Mapping and Localization with RFID Technology

Dirk Hähnel Wolfram Burgard Dieter Fox Ken Fishkin Matthai Philipose
University of Freiburg University of Washington Intel Research Seattle
Department of Computer Science Computer Science and Engineering Seattle, WA, USA
Freiburg, Germany Seattle, WA, USA

the toothpaste has very different meanings depending on

whether it happens in the storage room or in the bathroom.
In this paper, we investigate how RFID technology can be
enhanced by location information. We use a mobile robot
equipped with RFID antennas to determine the locations of
RFID tags attached to objects in an indoor environment.
Figure 2 (left) depicts the robot built to carry out this re-
search. The robot consists of an off-the-shelf Pioneer 2 robot
equipped with a laser range scanner and two RFID antennas.
Fig. 1. Typical RFID tags used to label objects. The size of the tag depicted The antennas are mounted on top of the robot and point
in the center is 11 × 5 cm. approximately 45 degrees to the left and to the right with
respect to the robot. To use these antennas for estimating
the locations of objects, we first learn a sensor model that
Abstract— In this paper we analyze whether recent Radio describes the likelihood of detecting an RFID tag given its
Frequency Identification (RFID) technology can be used to location relative to one of the antennas. Since the noise of these
improve the localization of mobile robots and persons in their
environment. In particular we study the problem of localizing sensors is highly non-Gaussian, we represent the measurement
RFID tags with a mobile platform that is equipped with a likelihood model by a piecewise constant approximation. Then
pair of RFID antennas. We present a probabilistic measurement we describe a technique to estimate the locations of RFID tags
model for RFID readers that allow us to accurately localize using a mobile robot equipped with RFID antennas to detect
RFID tags in the environment. We also demonstrate how such tags. This process uses a map previously learned from laser
maps can be used to localize a robot and persons in their
environment. Finally, we present experiments illustrating that range data. We then apply Monte Carlo localization [4], [7]
the computational requirements for global robot localization can to estimate the pose of the robot and even of persons in this
be seriously reduced by fusing RFID information with laser data. environment. Experimental results suggest that it is possible to
accurately localize moving objects based on this technology.
Further experiments demonstrate that RFID tags greatly reduce
I. I NTRODUCTION the time required for global localization of a mobile robot in
Recent advances in the field of radio frequency identification its environment. Additionally, this technology can be used to
techniques have reached a state that will allow us within the drastically reduce the number of samples required for global
next years to equip virtually every object in an environment localization.
with small, cheap Radio Frequency Identification (RFID) This paper is organized as follows. After discussing related
tags [6]. Such tags contain circuitry that gain power from work we will present the sensor model for RFID receivers in
radio waves emitted by readers in their vicinity. They use this Section III. Then we describe how this model can be used
power to reply their unique identifier to the reader. Figure 1 in combination with a laser-based FastSLAM [8] approach to
depicts three different RFID tags that were used to carry out effectively determine the locations of RFID tags. In Section V
the experiments described in this paper. The detection range we describe how the resulting beliefs about the locations
of these tags is approximately 6 m. of the tags can be utilized to determine the position of the
RFID tags open up a wide variety of applications. For robot and of persons in the environment. Finally, we present
example, an important problem in the health-care sector is experimental results illustrating the advantages of RFID tags
the recognition of daily activities a home patient is engaged for robot localization and person tracking.
in. The Guide project [13] uses small RFID readers worn by a
person to identify the objects the person touches. The sequence II. R ELATED W ORK
of touched objects is used by a Bayesian reasoning system to In the last of years RFID sensors [6] have started to enter the
estimate the activity of the person and to provide support if field of mobile robotics. Nowadays RFID readers can detect
needed. Location context can provide important information low-cost passive tags in the range of several meters. These
for the interpretation of RFID readings. For example, touching improvements in the detection range of passive tags makes this
Fig. 2. Pioneer 2 with Sick Laser Range Finder, RFID reader and two
antennas (left). Experimental setup used for learning the likelihood function
of tag detections (right).

