Square Root SAM

Frank Dellaert
College of Computing
Georgia Institute of Technology

Abstract Solving the SLAM problem is one way to enable

a robot to explore, map, and navigate in a previously unknown
environment. We investigate smoothing approaches as a viable
alternative to extended Kalman filter-based solutions to the
problem. In particular, we look at approaches that factorize either
the associated information matrix or the measurement matrix
into square root form. Such techniques have several significant
advantages over the EKF: they are faster yet exact, they can be
used in either batch or incremental mode, are better equipped
to deal with non-linear process and measurement models, and
yield the entire robot trajectory, at lower cost. In addition,
in an indirect but dramatic way, column ordering heuristics
automatically exploit the locality inherent in the geographic
nature of the SLAM problem.
In this paper we present the theory underlying these methods,
an interpretation of factorization in terms of the graphical model
associated with the SLAM problem, and simulation results that
underscore the potential of these methods for use in practice.

I. I NTRODUCTION
The problem of simultaneous localization and mapping
(SLAM) [1][3] has received considerable attention in mobile
robotics as it is one way to enable a robot to explore, map, and
navigate in previously unknown environments. The traditional
approach in the literature is to phrase the problem as an
extended Kalman filter (EKF), with the robot pose and static
landmarks as the evolving filter state [2], [4], [5]. It is well
known that the computational requirements of the EKF become impractical fairly quickly, once the number of landmarks
in the environment grows beyond a few hundred. As a result,
many authors have looked at ways to reduce the computation
associated with the EKF, using both approximating [6][8] and
non-approximating [9] algorithms.
In this paper we propose that square root information
smoothing (SRIS) is a fundamentally better approach to the
problem of SLAM than the EKF, based on the realization that,
in contrast to the extended Kalman filter covariance or
information matrix, which both become fully dense over
time [7], [8], the information matrix I associated with
smoothing is and stays sparse;
in typical mapping scenarios this matrix I or, alternatively, the measurement matrix A, are much more
compact representations of the map covariance structure
I or A, both sparse, can be factorized efficiently using
sparse Cholesky or QR factorization, respectively, yielding a square root information matrix R that immediately
yields the optimal robot trajectory and map;
Factoring the information matrix is known in the sequential estimation literature as square root information filtering (SRIF).

The SRIF was developed in 1969 for use in JPLs Mariner

10 missions to Venus (as recounted by [10]), following the
development of covariance square root filters earlier that
decade. The use of square roots of either the covariance or
information matrix results in more accurate and stable algorithms, and, quoting Maybeck [11] a number of practitioners
have argued, with considerable logic, that square root filters
should always be adopted in preference to the standard Kalman
filter recursion. Maybeck briefly discusses the SRIF in a
chapter on square root filtering, and it and other square root
type algorithms are the subject of a book by Bierman [10].
However, as far as this can be judged by the small number
of references in the literature, the SRIF and the square root
information smoother (SRIS) are not often used.
In this paper we investigate the use of factorizing either
the information matrix I or the measurement matrix A into
square root form, as applied to the problem of simultaneous
smoothing and mapping (SAM). Because they are based on
matrix square roots, we will
refer to this family of approaches
as square root SAM, or SAM for short. They have several
significant advantages over the EKF:
They are much faster than EKF-based SLAM
They are exact, rather than approximate
They can be used in either batch or incremental mode
If desired, they yield the entire smoothed robot trajectory
They are much better equipped to deal with non-linear
process and measurement models than the EKF
When using QR, they are more accurate and stable
They automatically exploit locality in the the way that
sub-map [6] or compressed filter [9] SLAM variants do
However, there is also a price to pay:
Because we smooth the entire trajectory, computational
complexity grows without bound over time, for both
Cholesky and QR factorization strategies. In many typical
mapping scenarios, however, the EKF information matrix
will grow much faster.
As with all information matrix approaches, it is expensive
to recover the covariance matrix governing the unknowns.
We also present an interpretation of the resulting algorithms
in terms of graphical models, following [8], [12][14]. Doing
so yields considerable insight into the workings of otherwise
opaque black box algorithms such as Cholesky or QR
factorization. It exposes their kinship with recently developed inference methods for graphical models [15], such as
the junction tree algorithm. In particular, in Section II we
introduce the SLAM problem in terms of a directed graph

the state of the robot at the ith time step by xi , with i 0..M ,
a landmark by lj , with j 1..N , and a measurement by zk ,
with k 1..K. The joint probability model corresponding to
this network is
P (X, M, Z) =
P (x0 )

