The 2009 IEEE/RSJ International Conference on

Intelligent Robots and Systems

October 11-15, 2009 St. Louis, USA

NDT Scan Matching Method for High Resolution Grid Map

Tomohito Takubo, Takuya Kaminade, Yasushi Mae, Kenichi Ohara, Tatsuo Arai

Abstract— A new convergence calculation method of the is expanded according to the average distance between the
Normal Distributions Transform (NDT) scan matching for initial input scanning points and the corresponding reference
high resolution of grid maps is proposed. NDT scan matching grids. After the input points come close to the reference
algorithm usually has a good effect on large grids, so it
is difficult to generate the detailed map with small grids. grids, the input points are evaluated and adjusted precisely
The proposed method employs Interactive Closest Point(ICP) by using the original normal distribution.
algorithm to find corresponding point, and it also enlarges This paper proposes additional scan matching techniques
the convergence area by modifying the eigenvalue of normal for detailed mapping. Since the resolution of Laser Range
distribution so that the evaluation value is driven effectively Finder(LRF) is usually limited, the small object is not only
for the pairing data. In addition, outlier elimination process
is implemented to the scanning for sub-grid scale object. The ignored but also recognized in the false location by the
scanning data fromLeaser Renge Finder(LRF) have error but insufficient reflection. The false recognition is divided into 3
its set of detected small object can be clustered to determine cases: edge of object, locating in front of large object, surface
the Center of Mass(CoM) and the outlier data. The outlier of wall. Each recognition is corresponding to the location
commonly locates behind true points and it can be eliminated from the sensor. According to the scanning resolution of
when the robot observes from other point. Experimental result
shows the effectiveness of the proposed convergence algorithm LRF, the reliability of scanning area is decided to 2 cases:
and outlier elimination method. ”reliable”, ”suspicious”. Measurement points are judged by
the false recognition, and outlier is deleted. These techniques
I. INTRODUCTION are implemented to the grid mapping and the experimental
results show the feasibility of proposed method.
Simultaneous Localization and Mapping (SLAM) is im-
This paper is organized as follows. Section II discusses
portant technique to know unknown environment by the mo-
related works about environmental mapping. In section III,
bile robot. Normal Distributions Transform (NDT) algorithm
we describe the basic technique of grid mapping with NDT
[2] is one of the scan matching methods for the convergence
algorithm. Section IV explains proposed high resolution grid
calculation. In our previous study[1], we employs the grid
mapping and the difference from last study. Section VI shows
map with NDT scan matching; The method has the features
experimental results and the effectiveness of the proposed
of fast convergence speed and memory saving for data
method. Finally, we conclude this paper in section VII.
administration. When we implement grid map for navigation
of small robots, the grid size should be small and the detailed II. PREVIOUS WORK
map is necessary according to the robot size. However, there
Scan matching is necessary to prevent the accumulation
are problems to apply the NDT to the detailed grid map.
of the error of input data. Iterative Closest Point (ICP)
Since the grid size is small, it can capture only a few input
algorithm [4] is one of the scan matching algorithms. The
points in one scanning, and there is only a few grid having
algorithm estimates the optimal pose by minimizing the sum
the normal distribution value. Furthermore, a small error
of squared distance between corresponding points. To find
makes the input data be in the outside of the reference grid
the nearest corresponding point of each input point, the
area.
calculation load takes O(MN). Since the mapping data is
To solve the problems, we proposed two techniques for managed by collecting each coordinate value, the volume
the high resolution grid map with NDT scan matching. First, of data is increasing as the number of input scan is in-
the Interactive Closest Point(ICP) algorithm is implemented creasing. The major calculation load of the ICP spends for
to find corresponding point even if the input scan points the searching the nearest corresponding point. By using k-d
are in the outside of the reference grid area. Secondly, tree space quantization algorithm, it takes O(N log(M)) [5].
the convergence area adjustment method is proposed to Furthermore, additional calculation O(n log n) is required for
improve the evaluation value with all the pairing data. The building the k-d tree structure. Elias method [6] is another
normal distribution area enlarges as its eigenvalue becomes scan matching algorithm to find a nearest corresponding
large. By using the characteristics, the phased adjustment point. The method divides the scanning area into grids and
of convergence area was proposed. The convergence area checks them sequentially so that the corresponding point is
This research was partially supported by the Ministry of Education, found fast. In this case, ICP needs linear time. It depends
Science, Sports and Culture, Grant-in-Aid for Scientific Research (B), on the shape of the map data whether it is better to use an
18360123, 2008. Elias method or a k-d tree method.
T. Takubo is with Department of Systems Innovation, Graduate School of
Engineering Science, Osaka University, Toyonaka, Osaka, 560-8531, Japan Biber proposed the Normal Distributions Transform
[email protected] (NDT) algorithm as a new scan matching method [2]. This