technology more and more attractive for robotics applications

since the information provided by tags can be used to support
various tasks like navigation, localization, mapping, and even Fig. 3. Detection field for the left (upper/pink histogram) and right
service applications such as people tracking. (lower/yellow histogram) antenna
Most of the applications of RFID technology, however,
assume that the readers are stationary and only the tags that
are attached to objects or persons move. The main focus is Following aspects need to be considered when designing an
to trigger events if a tag is detected by a reader or entering observation model for RFID tags.
the field of range (for example, to keep track of the contents 1) There are plenty of false-negative readings, i.e., situa-
of storage places [2]). Recently Kantor and Singh used RFID tions in which the tag is not detected although it is in
tags for mapping. Their system relies on active beacons which the vicinity of the antenna
provide distance information based on the time required to 2) Additionally, we obtain false-positive readings. In such
receive the response of a tag. Additionally, the positions of a case the antenna detects a tag that is not in range spec-
the tags have to be known more or less accurately [14], [9]. ified by the manufacturer. This also includes detection
Also Tsukiyama [16] requires given positions of the RFID of the RFID tag with the wrong antenna.
tags. Their system assumes perfect measurements, therefore
There are several reasons for this. For example, the orientation
they are not using any technique to deal with the uncertainty
of the tag with respect to the RFID receiver influences the
of the sensor.
energy absorbed by its own antenna. Depending on this angle,
The problem considered here is closely related to the
the energy will vary and sometimes not be high enough to
simultaneous localization and mapping (SLAM) problem, in
power the chip inside the tag. In such a case the tag will simply
which a robot has to generate a map while simultaneously
not respond. Furthermore, the shape and size of detection
estimating its pose relative to this map. However, due to the
range largely depends on the environment. For example, metal
limited accuracy of the RFID sensors SLAM-techniques for
typically absorbs the energy sent by the RFID reader and
range-only [14], [9], bearing-only [3] or range and bearing [5],
therefore tags attached to metallic objects will be detected
[11], [15] cannot be applied directly to the data provided by
only in a short range. But even other, non-metallic objects
the RFID system. Our algorithm instead uses a variant of Fast-
greatly influence the detectability of tags. For example, a tag
SLAM [12] to learn the geometric structure of the environment
attached to a concrete wall will result in a different detection
using laser data [8] and then estimates the positions of the tags
statistics. Furthermore, the radio frequency waves emitted by
based on the trajectory computed by the FastSLAM algorithm.
the antenna can be reflected by objects such that the antenna
III. L EARNING A P ROBABILISTIC S ENSOR M ODEL FOR even detects objects outside the specified detection range. Note
THE RFID A NTENNA that the observation model for the RFID antennas should be
To localize an RFID tag we need to know the posterior able to cover this wide range of situations and should not make
p(x | z1:t ) where x is the pose of the tag and z1:t are the the robot overly confident in the location of a particular tag
data gathered in the time steps 1, . . . , t. According to Bayes or even in its own location during localization.
rule and under the assumption of independence of consecutive To determine the observation model for the RFID antennas
measurements given we know the relative pose x of a tag we we generated a statistics by counting frequencies. We pro-
obtain the following recursive update rule. ceeded in the following way. We attached an RFID tag to a
box and rotated the robot in front of the box. We repeated
p(x | z1:t ) = αp(zt | x)p(x | z1:t−1 ) (1)
this for different distances and counted for every point in a
According to this equation, the key term is the quantity p(zt | discrete grid the frequency of detections of the antenna given
x) which specifies the likelihood of the observation zt given the tag was placed at a pose covered by this grid cell relative
the pose x of the tag relative to the robot. to the robot.
 To each of the randomly chosen potential positions we
assign a numerical value storing the posterior probability
0.5
p(x | z1:t ) that this position corresponds to the true pose of
0.7 the tag. Whenever the robot detects a tag, the posterior is
0.9 updated according to Equation (1) and using the sensor model
described in the previous section.
V. L OCALIZATION WITH RFID TAGS
Given the posterior distribution p(x | z1:t ) about the poten-
