See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/220806531

Adaptive Full Scan Model for Range Finders in Dynamic Environments

Conference Paper · January 2008

DOI: 10.1007/978-3-642-00196-3_51 · Source: DBLP

CITATIONS READS

0 32

4 authors, including:

Tinne De Laet
KU Leuven
123 PUBLICATIONS   731 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

ABLE - Achieving Benefits from Learning Analytics View project

readySTEMgo View project

All content following this page was uploaded by Tinne De Laet on 12 August 2014.

The user has requested enhancement of the downloaded file.

Adaptive Full Scan Model for Range Finders in
Dynamic Environments

Tinne De Laet, Joris De Schutter and Herman Bruyninckx

Department of Mechanical Engineering,

Katholieke Universiteit Leuven,
Celestijnenlaan 300B, B-3001 Leuven, Belgium.
Tinne.DeLaet@mech.kuleuven.be
http://people.mech.kuleuven.be/~ tdelaet

Summary. Sensor models directly influence the efficiency and robustness of the
estimation processes used in robot and object localization. Therefore this paper
focusses on a probabilistic range finder sensor model for dynamic environments.
The dynamic nature results from the presence of unmodeled and possibly moving
objects and people.
The goal of this paper is twofold. First of al we present experiments to validate
the Rigorously Bayesian Beam Model (RBBM), a new model we proposed in a
previous paper. Second, we propose a sample-based full scan model to improve the
state of the art models. In contrast to these Gaussian-based state of the art full scan
models, the proposed model is able to handle the multi-modality of the range finder
data, which is shown here to occur even in simple static environments.

1 Introduction
Intelligent robots require sensors to perceive their environment. To translate
sensor measurements into intelligent behavior, the measurements have to be
interpreted in the context of a physical sensor model. The design of a sensor
model is a trade-off between accuracy (hence, increased intelligence) on the
one hand, and simplicity and robustness on the other hand. While simplicity
and robustness are improved by only using physically interpretable parame-
ters, accuracy requires that: (i) the mathematical sensor model approximates
the physical sensor well, and (ii) conditional probabilities on a well-chosen sub-
set of the model parameters represent the stochastic nature of the involved
physics.
In a probabilistic approach, inaccuracies are embedded in the stochastic
nature of the model, particularly in the conditional probability density rep-
resenting the measurement process. This way, good results and intelligent
behavior are achieved using a limited modeling effort. Nevertheless, it is of
vital importance that all types of inaccuracies affecting the measurements are
2 Tinne De Laet, Joris De Schutter and Herman Bruyninckx

incorporated in the probabilistic sensor model. Inaccuracies arise from sensor

limitations, noise, and the fact that most complex environments can only be
represented and perceived in a limited way. The dynamic nature of the envi-
ronment in particular is an important source of inaccuracies. This dynamic
nature results from the presence of unmodeled and possibly moving objects
and people.
This paper focuses on probabilistic range finder sensor models for dynamic
environments. Range finders, which are widely used in mobile robotics, mea-
sure the distances z to objects in the environment along certain directions
θ relative to the sensor. The sensor model is used in a form suitable for
mobile robot localization, i.e.: P (Z = z | X = x, M = m)1 , where Z indicates
the measured range, X the position of the mobile robot (and of the sensor
mounted on it), and M the environment map. The models are however useful
in other applications of range sensors as well.
This paper is organized as follows. After discussing related work in the
next section, Section 3 discusses the RBBM as it was proposed in [1] and
presents new experimental results to validate this model. Section 4 presents
the sample-based full scan model which is able to handle the multi-modality
of the range finder data. Finally, Section 5 summarizes the results of the
paper.

2 Related Work

The range finder sensor models available from literature are traditionally di-
vided in three main groups: feature-based approaches, beam-based models and
correlation-based methods. Feature-based approaches extract a set of features
from the range scan and match them to features contained in an environ-
mental model. Beam-based models, also known as ray cast models, consider
each distance measurement along a beam as a separate range measurement.
These models represent the one-dimensional distribution of the distance mea-
surement by a parametric function, which depends on the expected range
measurement in the respective beam directions. In addition, these models are
closely linked to the geometry and the physics involved in the measurement
process. They often result in overly peaked likelihood functions due to the
underlying assumption of independent beams. The last group of range finder
sensor models, correlation-based methods, build local maps from consecutive
scans and correlate them with a global map. The simple and efficient likeli-
hood field models or end point model [2] are related to these correlation-based
methods. A nice summary of the advantages and drawbacks of the different
range finder sensor models is given in [3].

