JAMRIS 2010 Vol 4 No 2

Page 1


JOURNAL of AUTOMATION, MOBILE ROBOTICS & INTELLIGENT SYSTEMS

Editor-in-Chief Janusz Kacprzyk

Executive Editor: Anna Ładan aladan@piap.pl

(Systems Research Institute, Polish Academy of Sciences; PIAP, Poland)

Associate Editors: Mariusz Andrzejczak (PIAP, Poland) Katarzyna Rzeplińska-Rykała (PIAP, Poland)

Co-Editors: Dimitar Filev (Research & Advanced Engineering, Ford Motor Company, USA)

Kaoru Hirota

Webmaster: Tomasz Kobyliński tkobylinski@piap.pl

(Interdisciplinary Graduate School of Science and Engineering, Tokyo Institute of Technology, Japan)

Witold Pedrycz

Proofreading: Urszula Wiączek

(ECERF, University of Alberta, Canada)

Roman Szewczyk (PIAP, Warsaw University of Technology, Poland)

Editorial Office: Industrial Research Institute for Automation and Measurements PIAP Al. Jerozolimskie 202, 02-486 Warsaw, POLAND Tel. +48-22-8740109, office@jamris.org

Copyright and reprint permissions Executive Editor

Editorial Board: Chairman: Janusz Kacprzyk (Polish Academy of Sciences; PIAP, Poland) Plamen Angelov (Lancaster University, UK) Zenn Bien (Korea Advanced Institute of Science and Technology, Korea) Adam Borkowski (Polish Academy of Sciences, Poland) Wolfgang Borutzky (Fachhochschule Bonn-Rhein-Sieg, Germany) Oscar Castillo (Tijuana Institute of Technology, Mexico) Chin Chen Chang (Feng Chia University, Taiwan) Jorge Manuel Miranda Dias (University of Coimbra, Portugal) Bogdan Gabryś (Bournemouth University, UK) Jan Jabłkowski (PIAP, Poland) Stanisław Kaczanowski (PIAP, Poland) Tadeusz Kaczorek (Warsaw University of Technology, Poland) Marian P. Kaźmierkowski (Warsaw University of Technology, Poland) Józef Korbicz (University of Zielona Góra, Poland) Krzysztof Kozłowski (Poznań University of Technology, Poland) Eckart Kramer (Fachhochschule Eberswalde, Germany) Andrew Kusiak (University of Iowa, USA) Mark Last (Ben–Gurion University of the Negev, Israel) Anthony Maciejewski (Colorado State University, USA) Krzysztof Malinowski (Warsaw University of Technology, Poland)

Andrzej Masłowski (PIAP, Poland) Tadeusz Missala (PIAP, Poland) Fazel Naghdy (University of Wollongong, Australia) Zbigniew Nahorski (Polish Academy of Science, Poland) Antoni Niederliński (Silesian University of Technology, Poland) Witold Pedrycz (University of Alberta, Canada) Duc Truong Pham (Cardiff University, UK) Lech Polkowski (Polish-Japanese Institute of Information Technology, Poland) Alain Pruski (University of Metz, France) Leszek Rutkowski (Częstochowa University of Technology, Poland) Klaus Schilling (Julius-Maximilians-University Würzburg, Germany) Ryszard Tadeusiewicz (AGH University of Science and Technology in Kraków, Poland)

Stanisław Tarasiewicz (University of Laval, Canada) Piotr Tatjewski (Warsaw University of Technology, Poland) Władysław Torbicz (Polish Academy of Sciences, Poland) Leszek Trybus (Rzeszów University of Technology, Poland) René Wamkeue (University of Québec, Canada) Janusz Zalewski (Florida Gulf Coast University, USA) Marek Zaremba (University of Québec, Canada) Teresa Zielińska (Warsaw University of Technology, Poland)

Publisher: Industrial Research Institute for Automation and Measurements PIAP

If in doubt about the proper edition of contributions, please contact the Executive Editor. Articles are reviewed, excluding advertisements and descriptions of products. The Editor does not take the responsibility for contents of advertisements, inserts etc. The Editor reserves the right to make relevant revisions, abbreviations and adjustments to the articles.

