1

Exploration and Mapping with Autonomous Robot Teams
Results from the MAGIC 2010 Competition
Edwin Olson Johannes Strom Rob Goeddel Ryan Morton Pradeep Ranganathan Andrew Richardson
University of Michigan
{ebolson,jhstrom,rgoeddel,rmorton,rpradeep,chardson}@umich.edu

Abstract— The potential impact of autonomous robotics is
magnified when those robots are deployed in teams: a team of
cooperating robots can greatly increase the effectiveness of a
human working alone, making short work of search-and-rescue
and reconnaissance tasks. To achieve this potential, however,
a number of challenging problems ranging from multi-robot
planning, state estimation, object detection, and human-robot
interfaces must first be solved. The MAGIC 2010 competition,
like the DARPA grand challenges that preceded it, presented
a formidable robotics problem designed to foster fundamental
advances in these difficult areas. MAGIC asked teams of robots
to collaboratively explore and map a 500 × 500m area, detect
and track benign and dangerous objects, and collaborate with
human commanders while respecting their cognitive limits.
This paper describes our winning entry in the MAGIC
contest, where we fielded a team of 14 autonomous robots
supervised by two human operators. While the challenges in
MAGIC were diverse, we believe that cooperative multi-robot
state estimation is ultimately the critical factor in building
a successful system. In this paper, we describe our system
and some of the technological advances that we believe were
responsible for our success. We also contrast our approach to
those of other teams.

Fig. 1. Team Michigan Robots. We deployed fourteen custom-made robots
that cooperatively mapped a 500×500m area. Each robot had a color camera
and a laser range finder capable of producing 3D point clouds.

a) Keywords: Multi-Agent Systems, Human-Robot Interaction, SLAM
I. I NTRODUCTION
Urban reconnaissance and search-and-rescue are ideal
candidates for autonomous multi-robot teams due to their
inherent parallelism and to the danger they present to
humans. However, this domain presents many challenging
problems which arise from working in complex, stochastic
and partially observable environments. In particular, nonuniform and cluttered terrain in unknown environments
presents challenges for both state-estimation and control,
resulting in complicated planning and perception problems.
Limited and unreliable communications further complicates
coordination amongst the individual agents and with their
human commanders.
To help address these difficult problems, the MultiAutonomous Ground robot International Challenge
(MAGIC) was conducted in November of 2010, where
five teams comprising nearly 40 robots competed for over
a million dollars in prize money. Teams were instructed
to explore and map a large indoor-outdoor area while
recognizing and neutralizing threats such as simulated
bombs and enemy combatants. Although the contest
showcased the abilities of teams to effectively coordinate
autonomous agents in a challenging environment it also

showed the limitations of the current state-of-the-art in state
estimation and perception (e.g. map building and object
recognition).
The MAGIC competition was the most recent of the
robotics grand challenges, following in the tradition of the
well-known competitions sponsored by the Defense Advanced Research Projects Agency (DARPA). These competitions ultimately trace back to a congressional mandate
in 2001 requiring one-third of all ground combat vehicles
to be unmanned by 2015. Over the course of the three
DARPA challenges, teams developed technologies for fully
autonomous cars, including the ability to drive in urban
settings, navigating moving obstacles and obeying traffic
laws [1], [2]. These contests fostered the development of new
methods for planning, control, state estimation, and perhaps
most importantly, robot perception and sensor fusion.
Unfortunately these advances were not mirrored in smaller
robots, such as those used by soldiers searching for and neutralizing improvised explosive devices (IEDs) or for robots
intended to help first responders with search and rescue
missions. Instead, tele-operation (remote joystick control by
a human) remains the dominant mode of interaction. These
real-world systems pose challenges that were not present in
the DARPA grand challenges which has held them back:

2

Fig. 2. Finalist robots. Each team used a unique robot platform (left-to-right, in ranked order): Team Michigan used 14 custom-built robots; University of
Pennsylvania fielded 7 custom robots; RASR based its 7 robots on the Talon commercial platform; Magician adapted a commercial base for its 5 robots;
Cappadocia built 6 tailored vehicles.

