A Supervised Learning Approach to Robot

IEICE TRANS. INF. & SYST., VOL.E90–D, NO.11 NOVEMBER 2007

1762

PAPER

A Supervised Learning Approach to Robot Localization Using a Short-Range RFID Sensor Kanji TANAKA†a) , Member, Yoshihiko KIMURO†† , Kentaro YAMANO† , Mitsuru HIRAYAMA† , Eiji KONDO† , and Michihito MATSUMOTO†† , Nonmembers

SUMMARY This work is concerned with the problem of robot localization using standard RFID tags as landmarks and an RFID reader as a landmark sensor. A main advantage of such an RFID-based localization system is the availability of landmark ID measurement, which trivially solves the data association problem. While the main drawback of an RFID system is its low spatial accuracy. The result in this paper is an improvement of the localization accuracy for a standard short-range RFID sensor. One of the main contributions is a proposal of a machine learning approach in which multiple classifiers are trained to distinguish RFID-signal features of each location. Another contribution is a design tool for tag arrangement by which the tag configuration needs not be manually designed by the user, but can be automatically recommended by the system. The effectiveness of the proposed technique is evaluated experimentally with a real mobile robot and an RFID system. key words: robot localization, RFID, Support Vector Machine, landmark arrangement

1. Introduction This work is concerned with the problem of robot localization using a short-range RFID (Radio Frequency Identification) sensor. In a few years, RFID tags may be attached to various objects, e.g. furniture, pillars and walls, in office environments. If a mobile robot is equipped with an RFID tag reader to communicate with them, the RFID tags can be viewed as low-cost landmarks for self-localization. A main advantage of such an RFID system is its robustness. The availability of landmark ID measurements makes trivial the data association problem, i.e. the problem of associating N observed features to N landmarks. While the main drawback of an RFID system is its low spatial accuracy. Methods that overcome this accuracy problem would be useful for a robust and efficient self-localization system. The result in this paper is an improvement of the localization accuracy for a short-range RFID sensor by using a machine learning technique. Self-localization is a fundamental problem in mobile robotics and has been intensively studied in the last decades [1]–[44], on several types of sensors includManuscript received January 17, 2007. Manuscript revised June 1, 2007. † The authors are with the Faculty of Departments of Mechanical Science and Engineering and Intelligent Machinery and Systems, Graduate School of Engineering, Kyushu University, Fukuoka-shi, 819–0395 Japan. †† The authors are with the Faculty of Systems & Information Technologies / KYUSHU, Fukuoka-shi, 814–0001 Japan. a) E-mail: [email protected] DOI: 10.1093/ietisy/e90–d.11.1762

ing sonars [45], vision sensors [21], laser range finders (LRF) [22], as well as on several types of recognition algorithms from batch algorithms [23] to online algorithms [22]. Also, in the frameworks of Simultaneous Localization And Mapping (SLAM) [24], the self-localization has been considered a basic method and utilized for various purposes, such as data association [25], loop-closing [26] as well as kidnapped robot problem [27]. The localization techniques have been widely applied to a variety of vehicles, not only office robots, but also search & rescue robots, intelligent wheelchair, humanoids, floor-cleaning and patrol robots [46]. What most of the existing systems have in common is some degree of data association errors. No landmark ID information is provided by standard sensors, such as vision sensors, LRF, etc. The landmark features are often misidentified even by the current best pattern-recognition techniques. As an example, SIFT features [28] that are widely used as robust landmarks in vision-based localization community, can be indistinguishable from one another when applied to a cluttered office [29]. Unfortunately, data association error is one of most serious errors in localization problems and might even cause structural errors in a self-localization system. Development of robust recognition techniques to deal with the data association error is an ongoing research topic. On the other hand, an RFID-based localization system trivially solves the data association errors. Nearby landmarks within the robot’s sensor range are detected and identified by their tag IDs [30]. Another desirable property is that no line-of-sight constraint between landmarks and the robot is required. These properties of RFID systems have been utilized for the localization and SLAM purposes in several studies [31]–[33]. However, most of them are a kind of multi-sensor fusion: the RFID sensor system is not used alone but combined with more accurate sensors like LRF. Such a multi-sensor fusion approach is effective but expensive in some applications, e.g. a low-price home robot application. Early success in purely RFID-based localization can be found in long-range applications. Suppose a number of tags sparsely distributed in a large environment, an RFID reader can be viewed as a range sensor that provides distance-to-tag information. A representative example is the “range-only localization” system by Kurth [34]. In their work, an RFID reader is modeled as a time-of-flight range sensor with a

c 2007 The Institute of Electronics, Information and Communication Engineers Copyright 