tial positions of an RFID tag we are now ready to compute
the likelihood of an observation y during localization, given
the robot or a person is placed at a location l. According to
Fig. 4. Simplified sensor model for the left antenna
the law of total probability and after transforming the global
coordinate system into the local reference system of the robot
we obtain
The resulting histogram is shown in Figure 3. This figure
X
p(y | l) = p(y | r(x, l))p(x | z1:t ). (2)
contains the detection statistics for both sensors. The his- x
tograms were built from 12,822 measurements. As can be seen
from the figure, both antennas show quite different behaviors In this equation the term r(x, l) represents the position of the
although they were measuring the same RFID tag. tag relative to the robot given the pose l of the robot and the
The resulting sensor model used to conservatively approx- location x of the tag sample. Thus, to determine the likelihood
imate the histograms depicted in Figure 3 is depicted in of a tag detection given the robot is at location l, we have to
Figure 4. This model consists of three components. The major integrate over the posterior probability of the tag’s location
detection range for each antenna consists of an arc with an given the data obtained during the mapping process. Note that
opening angle of 95 degrees in the direction of the antenna. the quantity p(y | r(x, l)) is the sensor model described in
Additionally, an antenna always detects RFID tags that are Section III. It specifies the likelihood of measuring y given
close to it even if they are behind the antenna. This is modeled the detected RFID tag is at the position r(x, l) relative to the
by a circular region around the center of the receiver. The robot.
corresponding likelihood for the two detection ranges are also To estimate the pose l of the robot or of persons in the
depicted in Figure 4. For locations outside these areas we environment, we apply the well-known recursive Bayesian
assume a constant likelihood of 0.5. filtering scheme:
p(lt | y1:t , u0:t−1 ) = α · p(yt | lt )
IV. M APPING RFID TAGS Z
0 0 0
The first application of the sensor model described in the · p(lt | ut−1 , lt−1 ) · p(lt−1 | y1:t−1 , u0:t−2 ) d lt−1 (3)
lt0
previous section is estimating the location of RFID tags in
the environment using a mobile robot. To learn the positions Here α is a normalization constant ensuring that p(lt |
of the tags our system proceeds in two steps. First it learns y1:t , u0:t−1 ) sums up to one over all lt . The term p(lt |
0
the geometric structure of the environment using a laser range ut−1 , lt−1 ) describes the probability that the object is at
sensor. Afterwards we estimate the positions of the tags based position lt given it executed the movement ut−1 at position
0
on the path of the robot. lt−1 . This quantity is computed depending on the object
Since our robot is equipped with a laser range scanner, we we are tracking. In the case of the robot we compute this
apply the FastSLAM algorithm [8] to learn the geometrical quantity based on the odometry measurements [7]. If we
structure of the environment. The resulting map used for the are tracking persons, we simply represent this density by a
experimental results is depicted in Figure 5. Given this map Gaussian centered around lt . Furthermore, the quantity p(yt |
and the maximum likelihood path of the robot computed by lt ) denotes the likelihood of the observation yt according to
the FastSLAM algorithm we can now estimate the locations of our observation model, which is computed using Equation (3).
the RFID tags. Here we apply the recursive Bayesian filtering To represent the posterior about the pose of the object being
scheme given in Equation 1. tracked we apply Monte-Carlo localization [4], [7]. In Monte-
To represent the belief about the pose of an RFID tag we use Carlo localization, the belief of the robot is represented by a set
a set of 1000 randomly chosen positions uniformly distributed of random samples [1]. Each sample consists of a state vector
in a 25 square meter wide area around the current pose of the of the underlying system, which is the pose l of the robot in
robot. This area is independent of the antenna that detected our case, and a weighing factor ω. The latter is used to store
the tag in order to avoid that a detection failure of an antenna the importance of the corresponding particle. The posterior
results in a suboptimal placement of the sampled positions. is represented by the distribution of the samples and their
It is initialized at the first detection of the RFID tag by the importance factors. The particle filter algorithm used by our
robot. system is also known as sequential importance sampling [1].
 Fig. 8. Places where the robot has detected the RFID tag with the left
(unfilled circle) or right antenna (filled circle)