1) Limited/Unreliable GPS. GPS is often unreliable or
inaccurate in dense urban environments or indoors.
GPS can also be jammed or spoofed by an adversary.
The winning DARPA vehicles relied extensively on
GPS.
2) Multi-robot cooperation. Individually, robots are generally less capable than humans. Their potential arises
from multi-robot deployments that explicitly coordinate.
3) Humans-in-the-loop. By allowing a human to interact
with a robot team in real-time, the system becomes
more effective and can adapt to changes in the mission objectives or priorities. This entails developing
visualization methods and user interface abstractions
that allow the human to understand and manipulate
the state of the team.
The MAGIC contest focused on increasing the effectiveness of multi-robot systems by increasing the number of
robots that a single human commander could effectively
manage. This is in contrast to current robot systems, which
typically have one or more operator per robot. The contest
was jointly organized by the United States Army and the
Australian Defence Science and Technology Organisation
and required participants to deploy a team of cooperating
robots to explore and map a hostile area, recognize and
catalog the location of interesting objects (people, doorways,
IEDs, cars, etc.), and perform simulated neutralization of
IEDs using a laser pointer. Two human operators were
allowed to interact with the system, but the interaction time
was measured and used to calculate a penalty to the team’s
final score.
The contest attracted 23 teams from around the world, and
through a series of competitive down-selects, was reduced
to five finalists who were invited to Australia for the final
competition. The venue was the Adelaide Showgrounds, a
500x500m area including a variety of indoor and outdoor
spaces. Aerial imagery provided by the contest organizers constituted the only prior knowledge. While DARPA
challenges provided detailed GPS waypoints describing the
location and topology of the safe roads, MAGIC robots
would have to figure this out on their own. Whereas
other search-and-rescue robotics contests typically focus on
smaller environments with significant mobility and manipulation challenges, (e.g. RoboCup Rescue league), MAGIC
was conducted at a much larger scale with an increased focus

on autonomous multi-robot cooperation [3].
To succeed in MAGIC, a team needed to combine robot
perception, mapping, planning, and human interfaces. This
paper highlights some of the key decisions and algorithmic choices which led to our team’s first place finish [4].
Additionally, we will highlight how our mapping and state
estimation system differed from other competitors, one of the
key differences which we believe set our team apart from our
competitors.
II. S YSTEM D ESIGN
We begin by describing how our system worked at a highlevel. Fundamentally, most teams pursued a similar strategy.
Our system was largely centralized: a ground control
station collected data from individual robots, fused it to
create an estimate of the current state of the system (the
position of the robots, the location of important objects, etc.),
then used this information to assign new tasks to the robots.
Most robots focused on exploring the large competition area,
a task well-suited to parallelization. However, other robots
could perform additional tasks, such as the neutralization
of an improvised explosive device. The discovery of such a
device would cause a “neutralize” task to be assigned to a
nearby robot.
The human operators were located at the ground control
station and were able to view the current task assignments,
a map of the operating area, and (perhaps most importantly)
guide the system by vetting sensor data or overriding task
assignments.
The robots received their task assignments via radio and
were responsible for executing that task without additional
assistance from the ground control station. For example,
robots used their 3D laser range-finder to identify safe
terrain and avoided obstacles on their own. They were
also responsible for autonomously detecting IEDs and other
objects. The information gathered by the robots (including
object detection data and a map of the area immediately near
the robot) was heavily compressed and transmitted back to
the ground control station. (In practice, these messages were
often relayed by other robots in order to overcome the limited
range of our radios.)
With the newly-collected information, the ground control
station updated its map, user interfaces, and computed new
(and improved) tasks for each of the robots. This process
continued until the mission was completed.