TANAKA et al.: A SUPERVISED LEARNING APPROACH TO ROBOT LOCALIZATION USING A SHORT-RANGE RFID SENSOR

1763

maximum range of around 20 m and a standard deviation of 1 m. This problem is quite different from the conventional “bearing-range localization” problems [21], [22] where not only the range but also the bearing to a landmark is provided, e.g. by LRF. Their range-only localization technique is based on Kalman-filters and successful for a real robot system [34], [35]. However, the RFID sensor is corrupted by very large noises when used at short range. For example, the noise deviation can be 3 m or more when the distance-to-tag is less than 3 m. Such erroneous short-range measurements are simply ignored by their system [34]. In this paper, we address the problem of RFID-based localization in a short range application. Suppose a small number of tags are densely distributed in a small office (Figs. 1 and 2), e.g. an office sized 5 m × 5 m, only shortrange measurements may be obtained. Our system is composed of active tags as well as a short-range RFID reader, which detects nearby tags based on their signal intensity. A difficulty arises from the fact that the intensity signals at neighboring locations are similar to each other and not simply distinguishable. Moreover, signals are affected by the environment condition as well as by the electric field. This makes it difficult for a system designer to model in advance the signal noise characteristics. Thus, some model or internal representation should be learned online by the robot-self based on its history of exploration and sensing. One of our main contribution is a proposal of a machine learning approach [36]–[38] to such a short-range application. This approach is based on a machine learning technique, Support Vector Machine (SVM) [40]. A binary SVM classifier is trained in order to distinguish between signals at a certain location and signals at the other locations. The multi-class problem of the robot localization is reformulated as a set of binary classification problems. Then, multiple independent classifiers are trained for each candidate location. In addition, variety of virtual training examples are synthesized by a novel technique from a limited number of real sensor measurements. As a result, an internal representation of the transformation from a sensor measurement to the corresponding location ID is obtained. Although a lot of research about the pattern recognition for mobile robot localization have been performed in recent years e.g. [9], [12], [15], [22], there have been yet few studies on such an SVMbased localization problem. A most relevant work is visionbased localization technique by Morita et al. [39]. In their work, a key advantage of an SVM-based learner is described that it is almost free from the manual setting of threshold values and parameters. This is contrastive to many existing learning techniques that require such manual settings. In addition, a two stage SVM method is proposed that employ two types of SVMs, one for object recognition and another for robot localization. Their vision-based localization was successful in complex outdoor environments. Our system is also based on SVM techniques but using a very different type of sensor, a short range RFID sensor. Our fundamental difference from the above cited work lies in our ability to deal with the low-resolutional time-varying RF measure-

ments. Our second contribution is a design tool for tag arrangement. In the previous studies [31]–[34], the configuration of tags with respect to the environment is manually designed. However, there is no guarantee that such a manually designed configuration is an optimal one. This is problematic since the localization performance depends strongly on the tag configuration, the obstacle configuration as well as the electric field. This problem could be more serious in practice, since many potential users would have no expert knowledge of the RFID technology or the robot localization. A brute-force strategy such as distributing many tags densely over the entire workspace might be costful and not practical. Hence, a function of recommending a best tag configuration to its users would be very useful. Our solution is a kind of hypothesize-and-test approach: (1) a set of hypothesized tag-configurations are randomly generated in a simulated environment, (2) utility of each configuration is evaluated according to the aforementioned SVM-based localization technique, then (3) a best hypothesis with the highest utility score is recommended to the user. A simple objective function can be designed by the user in order to establish interactive decision making. This paper is organized as follows. In Sect. 2, the main architecture of the proposed system is illustrated and some assumptions are made. In Sect. 3, the basic usage of SVM in the context of the localization problem is described. In Sect. 4, the technique of tag arrangement is presented. In Sect. 5, experimental results of the localization and the tag arrangement tasks are reported and discussed. In Sect. 6, the conclusions are provided. 2. Overview of the System The purpose of self-localization is to estimate the robot’s location as accurate as possible. A set of candidate locations are given in advance by the user. The robot is initially placed at an unknown viewpoint, and then, the robot is required to narrow the candidates given a set of sensor measurements. This is a basic localization problem, which is considered an element of other localization frameworks such as global localization [23] as well as kidnapped robot problem [27]. The mobile robot is equipped with an RFID reader that is used as an RFID sensor by our system (Fig. 1). The RFID sensor is based on active tags, which have its own power

(a)

(b)

Fig. 1 Mobile robot with RFID System. (a) A mobile robot with an RFID sensor. (b) An RFID tag.

IEICE TRANS. INF. & SYST., VOL.E90–D, NO.11 NOVEMBER 2007

1764

source. Signals from multiple tags can be concurrently read by an RFID reader. The signal strength diminishes with respect to the distance between the reader and the tag. The signal is also affected by obstacles between the reader and the tag. The signal intensity has a strong relationship with the sensor range. The sensor range can be several mili-meter, several centi-meter or several meter according to the ISO specification. In our case, the sensor range is up to around 3 m. The unique ID signal is sent by a tag in random with an average of 7 seconds. The send-cycle depends on the RFID system, and there is usually a tradeoff between the sendcycle and the battery-life. The send-cycle is unique for the system and known in advance. From the user’s point of view, some practical constraints should be imposed on the tags. For example, a bruteforce strategy like distributing many tags densely over the entire workspace might be costful and not practical. In addition, standard RFID tags may not be customized for the robot localization purpose, but should be used for various purposes. Hence, the following constraints are put on the tags.

(a)

• Any tag should not be placed on the floor. • Only the standard ID data is available for every tag. Figure 2 (a) describes an instance of tag arrangement, which is actually used in our experiments. In this figure, the small black rectangles indicate the candidate locations, while the small white rectangles with 3 characters indicate locations of the individual tags. Although the tags are almost uniformly distributed in this example, tag arrangements in nonuniform distribution will be also considered in the experiments. Figure 2 (b) illustrates some snapshots of the actual environment. The localization procedure is composed of two main phases, training phase and classification phase. The training phase is aimed to learn characteristics of signals that would be received at each location. The robot is initially placed at each location and allowed to explore several viewpoints nearby the start location for a certain period of time. After the exploration, features are extracted from the raw sensor signals, and a set of training examples are generated for each location. Then, a binary SVM classifier is trained for each location in order to distinguish between signals at the location and signals at the other locations. When the training phase has finished, it becomes possible to estimate the location given a set of sensor measurements. The classification phase is aimed to identify at which locations the robot is currently located. In this phase, a set of features are extracted in the same way as in the training phase, then the samples are input to the SVMs. As a result, a classification result is output by the SVM for each input sample, representing whether the input sample is or is not similar with the learned samples. Based on the similarity values, the final decision will be made by the system.

(b) Fig. 2 An example of tag arrangement. (a) Plan view (small white rectangle: a tag location, small black rectangle: a candidates location). (b) From four different camera positions.

