Coordinated Multi-Robot Exploration
using a Segmentation of the Environment
Kai M. Wurm Cyrill Stachniss Wolfram Burgard
Abstract— This paper addresses the problem of exploring
an unknown environment with a team of mobile robots. The
key issue in coordinated multi-robot exploration is how to
assign target locations to the individual robots such that the
overall mission time is minimized. In this paper, we propose a
novel approach to distribute the robots over the environment
that takes into account the structure of the environment.
To achieve this, it partitions the space into segments, for
example, corresponding to individual rooms. I nstead of only
considering frontiers between unknown and explored areas as
target locations, we send the robots to the individual segments
with the task to explore the corresponding area. Our approach
has been implemented and tested in simulation as well as
in real world experiments. The experiments demonstrate that
the overall exploration time can be significantly reduced by
considering our segmentation method.
I. INTRODUCTION
Autonomous robots that are designed to create a map
of their environment require the capability to effectively
cover the space. There are several applications in which
robots have been designed to autonomously explore their
environment such as planetary exploration or in disaster
missions. Using a coordinated team of robots instead of a
single robot has often been suggested to be advantageous [4],
[7] and cooperating robots have the potential to accomplish
a task faster than a single robot [11]. By using several
robots, redundancy can be explicitely introduced so that such
a team can be expected to be more fault-tolerant than a
single robot. Another advantage of robot teams arises from
merging overlapping sensor information, which can help to
compensate for sensor uncertainty. However, when robots
operate in teams there is the risk of interference between
them [10], [20]. For example, if the robots have the same
type of active sensors such as ultrasound sensors, the overall
performance can be reduced due to cross-talk. The more
robots are used, the more time each robot may spend on
detours in order to avoid collisions with other members of
the team.
In this paper, we consider the problem of efficient explo-
ration with teams of mobile robots that seek to minimize
the overall time required to complete the mission. The entire
task of coordinating a team of robots during exploration can
roughly be separated into two subsequent tasks. First, one
needs to identify potential exploration targets for the robots.
Second, one needs to assign the individual robots to the target
locations calculated in the previous step.
All authors are with the University of Freiburg, Department of Computer
Science, D-79110 Freiburg, Germany
Fig. 1. Typical coordination of robots obtained by assigning them to
different segments of the partial map.
A popular method for generating potential exploration
targets has been proposed by Yamauchi et al. [26]. In
this approach, robots are sent to so-called frontiers, which
are given as the borders between the explored and the
unexplored space. In a multi-robot context, it is important
to carefully assign robots to targets so that redundant work
and interference between robots is minimized. Therefore, it
is the exploration strategy which affects the efficiency of the
robot team the most. In many approaches, the robots are
assigned directly to frontier targets based on a cost function
that takes into account the expected path costs or travel time
as well as a utility function that covers aspects such as the
expected gain in information [3], [9], [21], [24], [28].
The coordination strategies described there consider in-
dividual locations rather than segments of the environment.
Segmentation approaches which have recently received an
increased amount of attention [2], [8], [23], [27] have origi-
nally been designed to facilitate topological localization and
loop closing or have been used to reduce planning costs. In
this paper, we introduce a new online coordination strategy
for multi-robot exploration. It uses a segmentation of the
already explored area to assign robots to segments instead
of directly assigning them to frontier targets. Based on this
segmentation, the robots are distributed over the environment
more effectively which leads to a reduction of redundant
work and the avoidance of interference between robots. As
a result, the exploration time is significantly reduced.
This paper is organized as follows. After discussing re-
lated work, we describe the Hungarian Method for target
assignment in Section III. In Section IV, we introduce a
graph-based method for map segmentation while in Section
V we present our coordination approach. Finally, we describe
simulated and real world experiments conducted to evaluate
our approach.