1
To simplify notation, the explicit mention of the random variable in the proba-
bilities is omitted whenever possible, and replaced by the common abbreviation
P (x) instead of writing P (X = x).
 Adaptive Full Scan Model for Range Finders 3

Range finder sensor models can also be classified according to whether

they use discrete geometric grids [4–8] or continuous geometric models [9–11].
Moravec [8] proposed non-Gaussian measurement densities over a discrete grid
of possible distances measured by sonar; the likelihood of the measurements
has to be computed for all possible positions of the mobile robot at a given
time. Even simplified models [7] in this approach turned out to be computa-
tionally too expensive for real-time application. Therefore, Fox et al. proposed
a beam model consisting of a mixture of two physical causes for a measure-
ment: a hit with an object in the map, or with an object not yet modeled in the
map [6]. The last cause accounts for the dynamic nature of the environment.
An analogous mixture adds two more physical causes: a sensor failure and
an unknown cause resulting in a ‘max-range’ measurement and a ‘random’
measurement, respectively [9, 10]. While [9, 11] use a continuous model, [10]
presents the discrete analog of the mixture, taking into account the limited
resolution of the range sensor. [11] extend the basic mixture model for use in
Monte Carlo localization. To overcome problems due to the combination of
the limited representational power and the peaked likelihood of the accurate
range finder, they propose an adaptive likelihood model. The likelihood model
is smooth during global localization and more peaked during tracking.
Recently, different papers tried to tackle the problems associated with
beam-based models, caused by the independence assumptions between beams.
[3] propose a sensor model for the full scan. The model treats the sensor mod-
eling task as a non-parametric Bayesian regression problem, and solves it
using Gaussian processes. It is claimed that the Gaussian beam processes
combine the advantages of the beam-based and the correlation-based models.
Due to the underlying assumption that the measurements are jointly Gaus-
sian distributed, the Gaussian beam processes are not suited to take into
account the non-Gaussian uncertainty due to the dynamic nature of the en-
vironment. An alternative approach to handle the overly-peaked likelihood
functions resulting from the traditional beam models is proposed in [12]. A
location-dependent full scan model takes into account the approximation er-
ror of the sample-based representation, and explicitly models the correlations
between individual beams introduced by the pose uncertainty. Despite the
modeled correlation between beams, the measurements are still assumed to
be jointly Gaussian distributed, which again limits the applicability in dy-
namic environments. Pfaff et al. [11] extend the basic mixture model for use
in Monte Carlo localization. To overcome problems due to the combination of
the limited representational power and the peaked likelihood of the accurate
range finder, they propose an adaptive likelihood model. The likelihood model
is smooth during global localization and more peaked during tracking.
Recently, different papers tried to tackle the problems associated with
beam-based models, caused by the independence assumptions between beams.
[3] propose a sensor model for the full scan. The model treats the sensor mod-
eling task as a non-parametric Bayesian regression problem, and solves it using
Gaussian processes. It is claimed that the Gaussian beam processes combine
4 Tinne De Laet, Joris De Schutter and Herman Bruyninckx

the advantages of the beam-based and the correlation-based models. Due to

the underlying assumption that the measurements are jointly Gaussian dis-
tributed, the Gaussian beam processes are not suited to take into account
the non-Gaussian uncertainty due to the dynamic nature of the environment.
An alternative approach to handle the overly-peaked likelihood functions re-
sulting from the traditional beam models is proposed by Pfaff et al. [12]. A
location-dependent full scan model takes into account the approximation er-
ror of the sample-based representation, and explicitly models the correlations
between individual beams introduced by the pose uncertainty. Despite the
modeled correlation between beams, the measurements are still assumed to
be jointly Gaussian distributed, which again limits the applicability in dy-
namic environments.

3 Rigorously Bayesian Beam Model

This section gives first a brief description of the RBBM and the differences
with respect to the state of the art models. Next, additional experiments are
presented that show that the RBBM has the same representational power as
the state of the art models.