3. Localization Method 3.1 A Multi-Class SVM Support Vector Machine (SVM) is a model of binary classifier. SVM is based on structural risk minimization principle in computational learning theory [40]. Given a pair of separable sample sets, a positive and a negative example sets, SVM determines a hyperplane of separating band that would separate even unseen positive and negative samples in the high-dimensional feature space. In the case of non-separable sample sets, a penalty on the objective function is introduced for misclassified samples. SVM has some desirable characteristics in real world applications. Firstly, the hyperplanes are determined so as to minimize the structural risk, and give good generative performance. Secondly, optimality of the training results is guaranteed, in the sense that its training is a convex function optimization. Thirdly, a complex hyperplane can be achieved by using kernel trick technique.

TANAKA et al.: A SUPERVISED LEARNING APPROACH TO ROBOT LOCALIZATION USING A SHORT-RANGE RFID SENSOR

1765

The formulation of SVM is a binary classification problem, hence SVM is basically a binary classifier. Some extension is needed in the case of multi-class problems. The self-localization is a multi-class problem of classifying multiple candidates of location, and is reformulated as a set of binary classification problems. Our multi-class extension is a one-versus-rest (or one-against-all) output encoding scheme. One SVM classifier is trained for each candidate location and used for distinguishing the location’s features from the other locations’ features. Thus, if there are NL locations L0 , L1 , · · ·, LNL −1 to be identified, NL independent SVMs are needed. The sign of the SVM outputs {−1, +1} is used as the classification result. This output value gives a Boolean value corresponding that the robot is or is not at the corresponding location.

Fig. 3

Synthesizing training data.

3.2 Generating Synthesized Sensor Data In order to efficiently learn the relationship between sensor data and robot’s location, a variety of training data is required by SVM. Specifically, various sensor data is expected because sensor data is influenced not only by the robot’s viewpoint, but also by time, by material of occluding objects as well as by magnetic or electromagnetic field. This is problematic since the time allowed in the training phase is typically very limited to obtain sufficient amount of sensor data. To deal with this problem, a various of sensor data is synthesized from real sensor data. Let T denote the period of RFID signals, NT the number of observed tags, S (t) a real vector of NT dimension representing the signal received at a time t. If any signal is received from the tag #i, the i-th element S i (t) of S (t) is set as received signal’s intensity, otherwise, set as 0. Regarding a signal S (t) at t will be similar with the signal S (t + T ) in the next period, one period T of signal sequence S t = (S (t), S (t + 1), · · · , S (t + T − 1))

(1)

is defined as minimum unit of training data. Now, NS synthesized signal sequences are generated by iterating the following two phases alternatively. 1. Randomly sample one period of signal sequence S t from the real signal sequence. 2. Randomly sample T integer values k0 , k1 , · · ·, kT −1 , and generate a synthesized signal sequence   S t = S (t + k0 T ), S (t + 1 + k1 T ),  · · · , S (t + T − 1 + kT−1 T ) . (2) Figure 3 illustrates the procedure of generating a synthesized signal sequence and a feature vector. Let t s and te respectively denote the start and the end points in time of the real sequence, the range of ki is determined as t s ≤ t + ki T ≤ te .

Fig. 4

Dividing a training feature set into positive and negative sets.

3.3 Localization with a Set of SVM Each SVM is trained so as to identify each Ln (n ∈ [0, NL − 1]) using two sets of feature vectors, the positive examples that are likely acquired at the location Ln , and the negative examples that are likely acquired at the other locations {Ln |n  n, n ∈ [0, NL − 1]}. Figure 4 illustrates how the entire set is divided into the positive and the negative sets. The training procedure is as follows. First, a feature vector is generated from each synthesized signal sequence, as a real vector of NT dimension. This feature vector is computed as an average of signal intensity: t+T −1 1  S (u). T u=t

After that, NE positive examples are simply sampled from the set of feature vectors for Ln , while NE negative examples are sampled from the feature vectors for another location Ln (n  n) selected randomly. In the above procedure, a robust feature is extracted by averaging a fixed-length synthesized signal sequence (SSS). The length of SSS is fixed to one period of signals, e.g. 7 s. In general, there is a tradeoff between length of the SSS and robustness of the extracted feature. If the SSS length is set longer than one signal period, the signal feature could be

IEICE TRANS. INF. & SYST., VOL.E90–D, NO.11 NOVEMBER 2007

1766

more robust. However, such a strategy requires the robot to stay at one location for more than one period. Hence, the desirable length of SSS should depend on the individual applications. Our case can be viewed as a most simple case, one period length SSS. In the classification phase, every time new data arrives, the data is input to every SVM. Then, a Boolean value ∈ {−1, +1} is given by each SVM. If the output value is positive, it can be said likelihood of the robot being located at the corresponding location is high, otherwise, low. A positive value may be output by more than one SVM. For example, if there are ten SVMs and positive values are output by only two SVMs, and if either of these two SVMs corresponds to the true robot’s location, it can be said that the localization is successful with 80% accuracy. Our decoding scheme for the multiple SVM is motivated by the general self-localization framework of multiple hypothesis tracking (MHT), allowing multiple hypotheses (of the robot location) to coexist until enough evidence is available [22]. In the MHT framework, the robot sometimes does not really know where it is exactly, but is able to narrow the location candidates (hypotheses) if it moves to a novel viewpoint and makes additional observations. Our scheme of narrowing the location candidates would be effective in such a general self-localization framework. 4. Tag Arrangement Method Success of localization depends largely on arrangement of the RFID tags. On the other hand, it is often difficult for the users to arrange tags in effective locations. Thus, there is a tradeoff between the cost of the robot localization task and the cost of the tag arrangement task. This problem is partially similar with the traditional problems of landmark selection [18], [41], [42], landmark design [43], as well as landmark arrangement [44]. In general, there are various candidates of the landmark arrangement. Each candidate C j of the arrangement should be evaluated with respect to its localization utility. A most simple way of this evaluation is that the user actually creates the landmark arrangement in the real environment and then tests the localization utility in the created environment. Unfortunately, this method is not practical in terms of the user’s cost, since the number of candidates is exponential to the number of landmarks. A more practical way is that the system creates a set of virtual arrangements {C j } and tests the localization utility of each candidate C j in a computer-simulated environment. There are two possible approaches to create such a virtual environment as well as to simulate sensor measurements. One is a model-based approach where an intermediate model (e.g. 3D geometric model) is reconstructed from actual sensor measurements and then can be used to simulate sensor measurements at arbitrary viewpoints. Another method is a sensor-based approach where actual sensor measurements are directly used to simulate sensor measurements at the same viewpoints. Most of previous techniques