M
Y

P (xi |xi1 , ui )

i=1

Fig. 1. A synthetic environment with 20 landmarks in which we simulated

a robot taking 156 bearing and range measurements along a trajectory of 21
poses. Also shown are the mean and covariance matrices as estimated by an
EKF. Note the effects of process and measurement noise.

x0

z1

x1

z2

l1

x2

z3

...

...

z4

l2

...

xM

zK

lN

Fig. 2.
Bayesian belief network representation of the SLAM problem.
The state x of the robot is governed by a Markov chain, on top, and the
environment of the robot is represented at the bottom by a set of landmarks
l. The measurements z, in the middle layer, are governed both by the state
of the robot and the parameters of the landmark measured.

or belief net. However, to understand factorization, the view

of smoothing in terms of an undirected Markov random field
(MRF), introduced in Section V, is more appropriate. By
exploiting the SLAM-specific graph structure, we are able
to immediately speed up factorization by a factor of 2. We
believe that even more efficient algorithms can be developed
by viewing the problem as one of computation on a graph.
II. SLAM BACKGROUND

K
Y

P (zk |xik , ljk ) (1)

k=1

where P (x0 ) is a prior on the initial state of the robot,

P (xi |xi1 , ui ) is the motion model, parameterized by a control
input ui , and P (z|x, l) is the landmark measurement model.
The above assumes a uniform prior over the landmarks l.
Furthermore, it assumes that the data-association problem has
been solved, i.e., that the indices ik and jk corresponding to
each measurement zk are known.
As is standard in the SLAM literature [2], [4], [5], we assume Gaussian process and measurement models [11], defined
by
xi = fi (xi1 , ui ) + wi
(2)
where fi (.) is a process model, and wi is normally distributed
zero-mean process noise with covariance matrix i , and
zk = hk (xik , ljk ) + vk

(3)

where hk (.) is a measurement equation, and vk is normally

distributed zero-mean measurement noise with covariance k .
The equations above model the robots behavior in response
to control input, and its sensors, respectively.
For the prior P (x0 ), we will assume that x0 is given and
hence it is treated as a constant below. This considerably
simplifies the equations in the rest of this document. In
addition, this is what is often done in practice: the origin of
the coordinate system is arbitrary, and we can then just as well
fix x0 at the origin. The exposition below is easily adapted to
the case where this assumption is invalid.
Below we also need the first-order linearized version of the
process model (2), given by
x0i + xi = fi (x0i1 , ui ) + Fii1 xi1 + wi

(4)

where Fii1 is the Jacobian of fi (.) at the linearization point

x0i1 , defined by

fi (xi1 , ui ) 
Fii1 =
 0
xi1
x
i1

SLAM refers to the problem of localizing a robot while

simultaneously mapping its environment, illustrated in Figure
1. Below we assume familiarity with EKF-based approaches
to SLAM [2], [4], [5]. In this section we introduce the SLAM
problem and the notation we use, but we do not re-derive the
extended Kalman filter. Rather, in Section III we immediately
take a smoothing approach, in which both the map and the
robot trajectory are recovered.
Following the trend set by FastSLAM and others [12], we
formulate the problem by referring to a belief net representation. The model we adopt is shown in Figure 2. Here we denote

Note that ui is given and appears as a constant above. The

linearized measurement equations are obtained similarly,
zk = hk (x0ik , lj0k ) + Hkik xik + Jkjk ljk + vk
Hkik

(5)