out in the Intel Research Lab, Seattle, WA. Figure 5 shows

Fig. 5. Map of the Intel Research Lab Seattle generated by our FastSLAM a two-dimensional occupancy grid map generated with our
routine. FastSLAM routine. The size of the environment is 28m by
28m. We installed 100 tags in this environment (see Figure 6).
The tags were of the types depicted in Figure 1 and all of them
were able to communicate with the robot. Most of them were
installed along the circular corridor of the environment.
A. Mapping RFID tags
As already mentioned above, we use the trajectory estimated
by our FastSLAM routine to determine the posteriors about
the locations of the tags. When a tag is detected for the first
time, we initialize a discrete set of randomly chosen points
around the robot and use a uniform distribution to initialize
the belief. Whenever a tag is detected, the posterior probability
of each sample in that set is multiplied with the likelihood of
Fig. 6. RFID tags attached to walls. the observation given the tag is at the position corresponding
to that sample. Afterwards we normalize the belief over all
samples.
It updates the belief about the pose of the robot according to Figure 7 shows a typical example for the evolution of the
the following two alternating steps: belief of an RFID tag. The leftmost image shows the initial
1) In the prediction step, we draw for each sample a sample set after the first detection of an RFID tag. The remain-
new sample according to the weight of the sample and ing images illustrate how the belief focuses on the true position
0
according to the model p(lt | ut−1 , lt−1 ) of the robot’s of the tag as more measurements are obtained. They show the
dynamics given the movement ut−1 executed since the corresponding beliefs after 6, 17, 65, and 200 measurements.
previous update. Note that the diameter of each circle representing a particle
2) In the correction step, the new observation yt is inte- corresponds to it the posterior probability of that pose. As
grated into the sample set. This is done by bootstrap can be seen from the figure, the belief quickly converges
resampling, where each sample is weighted according to a unimodal distribution. Note that this is not necessarily
to the observation likelihood p(yt | lt ). the case. In principal, our representation can also handle
To globally localize the object, we initialize the particle set ambiguities in which the location of an RFID tag cannot be
with a uniform distribution. In the case of RFID sensors, we determined uniquely, for example, because the robot cannot
fortunately can efficiently sample potential locations of the reach locations which are required to resolve the ambiguity.
object. We simply place samples only in the potential detection Figure 8 depicts the positions of the robot when it detected
range of the RFID sensor. Such an approach has been applied the tag, for which the beliefs are plotted in Figure 7. Detections
successfully in the past, for example by Lenser et al. [10]. of the right antenna are displayed by filled circles and for each
detection of the left antenna we draw an unfilled circle. As
VI. E XPERIMENTAL R ESULTS can be seen from the figure, the measurement noise is quite
Our approach described above has been implemented and high and there are several false detections. Nevertheless, our
tested using a Pioneer 2 robot equipped with a SICK LMS algorithm is able to accurately localize the tag at the wall close
laser range-finder and an Alien Technology’s 915 MHz RFID to the entrance.
reader with two circularly polarized antennas (see left image After traveling 791.93m with an average speed of 0.225m/s
of Figure 2). The experiments described here were carried the robot had processed 50,933 detections of an RFID tag.
Fig. 7. Evolution of the posterior about the localization of an RFID tag over time. The width of the circles represent the posterior probability of the
corresponding positions. It is drawn proportional to ratio between the corresponding sample and the maximum likelihood sample.