II. RELATED WORK
The problem of exploring unknown terrains with teams
of mobile robots has received considerable attention in the
past. Yamauchi [26] presented a technique to learn maps
with a team of mobile robots. He introduced the concept of
frontiers between known and unknown areas in a grid map,
which are widely used to select potential target locations
during exploration. In this paper, we also consider frontiers
but additionally utilize the structure of the environment for
defining potential target locations. Koenig et al. [14] analyze
different terrain coverage methods for small robots with
limited sensing and computational capabilities. Furthermore,
there has been research on how to deal with limited commu-
nication in the context of multi-robot exploration [3], [19].
An approach towards cooperation in heterogeneous robot
systems has been presented by Singh and Fujimura [21].
If a robot is too big to pass through a narrow passage, it
informs other robots about this task. Howard et al. [12]
presented an incremental deployment approach that explicitly
deals with obstructions, i.e., situations in which the path of
one robot is blocked by another. Zlot and colleagues [28]
proposed an architecture for mobile robot teams in which the
exploration is guided by a market economy. They consider
sequences of potential target locations for each robot and
trade tasks between the robots using single-item first-price
sealed-bid auctions. Such auction-based techniques have also
been applied by Gerkey and Matari
´
c [9] to efficiently solve
the task allocation problem with a group of robots.
Matari
´
c and Sukhatme [17] consider different strategies for
task allocation in robot teams and analyze the performance
of the team in extensive experiments. Ko et al. [13] present
an approach that uses the Hungarian method [15] to compute
the assignments of frontier cells to robots. In contrast to our
work, Ko et al. mainly focus on finding a common frame
of reference in case the start locations of the robots are not
known.
In a previous work [24], we considered the problem
of integrating semantic background information into the
coordination procedure. This technique is related to the
method proposed in this paper, even if the methodology
is substantially different. Compared to our previous ap-
proach [24], we obtain a significantly reduced exploration
time also for small teams of robot. The map segmentation
technique used throughout this work is related to the spatial
semantic hierarchy introduced by Kuipers and Byun [16].
The difference lies in the fact that we do not learn a model
based on distinct places but utilize this information for a
better coordination. Learning topological maps is itself a
research field on its own and different methods have been
proposed [2], [8], [25], [27]. These approaches are related to
the technique described in this paper as they can be applied
to separate the environment into appropriate regions that are
then assigned to the individual robots.
III. TARGET ASSIGNMENT
USING THE HUNGARIAN METHOD
In 1955, Kuhn [15] presented a general method, which is
often referred to as the Hungarian method, to assign a set of
jobs to a set of machines given a fixed cost matrix. Consider
a given n × n cost matrix which represents the cost of all
individual assignments of jobs to machines. The Hungarian
method, which is able to find the optimal solution with the
minimal cost given this matrix, can be summarized by the
following three steps:
1) Compute a reduced cost matrix by subtracting from
each element the minimal element in its row. After-
wards, do the same with the minimal element in each
column.
2) Find the minimal number of horizontal and vertical
lines required to cover all zeros in the matrix. In case
exactly n lines are required, the optimal assignment is
given by the zeros covered by the n lines. Otherwise,
continue with Step 3.
3) Find the smallest nonzero element in the reduced cost
matrix that is not covered by a horizontal or vertical
line. Subtract this value from each uncovered element
in the matrix. Furthermore, add this value to each
element in the reduced cost matrix that is covered by
a horizontal and a vertical line. Continue with Step 2.
The computationally difficult part lies in finding the mini-
mum number of lines covering the zero elements (Step 2).
The overall algorithm has a complexity of O(n
3
). The
method described above can directly be applied to assign
a set of target locations (tasks) to the individual robots
(machines). Here, each entry in the cost matrix can be the
length of the path the corresponding robot has to travel to
reach the designated target point.
Since the implementation of the Hungarian method de-
scribed above requires the number of jobs and the number
of machines to be equal, we need to slightly adapt the
cost matrix computed in that way. This can be achieved
by expanding the cost matrix using “dummy machines”
(which will result in target locations that are not approached
by any of the robots) and by duplicating existing targets.
The Hungarian Method is then able to compute the optimal
assignment, given the cost matrix.