3.1 Description

We recently proposed a rigorously Bayesian modeling of the probabilistic

range sensor beam model for dynamic environments, referred to as RBBM [1].
Similar to [9,11] the sensor model is derived for a continuous geometry. Unlike
previous models [6,9–11], the mixture components are founded on a Bayesian
modeling. This modeling makes use of probabilistic graphical models, in this
case Bayesian networks.
The obtained RBBM can be written as a mixture of four possible causes
of a range measurement: a hit with the map, a hit with an unmodeled object,
an unknown cause resulting in a random measurement and a sensor failure
resulting in a maximum reading measurement:

P (z | x, m) = π1 Phit (z | x, m) + π2 Poccl (z | x, m) +
π3 Prand (z | x, m) + π4 Pmax (z | x, m) , (1)

where,
 Adaptive Full Scan Model for Range Finders 5

π1 = (1 − p′ ) (1 − π3 − π4 ), (2)
′
π2 = p (1 − π3 − π4 ), (3)
Phit (z | x, m) = N (z; z ⋆, σm ) , (4)

 1⋆ 1−p′
z [1−( z ⋆ −z 2 if 0 ≤ z ≤ z ⋆
Poccl (z | x, m) = z⋆
p′ )] , (5)
0 otherwise.
(
1
if 0 ≤ z ≤ zmax ,
Prand (z | x, m) = zmax , and (6)
0 otherwise.
(
1 if z = zmax ,
Pmax (z | x, m) = I (zmax ) = , (7)
0 otherwise.

with z ⋆ is the distance the range sensor would ideally measure, i.e. the distance
to the closest object in the map.
The RBBM, Eq. (1-7), depends on four independent model parameters:

Θ = [σm , p′ , π3 , π4 ] , (8)

while zmax is a known sensor characteristic. This set of parameters has a clear
physical interpretation; σm is the standard deviation of the zero mean Gaus-
sian measurement noise governing Phit (z | x, m); p′ is the probability that the
map is occluded; π3 and π4 are the probabilities that the range finder re-
turns an unexplainable measurement (unknown cause) and a maximum read-
ing (sensor failure), respectively. We presented a maximum-likelihood (ML)
estimation and a variational Bayesian (VB) estimation algorithm both based
on the expectation-maximization approach to learn the RBBM model param-
eters [1].
With respect to the state of the art beam model of Thrun et al. [9], the
RBBM has: (i) a different functional form for the probability of range mea-
surements caused by unmodeled objects, (ii) an intuitive explanation for the
discontinuity encountered in the cited paper, and (iii) a reduction in the num-
ber of model parameters. The presented ML and VB estimation algorithms
directly benefit from the reduced number of model parameters.

3.2 Experiment

In the learning experiment, the experimental data reported by Thrun et al. [9]
is used. The data consists of two series of measurements obtained with a
mobile robot traveling through a typical office environment. From the set of
measurements, 10000 measurements that are centered around two different
expected ranges, are selected. The two obtained sets with different expected
ranges are shown in Fig. 1.
The goal of this section is threefold: (i) to learn the model parameters
(8) of the RBBM (1) from experimental data, (ii) to compare the results of
 6 Tinne De Laet, Joris De Schutter and Herman Bruyninckx
N N
1.2 0.8

0.7
1.0

0.6

0.8
0.5

0.6 0.4

0.3
0.4

0.2

0.2
0.1

0 0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
z[m] z[m]

(a) Short range (b) Long range

Fig. 1. Data for second learning experiment reported in [9]. These data consist
of two series of measurements obtained with a mobile robot traveling through a
typical office environment. From the set of measurements 10000 measurements that
are centered around two different expected ranges are selected.

the ML-EM and VB-EM estimator, and (iii) to compare the results of the
proposed estimators with the learning approach of Thrun’s model [9].
To see how well the learned model explains the experiment, the learned
continuous pdf P (z | x, m, Θ) of Eq. (1) has to be compared with the discrete
pdf of the experimental data (histogram) H (z). To this end, the learned pdf
is first discretized using the same bins {zf }f =1:F as the experimental pdf.
To quantize the difference between the learned and the experimental pdf two
‘distance’ measures are used: the discrete KL-divergence:
F
X H (zf )
d1 ≈ KL (H||P ) ≈ H (zl ) log , (9)
P (zf | x, m, Θ)
f =1

and the square root of the discrete Hellinger distance:

v
u F  2
uX 1 1
d2 ≈ DH (H||P ) ≈ t H (zf ) 2 − P (zf | x, m, Θ) 2 . (10)
f =1

The latter is known to be a valid symmetric distance metric [13].