3

Such a system poses many challenges: How does the
ground control station compute efficient tasks for the robots
in a way that maximizes the efficiency of the team? How can
a human be kept informed about the state of the system?
How can the human contribute to the performance of the
system? How do the robots reliably recognize safe and
unsafe terrain? How do they detect dangerous objects? How
can the information collected by the robots be compressed
sufficiently to enable it to be transmitted over a limited and
unreliable communications network? How does the ground
control station combine information from the robots into a
single globally-consistent view?
Recognizing that many of these tasks rely on a highquality map of the world, our team focused on the challenge
of fusing robot data into a globally-consistent view. Not
only was the accuracy of this map a primary evaluation
criterion in the MAGIC competition, but it was also a critical
component of effective multi-agent planning and the humanrobot interface. For example: it is difficult to know where
to send the robots next if one does not know where they
are now, or if one does not know where they have already
explored.
One of the more obvious differences between our team and
other teams was the accuracy of the maps that we produced.
Map quality pays repeated dividends throughout our system,
with corresponding improvements in human-robot interfaces,
planning, etc. The variability in map-quality between different teams is a testament to the difficulty and unsolved
nature of multi-robot mapping. Our team began with a stateof-the-art system, but these methods were inadequate both in
terms of scaling to large numbers of robots, and in terms of
dealing with the errors that inevitably occur. New methods,
both automatic and human-in-the-loop, were needed in order
to achieve an adequate level of performance. The following
section explores a few of these methods.
III. T ECHNICAL C ONTRIBUTIONS
While MAGIC posed many technical challenges, mapping
and state estimation were arguably the most critical. Using
the Global Positioning System (GPS) may seem like an obvious starting point. However, even under best-case conditions,
GPS can not provide a navigation solution for the significant
fraction of the time that robots spend indoors. Outdoors, GPS
data (particularly from consumer grade equipment) is often
fairly good—within a few meters, perhaps. But GPS can
also be wildly inaccurate due to effects like multi-path. In a
combat situation, GPS can be easily jammed or even spoofed.
Consequently, despite having GPS receivers on each robot,
we ultimately opted not to use GPS data, instead relying on
the robots’ sensors to recognize landmarks. This strategy was
not universally adopted, however; most teams did use GPS
to varying degrees.
A. Overview of Mapping and State Estimation
Conceptually, map-building can be thought of as an alignment problem: robots periodically generate maplets of their
immediate surrounds using a laser scanner. The challenge is

to determine how to arrange the maplets so that they form a
large coherent map, much like the process of assembling a
panoramic photo from a number of overlapping photos (see
Fig. 3). Not only can we recover a map this way, but the
position of each of the robots is also known, since they are
at the center of their maplets.
Our team’s state-estimation system was based on a standard probabilistic formulation of mapping in which the desired alignment can be computed by performing inference on
a factor graph. (See [5], [?] for a survey of other approaches.)
Our factor graph contains nodes for unknown variables (the
location of each maplet) and edges connecting nodes when
something is known about the relative geometric position
of the two nodes. Loosely speaking, an edge encodes a
geometrical relationship between two maplets, i.e., “maplet
A is six meters east and rotated thirty degrees from maplet
B.” Of course, none of these relationships are known with
certainty, so edges are annotated with a covariance matrix.
It is common for a map to contain many of these edges, and
for those edges to subtly disagree with one another.
More formally, let the position of all the maplets be
represented by the state vector x. This vector can be quite
large: it contains two translation and one rotation component
for every maplet, and there can be thousands of maplets.
Edges convey a conditional probability distribution p(zi |x),
where zi is a sensor measurement. This quantity is the
measurement model: given a particular configuration of the
world, it predicts the distribution of the sensor. For example,
a range sensor might return the distance between two variable
nodes plus some Gaussian noise whose variance can be
empirically measured.
Our goal is to compute p(x|z), or the posterior distribution
of the maplet positions given all of the sensor observations.
Using Bayes’ rule, and assuming that we have no a priori
knowledge of what the map should look like (i.e., p(x) is
uninformative), we obtain:
p(x|z) ∝

p(zi |x)

