). ACM Press, New York,
1999, 129–139.
5. Goldberg, D. and Mataric, M. Coordinating mobile robot group
behavior using a model of interaction dynamics. In Proceedings of the
3rd International Conference on Autonomous Agents, Agents’99 (Seattle
May 1–5). ACM Press, New York, 1999, 100–107.
6. Guibas, L., Lin, D., Latombe, J.-C., LaValle, S., and Motwani, R. Vis-
ibility-based pursuit evasion in a polygonal environment. Int. J. Com-
put. Geom. Appl. (2000).
7. Matarić, M. Integration of representation into goal-driven behavior-
based robots. IEEE Trans. Robot. Autom. 8, 3 (June 1992), 304–312.
8. Matarić, M. Designing and understanding adaptive group behavior.
Adapt. Behav. 4, 1 (Dec. 1995), 50–81.
9. Matarić, M. Reducing locality through communication in distributed
multiagent learning. J. Exper. Theor. Artif. Intel. 10, 3, special issue on
learning in distributed artificial intelligence systems, G. Weiss, Ed.
(1998), 357–369.
10. Shatkay, H. and Pack Kaelbling, L. Learning topological maps with
weak local odometric information. In Proceedings of the 15th Interna-
tional Joint Conference on Artificial Intelligence, IJCAI’97 (San Francisco,
Aug. 23–29). Morgan-Kaufmann Publishers, San Francisco, 1997,
11. Vaughan, R., Stoy, K., Sukhatme, G., and Matarić, M. Whistling in the
dark: Cooperative trail following in uncertain localization space. In Pro-
ceedings of the 4th International Conference on Autonomous Agents
(Barcelona, Spain, June 3–7, 2000).
12. Weiser, M. Some computer science problems in ubiquitous computing.
Commun. ACM 36, 7 (July 1993).

This work is supported by the National Science Foundation under grant ANI-
9979457, by DARPA under contract DAAE07-98-C-L026 and grant DABT63-99-1-
0015, and by the Office of Naval Research under grants N0014-99-1-0162 and
N00014-95-1-0759. Our work on resource transport using a large robot group is sup-
ported under the DARPA Mobile Autonomous Robot Software (MARS) program (see

Gaurav S. Sukhatme ( is a research

assistant professor in the Computer Science Department of the
University of Southern California in Los Angeles and associate director
of the USC Robotics Laboratories.
Maja J. Matari´c ( is an assistant professor in
the Computer Science Department of the University of Southern
California in Los Angeles and director of the USC Robotics

© 2000 ACM 0002-0782/00/0500 $5.00

COMMUNICATIONS OF THE ACM May 2000/Vol. 43, No. 5 73