978-1-4244-3804-4/09/$25.00 ©2009 IEEE 1517

 A. Normal Distributions Transform (NDT)
The NDT algorithm is one of the scan matching methods.
In this algorithm, the scanned space is divided into cells,
and the input points in each cell are converted into normal
distribution which characterizes the distribution of points.
The latest input scan is evaluated in each cell by the normal
distribution to converge them to the feasible shape. Since the
evaluation is performed only in the cell including both input
Fig. 2. Flow of NDT scan matching scan points and normal distribution, the calculation load is
Fig. 1. NDT converts for refer-
ence scan O(N).
The NDT process is as follows. First, normal distribution
is calculated by using point data in each cell(Fig.1). The
method divides the scanning space into grids at regular cell having normal distribution is defined ”reference scan”.
intervals, and the distributions of scanning data in each Second, the input scan points are matched to the reference
grid are approximated by the normal distribution. The ICP scan (Fig.2). For the process, it is necessary to store and
algorithm matches the input point with the reference point, update an average coordinate and a covariance matrix of
but the NDT algorithm matches the input point with the reference scan in each grid.
normal distributions of the reference grid. NDT algorithm Now, we define following variables.
do not need to seek the reference grids for the input scan, • p = (pi )ti=1,2,3 = (t x , ty , θ)t : A coordinate transformation
so the calculation cost is only O(N) and the matching parameter between the reference scan and the input
speed is fast. Martin [8] and Takeuchi [9] apply NDT scan, where (t x , ty )t describes the translation and θ is
to three dimensions. Takeuchi proposed new convergence the rotation.
algorithm which widens the size of grid locating in long- • xi : A point coordinate of the input scan.
′
distance to catch the large error input. Nora [10] applied • xi :The input scan data which were transformed by
a multi grid size algorithm for matching, and an iterative coordinate transformation parameter p
calculation algorithm with Lievenberg-Marquardt method for • qi , Σi : The mean coordinate and the covariance matrix
′
effective convergence. Martin also recently proposed new of the normal distribution corresponding to the point xi
approach to appearance based place recognition exploiting • Mi : The total number of the points in the cell.
NDT [11]. Furthermore, they presented a comparison of 3D The evaluation function is defined as,
scan registration algorithms: ICP, NDT[12]. ′ ′
∑ −(xi − qi )t Σ−1
i (xi − qi )
N−1
The advantage of the NDT algorithm is that the calculation
E(X, p) = exp . (1)
amount is few. However, the convergence performance de- 2
i
pends on the grid size because the outlier from the reference
′
grid is not used. In the previous study of NDT, the grid is The xi is calculated using the input data xi and the transfor-
only used for matching input points with reference grids; the mation parameter p as follow:
grid size is usually larger than several tens of centimeters,
so it’s not useful to display map. Thus, the point data is  ′       
 