are based on the model based approach, requiring some accurate sensor (e.g. LRF, a vision sensor) for the model reconstruction. Although this approach has an advantage that actual sensor measurements at only a few viewpoints are required, this approach is difficult in the case of inaccurate sensors like an RFID sensor where the model reconstruction itself becomes a costful and challenging task due to the large amount of disturbances. On the other hand, although the sensor-based approach requires sensor measurements at many viewpoints, this approach does not assume the reconstruction of any intermediate model nor the accurate sensors. It can be concluded that the sensor-based approach is a most relevant way to simulate sensor measurements in our case. Given actual measurement Z for the actual tagarrangement C, a virtual measurement Z j for a virtual tagarrangement C j (⊂ C) can be created by the synthesis technique described in Sect. 3.2. Thus, Z j is created by replacing its i-th element Zi of Z with 0 if i  C j . With above technique, the localization utility can be evaluated for arbitrary subset C j (⊂ C) of the actual tag set C. Taking this into account, the tag arrangement task is performed in the following procedure. 1. A user arranges as many tags C as possible in a random configuration with respect to the real environment. 2. The robot collects sensor data by exploration and sensing in the real environment, evaluates utility of each possible arrangement C j (⊂ C) in the virtual environment, and based on the results, recommends to the user the best arrangement with a highest evaluation value. 3. The user removes from the real environment the tags that are not included in the recommended tags C j . With this procedure, the cost of arranging and removing the tags is up to proportional to the number of tags and sufficiently low. Since there are impractical number of candidates of the tag arrangement, a practical number NC of candidates {C j } should be sampled. In addition, a simple objective function is designed by the user in order for an interactive decision making. This objective function is a form of penalty function, defined as a weighted sum of three kinds of penalty values. The first penalty value is represented by the “success rate” r s , measuring how much the localization being successful, determined as 1.0 if the SVM of the true location output any positive value or 0.0 otherwise. The second penalty value is “ambiguity” ra , measuring how many the NL candidate locations are narrowed down to, determined as the ratio of SVMs that output positive values. The third penalty is represented by the number NT of tags required for the tag arrangement. Based on these terminologies, the penalty function of a tag arrangement L is defined as p(L ) = w s (1 − r s ) + wa ra + (1 − w s − wa )NT /NT o , (3) where w s , wa (∈ [0, 1]) are weighting coefficients and NT o is a normalizing constant.

TANAKA et al.: A SUPERVISED LEARNING APPROACH TO ROBOT LOCALIZATION USING A SHORT-RANGE RFID SENSOR

1767

5. Experiments In order to evaluate the effectiveness of the proposed techniques, an experiment with a real mobile robot system is conducted. The experimental environment is an office room, sized 6 m × 4.5 m, containing furniture, PCs, pillars, etc. An overview of the experimental environment is shown in Fig. 2. The normalizing constant NT o is empirically set 100. A preliminary experiment is performed to determine the SVM parameters. The SVM parameters are optimized with respect to the localization accuracy obtained in the preliminary experiments. Specifically, the Radial Basis Function (RBF) is chosen as the kernel function since the RBF kernel (with σ2 = 5,  = 15) has found to be most accurate. In the following experiments, the SVM parameters are fixed to these predetermined values. 25 active RFID tags are used at most in this experiment. The timings of sending signals are not controlled by the RFID sensor, but by the tag itself. The send-cycle between two successive signals is fixed. The timings are not synchronous between different tags. The sensing range of the RFID sensor is almost fixed and slightly depending on the surrounding conditions, e.g. occluding objects. Our RFID system is SPIDER provided by RF-CODE, Inc† . The maximum sensing range of the RFID sensor is around 3 m. The time interval between two successive signals is T = 7 s. The function of multi-tag reading is available. Sensitivity of the reader is configurable in 8 levels. Level #1 corresponds to the highest intensity level, while level #8 corresponds to the lowest level. 5.1 Self-Localization The purpose of our self-localization technique is to estimate the robot’s location as accurate as possible. The number of candidate locations is NL = 10. Each candidate location is shown as a black rectangle with the corresponding tag ID number in Fig. 2. The robot is initially placed at each of the 10 candidate locations, then the robot follows a randomlydetermined trajectory in a small area (0.5 m×0.5 m) centered at the candidate location for a short period of time (60 s). A set of sensor measurements from the RFID sensor are collected by the robot at a randomly determined viewpoint on the trajectory. The collected sensor measurements are used for training phase or classification phase, as described in Sect. 2. For this aim, various of training or classification dataset are synthesized from the input sensor measurements in the procedure described in Sect. 3.2. A feature vector is extracted from each synthesized signal sequence by averaging the signal intensity as described in Sect. 3.3. Figures 5 (a) and (b) describe two examples of the extracted signal features at the locations #4 and #9. Either figure shows a histogram of 100 signal features at the corresponding candidate location. It can be seen that even though these features are obtained at the same location, variety of signal features are

(a)