IV. MAP SEGMENTATION
Several researchers investigated the problem of segment-
ing maps based on the partitioning of a graph [1], [8],
[16], [25], [27]. A very popular graph-based representation
in this context are Voronoi Graphs (VGs) [5]. To compute
the Voronoi Graph G(m) = (V, E) of a given map m, we
consider the set O
p
(m) which contain for each point p in
the free-space C of m the set of closest obstacle points. The
Voronoi Graph then is given by the set of points in O
p
(m)
for which there are at least two obstacle points with an equal
minimal distance:
V = {p ∈ C | |O
p
(m)| ≥ 2} (1)
E = {(p, q) | p, q ∈ V, p adjacent q in m} (2)
For each pair of nodes in G(m) we add an edge if their
corresponding points in m are adjacent. The Voronoi Graph
can be generated from metric maps of the environment such
as occupancy grid maps [6], [25]. In a practical implemen-
tation this can be efficiently done by applying the Euclidean
distance transformation [18] to an occupancy grid map. This
transformation results in a distance map which holds for each
grid cell the distance to the closest obstacle. A Voronoi Graph
can then be constructed using skeletonization on the distance
map. Figure 2 illustrates the process of generating a Voronoi
Graph for an example occupancy grid map.
After generating the Voronoi Graph we are now interested
in creating a partitioning of the graph into k disjoint sets
V
1
, V
2
, . . . , V
k
with
V =
k
[
i=1
V
i
(3)
such that each cluster of nodes V
i
corresponds to a segment
we can assign robots to. Thrun et al. suggest the graph to
be separated at so-called critical points [25]. Here, critical
points are those nodes in the Voronoi Graph at which
the distance to the closest obstacle in the map is a local
minimum. This is usually the case in doorways or other
narrow passages.
Whereas this approach is able to reliably find doorways, it
also generates a lot of false positive candidates in cluttered
environments. To eliminate these false positives, we constrain
them in the following way: First, critical points have to be
nodes of degree 2 (two edges) and second, need to have
a neighbor of degree 3 (a junction node). In addition, we
require the points to lead from known into unknown areas,
since segments which do not contain unknown areas can
safely be ignored in an exploration task. To verify this
constraint, we compute the distance to the closest reachable
unknown cell for each point. This can be done efficiently
in a similar way as the computation of the distance map.
Figure 3 shows a pruned version of the Voronoi graph and
the critical points found by our algorithm. All doorways
have been selected as candidates and the number of false
positives is much smaller than the number of critical points
according to the definition of Thrun et al. [25] which includes
distance minima in the Euclidean distance transformation
within corridors and rooms.
In the practical experiments described in this paper we
found that this segmentation technique yields sufficient re-
sults and allows to nicely distribute the robots. In unmodified
office environments, we can typically reliably separate rooms
and segments of a corridor. Other, more complex environ-
ments may however suggest more sophisticated segmentation
algorithms which rely on hand-labeled training data [2], [8]
or more complex reasoning [1], [27].
Fig. 2. Generation of the Voronoi Graph. Left: Example grid-map. Center:
Map plus distance transform (the darker a point the larger the distance to
the closest obstacle). Right: Map and Voronoi Graph generated from the
distance transform using skeletonization.
Fig. 3. Example segmentation of a small fraction of an environment. The
marked nodes are the candidates for the partitioning of the graph calculated
by our approach.
V. ASSIGNMENT OF ROBOTS TO TARGET AREAS
Typical approaches to coordinated exploration seek to
minimize the time needed to cover the whole environment
with the robot’s sensors. Therefore, it is often sub-optimal to
explore the same (local) area with more than one robot. A
cluster of robots which has a serious overlap in the field of
view of the robots’ sensors does not exploit its full potential.
In practice, it is generally much more efficient to explore
separate regions of the environment instead. For this reason,
it is important to assign robots to exploration targets such
that the robots do not get too close to each other during
exploration.