The parameters of the learning algorithms are listed in Table 1. Fig. 2 and
Table 2 show the results of the ML-EM and VB-EM estimators for the RBBM
compared to the results of the ML estimator for Thrun’s model [9] for these
two sets. The results are obtained by running the learning algorithms for 30
iteration steps. The proposed ML-EM and VB-EM estimator outperform the
ML-EM estimator for Thrun’s model [9] for the studied data sets. Despite the
reduced number of parameters of the RBBM compared to Thrun’s model, the
RBBM has at least the same representational power.
 Adaptive Full Scan Model for Range Finders 7

ML-EM RBBM VB-EM RBBM ML-EM Thrun’s model

σm,init = 0.5 p′init = 13 σm,init = 0.5
p′init = 0.4 βinit = 5000 zhit,init = 0.4
π3,init = 0.2 Winit = 12 zshort,init = 0.3
π4,init = 0.1 µ̄init = xmp zmax,init = 0.1
νinit = 100 zrand,init = 0.2
α1,init = 58 λshort,init = 14
α2,init = 18
α3,init = 18
α4,init = 18
β0 = 5
W0 = 50
µ̄0 = xmp
ν0 = 100
α0 = 1
Table 1. EM-parameters for the learning experiment (all in SI-units). In the ML
approaches, the mean of Phit (z | x, m) is set to xmp , i.e. the most probable bin of
the histogram of the training set H (z)

(a) Discrete KL-divergence

Experiment d1 (Eq. (9))
ML-EM VB-EM ML-EM
RBBM RBBM Thrun’s model
short range 0.5295 0.5127 0.7079
long range 0.4366 0.4368 0.5852
average 0.4830 0.4747 0.6465
(b) square root Hellinger distance
Experiment d2 (Eq. (10))
ML-EM VB-EM ML-EM
RBBM RBBM Thrun’s model
short range 0.3166 0.2971 0.5629
long range 0.1683 0.2100 0.3481
average 0.2425 0.2535 0.4555

Table 2. Discrete KL-divergence and square root Hellinger distance for the second
learning experiment between the training set and the results of the ML-EM and
VB-EM estimators for the RBBM and the ML-EM estimator for Thrun’s model [9]
8 Tinne De Laet, Joris De Schutter and Herman Bruyninckx

P (z | x, m) P (z | x, m)
1.2 0.8
ML-EM RBBM
ML-EM RBBM
0.7 VB-EM RBBM
1.0 VB-EM RBBM ML-EM Thrun’s model
ML-EM Thrun’s model 0.6 Histogram Training set
Histogram Training set
0.8
0.5

0.6 0.4

0.3
0.4
0.2
0.2
0.1

0 0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 z[m] 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 z[m]
(a) Short range (b) Long range

Fig. 2. Comparison of the results of the ML-EM and VB-EM estimators for the
RBBM and the results of a maximum likelihood estimator for Thrun’s model [9] for
the data of Fig. 1.

4 Sample-based Full Scan Model

This section extends the RBBM (1)to an adaptive full scan model for dynamic
environments; adaptive, since it automatically adapts to the local density of
samples when using sample-based representations; full scan, since the model
takes into account the dependencies between individual beams.
In many applications using a range finder, the posterior is approximated
by a finite set of samples (histogram filter, particle filters). The peaked likeli-
hood function associated with a range finder (small σm due to its accuracy) is
problematic when using such finite set of samples. The likelihood P (z | x, m)
is evaluated at all samples, which are approximately distributed according to
the posterior estimate. Basic sensor models typically assume that the estimate
x and the map m are known exactly, that is, they assume that one of the sam-
ples corresponds to the true value. This assumption, however, is only valid in
the limit of infinitely many samples. Otherwise, the probability that a value
exactly corresponds to the true location is virtually zero. As a consequence,
these peaked likelihood functions do not adequately model the uncertainty
due to the finite, sample-based representation of the posterior [12]. Further-
more, the use of a basic range finder model typically results in even more
peaked likelihood models, especially when using a large number of beams per
measurement, due to multiplication of probabilities. In practice, the problem
of peaked likelihoods, is dealt with in various ways: sub-sampling the mea-
surement (fewer beams); introducing minimal likelihoods for beams; inflating
the measurement uncertainty; or other means of regularization of the resulting
likelihoods. These solutions are not satisfactory however, since the additional
uncertainty due to the sample-based representation is not known in advance.
The additional uncertainty strongly varies with the number of samples and
 Adaptive Full Scan Model for Range Finders 9