(b) Fig. 5 Signals received at each candidate location (horizontal axis: tag ID, vertical axis: frequency). Each plot indicates each level of intensity, quantized into eight levels, (a) location #4. (b) location #9.

obtained. It can be also seen that even though the two histogram correspond to different locations, the histograms are partially similar with one another, and not simply differentiable. These similarity and dissimilarity make it a difficult problem to learn and classify the candidate locations. The localization performance is evaluated in terms of the success rate r¯s (averaged) and the ambiguity r¯a (averaged). The definition of the success rate and the ambiguity are same as the ones shown in Sect. 4. According to the definitions, the localization is most successful if a small number of classifiers (SVMs) output positive value and if one of them corresponds to the true robot location. For statistical investigation, 1,000 different tag arrangements L are generated in a computer simulation. For each arrangement, synthesized training and classification datasets are generated, then the localization experiment is conducted to evaluate the localization performance. Figure 6 illustrates the performance results. The two graphs describe the relationships among the number of tags NT , the success rate r¯s and the ambiguity r¯a . Either graph †

http://www.rfcode.com/

IEICE TRANS. INF. & SYST., VOL.E90–D, NO.11 NOVEMBER 2007

1768

(a)

(b) Fig. 6 The relationship among the success rate, ambiguity, and the number of tags.

is the average of the results for 1,000 different tag arrangements. The error bar indicates the standard deviation. It can be seen from Fig. 6 (a), the success rate increases as the number of tags increases. The success rate exceeds a high value 0.95 when the number of tags is larger than 14. From Fig. 6 (b), the ambiguity decreases as the number of tags increases. The ambiguity fall below 0.2 when the number of tags is larger than 10. It can be said that the localization performance is sufficiently high when more than 10 tags are used, in this experiment. This result is significant, since the localization is successful with a high localization accuracy, 0.5 m × 0.5 m, even though the range of the RFID sensor is short and the spatial resolution is low, 3 m. Although in this experiment the robot is required to stay at one location for at least one period of time T = 7 s, recent years variety of low-cost tags with shorter T have been developed, thus our system could be extended to faster moving robots. In this experiment, a relatively small sized (6 m × 4.5 m) environment is considered. It is possible to extend our system to a large-scale environment, since the ID data provided by the RFID system could be used to constrain the robot location within an environment region as small as the one considered in this experiment. In some ap-

plications, the achieved localization accuracy 0.5 m × 0.5 m may not be satisfactory. In such a case, our system should be extended to the framework of multiple hypothesis tracking (MHT) [22], as described in Sect. 3.3, and there exist effective methods to improve the localization accuracy in the MHT framework. Finally, our approach regards as relevant every location that is classified as relevant by the corresponding SVM. An alternative approach might be simply regarding the nearest location to the highest tag signal as the winner. Such an approach would be effective if RF signals are sufficiently reliable and almost uniform at every point in time. On the other hand, our approach is rather conservative in only eliminating false candidates that contradict to the features that are learned over periods of time. Such an approach is suitable for the case where individual measurements are less reliable, and is applicable to the MHT framework. It can be concluded that the proposed method of feature extraction and learning SVM classifiers is effective when dealing with low-resolutional time-varying RF signals. We have also investigated the computational efficiency of the proposed methods. The cost for sensor acquisition and feature extraction was sufficiently low compared with the period time T of tag signals. On the other hand, the cost for learning and classification depends not on the spatial resolution of the tag reader, but on the dimension and the number of the feature vector as well as on the SVM algorithm. The classification cost, using a normal PC (CPU Pentium 4 1.8 GHz, Memory 256 Mbyte) with a large number (1,000) of high (32) dimensional feature vectors, was less than 0.01 s. The learning cost was less than 5 minutes, which is sufficiently low compared with the cost spent by the robot to explore and observe the entire locations. 5.2 Tag Arrangement An experiment of the tag arrangement task is also conducted. In this task, various candidates of the arrangement are generated by the system, and the best arrangements are selected in terms of the user-defined objective (penalty) function (3). For computational feasibility, a practical number (1,000) of candidates are randomly sampled from the huge number (225 ) of potential candidates. Although the computational complexity of the method is basically exponential to the number of tags, the actual cost is typically feasible. For example, in this experiment, the time for tag arrangement task is less than 10 minutes, including the time for synthesizing sensor data, feature extraction, learning and classification. It can be said the computational cost is sufficiently low compared with the cost that the user actually creates the tag arrangement in the real environment. In real applications, the objective function (3) should be set by the user considering the purpose of the individual self-localization tasks. For example, in some applications, increase of the localization accuracy might be most important purpose, and in such a case the weighting coefficient w s should be set higher than the other weights. In some other

TANAKA et al.: A SUPERVISED LEARNING APPROACH TO ROBOT LOCALIZATION USING A SHORT-RANGE RFID SENSOR

1769

(a)

(b)

(c)

Fig. 7 Best tag arrangement for three different penalty weights (wa , wb ) of (a) (1/4, 1/4), (b) (1/2, 1/4), and (c) (1/4, 1/2).