Indoor environments are in general structured environ-
ments. Buildings are usually divided into rooms which can be
reached via corridors. In many cases, it can be a disadvantage
to assign more than one robot to one room. The room might,
for example, be too small for a second robot to speed up
it’s exploration even though there initially is more than one
frontier in the room. When the room is fully explored, robots
might even block each other while trying to leave the room
which will result in an increase in exploration time.
In our approach, we assign individual robots to different
segments of unexplored space. Segments could be separate
rooms, corridors, or parts of larger corridors or rooms. This
takes into account the structure of the environment and
prevents the forming of inefficient clusters of robots.
Algorithm 1 Target Assignment Using Map Segmentation.
1: Determine segmentation S = {s
1
, ..., s
n
} of map.
2: Determine the set of frontier targets for each segment.
3: Compute for each robot i the cost C
i
s
for reaching each
map segment s ∈ S.
4: Discount cost C
i
s
if robot i is already in segment s.
5: Assign robots to segments using the Hungarian Method.
6: for all segments s do
7: Assign robot(s) to frontier targets in s w.r.t. path costs
using the Hungarian Method.
8: end for
Our assignment algorithm is summarized in Algorithm 1.
An assignment is determined whenever one of the robots
requests a new exploration target. First, a partition of the
partial map of the environment is created using the graph-
based method described in Section IV. To generate targets
within the segments, we then determine the set of frontier
cells. The cost C
i
s
for reaching segment s with robot i is
defined as the expected path cost to the nearest frontier
cell within s. This cost is discounted by a constant factor
if robot i is already located in segment s. This has the
effect that the robots stay in their assigned segment until
it is completely explored. After computing the costs of
a segment, an assignment is calculated by applying the
Hungarian method (see Section III) based on the cost matrix.
The Hungarian method does not assign more than one
robot to the same segment unless there are more robots avail-
able than there are unexplored segments. To appropriately
handle those cases in which multiple robots are assigned to
a single segment, we apply a local assignment based on the
cost-optimal frontier within a segment. For this reason, our
algorithm is equivalent to a purely frontier-based assignment
if the environment cannot be partitioned, i.e., there is only
one segment.
By assigning robots to separate segments, an appropriate
distribution of the robots can be achieved. As we will
demonstrate in the experiments, this leads to a significant
reduction in exploration time. Instead of aiming at the closest
frontier, robots share work more efficiently. A typical office
environment, for example, contains corridors and rooms.
Using our approach, each of the corridors is explored com-
pletely by one of the robots. In this way, the rough structure
of the building will quickly be revealed. Meanwhile other
robots will be assigned to the rooms reachable from the
corridors, one at a time. This behavior does not only appear
to be a natural way of exploring an unknown environment,
our experiments also revealed that it significantly increases
the efficiency of the robot team compared to approaches
which ignore the structure of the building.
Fig. 4. Maps used in our simulated experiments: Building 079 of the
Freiburg University (top) and Bremen University Cartesium (bottom).
Note that our algorithm is not limited to homogenous
teams of robots. Consider the situation in which one par-
ticular robot cannot enter a certain part of the environment
while another robot can. The assignment algorithm described
above can be applied in this case by using modified segment
costs
¯
C
i
s
defined as:
¯
C
i
s
=
C
i
s
, if robot i can enter segment s
∞ , otherwise.
(4)
VI. EXPERIMENTAL RESULTS
Our approach has been implemented and evaluated using
simulated as well as with real teams of robots. The real
world experiments were conducted using two ActivMedia
Pioneer II robots equipped with a laser range finder with
a 180 degrees field of view. For generating the simulation
results, we used the Carnegie Mellon Robot Navigation
Toolkit. In all our experiments we assumed that the robots
share a joint occupancy grid map, which is generated based
on the sensor readings of all robots and under the assumption
that all positions of the vehicles are known. This map is used
for coordination, path planning, and path execution. We also
assume that there is a central planning component which
can communicate with all robot and can assign exploration
targets to them. If there is only a limited communication
range, then clusters of robots can be coordinated if one
selects one individual planning agent per cluster [13], [22].