Jkjk

where
and
are respectively the Jacobians of hk (.)
with respect to a change in xik and ljk , and are evaluated at
the linearization point (x0ik , lj0k ):




jk hk (xik , ljk ) 
ik hk (xik , ljk ) 
Hk =
Jk =

 0 0
xik
ljk
(x0 ,l0 )
(x ,l )
ik

jk

ik

jk

III. S MOOTHING SLAM AND L EAST S QUARES

We investigate smoothing rather than filtering, i.e., we are
interested in recovering the maximum a posteriori (MAP)

estimate for the entire trajectory X = {xi } and the map

L = {lj }, given the measurements Z = {zk } and control

inputs U = {ui }. Let us collect all unknowns in X and L in

the vector = (X, L). Under the assumptions made above,

we obtain the MAP estimate

= argmax P (X, L|Z) = argmax P (X, L, Z)

= argmin log P (X, L, Z)

by solving the following non-linear least-squares problem:

M
X

kxi fi (xi1 , ui )k2i +

i=1

K
X

kzk hk (xik , ljk )k2k (6)

k=1

Here kek2 = eT 1 e is the squared Mahalanobis distance.

In practice one always considers a linearized version of
this problem. If the process models fi and measurement
equations hk are non-linear and a good linearization point
is not available, non-linear optimization methods such as
Gauss-Newton iterations or Levenberg-Marquardt will solve a
succession of linear approximations to (6) in order to approach
the minimum [16]. This is similar to the extended Kalman
filter approach to SLAM as pioneered by [4], [17], [18], but
allows for iterating multiple times to convergence.
In what follows, we will assume that either a good linearization point is available or that we are working on one
of these iterations. In either case, we have a linear leastsquares problem that needs to be solved efficiently. Using the
linearized process and measurement models (4) and (5):
= argmin

M
X

kHkik xik + Jkjk ljk ck k2k

(7)

k=1

where we define ai = x0i fi (x0i1 , ui ) and ck = zk

hk (x0ik , lj0k ). To avoid treating xi in a special way we also
introduce the matrix Gii = Idd , with d the dimension of
the robot state, and we drop the notation as implied.
Below we assume, without loss of generality, that the
covariance matrices i and k are all unity. Because of
2

kek2 = eT 1 e = (T /2 e)T (T /2 e) = T /2 e
2

1/2

(8)

i=1
K
X

Finally, after collecting the Jacobian matrices into a matrix

A, and the vectors ai and ck into a right-hand side (RHS) vector b, we obtain the following standard least-squares problem,
= argmin kA bk2

kFii1 xi1 + Gii xi ai k2i

+

Fig. 3. Measurement matrix A associated with the problem in Figure 1,

along with the information matrix I = AT A, and its Cholesky triangle R.

Here the state is 3D and both landmarks and measurements are 2D, hence the
size of A is (20 3 + 156 2) (20 3 + 20 2) = 372 100.

with
the matrix square root of , we can always
eliminate i from (7) by pre-multiplying Fii1 , Gii , and ai in
T /2
each term with i
. A similar story holds for the matrices
k , where for scalar measurements this simply means dividing
each term by the measurement standard deviation. Below we
assume that this has been done and drop the Mahalanobis norm
in favor of the regular 2-norm.

which is our starting point below. A can grow to be very large,

but is quite sparse, as shown in Figure 3. If dx , dl , and dz
are the dimensions of the state, landmarks, and measurements,
As size is (N dx +Kdz )(N dx +M dl ). In addition, A has a
typical block structure, e.g., with M = 3, N = 2, and K = 4:

G11
F21 G22

2
3

F
G
3
3

1
1

H
J
A=
1
1

H21
J
2

H32
J31
H43
J42
Above the top half describes the robot motion, and the
bottom half the measurements. A mixture of landmarks and/or
measurements of different types (and dimensions) is easily
accommodated. Note that the non-zero blocks pattern of
the measurement part is also the adjacency matrix for the
measurement part of a belief net like the one in Figure 2.