applications, reduction of the ambiguity might be most important, and wa should be set higher than the other weights. In our experiment, three typical cases are considered that correspond to three different purposes of self-localization, (a) reduction of the tags, (b) increase of the accuracy, (c) reduction of the ambiguity. These three cases respectively correspond to the following parameter settings, (a) (1/4, 1/4), (b) (1/2, 1/4), (c) (1/4, 1/2). The tag arrangement results are summarized in Table 1. For each case (a), (b) or (c), 5 best tag arrangements are indicated. In this table, the first four columns are the evaluation values r¯s , r¯a , NT as well as their weighted sum p(L ) defined in (3). The last (fifth) column is a set of tag IDs of the corresponding tag arrangement. For each case, the spatial configurations of the best tag arrangement are described in Fig. 7. The tags appearing in this figure are emphasized in bold in Table 1. It can be concluded that not all the 25 tags are required in each case. For case (a), where reduction of tags is considered most important purpose, spatial density of tags is significantly low, especially any tag is not placed in the left upper region of the room. On the contrary, for case (c), where reduction of the ambiguity is considered most important, tags are placed along the outside wall more densely and in virtually equal intervals. Finally, for case (b), where increase of the localization accuracy is considered most important, although the ambiguity is slightly larger compared to the case (c), the number of tags is significantly lower, 8. In consequence, the proposed technique is able to reduce the number of tags while keeping high-level localization accuracy, thus yield highly efficient tag arrangements as a result. 6. Conclusions A solution to purely RFID-based robot localization using a short-range RFID sensor is presented. In the case of short range applications, RFID signals are prone to noises and not always predictable. To deal with such ambiguous sensor measurements, a supervised machine learning technique based on Support Vector Machine is introduced. Our solution is a sensor-based approach, in which virtually no

Table 1 5 best tag arrangements for three different penalty weights. (a) (ws , wa ) = (1/4, 1/4). (b) (ws , wa ) = (1/2, 1/4). (c) (ws , wa ) = (1/4, 1/2). (a) p(L ) 0.45 0.45 0.45 0.45 0.45

r¯s 1.0 1.0 1.0 1.0 1.0

r¯a 0.21 0.14 0.14 0.18 0.18

NT 4 8 8 6 6

Tags FLZ, HEJ, MDM, UJE ORB, CEO, EJA, YGQ, HEJ, RZG, TFC, EBM GCJ, CEO, ZJW, YGQ, HEJ, GPF, KJF, RZS DUR, HEJ, IGU, KJF, RZG, EBM TSE, DUR, YGQ, RZG, TFC, EBM

(b) p(L ) 0.24 0.24 0.25 0.25 0.25

r¯s 1.0 1.0 1.0 1.0 1.0

r¯a 0.14 0.14 0.18 0.18 0.17

NT 8 8 6 6 7

ORB, CEO, EJA, YGQ, HEJ, RZG, TFC, EBM GCJ, CEO, ZJW, YGQ, HEJ, GPF, KJF, RZS DUR, HEJ, IGU, KJF, RZG, EBM TSE, DUR, YGQ, RZG, TFC, EBM TSE, DUR, YGQ, IGU, MDM, KJF, UJE

p(L ) 0.28

r¯s r¯a 1.0 0.11

NT 13

Tags

0.28 0.28 0.28

1.0 0.14 1.0 0.14 1.0 0.14

8 8 10

0.28

1.0 0.14

11

Tags

(c) OXY, TSE, ORB, GCJ, CEO, DUR, PYY, FLZ, HEJ, IGU, RZG, UJE, EBM ORB, CEO, EJA, YGQ, HEJ, RZG, TFC, EBM GCJ, CEO, ZJW, YGQ, HEJ, GPF, KJF, RZS TSE, ORB, GCJ, ZJW, YGQ, FLZ, ZEV, HEJ, IGU, UJE TSE, ORB, CEO, YGQ, PYY, HEJ, IGU, KJF, UJE, TFC, RZS

sensor model is defined in advance, but more robust sensor model adapting to individual environments is created by the robot-self in online. This sensor-based approach is then also applied to the tag-arrangement task, in which multiple hypotheses of the arrangement are generated and scored in terms of the localization utility. To enhance the efficiency of the proposed approach, a technique of synthesizing variety of sensor measurements from limited amount of real measurements is proposed. From the experimental results, it can be concluded that the localization accuracy (0.5 m × 0.5 m) achieved by the proposed technique can be higher than the sensor accuracy (3 m) provided by the RFID sensor. This result is sufficient for narrowing the candidates of the robot’s location. If we want to uniquely determine the robot’s viewpoint, a multi-hypothesis approach would be useful for in-

IEICE TRANS. INF. & SYST., VOL.E90–D, NO.11 NOVEMBER 2007

1770

tegrating sensor measurements at multiple viewpoints in the framework of the global localization [22]. Even in such a framework, the proposed technique can be viewed as a robust hypothesis tester for detecting true robot position hypotheses. There remains some issues to be discussed. Firstly, we focused on active RFID tags. Therefore, the timings of signal sending and receiving could not be controlled by the robot. In the case of passive tags, the timings may be controlled by the robot, which would reduce complexity of the self-localization problem. Secondly, we used homogeneous tags. In real applications, there might be heterogeneous tags with different signal intensities or with different signal intervals. Even in such a case, the various of signal features are expected to be learned adaptively by the proposed machine learning. Thirdly, we chose a sensor-based approach to evaluate utility of tag-arrangement candidates. With comparison to conventional model-based approach, this approach is robust, especially in the case of low accurate sensors like RFID sensor. In the case of RFID sensor, sensor measurements are prone to noises, and not always predictable due to signal occlusions and other many causes. Even in such a case, our sensor-based approach is able to have a realistic prediction since this approach is based on the real sensor data. Finally, we experimented in a static environment. In dynamic office environments, the configuration of tags attached on movable objects could be modified. For consistency, such modifications of tag arrangement have to be detected by the robot-self, and if some modifications are detected, the SVMs should be retrained. References [1] M. Drumheller, “Mobile robot localization using sonar,” IEEE Trans. Pattern Anal. Mach. Intell., vol.PAMI-9, no.2, pp.325–332, 1987. [2] J. Leonard and H. Durrant-White, “Mobile robot lacalization by tracking geometric beacons,” IEEE Trans. Robot. Autom., vol.7, no.3, pp.89–97, 1991. [3] S. Atiya and G.D. Hager, “Real-time vision-based robot localization,” IEEE Trans. Robot. Autom., vol.9, no.6, pp.785–800, 1993. [4] J.J. Leonard, B.A. Moran, I.J. Cox, and M.L. Miller, “Underwater sonar data fusion using an efficient multiple hypothesis algorithm,” Proc. IEEE Int. Conf. Rob. and Aut., pp.2995–3002, 1995. [5] R. Talluri and J.K. Aggarwal, “Mobile robot self-location using model-image feature correspondence,” IEEE Trans. Robot. Autom., vol.12, pp.63–77, 1996. [6] K.T. Simsarian and T.J. Olson, “View-invariant regions and mobile robot self-localization,” IEEE Trans. Robot. Autom., vol.12, pp.810–815, 1996. [7] M. Betke and L. Gurvits, “Mobile robot localization using landmarks,” IEEE Trans. Robot. Autom., vol.13, no.2, pp.251–263, 1997. [8] D. Fox, W. Burgard, and S. Thrun, “Markov localization for mobile robots in dynamic environments,” Artif. Intell. Res., vol.11, pp.391– 427, 1999. [9] U. Gerecke, N.E. Sharkey, and A.J.C. Sharkey, “Reliable robot localization with an ensemble approach,” Proc. Int. Symposium on Robotics and Automation (ISRA-2000), 2000. [10] P. Jensfelt, D. Austin, O. Wijk, and M. Anderson, “Feature based condensation for mobile robot localization,” Proc. IEEE Int. Conf.