All rights reserved ©

1


JOURNAL of AUTOMATION, MOBILE ROBOTICS & INTELLIGENT SYSTEMS VOLUME 4, Nツー 2, 2010

CONTENTS REGULAR PAPER 55

3

Two dimensional model of CMM probing system S.H.R. Ali

The study of bio-inspired robot motion control system T. Matsuo, T. Yokoyama, D. Ueno, K. Ishii 62

8

New approach to the accuracy description of unbalanced bridge circuits with the example of Pt sensor resistance bridges Z.L. Warsza

Development of antagonistic wire-driven joint employing kinematic transmission mechanism T. Sonoda, Y. Nishida, A.A.F. Nassiraei, K. Ishii

16

Data fusion in measurements of angular position S. ナ「czak

SPECIAL ISSUE SECTION

DEPARTMENTS 71

23

Brain-like Computing and Applications Guest Editors: Tsutomu Miki, Takeshi Yamakawa

IN THE SPOTLIGHT 73

EVENTS

25

The spike-timing-dependent plasticity function based on a brain sequential learning system using a recurrent neuronal network G. Ogata, K. Natsume, S. Ishizuka, H. Hayashi 31

A Simple local navigation system inspired by hippocampal function and its autonomous mobile robot implementation T. Miki, H. Hayashi, Y. Goto, M. Watanabe, T. Inoue 39

Building a cognitive map using an SOM2 K. Tokunaga, T. Furukawa 48

A human robot interaction by a model of the emotional learning in the brain S. Sonoh, S. Aou, K. Horio, H. Tamukoh, T. Koga, T. Yamakawa

2


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N째 2

2010

TWO DIMENSIONAL MODEL OF CMM PROBING SYSTEM Received 2nd December 2009; accepted 13th January 2010.

Salah H. R. Ali

Abstract: A coordinate measuring machine (CMM) as an automation technology is playing the key role in the modern industry to improve the measurement accuracy. Accurate probing that is computer controlled is the current trend for the next generation of coordinate metrology. However, the CMM probing system is limited by its dynamic root errors that may markedly affect its response characteristics. In this paper, dynamic response errors of CMM measurements have been analyzed. The adopted probe stylus sizes throughout the course of measurements are found to cause some waviness errors during CMM operations due to each of the prescribed angle of the probe tip contact point with the specimen surface and the radius of the stylus tip. Variations in the geometry of the stylus have their consequent effects on its inherent intrinsic dynamic characteristics that in turn would cause relevant systematic root errors in the resulted measurements. Unforeseeable geometrical errors of a CMM using a ductile touch-trigger probing system have been characterized theoretically. These results are analyzed in order to investigate the effect of the dynamic root errors in the light of six probe stylus tip of the situation into account when assessing the accuracy of the CMM measurements. Analytical approaches have been applied on a developed two dimensional model (2DM) of stylus tip to demonstrate the capability of such approaches of emphasizing the root error concept using the strategy of CMM ductile trigger type of probe. Keywords: CMM, trigger probe, stylus tip, tip root errors, and two-dimensional-model (2DM).

1. Introduction One challenge for advanced coordinate metrology is the accurate dimensional measurements on modern engineering objects, especially in aerospace and automotive industries. The probe is one of the most important systems of CMM measurement accuracy. However, studies on CMMs cannot separate the characteristic performance of the probing system from other CMM error sources [1-9]. The stylus tip contact with the detected surface is the source of electronic signals that will develop the pattern on the working objects. So, the performance of the CMM overall system is very much dictated by the motion precision of the probe stylus tip and its actuator. Therefore, the probe stylus tip is literally at the center of the CMM operation and a key element of coordinate measurements. A variety of probe designs in CMM are available today, although most probes are compatible with most CMMs [1,