IV. C HOLESKY AND QR FACTORIZATION

In this section we briefly review Cholesky and QR factorization and their application to the full rank linear leastsquares (LS) problem in (8). The exposition closely follows
[19], which can be consulted for a more in-depth treatment.
For a full-rank m n matrix A, with m n, the unique LS
solution to (8) can be found by solving the normal equations:
AT A = AT b

(9)

This is normally done by Cholesky factorization of the information matrix I, defined and factorized as follows:

I = AT A = RT R

(10)

The Cholesky triangle R is an upper-triangular n n matrix1

and is computed using Cholesky factorization, a variant of LU
factorization for symmetric positive definite matrices. It runs
in n3 /3 flops. After this, can be found by solving
first RT y = AT b and then R = y
by back-substitution. The entire algorithm, including computing half of the symmetric AT A, requires (m + n/3)n2 flops.
For the example of Figure 1, both I and its Cholesky
triangle R are shown alongside A in Figure 3. Note the very
typical block structure of I when the columns of A are ordered
in the canonical way, i.e., trajectory X first and then map L:
 T

IXL
AX AX
I=
T
IXL
ATL AL

Here IXL = ATX AL encodes the correlation between robot

states X and map L, and the diagonal blocks are band-limited.
An alternative to Cholesky factorization that is both more
accurate and numerically stable is to proceed via QRfactorization without computing the information matrix I.
Instead, we compute the QR-factorization of A itself along
with its corresponding RHS:




R
d
QT A =
QT b =
0
e
Here Q is an m m orthogonal matrix, and R is the
upper-triangular Cholesky triangle. The preferred method for
factorizing a dense matrix A is to compute R column by
column, proceeding from left to right, using Householder reflections to zero out all non-zero elements below the diagonal.
The orthogonal matrix Q is not usually formed: instead, the
transformed RHS QT b is computed by appending b as an extra
column to A. Because the Q factor is orthogonal, we have:

2
2
2
2
kA bk2 = QT A QT b 2 = kR dk2 + kek2
2

Clearly, kek2 will be the least-squares residual, and the LS

solution can be obtained by solving the square system
R = d

(11)

1 Some treatments, including [19], define the Cholesky triangle as the lowertriangular matrix G = RT , but the other convention is more convenient here.

via back-substitution. The cost of QR LS is dominated by the

cost of the Householder reflections, which is 2(m n/3)n2 .
Comparing QR with Cholesky factorization, we see that
both algorithms require O(mn2 ) operations when m  n, but
that QR-factorization is a factor of 2 slower. However, this is
only for dense matrices: if A is sparse, as is the case in the
SLAM problem, QR factorization becomes quite competitive.
V. A G RAPHICAL M ODEL P ERSPECTIVE
Cholesky or QR factorization are most often used as black
box algorithms, but in fact they are surprisingly similar
to much more recently developed methods for inference in
graphical models [15]. Taking a graphical model view on
SLAM exposes its sparse structure in full, and shows how
sparse factorization methods in this context operate on a graph.
When examining the correlation structure of the problem
it is better to eliminate Z and consider the undirected graph
that encodes the correlations between the unknowns only. In
[7], [8] this view is taken to expose the correlation structure
inherent in the filtering version of SLAM. It is shown there that
inevitably, when marginalizing out the past trajectory X1:m1 ,
the information matrix becomes completely dense. Hence, the
emphasis in these approaches is to selectively remove links to
reduce the computational cost of the filter, with great success.
In contrast, in this paper we consider the graph associated