Robotics and Automation, pp.2531–2537, 2000. [11] K. Arras, J. Castellanos, M. Schilt, and R. Sieqwart, “Towards feature-based multi-hypothesis localization and tracking,” Proc. 4th European Workshop on Advanced Mobile Robots, pp.57–65, 2001. [12] U. Gerecke, N.E. Sharkey, and A.J.C. Sharkey, “Common evidence vectors for reliable localization with SOM ensembles,” Proc. Engineering Application of Neural Networks, 2001. [13] A. Howard, M.J. Mataric, and G.S. Sukhatme, “Localization for mobile robot teams: A maximum likelihood approach,” Technical Report IRIS-01-407, USC, 2001. [14] J.-S. Gutmann and D. Fox, “An experimental comparison of localization methods continued,” Proc. IEEE/RSJ Int. Conf. Intelligent Robots and Systems, pp.736–743, 2002. [15] J. Matjaˇz, W. Horst, and B. Horst, “Mobile robot localization under varying illumination,” Proc. Int. Conf. Pattern Recognition, pp.741– 744, 2002. [16] H.-M. Gross, A. Koenig, H.-J. Boehme, and C. Schroeter, “Visionbased Monte Carlo self-localization for a mobile service robot acting as shopping assistant in a home store,” Proc. IEEE/RSJ Int. Conf. Intelli. Rob. and Sys., pp.256–262, 2002. [17] S. Lenser and M. Velose, “Sensor resetting localization for poorly modeled mobile robots,” IEEE Int. Conf. Rob. and Aut., pp.1225– 1232, 2002. [18] C.F. Olson, “Selecting landmarks for localization in natural terrain,” Autonomous Robots, vol.12, pp.201–210, 2002. [19] M. Bestelli, D.G. Sorrenti, and F.M. Marchese, “A robot localization method based on evidence accumulation and multi-resolution,” Proc. IEEE/RSJ Int. Conf. Intelli. Rob. and Sys., pp.415–420, 2002. [20] L.M. Paz, P. Pinies, J. Neira, and J.D. Tardos, “Global localization in slam in bilinear time,” Proc. IEEE/RSJ Int. Conf. Intelligent Robots and Systems, pp.655–661, 2005. [21] D.C.K. Yuen and B.A. MacDonald, “Vision-based localization algorithm based on landmark matching, triangulation, reconstruction, and comparison,” IEEE Trans. Robotics, vol.21, no.2, pp.217–226, 2005. [22] S. Thrun, D. Fox, W. Burgard, and F. Dellaert, “Robust Monte Carlo localization for mobile robots,” Artificial Intelligence Journal, vol.128, no.1, pp.99–141, 2001. [23] J. Neira, J.D. Tardos, and A. Castellanos, “Linear time vehicle relocation in slam,” Proc. 2005 IEEE Int. Conf. Robotics and Automation, pp.427–433, 2003. [24] S. Thrun, D. Koller, Z. Ghahmarani, and H. Durrant-White, “Slam updates require constant time,” Technical Report, School of Computer Science, Carnegie Mellon University, 2002. [25] D. Hahnel, W. Burgard, B. Wegbreit, and S. Thrun, “Towards lazy data association in SLAM,” Proc. 11th International Symposium of Robotics Research, pp.83–105, 2003. [26] P. Newman and K. Ho, “Slam-loop closing with visually salient features,” Proc. IEEE Int. Conf. Robotics and Automation, pp.635–642, 2005. [27] A. Lankenau and T. R¨ofer, “Mobile robot self-localization in largescale environments,” Proc. IEEE Int. Conf. on Robotics and Automation, pp.1359–1364, 2002. [28] D.G. Lowe, “Distinctive image features from scale-invariant keypoints,” Int. J. Computer Vision, vol.60, no.2, pp.91–110, 2004. [29] S. Se, D. Lowe, and J. Little, “Vision-based mobile robot localization and mapping using scale-invariant features,” Proc. IEEE International Conference on Robotics and Automation, pp.2051–2058, 2001. [30] M. Schneider, Radio frequency identification technology and its applications in the commercial construction industry, Master’s of Science in Civil Engineering Examination at the University of Kentucky, 2003. [31] D. Haehnel, W. Burgard, D. Fox, K.P. Fishkin, and M. Philipose, “Mapping and localization with RFID technology,” IEEE Int. Conf. Robotics and Automation, pp.1015–1020, 2004. [32] A. Kleiner, J. Prediger, and B. Nebel, “RFID technology-based ex-