10]. Metrology engineer should understand the behavior of each type of probe, where CMM probes are classified into two main categories; contact (tactile) probes and non-contact probes. As the name suggests, a contact probe gathers data by physically touching the artifact or specimen. CMM contacting probes are divided also into two specific families of hard probes and touch trigger or scanning probes, which maintain contact with the specimen surface during data collection [10]-[12]. The probing system in CMM machines includes stylus and stylus tip, which have their own dynamic characteristics during measuring processes [7], [13]. 1.1. CMM hard probes To use a hard probe, the CMM operator manually brings the probe into contact with the specimen, allows the machine to settle and manually signals of the CMM to record the probe position. The CMMs software treats the readings to compensate for the diameter of the probe stylus tip. CMM hard probes are available in a variety of configurations and continue to have a broad application in coordinate metrology. When used in conjunction with manual CMMs, they are most frequently used to measure curved surfaces, distances between specimen features, angles and the diameter and centerline location of bores in applications that require low to medium accuracy. Hard probes are simple in use and rugged also, but their repeatability quality depends upon his operator touch. Because every operator has a different touch when moving and bringing the probe into contact with the specimen, therefore this hard type probe is not commonly used in large mass production companies. 1.2. CMM touch trigger probe Recently, the touch trigger (scanning) probe is the most common type of probes used in CMM. Ductile trigger probes are precision-built, touch-sensitive devices that generate an electronic digital signal each time the probe contacts a point on the specimen surface, which is usually indicated by an LED and an audible signal. The probe head itself is mounted at the end of one of the CMM's moving axes. It can be rotated automatically, and can accommodate many different probe stylus tips and attachments. These features make the CMM trigger probe a versatile and flexible data-gathering device. CMM touch trigger probes eliminate influence of operator touch on measuring results compared to hard probe type. It can be fitted on direct computer numerical control (CNC-CMMs) and manual CMMs [10]-[11]. An improvement on the basic touch trigger probe design incorporates piezobased sensors to translate the deflection of the probe Articles

3


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

into a constant digital acoustic signal that is recorded by the CMM. This design improves the accuracy of measurements due to the elimination of the effect of stylus bending (caused by force variations when the touch trigger probe contacts the specimen) and inaccuracies caused by the probe's internal electromechanical parts. While, the last element in the probe styles is the tip. In practice, the part measured waviness deviate from the desired value owing to many quasi-static systematic errors as inherited intrinsic are geometric error of measuring probe tip, thermally induced distortions of machine probe elements, errors arising from the static deflection or stiffness of machine-fixture-specimen-probe system under the touch force and other errors [1], [9]. Measurement accuracy is commonly determined by the kinematic accuracy of the CMM probe and a big portion of machines used with low kinematic accuracy. Software-based error compensation is a method for anticipating the combined effect of all of these above factors on standard precise and accurate spherical artifact and suitably modifying the conventionally designed probe tip scanning trajectory. Considerable research works have been reported to improve the kinematic accuracy of the CMM machine probe, which is too sophisticated to implement, there are few programs that focus on changing the CNC program to compensate the probe error [6], [14]. While, generating a cylindrical surface, its profile often concavely deviates from the ideal profile, which initiates the necessity of measuring the surface profile during the measuring process and suitable CMM strategy. Therefore, it is very difficult to separate and study these types of errors in practice clearly. In such cases, CMM is a reliable tool for verification, dimensional measurement and geometrical waviness form deviation accuracy for selected surface profile needs to be carried out theoretically. Probe stylus tip path is updated during measuring based on the parameters measured by the newly developed techniques. Thus, a high-quality inner cylindrical surface measurement has been successfully generated by process using software compensation. The geometrical form and undulation of spare parts for machinery and mechanical equipment is an important and active role in the technologies applications of industrial metrology. During the stages of assembly and operation of mechanical systems, the ruby ball tip of the stylus that used to measure deviations in geometrical feature always requires a thorough test to achieve high measurement accuracy, specially when measure the deviations of difficult forms and its waviness in the three directions (X, Y, and Z). Recently, both theoretical analysis and experimental studies pointed at the triggering point being the main source of probe errors [1]-[6], [8], [12], [14-16]. But unfortunately, this area needs more dynamic analyses to understand stylus response error sources according to the design and construction of CMMs, especially new CMM machines [17]. The error caused by probe loping has become a significant component of the total system errors. However, most of the studies on scanning CMMs cannot separate the performance of the probing system from other error sources of the CMM [1], [8]. Since CMM trigger probes are precision equipments themselves, their performance should be studied separately from the rest of 4

Articles

N째 2

2010

the components of the CMM in order to characterize its behavior to improve the measurement accuracy. This principle of operation effectively triggers the probe at a constant force regardless the contact area between the probe stylus tip and the measured specimen.