′   
x cos θ − sin θ  x t x   x cos θ − y sin θ + t x 
    + = .
also stored to display detailed environmental map. In our x i = ′ =      (2)
y sin θ cos θ y ty x sin θ + y cos θ + ty 
last study[1], we implemented NDT algorithm and grid
mapping with small grid. Since the grid size is small, the Since the Newton’s Algorithm is a nonlinear function min-
map has enough resolution and the data volume for mapping imization algorithm, an optimizing function for NDT scan
doesn’t relate to the scanning number of times; the point matching must be;
data is not stored. We proposed high resolution grid mapping
method with NDT and showed effectiveness. In this paper, f (p) = −E(X, p). (3)
we compare the several grid sizes and the search area in our Parameter p is updated by following equation;
proposed method, and evaluate the accuracy of convergence.
In addition, we implement outlier elimination algorithm for ∆p = −H −1 g , pnew = pold + ∆p, (4)
small object sensing.
where g and H are calculated for each input points by
III. NDT Grid Map following equations;

This section briefly describes the NDT scan matching ∑

N−1 ∑
N−1
g= e
gi , H= ei.
H (5)
algorithm and the grid map. We will call ”NDT Grid Map”
i=0 i=0
that employs grid mapping and NDT scan matching. We
can expect fast convergence calculation and ease of data e
gi and He i are partial differential and second partial differen-
management. Since the grid map, robots can use directly tial of optimizing function f (p) with p. (For details, please
it for navigation. refer to [2])

1518
 The computational amount for q j and covariance Σ j is
reduced by the following equations;
0.41[m]
m j = m jold + x j , M j = M jold + N j , S j = S jold + x j x j (6)

mj S j − q j mtj 0.29[m]
qj = , Σj = (7)
Mj Mj (a)Grid Map(b)Distribution(c)Enlarged Distribution
where N j is number of input points for each cell at j scan. Fig. 3. Enlarged normal distributions of reference scans
Each cell has m j and S j , and they are updated by eq. (6).
When the scan matching is operated, the average q j and the
covariance matrix Σj are calculated by eq. (7). ReferenceScan

B. Grid Map
Grid map is one of the representation methods. In this Input Scan
method, space is divided into a lattice of congruent cells.
The cell is minimum resolution and the measured points
in the cell are destroyed. Thus, memory consumption does
not increase even if a new input scan was stored. To share (a)Elias Method (b)Enlarge Grid (c)Make Distribution
the data of NDT scanning match, each grid stores following
Fig. 4. Enlarged Reference Scan
elements:
• M : The count of inputted data
• m :A summation of the coordinate vectors
Let φ j be an eigenvector of a eigenvalue λ j , and let k
• S :A matrix which becomes the covariance matrix
be an expansion coefficient of the eigenvalue, a new
′
IV. The high resolution grid mapping with NDT covariance matrix Σj is given as follow;
A. new convergence algorithm ′ ′
Σj = VΛ V t (8)
Grid size is a key of NDT algorithm. If the grid size
is small, the map becomes high resolution, but the total [ ]
′ kλ1 0
computational amount increases. On the contrary, if a grid V = [φ1 φ2 ] , Λ =
0 kλ2
size is big, a computational amount becomes little, but the
resolution becomes low. 3) Fusion around minority reference cell
When the grid size is small, the pair of a reference scan If a cell had only one or two points of data, it
and a input scan will be easily changed, and the evaluation cannot make normal distribution. Small grid makes
value will be unsteadily changed, even if the input error is this problem frequently at the first scan time and
small. Furthermore, the NDT scan matching does not make far place. To solve the problem, we implement the
pair with the neighboring grid. Thus, the available pairs summing up process to the neighboring around the
decrease and the convergence calculation becomes difficult. cell which has three or less points.. Fig.4 shows the
To solve the problems, we propose following three oper- summing up process. When a reference cell for an
ations for NDT algorithm. input scan has points less than three like Fig.4 (a) and
1) Search for neighboring grids it cannot make normal distribution, The neighboring
The NDT algorithm doesn’t need to search a reference cells around the cell are checked like Fig.4 (b) so
scan for a input scan. However, when the grid size is that the summed up cell is made ike Fig.4 (c). Extra
small, even if the error of input is small, it reduces data space is not necessary to implement the algorithm,
the number of matching pair. The problem is solved since it is sufficient to calculate sum of eq.(6) about
by using Elias method searching for nearest reference the neighboring cells.
scan from neighboring grids.
2) Expansion of normal distribution B. Evaluation of proposed method
When the input scan is matched with the reference Fig.5(a) shows experimental data; gray points indicate
scan by using above process, if the distance of the two input scan, black points indicate reference scan. We compare
scans is too long, the degree of conformance becomes several grid size and the changes of the evaluation value with
small and it does not affect convergence process. As the error of input scan are compared by several grid sizes and
a solution of the problem, the normal distribution expansion coefficients. Fig.5(b)-(f) shows evaluation value
expansion method is proposed; it enlarged the skirts E(p) of eq.(1). The input error is given ±40 [cm] in each x,
of the normal distribution to give an effectual value y axis.
to a long-distance pair. Specifically, we enlarge two First, Fig.5(b), (c) show the evaluation value of the original
eigenvalues of the covariance matrix Σ of the eq. (7). NDT algorithm by using 2 and 20 [cm] grid. The distribution