with the smoothing information matrix I = AT A, which does

not become dense, as past states are never marginalized out. In
particular, the objective function in Equation 6 corresponds to
a pairwise Markov random field (MRF) [20], [21] through the
Hammersley-Clifford theorem [20]. The nodes in the MRF
correspond to the robot states and the landmarks, and links
represent either odometry or landmark measurements. The
resulting bipartite graph corresponding to the example of
Figure 1 is shown in Figure 4.
Both QR and Cholesky factorization eliminate one variable
at a time, starting with 1 , corresponding in the leftmost
column of either A or I. The result of the elimination is
that 1 is now expressed as a linear combination of all
other unknowns j>1 , with the coefficients residing in the
corresponding row R1 of R. In the process, however, new
dependencies are introduced between all variables connected
to 1 , which causes edges to be added to the graph. The next
variable is then treated in a similar way, until all variables have
been eliminated. This is exactly the process of moralization
and triangulation familiar from graphical model inference. The
result of eliminating all variables is a chordal graph, shown
for our example in Figure 4b.
The single most important factor to good performance is
the order in which variables are eliminated. Different variable
orderings can yield dramatically more or less fill-in, defined
as the amount of edges added into the graph. As each edge
added corresponds to a non-zero in the Cholesky triangle R,
both the cost of computing R and back-substitution is heavily
dependent on how much fill-in occurs. Unfortunately, finding
an optimal ordering is NP-complete. Discovering algorithms
that approximate the optimal ordering is an active research

(a)

(b)

(c)

Fig. 4. a) The graph of the Markov random field of the associated SLAM problem from Figure 1. b) The triangulated graph: each edge corresponds to a nonzero in the Cholesky triangle R. c) The corresponding elimination tree showing how the state and landmarks estimates will be computed via back-substitution:
the root is computed first - in this case a landmark near the top left - after which the computation progresses further down the tree.

area in linear algebra. A popular method for medium-sized

problems is colamd [22]. However, as we will show in Section
VII, using domain knowledge can do even better.
A data structure that underlies many of these approximate
column ordering algorithms is the elimination tree. It is defined
as a depth-first spanning tree of the chordal graph after
elimination, and is useful in illustrating the flow of computation during the back-substitution phase. The elimination tree
corresponding to our example, for a good column ordering,
is shown in Figure 4c. The root of the tree corresponds to
the last variable n to be eliminated, which is the first to
be computed in back-substitution (Equation 11). Computation
then proceeds down the tree, and while this is typically done
in reverse column order, variables in disjoint subtrees may
be computed in any order. In fact, if one is only interested
in certain variables, there is no need to compute any of the
subtrees that do not contain them. The elimination tree is also
the basis for multifrontal QR methods [23], which we have
also evaluated in our simulations below.
VI. S QUARE ROOT SAM
A batch-version of square root information smoothing and
mapping is straightforward and a completely standard way of
solving a large, sparse least-squares problem:
1) Build the measurement matrix A and the RHS b as
explained in Section III.
p
2) Find a good column ordering p, and reorder Ap A
2
3) Solve p = argmin kAp p bk2 using either the
Cholesky or QR factorization method from Section IV
r
4) Recover the optimal solution by p , with r = p1
In tests we have obtained the best performance with sparse
LDL factorization [24], a variant on Cholesky factorization
that computes I = LDLT , with D a diagonal matrix and L
a lower-triangular matrix with ones on the diagonal.

We have also experimented with finding better, SLAM

specific column re-orderings. A simple idea is to use a
standard method such as colamd, but have it work on the
sparsity pattern of the blocks instead of passing it the original
measurement matrix A. This amounts to working directly with
the bipartite MRF graph from Section V, making accidental
zeros due to the linearization invisible. Surprisingly, as we will
show, the symbolic factorization on this restricted graph yields
better column orderings.
Incremental

SAM

In a robotic mapping context, an incremental version of the

above algorithm is of interest. It is well known that factorizations can be updated incrementally. One possibility is to use
a rank 1 Cholesky update, a standard algorithm that computes
the factor R0 corresponding to a I 0 = I + aaT , where aT
is a new row of the measurement matrix A. However, these
algorithms are typically implemented for dense matrices only,
and it is imperative that we use a sparse storage scheme for optimal performance. While sparse Cholesky updates exist, they
are relatively complicated to implement. A second possibility,
easy to implement and suited for sparse matrices, is to use a
series of Givens rotations (see [19]) to eliminate the non-zeros
in the new measurement rows one by one.
A third possibility, which we have adopted for the simulations below, is to update the matrix I and simply use a
full Cholesky (or LDL) factorization. QR factorization is more
accurate and has better numerical properties, but a Cholesky
or LDL factorization can be corrected with one linear update
step to achieve the same accuracy, if required.
Because the entire measurement history is implicit in I,
one does not need to factorize at every time-step. At any time
during an experiment, however, the map and/or trajectory can
be computed by a simple factorization and back-substitution,
e.g., for visualization and/or path planning purposes.