(1)

Our goal is to find the maplet positions x that has maximum probability p(x|z). Assuming that all of the edges are
T −1
simple Gaussian distributions of the form e(zi −µ) Σ (zi −µ) ,
this computation becomes a non-linear least-squares problem. Specifically, we can take the logarithm of both sides,
which converts the right hand side into a sum of quadratic
losses. We maximize the log probability by differentiating
with respect to x, which results in a first-order linear system.
The key idea is that maximum likelihood inference on a
Gaussian factor graph is equivalent to solving a large linear
system; see [6] for a more detailed explanation. The solution
to this linear system yields the position of each maplet.
Critically, the resulting linear system is extremely sparse.
This is because each edge typically depends on only two
maplet positions. In our system, each maplet was, generally,
connected to between 2 and 5 other maplets. Sparse linear
algebra methods can exploit this sparsity, greatly reducing the
time needed to solve the linear system for x. Our method
was based on sparse Cholesky factorization [7]: we could
compute solutions for a graph with 4200 nodes and 6300

4

Fig. 3. Mapping Overview. Individual “maplets” (top left) are matched in a pair-wise fashion; the resulting network of constraints can be illustrated in
a factor graph similar to the bottom figure, in which circles represent robot positions and squares represent probabilistic constraints. The final map (top
right) is computed by reprojecting all of the sensor observations according to the maximum likelihood robot positions.

edges in about 250 ms on a standard laptop CPU. New data
is always arriving, and so this level of performance allows
the map to be updated several times per second.
An important advantage of using the factor graph formulation is that it is possible to retroactively edit the graph
to correct errors. For example, if a sensing sub-system
erroneously adds an edge to the graph (incorrectly asserting,
perhaps, that two robot poses are a meter apart), we can
“undo” the error by deleting the edge and computing a new
maximum likelihood estimate. This sort of editing is not
possible, for example, with methods based on Kalman filters.
In our case, we rely on human operators to correct these
relatively rare errors (see Section III-C).
B. Scan Matching & Loop Validation

Fig. 4. Brute-force search for best maplet alignments. The search space is
3D (two translation and one rotation component) which is illustrated above
as a series of 2D cross-sections. Bright areas indicate good alignments.
Finding the best match quickly is critical to a large-scale mapping system.
The resulting matches become edges in the factor graph.

Our mapping approach is dependent on identifying highquality edges. In general, more edges result in a better map
since the linear system becomes over-constrained, reducing
the effect of noise from individual edges.
Our system used a number of different methods to generate
edges including dead-reckoning (based on wheel encoder
odometry and a low-cost IMU) and visual detection of other
robots using their 2D “bar codes” (see Fig. 1) [8]. But by
far, the most important source of edges in our system was
our scan-matching system. This approach directly attempts
to align two maplets by correlating them against each other,
looking for the translation and rotation that maximize their
overlap. One such matching operation is illustrated in Fig. 4:
the probability associated with each translation and rotation
is computed in a brute-force fashion.
This alignment process is computationally expensive, and
in the worst-case, every maplet must be matched with every
other maplet. In practice, our dead-reckoning data can help
rule out many matches. But with fourteen robots operating

simultaneously and each one producing a new maplet every
1.4 seconds, hundreds or thousands of alignment attempts
per second are needed.
Our approach to mapping was based on an accelerated
version of a brute-force scan matching system [9]. The key
idea is to use a multi-resolution matching system: we generate low-resolution versions of the maplets and first attempt
to align these. Because they are smaller, the alignment is
much faster. Good candidate alignments are then attempted
at higher resolution.
While simple in concept, a major challenge is ensuring
that the low-resolution alignments do not under-estimate
the quality of an alignment that could occur using higherresolution maplets. Our solution relied on constructing the
low-resolution maplets in a special way. Rather than applying
a typical low-pass-filter/decimate process (which would tend
to obliterate structural details), we used a max-decimate