2. Mathematical Model Since the influence of some unforeseeable factors affecting probe inaccuracy could be small, so it requires accurate analytical model during analysis. Thus, for this investigation a new two-dimensional-model (2DM) has been used to present the root error due to ball tip size of the CMM probe stylus at measurement operations. 2.1. Stylus ball tip error During scanning all CMM touch probes in the coordinate measurement have a natural ball tip errors [14]. Supposing 2DM, where stylus ball is steady placed in horizontal position, thus only X-axis and Y-axis translation movement of the stylus is possible. Assuming no stylus tip ball deformation and no surface deformation under the test, following 2DM model can be presented in Figs. 1 and 2. Figure 1 shows the measurement principle of the proposed system which include contact points 1, 2 and 3 that are indicated on the vertical and horizontal plans of the probe stylus tip with l stylus length and ball tip radius of R with predicate angle q. Figure 2 presents that due to the finite size of the probe stylus ball tip, the contact point on a cylindrical surface will be along the stylus axis, but relatively at some point on the side of the ball where the test surface and the stylus tip ball slope angle match horizontally.

Fig. 1. Horizontal placed probe stylus ball tip radius (R). Because of the ball does not touch the test artifact specimen along the same stylus slope angle, there will be an error E in the measured length for any measurement point where the test part surface slope angle (q) is not zero degree. The error E and different possible positions of the ball tip are shown. Case d is at a higher surface slope (180째) and thus it has a larger measurement error E, while case a, is located at a lower surface slope and it has a smaller measurement error E. Therefore, to get exact location of n point on slope surface, an error E, reduced by value DY is made, due to fact, that position of point t on stylus tip ball every time is captured. From figure 2, values DY and E can be expressed as follows:


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 2

2010

Fig. 2. Size of error E related to the probe tip ball radius (R) and surface slope degree (q). DY = R - r

(1)

DY = R - (R cos q) = R (1 - cos q)

(2)

Where the distance between two points t and n indicates the E in Y direction, while DY is the relative distance between points m and t in the Y-direction, the large scale of the tip ball in case b. From appointed 2DM can be stated that measurement error E and DY values are made only in Y-axis and is dependant on surface slope. In point 2 or 3 (according to figure 1, where is matching angle q=180°) DY and error E would be maximal values, while error E and DY are equals zero only at scanning flat surfaces (q=0, 360°) when all points m, n and t are overlaying each other (m=n=t), as shown in case a.

3. Results and Discussions The root error due to CMM stylus ball tip cannot be neglected. Six different stylus balls with radii R=4.0, 3.0, 2.5, 2.0, 1.5 and 1.0 mm have been selected according to actually common using. A relative mutation in the Y-direction (DY) of the stylus ball can be observed according to the surface slope degree that called matching angle (q), Fig. 3. Through the application of accurate analytical 2DM

proposed in the measurement of cylindrical parts using CMMs, emerged two types of systematic unforeseeable errors. The first error resulted when increasing surface slope angle while the second error resulted when increasing the radius of the stylus tip ball. Fig. 3 shows how margin of errors are calculated theoretically and the output of climate surface slope degree (q) of the probe tip during contact through 360° at using complete cylindrical reference artifact for different six tips radii. It is observed that the amount of errors in the Y-direction starting from zero for each probe tip at the beginning of contact at 0°, while increasing incremental increase of inclination angle of a point contact of the tip with the artifact to reach its maximum relative value of 2R% at 180° and then come back to the decline to reach zero at 360°. It means that rotational motion that occurs during the probe tip scanning due to creeping of the tip at the base of the probe vibrates at the surface coming into contact with the cylindrical parts are also generate another error regularly. Hence, it can be concluded from this mathematicaltwo-dimensional model is capable to appear two different systematic errors with the movement of the probe during CMM scanning. The first is consequence of the creeping tip while the second is a result of increasing the radius of the tip became increasingly error rate to a maximum value

Fig. 3. Relative error of the probe ball tip at different surface slope angle. Articles

5


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

of 2R% during the measurement at the point of orthogonal (180°) of the cylindrical artifact. The maximum relative amount of these root errors are ranged from 8, 6, 5, 4, 3 to 2% at the same matching point (180°) for the ball tip radius of 4.0, 3.0, 2.5, 2.0, 1.5 and 1.0mm in respectively as shown in Fig. 3. In other words, figure 4 can help to conclude that; small probe tip of 1.0mm can be better used to diagnose the true state of the surface form of the specimens than that with bigger tip radius of 1.5, 2.0, 2.5, 3.0 and 4.0mm in reactively. This is because the probe tips of the large radii owing touch large contact area with the inner surface of the used standard artifact, and vice versa. In this case, the distortion of the measurement result using 1.0 mm probe tip become more visible and gives better estimate of the measured feature profile compared to the results of 4.0 mm probe tip.