Fig. 5. A synthetic environment with 500 landmarks along with a 1000-step

random walk trajectory, corresponding to 14000 measurements taken.

M
200

500

1000

N
180
500
1280
2000
180
500
1280
2000
180
500
1280
2000

none
0.031
0.034
0.036
0.037
0.055
0.062
0.068
0.07
0.104
0.109
0.124
0.126

ldl
0.062
0.062
0.068
0.07
0.176
0.177
0.175
0.181
0.401
0.738
0.522
0.437

chol
0.092
0.094
0.102
0.104
0.247
0.271
0.272
0.279
0.523
0.945
0.746
0.657

mfqr
0.868
1.19
1.502
1.543
2.785
3.559
5.143
5.548
10.297
12.112
14.151
15.914

qr
1.685
1.256
1.21
1.329
11.92
8.43
6.348
6.908
42.986
77.849
35.719
25.611

Fig. 6. Averaged simulation results over 10 tries, in seconds, for environments

with various number of landmarks N and simulations with trajectory lengths
M . The methods are discussed in more detail in the text. The none method
corresponds to doing no factorization and measures the overhead.

VII. I NITIAL F INDINGS AND S IMULATION R ESULTS

A. Batch SAM
We have experimented at length with different implementations of Cholesky, LDL, and QR factorization to establish
which performed best. All simulations were done in MATLAB
on a 2GHz. Pentium 4 workstation running Linux. Experiments were run in synthetic environments like the one shown
in Figure 5, with 180 to 2000 landmarks, for trajectories of
length 200, 500, and 1000. Each experiment was run 10 times
for 5 different methods:
none: no factorization performed
ldl: Davis sparse LDL factorization [24]
chol: MATLAB built-in Cholesky factorization
mfqr: multifrontal QR factorization [23]
qr: MATLAB built in QR factorization
The results are summarized in Figure 6. We have found that,
under those circumstances,
1) The freely available sparse LDL implementation by T.
Davis [24] beats MATLABs built-in Cholesky factorization by about 30%.

2) In MATLAB, the built-in Cholesky beats QR factorization by a large factor.

3) Multifrontal QR factorization is better than MATLABs
QR, but still slower than either Cholesky or LDL.
4) While this is not apparent from the table, using a good
column ordering is much more important than the choice
of factorization algorithm.
The latter opens up a considerable opportunity for original
research in the domain of SLAM, as we found that injecting
even a small amount of domain knowledge into that process yields immediate benefits. To illustrate this, we show
simulation results for a length 1000 random walk in a 500landmark environment, corresponding to Figure 5. Both I and
R are shown in Figure 7 for the canonical (and detrimental)
ordering with states and landmarks ordered consecutively. The
dramatic reduction in fill-in that occurs when using a good reordering is illustrated by Figure 8, where we used colamd [22].
Finally, when we use the block-oriented ordering heuristic
from Section VI, the fill-in drops by another factor of 2.