The experiments have been designed to verify that our ex-
ploration approach leads to significantly shorter exploration
time compared to a standard frontier-based approach.
A. Simulation Results
To evaluate our robot coordination algorithm, we simu-
lated teams of robots in various environments. We com-
pared our segmentation-based approach to a frontier-based
approach in which each robot is assigned to the closest
frontier which has not been assigned to another robot yet.
Since this strategy does not consider the structure of the
environment, it will in general also assign more than one
robot to one room or corridor if they contain more than one
frontier.
To eliminate influences from the segmentation algorithm
used in the real world experiment, we assumed a given
segmentation of the environment into rooms and corridors
in our simulation experiments. As mentioned above, such
a segmentation could also be reliably generated from the
partial map alone.
Figure 4 depicts two maps of real environments used
for the simulation (see also real world experiments). Both
of them are office environments, one at the University of
Freiburg and the other at the University of Bremen. To
make the maps more different, we added clutter to the map
representing the office environment located at the University
of Bremen.
We varied the size of the simulated team from two to six
robots (Freiburg map) respectively from two to eight robots
(Bremen map). Since the Bremen map is considerably bigger
than the Freiburg map, we simulated larger teams of robots
there. For each team size, we conducted a series of simulated
exploration runs starting from 20 different starting positions.
The results of our experiments can be seen in Figure 5.
We measured the runtime gain of our approach which
uses the assignment described in Section V compared to
the alternative assignment described above. We plotted the
runtime gain in percent of the total runtime against the size
of the robot team. The error bars in the plots indicate the
95% confidence level. It can be seen that our approach
significantly outperforms the approach which does not use a
segmentation based assignment.
The runtime gain is bigger for the Cartesium map since
this map features several large rooms. This observation can
be seen as an indicator as to when our approach will lead to
especially good results. Whenever the environment can be
divided into reasonably large and separated segments, our
technique substantially reduces the overall exploration time.
In general, our strategy assigns one robot to one segment.
As soon as there are more robots than segments multiple
robots may be assigned to the same segment as mentioned
in Section V. For this reason, the runtime gain of our
strategy will decrease for large teams of robots in small
environments. This can be seen in Figure 5. Note however,
that the overall time to complete the mission still gets smaller
the more robots are added to the task (the plot only shows
the improvement of our approach vs. the frontier-based
approach).
B. Real Robot Experiments
Our coordination algorithm has been evaluated using a
team of real robots. For this experiment, we used two
identical Pioneer II robots equipped with a laser range finder
and a standard laptop-computer. During the experiment both
robots were connected via a wireless network. The robot
localization was achieved using a standard scan-matching
approach. The relative starting poses of the robot were
manually set in the beginning. Figure 6 depicts the two robots
during their exploration mission.
The experiments were conducted in the lower floor of
building 079 of the Freiburg computer science campus. The
-2
0
2
4
6
8
10
12
14
1 2 3 4 5 6 7
runtime gain [%]
number of robots
segmentation based coordination
frontier based coordination
-5
0
5
10
15
20
25
1 2 3 4 5 6 7 8 9
runtime gain [%]
number of robots
segmentation based coordination
frontier based coordination
Fig. 5. Exploration time gain of our approach compared to a frontier-
based approach for the AIS lab in Freiburg (top) and the Bremen Cartesium
(bottom).
Fig. 6. Two robots exploring the AIS laboratory of the University of
Freiburg using our coordination approach.
building has a size of approximately 37m x 14m and consists
of numerous office rooms and two long corridors divided by
a door.
The team of robots was able to successfully explore the
environment using our coordination approach. The result of
one of the experiments can be seen in Figure 7. The figure
shows the combined map of both robots after the exploration
had finished. It also shows the trajectories of both robots
during the exploration. The total exploration time was less
than nine minutes, each of the robots traveled approximately
120m.
It can be seen that each of the rooms was explored by
exactly one of the robots. It can also be seen that both
corridors have been explored completely by one of the robots
while the other one was exploring rooms reachable from the
corridor. Another interesting effect is that the robots did not