AUTHOR Salah H. R. Ali - Engineering and Surface Metrology, Length and Precision Engineering Division, National Institute for Standards (NIS), Giza (12211), PO Box 136, Egypt. Mobile: 0020-126252561, E-mail: Dr_Salah@nis.sci.eg or SalahAli20@yahoo.com.

References [1]

[3]

[4]

[5]

6

Fig. 4. Scheme of the probe tips scanning path during measuring process.

[6]

4. Conclusion

[7]

Articles

2010

correction technique for probe performance accuracy of CMMs measurements can be developed, which is can be built upon both of surface form and probe stylus characteristics experimentally.

[2]

The application of the proposed accurate analytical two-dimensional-model (2DM) measurement technique can be used for its capability to present two types of systematic unforeseeable errors of the probe during CMM scanning process. The first error can be due to surface slope degree of the probe tip during contact with the detected surface of the cylindrical artifact in the Y-direction, with zero error value where the tip begin to start rotation and reach its maximum error value of 2R% at the point of orthogonal axis at 180°, then returns again to zero error at 360°. This error always occurs for the ball tip rotation during the scanning process as a result of creeping of the probe tip during touching the measured surface. The second error resulted at increased radius of the probe stylus tip ball. From carried out results analysis, are can conclude the following: • Increasing the probe tip radius decreases the averaged measured error signals of surface waviness; it may be due to large number of contact points of small tip on the artifact during scanning trajectory. It has been cleared that the probe stylus tip at scanning have a significant influence in the accuracy of CMM measurements using the strategy of touch trigger probe independently. • From results obtained, an easy calibration and

N° 2

[8]

[9]

[10]

[11] [12]

[13]

Hermann G., Geometric error correction in coordinate measurement, Acta Polytechnica Hungarica, vol. 4, no. 1, 2007, pp. 47-62. Krajewski G., Woźniak W., “One dimensional kinetic model of CMM passive scanning probes”, Journal of Automation, Mobile Robotics & Intelligent Systems, vol. 3, no. 4, 2009, pp. 172-174. Jae-jun Park, Kihwan Kwon, Nahmgyoo Cho, Development of a Coordinate measuring machine (CMM) touch probe using a multi-axis force sensor. Meas. Sci. Technology, vol. 17, 2006, pp. 2380-2386. Woźniak A., Dobosz M., “Influence of measured objects parameters on CMM touch trigger probe accuracy of probing”, Precision Engineering, Elsevier Inc., vol. 29, issue 3, 2005, pp. 290-297. Kasparaitis A., Sukys A., Dynamic errors of CMM probes, diffusion and defect data. solid state data. Part B, ISSN 1012-0394, vol. 113, 2006, pp. 477-482. Wu Y., Liu S., Zhang G., “Improvement of coordinate measuring machine probing accessibility”, Precision Engineering, vol. 28, 2004, pp.89-94. Yagüe J.-A., Albajez J.-A., Velázquez J., Aguilar J.-J., “A new out-of-machine calibration technique for passive contact analog probes“, Measurement, Elsevier Ltd., vol. 42, 2009, pp. 346-357. Woźniak A., Mayer J. R. R., Bałaziński M., “Stylus tip envelop method: corrected measured point determination in high definition coordinate metrology”, Int. Journal of Adv. Manuf. Technol., Springer, vol. 42, 2009, pp. 505-514. Ali S.H.R., “The Influence of fitting algorithm and scanning speed on roundness error for 50 mm standard ring measurement using CMM”, Int. Journal of Metrology & Measurement Systems, Polish Academy of Sciences, Warsaw, Poland, vol. 15, 2008, no. 1, pp. 31-53. Genest D. H., The right probe system adds versatility to CMMs, Available at: http://www.qualitydigest.com/jan97/probes.html Zeiss Calypso Navigator, CMM operation instructions and training manual. Revision 4.0, Germany, 2004. Dobosz M., Woźniak A., “CMM touch trigger probes testing using a reference axis”, Precision Engineering, Elsevier Inc, vol. 29, 2005, issue 3, pp. 281-289. Yagüe J.-A., Albajez J.-A., Velázquez J., Aguilar J.-J., A new out-of-machine calibration technique for passive contact analog probes", Measurement, Elsevier Ltd., vol. 42, 2009, pp. 346-357.