B. Incremental SAM
We also
compared the performance of an incremental version of SAM, described in Section VI, with a standard EKF
implementation by simulating 500 time steps in a synthetic
environment with 2000 landmarks. The results are shown in
Figure 10. The factorization of I was done using sparse LDL
[24], while for the column ordering we used symamd [22], a
version of colamd for symmetric positive definite matrices.
Smoothing every time step becomes cheaper than the EKF
when the number of landmarks N reaches 600. At the end,
with N = 1, 100, each factorization took about 0.6 s., and the
slope is nearly linear over time. In contrast, the computational
requirements of the EKF increase quadratically with N , and
by the end each update of the EKF took over a second.
As implementation independent measures, we have also
plotted N 2 , as well as the number of non-zeros nnz in the
Cholesky triangle R. The behavior of the latter is exactly
opposite to that of the EKF: when new, unexplored terrain
is encountered, there is almost no correlation between new
features and the past trajectory and/or map, and nnz stays
almost constant. In contrast, the EKFs computation is not
affected when re-entering previously visited areas -closing the
loop- whereas that is exactly when R fill-in occurs.
We have reason to belief that these results can be further improved: profiling of our implementation showed that roughly
2/3 of the time is spent updating I, which is an artifact of
our sparse matrix representation. A compressed row scheme
would probably increase the speed by another factor of 2. In
addition, incremental QR updating methods should be much
faster than doing full factorizations every time.
VIII. C ONCLUSION
In conclusion, we believe square root information smoothing to be of great practical interest to the SLAM community. It
recovers the entire trajectory and is exact, and even the decidedly sub-optimal incremental scheme we evaluated behaves

Fig. 7.

Original information matrix I and its Cholesky triangle. Note the dense fill-in on the right, linking the entire trajectory to all landmarks.

Fig. 8. Information matrix I after reordering and its Cholesky triangle. Reordering of columns (unknowns) does not affect the sparseness of I, but the
number of non-zeroes in R has dropped from approximately 2.8 million to about 250 thousand.

Fig. 9. By doing the reordering while taking into account the special block-structure of the SLAM problem, the non-zero count can be eliminated even
further, to about 130K, a reduction by a factor 20 with respect to the original R, and substantially less than the 500K entries in the filtering covariance matrix.

1200

12

x 10

1
0.9

1000

10

800

600

400

0.8
0.7
0.6
0.5
0.4
0.3

200

0.2
0.1

50

100

150

200

250

300

350

400

450

(a) number of landmarks N seen

500

50

100

150

200

250

300

350

(b) N 2 and nnz in R

400

450

500

50

100

150

200

250

300

350

400

450

500

(c) average time per call, in seconds

Fig. 10. Timing results for incremental SAM in a simulated environment with 2000 landmarks, similar to the one in Figure 5, but 10 blocks on the side. As
the number of landmarks seen increases, the EKF becomes quadratically slower. Note that the number of non-zeros nnz increases faster when large loops
are encountered around i = 200 and i = 350.

much better than the EKF as the size of the environments

grows. In addition, we conjecture that thepossibility of relinearizing the entire trajectory will make SAM cope better
with noisy measurements governed by non-linear measurement
equations. In contrast, non-optimal linearization cannot be
recovered from in an EKF, which inevitably has to summarize
it in a quadratic (Gaussian) approximation.
In this paper we only reported on our initial experiences
with this approach. In follow-up work we hope to say something about tight or amortized complexity bounds that predict
the algorithms performance on problems of a given size, and
we also would like to to establish that the algorithm performs
as well in practice as
it does in simulation. In addition, we
have only compared SAM against the EKF-based stochastic
map algorithm. Comparison of our approach to more recent
and faster SLAM variants, both approximate [6][8] and exact
[9], is the object of future work. Finally, we concentrated on
the large-scale optimization problem associated with SLAM.
Many other issues are crucial in the practice of robot mapping,
e.g. the data-association problem. In addition, the techniques
exposed here are meant to complement, not replace methods
that make judicious approximations in order to reduce the
asymptotic complexity of SLAM.
R EFERENCES
[1] R. Smith and P. Cheeseman, On the representation and estimation of
spatial uncertainty, Intl. J. of Robotics Research, 1987.
[2] J. Leonard, I. Cox, and H. Durrant-Whyte, Dynamic map building for
an autonomous mobile robot, Intl. J. of Robotics Research, vol. 11,
no. 4, pp. 286289, 1992.
[3] S. Thrun, Robotic mapping: a survey, in Exploring artificial intelligence in the new millennium. Morgan Kaufmann, Inc., 2003, pp. 135.
[4] R. Smith, M. Self, and P. Cheeseman, Estimating uncertain spatial
relationships in Robotics, in Autonomous Robot Vehicles, I. Cox and
G. Wilfong, Eds. Springer-Verlag, 1990, pp. 167193.
[5] J. Castellanos, J. Montiel, J. Neira, and J. Tardos, The SPmap: A
probabilistic framework for simultaneous localization and map building,
IEEE Trans. Robot. Automat., vol. 15, no. 5, pp. 948953, 1999.
[6] J. J. Leonard and H. J. S. Feder, Decoupled stochastic mapping, IEEE
Journal of Oceanic Engineering, pp. 561571, October 2001.