5

kernel. This ensures that low-resolutions of the maplets
are conservative: when aligning low-resolution maplets, we
never under-estimate the overlap that could result from
aligning the full-resolution maplets.
While our previous two-level scan matcher was fast (it
could perform around 50 matches per second), a faster
version of the algorithm was needed for MAGIC. We used a
full image pyramid of maplet resolutions; when alignments
at low-resolution still result in small amounts of overlap, it
eliminates large portions of the search space. Our improved
multi-resolution method achieved 500 match attempts per
second, a rate that was pivotal in being able to keep up with
the datarate of our robots.
Other teams used similar maplet matching strategies,
though they were not as fast. The Australian team “Magician”, for example, reports that their GPU-accelerated system
was capable of 7-10 matches per second.
This improvement in our matching speed allowed us to
consider a large number of possible matches in real-time to
support our global map. However, our state-of-the-art method
has a non-zero false positive rate. In short, it will align
maplets based on similar-looking structures, even if those
maplets are not actually near each other.
There is a fundamental trade-off between the number of
true positive and an increased risk of false positives. Increasing the threshold for what constitutes a “good enough”
match also increases the likelihood that similar looking,
but physically distinct locations will be incorrectly matched.
These types of false-positive matches can cause the inference
method to distort the map in order to explain the error.
To reduce the false positive rate to a usable level, we performed a loop-validation step on candidate matches before
they were added to the factor graph. The basic idea of loop
validation is to require that multiple matches “agree” with
each other [10], [11], [12]. Specifically, consider a topological “loop” of matches: a match between node A and B,
another match between B and C, and a third match between
C and A. If the matches are correct, then the composition of
their rigid-body transformations should approximately be the
identity matrix and can be added to the graph. Of course, it
is possible that two matches might have errors that “cancel”,
but this seldom occurs.
C. Human Robot Interfaces
In simple environments, such as an indoor warehouse, the
combination of loop-validation and automatic scan-matching
we have presented are sufficient to support completely
autonomous operation of our robot team (see Figure 5).
However, in less structured environments (like many of the
outdoor portions of the MAGIC 2010 competition), mapping
errors still occur. For example, the MAGIC venue contained
numerous cable conduits which caused robots to unknowingly get stuck, causing severe dead reckoning estimation
error. Our system was not able to handle these types of
problems autonomously.
However, these types of problems are relatively obvious
to a human operator. We developed a user interface that

Fig. 5. Indoor storage warehouse map. In uncluttered environments posing
few mobility challenges, our system can explore and map with very little
human intervention.

allowed a human operator to look for errors and intervene
when necessary. With new (validated) loop closures being
added to the graph at a rate of 2-3 per second, it would be
easy to overwhelm the human operator by asking for explicit
verification of each match.
Instead, the human operators would monitor the entire
map. When an error occurred (typically visible as a distortion
in the map), the operator could “roll-back” automatically
added matches until the problem was no longer present.
The operator could then ask the mapping system to perform
an alignment between two selected maplets near where the
problem was detected. This human-assisted match served as
a prior for future autonomous match operations, and so the
autonomous mapping system would be much less likely to
make the same mistake again.
We found that this approach, which required only a few
limited interactions to remove false positives, was a highly
effective use of humans to support the continued autonomy
of our planning system. We were the only team to build a
user interface that allowed direct supervision of the real-time
state estimate; other teams were required to handle failures
in automatic state estimation by requiring humans to track
the global state manually and then intervening at the task
allocation level. Early versions of our system lacked the
global mapping system— the human operators were instead
provided with separate map displays for each robot. Our
experience with this approach indicated that operators could
not effectively handle more than 5 or 6 robots in such a
fashion. Maintaining a global map is critical to scaling to
larger robot teams, and our user interface was a key part of
maintaining the consistency of that map.
IV. E VALUATION
The main evaluation metrics for an autonomous reconnaissance system are the quality of the final map produced and
the amount of human assistance required to produce it. These
were also the primary metrics the MAGIC organizers used to
determine the winner and subsequent ranking of the finalists
(see Figure 2). While the specific performance data used

mean

1

2

3

4

5

V. D ISCUSSION
sd

0

Actions per Minute

6

6

0

20

40

60

80

Time (min.)