1519
of evaluation value in (c) is smoother rather than (b), and
the area of local minimum are decreased. The original NDT 70

Evaluated Value
60
algorithm, which doesn’t match with neighboring grids, has 50
40

\>FP@
good performance by using large grid size. Next, Fig.5(d), 30
20
(e) show the evaluation value of the proposed NDT algorithm 10
by using 2[cm] grid with expansion coefficient k=10. (e) em- 0
40
20 40
ploys also the fusion process around minority reference cell. dy[cm] 0 0
20

The evaluation value of (d) is increased overall compared -20 -20 dx[cm]
[>FP@ -40 -40
with (b) in the same 2[cm] grid, and the local minimum
(a)2[cm], BA, Eigenvalue×1 (b)2[cm], BA, Eigenvalue×1
also decreases. The distribution of evaluation value of (e)
is expanded rather than (d), so that the function make easy

Evaluated Value
to converge large error inputs. Fig.5(f) shows the evaluation 180
250 160

Evaluated Value
140
value using 2[cm] grid with expansion coefficient k=100 and 200 120
150 100
fusion algorithm. The distribution of evaluation value of (f) 80
60
100 40
is also expanded rather than (e), thus the proposed algorithm 50 20
0
can be better as the expansion coefficient increases. Based on 0
40
40
20 40
40 20
the result, our last study[1] proposed an phased adjustment 20 20 dy[cm]0 0
dy[cm] 0 0 -20 -20 dx[cm]
algorithm of convergence area by reducing gradually the -20 -20 dx[cm] -40 -40
-40 -40
expansion coefficient k. However, the phased adjustment (d)2[cm], PA, Elias Method,
takes much time, and if the input error is not so big and the (c)20[cm], BA, Eigenvalue×1
Eigenvalue×10
expansion coefficient is fixed, the convergence result is not
so different. Thus, in this report, we employ fixed expansion
Evaluated Value

Evaluated Value
300 450
coefficient k=10. 250
400
350
200 300
250
C. Experiment using NDT grid map 150 200
100 150
100
Fig.6(a) shows experimental environment. The multi- 50 50
0 0
legged robot ”ASTERISK”[13] and LRF (URG-04LX : 40
40
40
20
40
20 20 20
Hokuyo Electric Co., Ltd.) are employed. In this paper, a dy[cm]0 0 dy[cm] 0 0
-20 -20 dx[cm] -20 -20 dx[cm]
grid size is defined 2[cm] and the size of a total grid space -40-40 -40 -40