the uncertainty of the estimate [11]. Fox et al. [14] propose to dynamically
adapt the number of samples by means of KLD sampling. For very peaked
likelihoods however, this might result in a huge number of samples. A possible
solution is to ensure that a critical mass of samples is located at the impor-
tant parts of the state space by sampling from the observation model [3, 12].
Sampling from the observation model however, is often only possible in an ap-
proximate and inaccurate way. Pfaff et al. [11] introduced an adaptive beam
model for dynamic environments, which explicitly takes location uncertainty
due to the sample-based representation into account. They compute the ad-
ditional uncertainty due to the sample-based representation, using techniques
from density estimation. When evaluating the likelihood function at a sample,
they consider a certain region around the sample, depending on the sample
density at that location. Then, depending on the area covered by the sample,
the variance of the Gaussian, σm , governing the beam model in Eq. (1), is
calculated for each sample. As a result, the beam model automatically adapts
to the local density of samples. Such a location dependent model results in
a smooth likelihood function during global localization and a more peaked
function during position tracking without changing the number of samples.
By considering a region around samples, the individual beams become
statistically dependent [3, 12]. The degree of dependency depends on the ge-
ometry of the environment and on the size and location of the considered
region. Beam models, such as the RBBM, implicitly assume however that the
beams are independent, that is:
B
Y
P (z | θ, x, m) = P (zb | θb , x, m) , (11)
b=1

where z = {zb }b=1:B and θ = {θb }b=1:B are the vectors containing the mea-
sured ranges and the angles of the different beams respectively; zb is the
range measured at the beam with angle θb ; B is the total number of beams
and P (zb | θb , x, m) is for instance the RBBM (Eq. (1)). By neglecting the de-
pendency between beams, the resulting likelihoods P (z | θ, x, m) are overly
peaked. Models taking into account the dependencies between beams con-
sider the full range scan and are therefore called full scan models further on.
The previously proposed full scan models [3, 12] both assume that the beams
of a range scan are jointly Gaussian distributed. The off-diagonal elements
of the covariance matrix associated with the Gaussian distribution represent
the dependency. To learn the model parameters, both methods draw samples
from the region around a sample and perform ray-casting using these sam-
ples. Plagemann et al. [3] train a Gaussian process which models the full scan,
while Pfaff et al. [12] directly provide a maximum likelihood estimate for the
mean and covariance of the Gaussian.
Section 4.1 shows that the dependency between beams may introduce
multi-modality, even for simple static environments. The multi-variate Gaus-
sian models [3,12] cannot handle this multi-modality. Therefore, a new sample-
10 Tinne De Laet, Joris De Schutter and Herman Bruyninckx

based method for obtaining an adaptive full scan model from a beam model,
able to handle multi-modality, is proposed. Section 4.2 extends the adap-
tive full scan model for dynamic environments by taking into account non-
Gaussian model uncertainty.

4.1 Sample-based Adaptive Full Scan Model for Static

Environments

Plagemann et al. [3] and Pfaff et al. [12] estimate the full scan model,
P (z | x, m)2 , based on a local environment U (x) of the exact estimate x:
Z
P (z | x, m) ≈ P (x̃ | x) Phit (z | x̃, m) dx̃, (12)

with P (x̃ | x) the distribution representing the probability that x̃ is an el-

ement of the environment U (x), i.e: x̃ ∈ U (x). Since this section does not
consider the dynamics of the environment, only one component of the RBBM
in Eq. (1) is used: Phit (z | x, m). The marginalization over the environment
U (x) in Eq.(12) introduces dependencies between the measurements zb of the
measurement vector z.
The environment U (x), as explained above, depends on the sample density
around the sample x under consideration. Pfaff et al. [11] propose to use a
circular region with diameter dU(x) , which is a weighted sum of the Euclidean
distance and the angular difference. Like Plagemann et al. [3] and Pfaff et
al. [12], an approximation of the above likelihood can be estimated online for
each sample x by simulating L complete range scans at locations drawn from
U (x) using the given map m of the environment. Contrary to the multivariate
Gaussian approximation proposed in [3, 12], we propose a sample-based ap-
proximation, able to handle multi-modality. Sampling from the environment
U (x) immediately results in a sample-based approximation of P (x̃ | x):
L
1X
P (x̃ | x) ≈ δx̃(l) , (13)
L
l=1