Fig. 7. Map Interaction Experiment. Our mapping operator re-enacted
supporting role for phase 2 dataset to measure the frequency of interaction
required to maintain a near-perfect state estimate. See Figure 6 for resulting
map. Overall, the human work-load was quite modest, averaging two
interactions per minute.

during the contest were not made public, we will present
selected results we obtained by processing our logs from
the contest. Additionally, we will compare with other teams’
published results where possible.
Lacking detailed ground truth for the MAGIC venue, the
best evaluation of map quality is necessarily subjective. Figure 6 shows post-processed maps for our team in comparison
to the mapping software of Magician (4th place) applied to
the data collected by UPenn’s team (2nd place). Additionally,
the actual map produced by our system during the competition is shown inset. These results show that high quality
maps can be produced in this domain – our competitionday results show that our state estimation was sufficiently
good to be used for support of online planning. This system
allowed us to completely explore the first two phases of the
magic competition, while simultaneously performing mission
objectives relating to dynamic and static dangers such as
IEDs and simulated mobile enemy combatants.
Ideally, we would also like to measure the frequency of
human interaction required to support our state estimation
system during the MAGIC contest. However, the data necessary to evaluate this metric was not collected during our
competition run, thus we replicated the run by playing back
the raw data from the competition log and having our operator re-enact his performance during the competition. These
conditions are obviously less stressful than competition, but
are still representative of human performance. The result,
shown in Figure 7 was an addition of 175 loop closures,
on average two interactions per minute, which generally
occurred in bursts. However, at one point, the operator did
not interact with the system for 5.17 minutes.
Our evaluation shows that we were able to support cooperative global state estimation for a team of autonomouslynavigating robots using a single part-time operator. Yet there
remain significant open problems including reducing human
assistance to even lower levels by improving the ability of the
system to autonomously handle errors. Additional evaluation
of our system, and technical descriptions of the other finalists
can be found in separate publications [4], [14], [15], [16],
[17].

The MAGIC competition’s focus was on increasing the
robot-to-human ratio and on efficiently coordinating the
actions of multiple robots. Key to reducing the cognitive load
on the operators is increasing the autonomy of the robots;
for a given amount of cognitive loading, more robots can be
handled if they simply require less interaction. We identified
global state estimation as a key technology to enable autonomy, and we believe that the mapping system we deployed
for MAGIC outperforms the systems of our competitors.
While this was one of the key factors differentiating us
from other finalists, it was not the only important point of
comparison. In fact, many of the other choices we made
while developing our system also had an important impact
on our performance.
In particular, we made a strategic choice early during
our development that our team would emphasize the use
of a large team of robots. This is reflected in the fact that
we brought twice as many robots to the competition as
the next largest team. This strategy ultimately affected the
design of all our core systems, including mapping, object
identification, and communication. Given that we had a
finite budget, it also forced us to deploy economical robot
platforms which had only the bare necessities in sensing to
complete the challenge. The result was that our robots were
also the cheapest of any of the finalists (by a significant
margin), costing only $11,500 USD each.
One approach to detecting dangerous objects, for example,
is to transmit video feeds back to the human operators and
rely on them to recognize the hazard. Given a design goal
of maximizing the number of robots, such a strategy is
unworkable: there is neither the bandwidth to transmit that
many images, nor could the humans be expected to vigilantly
monitor 14 video streams. Our system simply had to be
able to detect dangerous objects autonomously, whereas other
teams with smaller numbers of robots could be successful
with less automation. At the same time, however, handling
more tasks autonomously meant that our human operators
had more time to assist with mapping tasks.
VI. C ONCLUSION
The MAGIC 2010 competition showcases the progress
that has been made in autonomous, multi-agent robotics. Our
MAGIC experience suggests that competitions like these are
won by mastering a set of key technological competencies,
in this case collaborative state estimation. Our team’s focus
on global state estimation allowed us to make several contributions to the state of the art in autonomous map building.
We believe the quality of our maps was the most important
deciding factor that lead our team to win the contest, both
because map quality was explicitly an evaluation criterion
and because good state estimation supported high-level autonomy throughout our system, resulting in a net reduction
in human interaction.
However, MAGIC also highlights the shortcomings of
state-of-the-art methods. It remains difficult to maintain a
consistent map for large numbers of robots. While our