[7] M. Paskin, Thin junction tree filters for simultaneous localization and
mapping, in Intl. Joint Conf. on Artificial Intelligence (IJCAI), 2003.
[8] S. Thrun, Y. Liu, D. Koller, A. Ng, Z. Ghahramani, and H. DurrantWhyte, Simultaneous localization and mapping with sparse extended
information filters, Intl. J. of Robotics Research, vol. 23, no. 7-8, pp.
693716, 2004.
[9] J. Guivant and E. Nebot, Optimization of the simultaneous localization
and map building algorithm for real time implementation, IEEE Trans.
Robot. Automat., vol. 17, no. 3, pp. 242257, June 2001.
[10] G. Bierman, Factorization methods for discrete sequential estimation,
ser. Mathematics in Science and Engineering. New York: Academic
Press, 1977, vol. 128.
[11] P. Maybeck, Stochastic Models, Estimation and Control. New York:
Academic Press, 1979, vol. 1.
[12] K. Murphy, Bayesian map learning in dynamic environments, in
Advances in Neural Information Processing Systems (NIPS), 1999.
[13] U. Frese, P. Larsson, and T. Duckett, A multilevel relaxation algorithm
for simultaneous localisation and mapping, IEEE Trans. Robototics,
vol. 21, no. 2, pp. 196207, April 2005.
[14] J. Folkesson, P.Jensfelt, and H. I. Christensen, Graphical slam using
vision and the measurement subspace, in IEEE/RSJ Intl. Conf. on
Intelligent Robots and Systems (IROS), 2005.
[15] R. G. Cowell, A. P. Dawid, S. L. Lauritzen, and D. J. Spiegelhalter,
Probabilistic Networks and Expert Systems, ser. Statistics for Engineering and Information Science. Springer-Verlag, 1999.
[16] J. Dennis and R. Schnabel, Numerical methods for unconstrained
optimization and nonlinear equations. Prentice-Hall, 1983.
[17] R. Smith, M. Self, and P. Cheeseman, A stochastic map for uncertain
spatial relationships, in Int. Symp on Robotics Research, 1987.
[18] J. Leonard and H. Durrant-Whyte, Simultaneous map building and
localization for an autonomous mobile robot, in IEEE Int. Workshop
on Intelligent Robots and Systems, 1991, pp. 14421447.
[19] G. Golub and C. V. Loan, Matrix Computations, 3rd ed. Baltimore:
Johns Hopkins University Press, 1996.
[20] G. Winkler, Image analysis, random fields and dynamic Monte Carlo
methods. Springer Verlag, 1995.
[21] J. Yedidia, W. Freeman, and Y.Weiss, Generalized belief propagation,
in Advances in Neural Information Processing Systems (NIPS), 2000,
pp. 689695.
[22] P. R. Amestoy, T. Davis, and I. S. Duff, An approximate minimum
degree ordering algorithm, SIAM Journal on Matrix Analysis and
Applications, vol. 17, no. 4, pp. 886905, 1996.
[23] P. Matstoms, Sparse QR factorization in MATLAB, ACM Trans. Math.
Softw., vol. 20, no. 1, pp. 136159, 1994.
[24] T. A. Davis, Algorithm 8xx: a concise sparse Cholesky factorization
package, Univ. of Florida, Tech. Rep. TR-04-001, January 2004,
submitted to ACM Trans. Math. Software.