TANAKA et al.: A SUPERVISED LEARNING APPROACH TO ROBOT LOCALIZATION USING A SHORT-RANGE RFID SENSOR

1771

[33]

[34]

[35]

[36]

[37]

[38]

[39]

[40]

[41] [42]

[43]

[44]

[45]

[46]

ploration and slam for search and rescue,” Proc. IEEE/RSJ Int. Conf. Intelligent Robots and Systems, pp.4054–4059, 2006. B.K. Kim, N. Tomokuni, K. Ohara, T. Tanikawa, K. Ohba, and S. Hirai, “Ubiquitous localization and mapping for robots with ambient intelligence,” Proc. IEEE/RSJ Int. Conf. Intelligent Robots and Systems, 2006. D. Kurth, “Range-only robot localization and slam with radio,” Robotics Institute, Carnegie Mellon University, Technical Report CMU-RI-TR-04-29, 2004. J. Djugash, S. Singh, and P.I. Corke, “Further results with localization and mapping using range from radio,” Int. Conf. Field and Service Robotics, pp.231–242, 2005. M. Hirayama, K. Tanaka, N. Okada, and E. Kondo, “Vision-based self-localization in non-stationary environments by using support vector machines,” Proc. Int. Conf. Cybernetics and Intelligent System, pp.123–128, 2004. K. Yamano, K. Tanaka, M. Hirayama, E. Kondo, Y. Kimuro, and M. Matsumoto, “Self-localization of mobile robots with RFID system by using support vector machine,” Proc. IEEE/RSJ Int. Conf. Intelligent Robots and Systems, pp.3756–3761, 2004. K. Tanaka, M. Hirayama, N. Okada, and E. Kondo, “Location-driven image retrieval for images collected by a mobile robot,” JSME International Journal, vol.49, no.3, pp.1058–1066, 2006. H. Morita, M. Hild, J. Miura, and Y. Shirai, “Panoramic view-based navigation in outdoor environments based on support vector learning,” Proc. IEEE/RSJ Int. Conf. Intelligent Robots and Systems, pp.2303–2307, 2006. M.A. Hearst, B. Scholkopf, S. Dumais, E. Osuna, and J. Platt, “Trends and controversies - support vector machines,” IEEE Intelligent Systems, vol.13, no.4, pp.18–28, 1998. S. Thrun, “Bayesian landmark learning for mobile robot navigation,” Mach. Learn., vol.39, no.1, pp.41–78, 1998. F. Jason, M. Stephen, and S. Jonathan, “Sensory anticipation for autonomous selection of robot landmarks,” Anticipatory Behavior in Adaptive Learning Systems, pp.201–221, 2003. B. Amy, S. Daniel, B. Darius, D. Cristian, and W. Peter, “Mobile robot navigation using self-similar landmarks,” IEEE Int. Conf. Robotics and Automation, pp.1428–1434, April 2000. K. Tashiro, J. Ota, Y.C. Lin, and T. Arai, “Design of the optimal arrangement of artificial landmarks,” Proc. IEEE Int. Conf. Robotics and Automation, pp.407–413, 1995. E. Olson, J.J. Leonard, and S. Teller, “Robust range-only beacon localization,” IEEE J. Oceanic Engineering, vol.31, no.4, pp.949– 958, 2006. P. Jensfelt, S. Ekvall, D. Kragic, and D. Aarno, “Integrating slam and object detection for service robot tasks,” Proc. IEEE/RSJ IROS Workshop on Mobile Manipulators: Basic Techniques, New Trends and Applications, 2005.

Kanji Tanaka received the M.S and Ph.D. degrees in the graduate school of information science and electrical engineering, Kyushu University, in 1997 and 2000, respectively. He became a research assistant in the school of Systems Information Science in 2000, Future University-Hakodate. Since 2002, he has been a research associate at the Department of Intelligent Machinery and Systems, Kyushu University. His current research interests are intelligent robotics, machine learning and pattern recognition. He is a member of the IEEE, SICE, JSME, and RSJ.

Yoshihiko Kimuro received the BS and MS degrees in Electric Engineering from Kyushu University, Japan, in 1984 and 1986, respectively. He worked as a research associate at Kyushu University from 1987 to 1998. He received the Ph.D. degree in Information Science from Kyushu University in 1996. He has been with Institute of Systems & Information Technologies since 1998. His research interests include computer vision, robot systems and education of computers.

Kentaro Yamano received the BS and MS degrees in mechanical engineering from Kyushu University, in 2003 and 2005, respectively. He joined Nissan Motor Co., Ltd. in 2005. He has been engaged in the development of a navigation system of a low-price home robot.

Mitsuru Hirayama received the BS and MS degrees in mechanical engineering from Kyushu University, in 2004 and 2006, respectively. He joined Nissan Motor Co., Ltd. in 2006. He has been engaged in the development of machine learning technique for robot perception systems.

Eiji Kondo received the B.S, M.S., and Ph.D. degrees in mechanical engineering from Kyushu University, Fukuoka, Japan, in 1970, 1972, and 1978, respectively. He became an Assistant Professor and an Associate Professor in the Faculty of Engineering, Kyushu University. He is currently a Professor in the Department of Intelligent Machinery and Systems, Kyushu University. His current research interests are control theory and its applications. Dr. Kondo is a member of the JSME, SICE, and ISCIE.

Michihito Matsumoto received the B.E. degree in electrical engineering from Kyushu Institute of Technology. He received M.E. and Ph.D. degrees in electronics engineering from Kyushu University in 1977 and 1987, respectively. He joined NTT in 1977. Since 1987 he has been in the Project Group where he has been engaged in development of optical fiber splice technologies. He has been with Institute of Systems & Information Technologies since 1999.