where δx̃(l) denotes the delta-Dirac mass located in x̃(l) , and the samples are
distributed according to P (x̃ | x):
x̃(l) ∼ P (x̃ | x) . (14)
Using this sample-based approximation of P (x̃ | x) the likelihood of Eq. (12)
can be approximated as:
L
1X  
P (z | x, m) ≈ Phit z | x̃(l) , m . (15)
L
l=1
2
To simplify the notation θ and θ are omitted from P (z | θ, x, m) and
P (zb | θb , x, m), respectively.
 Adaptive Full Scan Model for Range Finders 11

Since this sample-based approximation has to be calculated online, the num-

ber of samples has to be limited. If the used environment U (x) is large, the
resulting approximation will be bad. To smooth the undesired bumpy behavior
due to the limited number of samples, the measurement noise σm governing
Phit (z | x, m) in Eq. (4), is artificially increased depending on the size of U (x)
by multiplying it with a factor:
q
1 + C dU(x) . (16)

In further simulations, C was set to 20.

Experiment

Fig. 3. Panorama taken from the Sick LMS 200 range finder mounted on the Kuka
361 industrial robot. The environment consists of a rectangular room with an object
(a Kuka KR 15/2 robot) in the middle. We show that even for this simple static
environment, the presented sample-based full scan model outperforms the Gaussian-
based state of the art models.

A simple environment consisting of a rectangular room with an object (a

Kuka KR 15/2 robot) in the middle (Fig. 3) is used to show that the marginal-
ization over (even small) U (x) to obtain the true likelihood not only introduces
dependencies between the beams but also multi-modality. The U (x) results
from a local uncertainty on the x- and y−position of 0.01m and a rotational
uncertainty of 5◦ . To obtain a reference, a Sick LMS 200 range finder is used
to take a large number of measurements (L = 1500) at random locations sam-
pled in U (x). To allow for exact positioning, the Sick LMS 200 is placed on a
Kuka 361 industrial robot. The Sick LMS 200 range finder is connected to a
laptop, communicating with the Kuka 361 industrial robot over the network
using Corba-facilities in the Open Robot Control Software, Orocos ( [15, 16]).
A simplified map of the environment (Fig. 3) is built to simulate the 150
complete range scans needed to construct a full scan model. The marginal
P (zb | x, m) of a selected beam (the middle one) is studied here in more detail.
The marginal likelihoods for the selected beam using the proposed sample-
based approximation of Eq. (15) and the Gaussian approximation proposed
12 Tinne De Laet, Joris De Schutter and Herman Bruyninckx

by [12], are compared in Fig. 4(a). The histogram of the measurements of

the selected beam in this figure clearly shows the multi-modality of the like-
lihood caused by the dependency between beams. Fig. 4(d) shows the differ-
ence for all beams between the experimentally obtained cumulative marginal
(L = 1500) and the Gaussian-based and sample-based approximation for all
beams. The mean difference with the experimental data for the sample-based
approximation is 2.8 times smaller than the difference for the Gaussian-based
approximation, even for the simple static environment of Fig. 3

4.2 Sample-based Adaptive Full Scan Model for Dynamic

Environments

The adaptive beam model proposed in [11] is suited for use in dynamic en-
vironments since it uses the four component mixture beam model [9, 10]. To
date however, the available adaptive full scan likelihood models of [3,12] have
not been adapted for dynamic environments. The assumption that the beams
are jointly Gaussian distributed, unable to capture the non-Gaussian uncer-
tainty due to environment dynamics, prevents the straightforward extension
for dynamic environments. In contrast, the sample-based approximation of
the full scan likelihood, as proposed in Section 4.1, can be extended to in-
clude environment dynamics. To this end, replace Phit (z | x, m) in Eq. (12)
and Eq. (15) by the full mixture of Eq. (1).