7

Fig. 6. Comparison of minimally post-processed maps from our team (left) and Magician’s mapping algorithm using UPenn’s data (right) from [13]. The
map we produced online during the challenge is inset top-left.

competition-day maps are fairly good, some distortions are
still evident. In particular, the perception systems still add
incorrect edges to the factor graph, and current inference
methods are highly sensitive to these errors. Our system
coped with these errors at the expense of greater operator
workload, but further improving these systems remains an
important goal for our team.
Ultimately, we feel that competitions like MAGIC 2010,
motivated by real-world problems, are invaluable in identifying important open problems and in promoting solutions
to them. These competitions serve as a reminder that there
are few truly “solved” problems.

[3]

[4]

[5]
[6]

ACKNOWLEDGMENTS

[7]

Team Michigan was a collaboration between the University of Michigan’s APRIL Robotics Laboratory and Soar
Technology. In addition to the authors of this paper, our core
team members included Mihai Bulic, Jacob Crossman, and
Bob Mariner. We were also supported by over two dozen
undergraduate researchers.
Our thanks also go to the MAGIC contest organizers who
mounted a massive effort to organize the competition and
an even larger effort to prepare the contest venue. A special
thanks goes to our liaison, Captain Chris Latham of the 9th
Combat Service Support Battalion in South Australia. Our
participation would not have been possible without the help
of our sponsors at Intel and Texas Instruments.

[10]

R EFERENCES

[14]

[1] S. Thrun, M. Montemerlo, H. Dahlkamp, D. Stavens, A. Aron,
J. Diebel, P. Fong, J. Gale, M. Halpenny, G. Hoffmann, K. Lau,
C. Oakley, M. Palatucci, V. Pratt, P. Stang, S. Strohband, C. Dupont,
L.-E. Jendrossek, C. Koelen, C. Markey, C. Rummel, J. van Niekerk, E. Jensen, P. Alessandrini, G. Bradski, B. Davies, S. Ettinger,
A. Kaehler, A. Nefian, and P. Mahoney, “Stanley: The robot that won
the darpa grand challenge,” in The 2005 DARPA Grand Challenge, ser.
Springer Tracts in Advanced Robotics. Springer Berlin / Heidelberg,
2007, vol. 36, pp. 1–43.
[2] C. Urmson, J. Anhalt, D. Bagnell, C. Baker, R. Bittner, M. N. Clark,
J. Dolan, D. Duggins, T. Galatali, C. Geyer, M. Gittleman, S. Harbaugh, M. Hebert, T. M. Howard, S. Kolski, A. Kelly, M. Likhachev,
M. McNaughton, N. Miller, K. Peterson, B. Pilnick, R. Rajkumar,
P. Rybski, B. Salesky, Y.-W. Seo, S. Singh, J. Snider, A. Stentz, W. .
Whittaker, Z. Wolkowicki, J. Ziglar, H. Bae, T. Brown, D. Demitrish,

[8]
[9]

[11]
[12]
[13]