Journal of Automation, Mobile Robotics & Intelligent Systems

[14]

[15]

[16]

[17]

VOLUME 4,

N° 2

2010

Lin Y. C., Sun W. I., “Probe radius compensated by the multi-cross product method in free form surface measurement with touch trigger probe CMM”, Int. Journal of Adv. Manuf. Technol., Springer, vol. 21, 2003, pp. 902–909. Li L., Jung J.-Y., Lee Ch.-M., Chung W.-J., “Compensation of probe radius in measuring free-formed curves and surface“, Int. Journal of the Korean Society of Precision Engineering, Springer, vol. 4, no. 3, 2003. Xiong Z., Li Z., “Probe radius compensation of workpiece localization”, ASME Transactions, vol. 125, February 2003, pp. 100-104. Zhao J., Fei Y. T., Chen X. H., Wang H. T., “Research on high-speed measurement accuracy of coordinate measuring machines“, Journal of Physics: 7th Int. Symposium on Measurement Technology and Intelligent Instruments, Conf. series 13, 2005, pp. 167-170.

Articles

7


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 2

2010

NEW APPROACH TO THE ACCURACY DESCRIPTION OF UNBALANCED BRIDGE CIRCUITS WITH THE EXAMPLE OF Pt SENSOR RESISTANCE BRIDGES Received 28th April 2009; accepted 24th September 2009.

Zygmunt Lech Warsza

Abstract: After short introduction transfer coefficients of the unloaded four arms bridge of arbitrary variable arm resistances, supplied by current or voltage source, are given in Table 1. Their error propagation formulas are find and two rationalized forms of accuracy measures, i.e. related to the initial bridge sensitivities and of double component form as sum of zero error and increment error of the bridge transfer coefficients are introduced. Both forms of transfer coefficient measures of commonly used bridge - of similar initial arm resistances in balance and different variants of their jointed increments, are given in Table 3. As the example limited errors of some resistance bridges with platinum Pt100 industrial sensors of class A and B are calculated Table 4 and analyzed. Presented approach is discussed and found as the universal solution for all bridges and also for any other circuits used for parametric sensors. Keywords: resistance bridge, sensor, measures of accuracy, error, uncertainty of measurements.

± ¥. In [1], [6] this obstacle was bypassed by relating the absolute value of any bridge accuracy measure to the initial sensitivity of the current to voltage or voltage-tovoltage bridge transfer function. These sensitivities are valuable reference parameters, as they do not change within the range of the bridge imbalance. In paper [7] the new double component approach to describing the bridge accuracy is developed. It has the form of sum of initial stage and of bridge imbalance accuracy measures. Such double component method of describing accuracy is commonly used for the broad range instruments, e.g. digital voltmeters. Relation of each component to accuracy measures of all bridge arm resistances have been developed. As the example formulas of accuracy measures of two bridges used for industrial Pt sensors will be presented and their limited errors calculated.

2. Basic formulas of bridge transfer functions Four resistance (4R) bridge circuit with terminals ABCD working as passive X type twoport of variable internal resistances is shown in Fig 1.

1. Introduction This paper is based on earlier author proposals given in papers [1], [6] - [8]. As it has been pointed there, the generalized accuracy description of the 4R bridge of arbitrary variable arm resistances was not existing in the literature. Some considerations of the bridge accuracy with sensors of very small increments only have been found in [9], [10]. Generalized description is urgently needed mainly for: - initial conditioning circuits of analogue signals from broadly variable immittance sensor sets, - identification of the changes of several internal parameters of the equivalent circuit of the object working as twoport X, when it is measured from its terminals for testing, monitoring and diagnostic purposes. Near the bridge balance state, application of relative errors or uncertainties is useless, as they are rising to

Fig. 1. Four arms bridge as the unloaded twoport of type X with the voltage or current supply source branch. In measurements the ideal supply of the bridge by current IAB ® J = const., if RG ® ¥ or by voltage UAB = E=const.; when RG=0; is commonly used. Also ¥ the output is unloaded, i.e.: RL ® ¥, U'DC ® UDC . For single variable measurements it is enough to know changes of one bridge terminal parameter and the output circuit voltage UDC is mostly used. With notations of Fig. ¥ 1 formulas (1), (2) of UDC are given in Table 1 [2],

Table 1. Open circuit bridge voltage and its transfer functions. a) current supply

8

Articles

b) voltage supply


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

where: IAB, UAB - current or voltage on bridge supply terminals A B, Ri º Ri0 + DRi º Ri0 (1+åi) - arm resistance of initial value Ri0 and absolute DRi and relative åi increments, r21, k21 - transfer functions of the bridge of opencircuited output, i.e.: current to voltage (transfer resistance) and ratio of two voltages. t0 =