Experiment

Fig. 5(a) and Fig. 5(b) compare the marginals for the selected beams (Fig. 4(a))
obtained from the adaptive full scan model for dynamic environments using
the proposed sample-based approximation and the Gaussian approximation
proposed by [12]. Fig. 5(c) shows a probability map of the adaptive full scan
model suited for dynamic environments for the example environment of Fig. 3.
The probability map shows that the marginalization over the environment
U (x) of a sample in Eq. (12) not only introduces dependency between beams
but also introduces multi-modality.

5 Conclusion

The contribution of this paper is twofold. First of al we presented experi-

ments to validate the Rigorously Bayesian Beam Model (RBBM) [?]. The
experiments showed that despite the reduced number of parameters of the
RBBM compared to state of the art model of Thrun et al. [9], the RBBM has
at least the same representational power.
Second, we proposed a sample-based full scan model to improve the state
of the art models. In constrast to the multi-variate Gaussian state of the art
 Adaptive Full Scan Model for Range Finders 13

4.5

4
4 Experimental data (1500 samples)
Sample-based approximation
3.5 Gaussian approximation [12]
3.5

3
3
y[m]

2.5
2.5

P (z | x, m)
2
2

1.5
1.5

1
1
Environment
0.5 Some measurements
0.5
Samples of U (x)
Example beams
0 0
−2.5 −2 −1.5 −1 −0.5 0 0.5 1 1.5 2 2.5 0 1 2 3 4 5 6
x[m] z[m]

(a) Environment model (b) Marginal likelihood for left beam

−3
beam130 x 10
7
Experimental data (1500 samples) Sample-based approximation
2.5 Sample-based approximation Gaussian-based approximation [12]
Gaussian approximation [12]
6

2
5
∆Pc (z | x, m)
P (z | x, m)

1.5 4

3
1

0.5
1

0 0
0 1 2 3 4 5 6 0 50 100 150 200 250 300 350 400
z[m] beam

(c) Marginal likelihood for right beam (d) Difference between experimental cu-
mulative marginal and Gaussian and
sample-based approximations

Fig. 4. Figure (a) indicates the model of the simple evnrionment of Fig. 3, consisting
of a room with an object in the middle, with black lines. The range finder is located
at (0.15m, 0.75m). Samples from U (x) (resulting from a local uncertainty on the x-
and y−position of 0.01m and a rotational uncertainty of 5◦ ) are shown with black
dots, and some simulated measurements are shown in grey. Figure (b) and Figure(c)
show the marginal likelihood P (zb | x, m) for a the two selected beams together with
the histogram of the experimentally recorded range finder data, the Gaussian-based
approximation (L = 150) of [12] and the sample-based approximation (L = 150).
In contrast to the Gaussian-based state of the art full scan model, the proposed
sample-based approximation is able to handle the multi-modality of the range finder
data. Figure (d) shows the difference for all beams between the experimentally
obtained cumulative marginal (L = 1500) and the Gaussian-based and sample-based
approximation. The mean difference with the experimental data for the sample-
based approximation is 2.8 times smaller than the difference for the Gaussian-based
approximation, even for the simple static environment of Fig. 3 and the small U (x).
14 Tinne De Laet, Joris De Schutter and Herman Bruyninckx
beam130
0.9 0.8
Large sample approximation Large sample approximation
Gaussian approximation [12] Gaussian approximation [12]
0.8 Sample-based approximation 0.7
Sample-based approximation

0.7
0.6

0.6
0.5
P (z | x, m)

P (z | x, m)
0.5
0.4
0.4

0.3
0.3

0.2
0.2

0.1 0.1

0 0
0 1 2 3 4 5 6 0 1 2 3 4 5 6
z[m] z[m]

(a) Marginal likelihood for left beam (b) Marginal likelihood for right beam

(c) Probability map

Fig. 5. Figure (a) and Figure(b) show the marginal likelihood P (zb | x, m) for
a the two selected beams 4(a) together with the Gaussian-based approximation
(L = 150) of [12] and the sample-based approximation (L = 150) extended for the
use in dynamic environments. In contrast to the Gaussian-based state of the art full
scan model, the proposed sample-based approximation is able to handle the multi-
modality of the range finder data. Figure (c) shows the probability map for a large
number of beams.

models, the proposed sample-based method is able to handle multi-modality.

