![](bg2.png)
2 ROBOT MAP BUILDING
The design of a mapping system is usually strongly
influenced by the specific sensors used. In the sys-
tem described in this paper we employ a laser range
scanner sensor (like (Gonz
´
ales-Ba
˜
nos and Latombe,
2002; Kwon and S.Lee, 1999; Lu and Milios, 1997)).
In general, the mapping activity is characterized by
the type of the map the robot produces, namely by
the way in which the acquired information about ob-
stacles and free space is represented. One can dis-
tinguish between geometrical and topological maps.
Geometrical maps can be composed of grids (Thrun,
2001; Burgard et al., 2000), points (Lu and Milios,
1997), or segments (Austin and McCarragher, 2001;
Gonz
´
ales-Ba
˜
nos and Latombe, 2002). Point and seg-
ment maps have usually two main advantages over the
grid-based approaches: firstly, they are a much more
compact representation of the environment and, sec-
ondly, they are easier to use. The mapping system
proposed in this paper uses a representation of the ob-
stacles as sets of points on the obstacle contour. This
approach allows a simple update of the map as new
information is acquired by the laser range scanner.
The global map of an environment is incrementally
built by integrating partial maps on the basis of the
(probabilistic) estimated positions of the robot (Thrun
et al., 2000), of the geometrical features of the maps
(Lu and Milios, 1997), or of a mixture of the two ap-
proaches. Thus, in almost all cases, it is important
that, during the exploration, the robot is able to lo-
calize itself not only by odometry, but also by de-
tecting landmarks or by matching data with an ex-
isting model of the environment. Indeed, techniques
for Simultaneous Localization And Mapping (SLAM)
(Dissanayake et al., 2001) and for Concurrent Map-
ping and Localization (CML) (Thrun et al., 1998)
have been extensively studied. In our approach the
robot matches the geometrical features of the maps to
correct the estimation of its pose provided by odome-
try.
The exploration strategies usually aim at reducing
the exploration time, by making a small number of
exploration steps, while trying to reduce the error, by
balancing the sensory information with the navigation
cost to reach the observation positions. Since one of
the main contributions of this paper is the implemen-
tation of an innovative theoretically well-founded so-
lution to the problem of determining the next best ob-
servation position, also known as the Next Best View
(NBV) problem, in the following we compare our ap-
proach with some significant techniques proposed in
literature. (Gonz
´
ales-Ba
˜
nos et al., 2000; Gonz
´
ales-
Ba
˜
nos and Latombe, 2002) describes an exploration
method based on segment maps. Candidate NBV po-
sitions are generated across the edge of the explored
regions, in which the robot is guaranteed to move
without collisions risk. To decide whether a candi-
date position in the free space is a good NBV position,
the amount of new information about the environment
that could be obtained from the candidate position is
estimated. More precisely, the evaluation of a candi-
date position q given by
g(q) = A(q) · exp(−λ · L(q))
where A(q) is an estimation of the unexplored area
visible from q, L(q) is the length of the path con-
necting the current robot position and q, and λ (set
to 20
−1
cm) weights the new information obtainable
from a candidate and the cost of travelling to reach
the position. The best candidate will minimize the
function g(q). Although extensive experimental re-
sults validated the effectiveness of the above criterion,
its theoretical bases are not well defined. For exam-
ple, the use of the exponential function in evaluating
the worthiness of a position and the value of the λ pa-
rameter are not theoretically justified. The criterion
proposed in this paper overcomes this limitation since
it has stronger mathematical bases.
The probabilistic approach proposed in (Burgard
et al., 2000) coordinates the exploration activities
of multiple robots in order to minimize the overall
exploration time. In this case, grid maps are em-
ployed and the cell in which a robot should move next
is the one that maximizes the difference between a
(probabilistic-based) measure of expected utility and
the path cost. Differently from our proposal, also
this criterion lacks of strong theoretical foundations;
moreover, it naturally applies to grid maps and its
extension to the more compact point maps is not
straightforward.
Finally, in (Sim and Dudek, 2003), a number of ex-
ploration policies (paths followed by the robot dur-
ing mapping) are evaluated for efficiency. Although
mathematically well-founded on Kalman filtering the-
ory, this work differs from our approach since it does
not evaluate only the next observation position but
a whole set of pre-determined observation positions
along a path.
3 THE PROPOSED MAPPING
SYSTEM
The mapping system proposed in this paper tries to
find an optimal exploration path by determining at
each iteration the next best observation position. The
robot performs the mapping process taking scans of
the environment and using them to update a world
model and to localize itself. A scan is composed of
a list of points that represent the outline of the world
objects within the range of the laser range scanner.
ICINCO 2004 - ROBOTICS AND AUTOMATION
72