R10R30 R10R30 , k0 = R ( R + R å i0 10 20)(R30+R40)

- initial bridge open circuit sensitivities of r21 and of k21, 4

åRi º i=1 åRi =(1+åÓR) åRi0

- sum of bridge

arm resistances; åÓR(åi), åRi0 - its increment and initial value, f(åi), fE(åi) - normalized bridge imbalance function of r21 and of k21, DL(åi), åÓR(åi) - increments of the function f(åi) numerator and of the denominator. Output voltage UDC may change its sign for some set of arm resistances. If transfer function r21= 0 or k21= 0, the bridge is in balance and in (3) and (4) its conditions of both supply cases are the same: R1R3=R2R4. The balance of the bridge can occurs for many different combinations of Ri, but the basic balance state is defined for all åi= 0, i.e. when: R10R30=R20R40

(5)

Formulas of the bridge terminal parameters are simplified by referencing all resistances to their initial valu-

N° 2

2010

es in the balance, i.e. Ri=Ri0(1+åi) and referencing initial resistances Ri0 to one of the first arm, i.e.: R20ºmR10, R40ºnR10 and from (3) R30=mnR10. Bridge transfer functions can also be normalized, as is shown in Table 1.

3. Accuracy description of broadly variable resistances The accuracy of measurements depends in complicated way on structure of the instrumentation circuit, values and accuracy of its elements and on various environmental influences of natural conditions and of the neighboring equipment. Two type of problems have been met in practice: - description of circuits and measurement equipment by instantaneous and limited values of systematic and random errors, absolute or related ones, as well by statistical measures of that errors, - estimation of the measurement result uncertainty, mainly by methods recommended by guide GUM. Measures of accuracy (errors, uncertainty) of the single value of circuit parameter are expressed by numbers, of variable parameter - by functions of its values. In both cases they depend on equivalent scheme of the circuit, on environmental and parameters of instrumentation used or have to be use in the experiment. The measures of broadly variable resistance Ri, e.g. of the stress or of the temperature sensor, could be expressed by two components: for its initial value and for its increment as it is shown by formulas of Table 2. Instantaneous absolute error Di and its relative value äRi referenced to Ri are given by formulas (6) and (7), relative limited error | äRi | of the poorest case of values and signs of | äi0 | and | Dåi | or | äåi | - by (8), and standard statistical measure äRi for random errors or uncertainties - by (9).

Table 2. Two-component formulas of the sensor resistance accuracy measures.

Articles

9


Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 4,

N° 2

2010

If errors of increment and of initial value of resistance are statistically independent then correlation coefficient ki = 0, but if they are strictly related each to the other then ki = ± 1. Exact ki value can only be find experimentally. From (8) follows that borders of the worse cases ±| äRi | of possible values of äRi nonlinearly dependent on åi even if | äi0 | and | Dåi | or | äåi | are constant [1], [6] - [8]. Distribution of the initial values and relative increments åi of the sensors' set resistances depends on their data obtained in the production process. Its actual values also depend on influences of the environmental conditions.

4. Description of the accuracy of bridge transfer functions Instantaneous values of measurement errors of bridge transfer functions r21 and k21 result from the total differential of analytical equations (3) and (4) from Table 1. After ordering all components of äRi absolute error of transfer function r21 is: (10) - weight coefficients of error äRi components

where:

- subscript i = 1,2,3,4 when j=3,4,1,2; - multiplier (–1)i+1=+1 if i is 1, 3 or –1 if i is 2, 4. If resistances are expressed as Ri = Ri0(1+åi), Rj = Rj0 (1+åj) formula (10) is (10a)

Absolute error of transfer function k21 has other forms, i.e.: (11) or

(11a) From (10) and (11) one could see that if errors äRi of the neighbouring bridge arms, e.g.: 1, 2 or 1, 4 have the same sign they partly compensate each other. If errors äRi of resistances Ri are expressed, as in (7), by their initial errors äi0 and incremental errors äiå, then (12)

where:

(12a)

If arm resistance Ri is constant, åi = 0, äRi = äi0, but weight coefficient wRi of its component in D r21 still depends on other arm increments åj. In initial balance state, i.e. when all arm increments åi = 0, the nominal transfer function r21(0) º r210 = 0, but real resistances Ri have some initial errors äi0 and usually D r210 ¹ 0, D k210 ¹ 0. Ä r210 = t0 ä210