is defined 8[m] based on the spec of LRF. The 2[cm] grid (e)2[cm], PA, Elias Method, (f)2[cm], PA, Elias Method,
is strict for the LRF, since it’s accuracy is ±10[mm]. The Eigenvalue×10, Enlarged Eigenvalue×100, Enlarged
dead reckoning has error since the calibration error of the grid grid
joint and the slipping on the floor. This sensor measures a 2
dimensional plane, and the range covering ± 120 degrees. It Fig. 5. The evaluation value
can get 683 points’ distances at one scan.
Fig.6(b)-(d) show experimental results. All maps consist of
2[cm] grids, and the black point represents scanned grids, the
gray line represents trajectory of robot. (b) is made by using
only dead reckoning, (c) employs original NDT algorithm:
it means k=1, (d) is generated by using proposed algorithm occurs frequently. Because of the detection principle of LRF,
with k=10. The error of dead reckoning is corrected by using the detection reaction is changed by the object condition:
NDT algorithm and more effective modification is conducted scale, surface shape, material etc., and it is a reason of false
by using our proposed method. detection. Fig.7 shows one shot scanning examples for a
V. Recognition for subgrid scale object small object. The LRF is located (x,y)=(0,0). The object is
The detectable scale of LRF is limited by the resolution iron stick whose diameter is 1.3[cm]. The black circle is true
performance, and the small object leads false recognition. position and the gray points are data from LRF. (a) is the
Our LRF has ±10[mm] resolution in radiation direction, result that only a metal stick scanned. The false detections
but the angular resolution capability is 0.36[deg]. Thus, the exist at the edge of object, and the error position tends to
resolution in the tangential direction decreases as the length be located behind the object. Thus, the nearest data cloud
in the radiation direction increases. If the detect point was is dependable. (b) is the result which scans a metal stick in
not so long, the 2[cm] grid is strict size for the small object front of wall. In this case, the small object is not completely
which smaller than 2[cm]. In this section, we propose subgrid detected: all scanned points shift to backward. The dotted
size object recognition method. line is setup position for the metal stick, so the detected
result has 5[cm] error. From fig.6, another current error is
A. Variation of false detection for subgrid object found in the surface of wall. Now, we categorize the error
The small object can be detected by using latest scanning pattern into three: edge of object, small object in front of
data, but it should be not stored because the false detection wall, surface of wall.

1520
 300
of object can be checked by clustering method in one shot
200 scanning data. The process is as follows;
a-1 Apply clustering process to the new input scan data and
100 select the cluster with small number.
a-2 Calculate Center of Mass(CoM) position of the cluster.

y[cm]
0
a-3 Remove the data in back of the CoM position.
-100
a-4 Convert the point data into grid data and store only the
grid being at the forefront.
-200 These processes are implemented in both Area 1 and 2.
(a)The Experimental 2) Small object in front of wall: In Area 2, when the
-300
Environment -200 -100 0 100 200 300 object is in front of wall/big object, the detected points move
x[cm] to backward. However, in Area 1, the detected points are true
(b)Without Scan Matching value in the same situation; they don’t move to backward.
300 300 Thus, we implement following process;
200
b-1 Delete the grid not observed in the Area 1.
200
b-2 If there are former grid data in the short of radiation
100 100 direction, don’t store the new scanned points; it might be
false detection moving to backward.
y[cm]
y[cm]

0 0 3) Surface of wall: The false detection grids at the surface

of wall/big object tends to be observed only a few times.
-100 -100
Thus, we can check the false detection by the number of
-200 -200 scanned times, delete them. In practical use, we define the
frequency of the checking once every three times.
-300 -300
-200 -100 0 100 200 300
4) Confirmation of detection result: There are still some
-200 -100 0 100 200 300
x[cm] x[cm] leavings after the above elimination process. When the
(c)NDT algorithm (BA) (d)NDT algorithm (PA) scanned grid is converted to the reference scan grid, the grids
are checked in the radiation direction and if there are former
Fig. 6. The Experiment grid data in short, eliminate the old grids. For the detection
of small object, this process is employed in Area 1 because
the detection in Area 2 is not reliable. Meanwhile, since the
25 big object is stably observed, we employ this process both
20
15
in Area 1 and 2, so that the mobile object can be eliminated.
2
10
VI. Environmental map with small object
y[cm]