14 14
12 12
10 10
8 8
6 6
4 4
2 2
0 0
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40

Fig. 9. Error (in m) during global localization with (green or light grey) and Fig. 11. Positioning error of the laser based global localization (in m) without
without (red or dark grey) odometry using RFID tags only. (red or dark grey) and with (green or light grey) RFID data.

The resulting map of the tags (at their maximum likelihood localization is depicted in the right image of the same figure.
position) is shown in Figure 10 (left). Thus, our sensor model As can be seen, even with such noisy sensors the estimated
allows to learn the positions multiple tags in a standard office trajectory is quite close to the ground truth.
environment.
C. Improving Global Localization with RFID Tags
B. Localization with RFID Tags The final experiment is designed to illustrate that the RFID
The next set of experiments is designed to illustrate that technology can be used to drastically improve the global lo-
the RFID map generated in the previous step can be used calization performance even in the case where highly accurate
to localize the robot and even persons equipped with RFID sensors such as laser range finders are used. To analyze this we
antennas. used a pre-recorded data set to figure out how efficiently the
In the first experiment we steered the robot through the robot can determine its global position in this map. Since the
environment and applied Monte-Carlo localization to globally RFID tags are only placed close to the corridor we generated
estimate the position of the vehicle. To simulate the situation samples only in the corridor of the environment. We compared
in which we localize a person instead of the robot we simple the time required for global localization using laser data
ignored the odometry information and changed the motion with the time needed when laser and RFID tags were used
model in the Monte Carlo localization procedure. As already simultaneously. Figure 11 shows the average localization error
mentioned above we used a standard motion model [7] to for a typical run for both cases. As the figure illustrates, global
estimate the pose of the robot. In order to localize and keep localization can be achieved much faster when laser and RFID
track of a person we simply replaced this motion model by a data are combined (green/light grey) compared to a situation
Gaussian distribution centered around the current pose. Note in which only laser data is used (red/dark grey).
that this is only a rough approximation of the motions of a Additionally, the use of RFID sensors can greatly reduce the
person. Better models therefore can be expected to result in number of samples required for global localization. Figure 12
more accurate estimates. shows the localization error depending on the number of
Figure 9 shows the localization error during a global lo- particles for the case in which only laser data is used as well as
calization run using RFID tags only. The two plots show for the situation in which the laser data is combined with RFID
the localization error for global localization without odometry information. It turns out that laser-based global localization is
(red/dark grey) and with odometry (green/light grey). efficient when at least 10.000 particles are used. On the other
The center image of Figure 10 shows the trajectory for hand, if we fuse the laser data with the information about the
the object being tracked when no odometry information is RFID tags, we can globally localize the object with as few as
used. The corresponding ground-truth obtained by laser-based 50 samples.
Fig. 10. Map of Intel Lab with most likely positions of the RFID tags (left), estimated trajectory (without odometry) (center) and the corresponding ground
truth (right).