This multi-modality was shown to occur even in simple static environments.
The sample-based approximation was shown to better explain experimen-
tally recorded range finder data. The mean difference with the experimental
data for the sample-based approximation is 2.8 times smaller than the differ-
ence for the Gaussian-based approximation. The proposed adaptive full scan
model was extended for use in dynamic environments by taking into account
non-Gaussian model uncertainty. To this end, the Rigorously Bayesian Beam
Model (RBBM) was extended to a full scan model.
 Adaptive Full Scan Model for Range Finders 15

Acknowledgment
All authors gratefully acknowledge the financial support by K.U.Leuven’s
Concerted Research Action GOA/05/10 and the Research Council K.U.Leuven,
CoE EF/05/006 Optimization in Engineering (OPTEC). Tinne De Laet is a
Doctoral Fellow of the Fund for Scientific Research–Flanders (F.W.O.) in Bel-
gium.

References
1. T. De Laet, J. De Schutter, and H. Bruyninckx, “Rigorously Bayesian range
finder sensor model for dynamic environments,” in Proceedings of the 2008 IEEE
International Conference on Robotics and Automation, Pasadena, California,
U.S.A., 2008.
2. S. Thrun, “A probabilistic online mapping algorithm for teams of mobile
robots,” The International Journal of Robotics Research, vol. 20, no. 5, pp.
335–363, 2001.
3. C. Plagemann, K. Kersting, P. Pfaff, and W. Burgard, “Gaussian beam pro-
cesses: A nonparametric bayesian measurement model for range finders,” in
Robotics: Science and Systems (RSS), Atlanta, Georgia, USA, June 2007.
4. D. Hähnel, D. Schulz, and W. Burgard, “Mobile robot mapping in populated
environments and sensor planning,” Journal of the Advanced Robotics, vol. 17,
no. 7, pp. 579–597, 2003.
5. D. Hähnel, R. Triebel, W. Burgard, and S. Thrun, “Map building with mobile
robots in dynamic environments,” in Proceedings of the 2003 IEEE International
Conference on Robotics and Automation. Taipeh, Taiwan: ICRA2003, 2003,
pp. 1557–1569.
6. D. Fox, W. Burgard, and S. Thrun, “Markov localization for mobile robots in
dynamic environments,” Journal of Artificial Intelligence Research, vol. 11, pp.
391–427, 1999.
7. W. Burgard, D. Fox, D. Hennig, and T. Schmidt, “Estimating the absolute
position of a mobile robot using position probability grids,” in Proc. of the
National Conference on Artificial Intelligence, 1996.
8. H. P. Moravec, “Sensor fusion in certainty grids for mobile robots,” AI Magazine,
vol. 9, pp. 61–74, 1988.
9. S. Thrun, W. Burgard, and D. Fox, Probabilistic Robotics. MIT Press, 2005.
10. H. Choset, K. M. Lynch, S. Hutchinson, G. A. Kantor, W. Burgard, L. E.
Kavraki, and S. Thrun, Principles of Robot Motion: Theory, Algorithms, and
Implementations. MIT Press, June 2005.
11. P. Pfaff, W. Burgard, and D. Fox, “Robust monte-carlo localization using adap-
tive likelihood models,” in European Robotics Symposium, H. Christensen, Ed.,
vol. 22. Palermo, Italy: Springer-Verlag Berlin Heidelberg, Germany, March
2006, pp. 181–194.
12. P. Pfaff, C. Plagemann, and W. Burgard, “Improved likelihood models for
probabilistic localization based on range scans,” in Proceedings of the 2007
IEEE/RSJ International Conference on Intelligent Robots and Systems. San
Diego, California: IROS2007, 2007.
 16 Tinne De Laet, Joris De Schutter and Herman Bruyninckx

13. C. M. Bishop, Pattern Recognition and Machine Learning. Springer, 2006.

14. D. Fox, “Adapting the Sample Size in Particle Filters Through KLD-Sampling,”
The International Journal of Robotics Research, vol. 22, no. 12, pp. 985–1003,
December 2003.
15. H. Bruyninckx, “Open RObot COntrol Software,” http://www.orocos.org/,
2001.
16. P. Soetens, “A software framework for real-time and distributed robot and
machine control,” Ph.D. dissertation, Department of Mechanical Engineering,
Katholieke Universiteit Leuven, Belgium, May 2006.

View publication stats