5
y[cm]

0
0
-5 By using the proposed NDT scan matching algorithm and
-2 -10 subgrid object error detection, the environmental map with
36 38 40 42 44 46 -15
x[cm]
-20
small objects is generated. Six wooden sticks whose diameter
(a)Edge Error 40 50 60 70 80 90 is 1.5[cm] and two iron sticks whose diameter is 1.3[cm] are
x[cm]
prepared in the experimental environment. The expansion
(b)Small object and Wall. coefficient of the proposed NDT algorithm is set k=10.
Fig. 7. Errors of LRF Fig.8 shows experimental result. Fig.8(a) is made without
error elimination algorithm. Fig.8(b) is using the proposed
elimination algorithm. The errors around the small objects
and the surface of wall are almost eliminated. The average
error of detected CoM positions is decreased from 4.1[cm] to
B. Elimination method for false detection
2.3[cm]. Fig.9 is the enlarged view of the surroundings of the
The catalog spec of LRF(URG-04LX) indicates that the iron stick at the center in Fig.8. Fig.9(a) shows the first map
smallest object detection is defined 1[deg]; 3 neighboring by using section V-B but the errors still remain; the robot is
laser spot. It means 2[cm] object can be detected near than moving around start position. Fig.9(b) shows the final map,
120[cm]. However, our preliminary experiments says that the it is confirmed that the errors are almost eliminated since the
reliable area is within 50[cm] for 2[cm] size objects and there robot is located opposite side of the iron stick.
are a lot of false detection in the area from 50 to 120[cm]. To
change the false detection algorithm, we divide the scanning VII. Conclusion
area into two areas; Area 1 is defined within 50[cm], Area High resolution grid mapping with NDT scan matching
2 is the other. and the subgrid object recognition are proposed. The con-
1) Edge of object: The false detection of edge can be vergence algorithm consists of three processes: search for
removed before inputting grid map, since the each data set neighboring cell, expansion of normal distribution, fusion

1521
 300 -40 -40
-42 -42
200 -44 -44
-46 -46

y[cm]

y[cm]
100 -48 -48
Robot Trajecotry -50 -50
y[cm]

Wooden Obstacles -52 -52

0
Metalic Obstacles
-54 -54
Grids
-100
-56 -56
26 28 30 32 34 36 38 26 28 30 32 34 36 38
x[cm] x[cm]

-200 (a)Errors are included (b)Errors are eliminated

Fig. 9. Efficacy of the Elimination Method
-300
-200 -100 0 100 200 300
x[cm]
(a)Map including errors [4] P.J.Besl and N.D.McKay, ”A method for registration of 3d shapes”,
300
IEEE Transactions on Pattern Analysis and Machine Intelligence,
1992, 14(2):239-256.
200 [5] M. Greenspan and M. Yurick: ”An Approximate K-D Tree Search for
Efficient ICP”, 3DIM03: 4th International Conference on 3-D Digital
Imaging and Modeling, Banff, Alberta, Canada, 2003, Oct. 6-10.
100 [6] M.A.Greenspan, G.Godin, J.Talbot: ”Acceleration of Binning Near-
Robot Trajecotry est Neighbor Methods”, In Proc. of Vision Interface 2000, Mon-
y[cm]

Wooden Obstacles treal,Canada, May 2000, pp. 337-344.