25
10000 samples
25
1000 samples
[2] J. Brusey, M. Harrison, Ch. Floerkemeier, and M. Fletcher. Reasoning
20 7500 samples 20 250 samples about uncertainty in location identification with RFID. In IJCAI-2003
5000 samples 100 samples Workshop on Reasoning with Uncertainty in Robotics, 2003.
50 samples
15 15 [3] M. Deans and M. Herbert. Experimental comparison of techniques
10 10 for localization and mapping using bearing-only sensor. In Seventh
Int. Symp. on Experimental Robotics, 2000.
5 5
[4] F. Dellaert, D. Fox, W. Burgard, and S. Thrun. Monte Carlo localization
0 0 for mobile robots. In Proceedings of the IEEE International Conference
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40
on Robotics and Automation (ICRA), 1999.
[5] G. Dissanayake, H. Durrant-Whyte, and T. Bailey. A computationally
Fig. 12. Localization error (in m) during global localization for different efficient solution to the simultaneous localisation and map building
numbers of particles and depending on whether only laser data is used (left (SLAM) problem. In ICRA’2000 Workshop on Mobile Robot Navigation
image) or whether the combination of laser data and RFID measurements is and Mapping, 2000.
used (right image). [6] Klaus Finkenzeller. RFID Handboook: Radio-Frequency Identification
Fundamentals and Applications. Wiley, New York, 2000.
[7] D. Fox, W. Burgard, F. Dellaert, and S. Thrun. Monte Carlo localization:
Efficient position estimation for mobile robots. In Proc. of the National
VII. C ONCLUSIONS Conference on Artificial Intelligence (AAA I), 1999.
[8] D. Hähnel, W. Burgard, D. Fox, and S. Thrun. An efficient fastslam
In this paper we presented an approach to generate maps of algorithm for generating maps of large-scale cyclic environments from
raw laser range measurements. In Proc. of the IEEE/RSJ International
RFID tags with mobile robots. We presented a sensor model Conference on Intelligent Robots and Systems (IROS), 2003.
that allows us to compute the likelihood of tag detections [9] George A Kantor and Sanjiv Singh. Preliminary results in range-only
given the relative pose of the tag with respect to the robot. localization and mapping. In Proceedings of the IEEE Conference on
Robotics and Automation (ICRA), 2002.
Additionally we described how to compute a posterior about [10] S. Lenser and M. Veloso. Sensor resetting localization for poorly
the position of a tag after the trajectory and the map has been modelled mobile robots. In Proc. of the IEEE International Conference
generated with a highly accurate FastSLAM algorithm for on Robotics & Automation (ICRA), 2000.
[11] J.J. Leonard and H.J.S. Feder. A computationally efficient method for
laser range scans. We furthermore present how the posteriors large-scale concurrent mapping and localization. In Proc. of the Ninth
can be used to localize a robot and persons in the environment. Int. Symp. on Robotics Research (ISRR), 1999.
The system has been implemented on a Pioneer 2 robot [12] M. Montemerlo, S. Thrun, D. Koller, and B. Wegbreit. FastSLAM:
A factored solution to the simultaneous localization and mapping
that was augmented by two RFID antennas. In practical problem. In Proceedings of the AAAI National Conference on Artificial
experiments we demonstrated that the system can build ac- Intelligence, Edmonton, Canada, 2002. AAAI.
curate maps of RFID tags. We furthermore illustrated that [13] M. Philipose, K. Fishkin, D. Fox, H. Kautz, D. Patterson, and
M. Perkowitz. Guide: Towards understanding daily life via auto-
the resulting maps can be used for accurate localization of identification and statistical analysis. In Proc. of the Int. Workshop
the robot and moving objects without odometry information. on Ubiquitous Computing for Pervasive Healthcare Applications (Ubi-
Finally we presented an experiment demonstrating that the health), 2003.
[14] Sanjiv Singh, George Kantor, and Dennis Strelow. Recent results in
combination of a laser-range scanner and RFID technology extensions to simultaneous localization and mapping. In International
can greatly reduce the computational demands for the global Symposium on Experimental Robotics, 2002.
localization of a moving mobile robot. [15] S. Thrun, D. Fox, and W. Burgard. A probabilistic approach to concur-
rent mapping and localization for mobile robots. Machine Learning and
Autonomous Robots (joint issue), 1998.
R EFERENCES [16] T. Tsukiyama. Navigation system for mobile robots using rfid tags.
In Proceedings of the International Conference on Advanced Robotics
[1] S. Arulampalam, S. Maskell, N. Gordon, and T. Clapp. A tutorial (ICAR), 2003.
on particle filters for on-line non-linear/non-gaussian bayesian tracking.
IEEE Transactions on Signal Processing, 50(2):174–188, 2002.