[15]
[16]

[17]

B. Litkouhi, J. Nickolaou, V. Sadekar, W. Zhang, J. Struble, M. Taylor,
M. Darms, and D. Ferguson, “Autonomous driving in urban environments: Boss and the urban challenge,” Journal of Field Robotics,
vol. 25, no. 8, 2008.
K. Saenbunsiri, P. Chaimuengchuen, N. Changlor, P. Skolapak,
N. Danwiang, V. pPoosuwan, R. Tienkum, P. anan Raktrajulthum,
T. Nitisuchakul, K. Bumrungjitt, S. Tunsiri, P. Khairid, N. Santi, and
S. yan Primee, “RobotCupRescue 2011 - robot league team iRAP
JUDY (thailand),” Tech. Rep., 2011.
E. Olson, J. Strom, R. Morton, A. Richardson, P. Ranganathan,
R. Goeddel, M. Bulic, J. Crossman, and B. Marinier, “Progress towards
multi-robot reconaissance and the MAGIC 2010 competition,” Journal
of Field Robotics, To appear.
H. Durrant-Whyte and T. Bailey, “Simultaneous localisation and
mapping (SLAM): Part I the essential algorithms,” Robotics and
Autonomous Systems, June 2006.
S. Thrun, W. Burgard, and D. Fox, Probabilistic Robotics. MIT Press,
2005.
F. Dellaert and M. Kaess, “Square root SAM: Simultaneous localization and mapping via square root information smoothing,” International Journal of Robotics Research, vol. 25, no. 12, pp. 1181–1203,
December 2006.
E. Olson, “AprilTag: A robust and flexible multi-purpose fiducial
system,” University of Michigan, Tech. Rep., 2010.
——, “Real-time correlative scan matching,” in Proceedings of the
IEEE International Conference on Robotics and Automation (ICRA),
Kobe, Japan, June 2009.
M. C. Bosse, “ATLAS: a framework for large scale automated mapping and localization,” Ph.D. dissertation, Massachusetts Institute of
Technology, Cambridge, MA, USA, February 2004.
E. Olson, “Robust and efficient robotic mapping,” Ph.D. dissertation,
Massachusetts Institute of Technology, Cambridge, MA, USA, June
2008.
——, “Recognizing places using spectrally clustered local matches,”
Robotics and Autonomous Systems, 2009.
R. Reid and T. Braunl, “Large-scale multi-robot mapping in MAGIC
2010,” in Robotics, Automation and Mechatronics (RAM), 2011 IEEE
Conference on. IEEE, 2011, pp. 239–244.
J. Butzke, K. Daniilidis, A. Kushleyev, D. D. Lee, M. Likhachev,
C. Phillips, and M. Phillips, “The university of pennsylvannia MAGIC
2010 mutli-robot team,” Journal of Field Robotics, To appear.
A. Lacaze, K. Murphy, M. D. Giorno, and K. Corley, “The reconnaissance and autonomy for small robots (RASR): MAGIC 2010
challenge,” Land Warfare Conference, 2010.
A. Boeing, M. Boulton, T. Brunl, B. Frisch, S. Lopes, A. Morgan,
F. Ophelders, S. Pangeni, R. Reid, and K. Vinsen, “WAMbot: Team
MAGICian’s entry to the multi autonomous ground-robotic international challenge 2010,” Journal of Field Robotics, To appear.
A. Erdener, E. O. Ari, Y. Ataseven, B. Deniz, K. G. Ince, U. Kazancioglu, T. A. Kopanoglu, T. Koray, K. M. Kosaner, A. Ozgur, C. C.
Ozkok, T. Soncul, H. O. Sirin, I. Yakin, S. Biddlestone, L. Fu, A. Kurt,
U. Ozguner, K. Redmill, O. Aytekin, and I. Ulusoy, “Team cappadocia
design for MAGIC 2010,” Land Warfare Conference, 2010.