0
Metalic Obstacles [7] Kazunori Ohno, Takafumi Nomura, Satoshi Tadokoro: ”Real-Time
Grids
Robot Trajectory Estimation and 3D Map Construction using Camera”,
-100
In Proc. of the 2006 IEEE International Conference on Intelligent
robots and Systems, 2006, pp.5279-5285.
[8] Martin Magnusson, Tom Duckett, Rolf Elsrud, and Lars-Erik Skager-
-200 lund: ”3D Modelling for Underground Mining Vehicles”, In Proc. of
SimSafe 2005 : modeling and simulation for public safety Linkping,
Sweden, 2005, May 30.
-300 [9] Eijiro Takeuchi and Takashi Tsubouchi: ”A 3-D Scan Matching using
-200 -100 0 100 200 300
Improved 3-D Normal Distributions Transform for Mobile Robotic
x[cm]
Mapping”, In Proc. of the 2006 IEEE/RSJ International Conference
(b)Errors are eliminated by proposed method. on Intelligent Robots and Systems, Beijing, China, 2006, October 9-15.
[10] Ripperda, N. and Brenner, C.: ”Marker-Free Registration of Terrestrial
Fig. 8. The Experimental Results Laser Scans Using the Normal Distribution Transform”, In Proc. of
the ISPRS Working Group V/4 Workshop 3D-ARCH 2005: ”Virtual
Reconstruction and Visualization of Complex Architectures”, Mestre-
Venice, Italy, 2005.
[11] Martin Magnusson, Henrik Andreasson, Andreas Nüchter, and Achim
around minority reference cell. Each process contributes to J. Lilienthal: ”Appearance-Based Loop Detection from 3D Laser Data
derive effective evaluation value, as a result the grid map with Using the Normal Distributions Transform”, In Proc. of 2009 IEEE
small cell is available to know the detailed environment. In International Conference on Robotics and Automation(ICRA 2009),
2009, pp.23-28.
addition, the subgrid size object, which is difficult to know [12] Martin Magnusson, Andreas Nüchter, Christopher Lörken, Achim J.
precisely depending on the sensor spec, is recognized by the Lilienthal and Joachim Hertzberg, ”Evaluation of 3D Registration Re-
error rejection algorithm for the vague input. Experimental liability and Speed - A Comparison of ICP and NDT”, In Proc. of 2009
IEEE International Conference on Robotics and Automation(ICRA
result shows the effectiveness of the detailed environmental 2009), 2009, pp.3907-3912.
map generated by the NDT high resolution grid mapping. [13] Tomohito TAKUBO, Tatsuo ARAI, Kenji INOUE, Hikaru OCHI,
The method will be used for navigation of small robot. Our Takeshi KONISHI, Taisuke TSURUTANI, Yasuo HAYASHIBARA,
Eiji KOYANAG: ”Integrated Limb Mechanism Robot ASTERISK”, In
next target is making of 3D environmental map. Journal of Robotics and Mechatoronics, Vol.18, No.2, 2006, pp.203-
214.
References
[1] Takuya Kaminade, Tomohito Takubo, Yasushi Mae and Tatsuo Arai:
”The Generation of Environmental Map Based on a NDT Grid Map-
ping”, In Proc. of 2008 IEEE International Conference on Robotics
and Automation(ICRA 2008), 2008, pp.1874-1879.
[2] Peter Biber and Wolfgang Straβer: ”The Normal Distributions Trans-
form:A New Approach to Laser Scan Matching”, In Proc. of the 2003
IEEE International Conference on Intelligent robots and Systems, pp.
2743-2748, 2003.
[3] Fumio Kanehiro, Takahashi Yoshimi, Shuuji Kajita, Mitsuharu Mori-
sawa, Kiyoshi Fujiwara, Kensuke Haruda, Kenji Kaneko, Hirohisa
Hirukawa, Fumiaki Tomita: ”Whole Body Locomotion Planning of
Humanoid Robots based on a 3D Grid Map”, In Proc. of the 2005
IEEE International Conference on Robotics & Automation, pp.1084-
1090,2005.

1522