(13a)

D k210 = k0 ä210

(13b)

where: ä210 = ä10 - ä20 + ä30 - ä40 Relative errors are preferable in measurement practice, but it is not possible to use them for transfer functions near the bridge balance as the ratio of absolute error D r21 ® D r210 ¹ 0 and the nominal r21® r210 = 0 (or for the voltage supplied bridge of D k21 and k21® k210 = 0) is rising to ± ¥. Then other possibilities should be applied. There are two possible ways to describe accuracy of the bridge transfer function r21 (or k21) in the form of one or of two related components: - absolute error of the bridge transfer function may be referenced to initial sensitivity factor t0 of r21 (or to k0 of k21) or to the range of transfer function r21max - r21min (or k21max - k21min); 10

Articles


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

-

initial error D r210 have to be subtracted from D r21 and then accuracy could be described by two separate terms: for zero and for transfer function increment, as it is common for digital instrumentation.

In the first type method it is preferable if error D r21 is referenced to the initial sensitivity factor t0 as constant for each bridge, then to full range of r21 as it could be change. Such relative error är21 could be presented as sum: (14) where:

ä210 = ä10 - ä210 + ä30 - ä40 - initial (or zero) relative error of r21=0; är21å (åi) relative error of normalized imbalance function f(åi) when r21¹0, also referenced to t0.

Error ä210 is similar for any mode of the supply source equivalent circuit of the bridge as twoport. Zero of the bridge may be corrected on different ways: by adjustment of the bridge resistances, by the opposite voltage on output or by the digital correction of converted output signal. In such cases from (14) it is (15) From (15) follows that related to t0 error är21å of r21 increment depends not only on increment errors äei of resistances Ri but also on their initial errors äi0 ¹ 0 even when initial error of the whole bridge d210=0, because after (12a) weight coefficients of äi0 in (15) depends on åi. The component of particular error äi0 despairs only when äi0=0. Functions of Dåi or äei may be approximated for some åi intervals by constant values. In the second type method absolute error of transfer function r21 (12) after subtracting its initial value is (16) And after referenced it to r21, and substitution wRi from (12a)

(17) where :

(17a,b)

Weight coefficients (17a,b) are finite for any value of r21 including r21= 0 because if all åi ® 0 also DL ® 0. Error är21r is equivalent to error äåi of the resistance Ri increment åi in formula (7). From ( 13a) and (17 ) is: for current to voltage transfer function Dr21= t0 ä210 + r21 är21r (18a) and similarly for voltage transfer function k21 Dk21= k0 ä210 + k21 äk21k (18b)

where:

N° 2

2010

t0 ä210 = Dr210, k0 ä210 = Dk210 - absolute errors of initial value r21 or k21 e.g. r210= 0 or k210= 0, är21r , äk21k - related errors of increments r21 - r210 or k21 - k210 from the initial stage.

Two component accuracy equation of k21 transfer function was funded by the same way as for r21. Actual values of instantaneous errors of r21 or k21 could be calculated only if signs and values of errors of all resistances are known. In reality it happens very rare. More frequently are used their limited systematic errors (of the worst case) and statistical standard deviation measures. Formulas of these accuracy measures of r21 or k21 could be obtained after transformation of error formulas (10) (18a,b). All these accuracy measures are possible to find in one component or two component forms. One-component formulas for arbitrary and main particular cases of 4R bridge are given in Tables in [1], [6], [7]. The two-component method of the bridge transfer function r21 accuracy representation, separately for its initial value (e.g. equal to zero) and for increment is similar like unified one used for digital instruments and of the broad range sensor transmitters. It is especially valuable if zero of the measurement track is set handily or automatically. Absolute measures could be transformed also by the sensor set linear or nonlinear function to the units of any particular measurand, e.g. in the case of platinum sensors - to °C [6], [8].

5. Description of accuracy measures of particular 4R bridges mostly used for sensors In the measurement practice the mostly used for sensors are four-arm bridges of resistances equal in the balance state. Formulas of accuracy measures for transfer functions r21 and k21 of these particular resistance bridges are much simpler then general ones [6]. They are presented in Table 3 with assumption that all correlation coefficients kij= 0. Formulas of k21 and its errors given are mainly for comparison as current supply is preferable one for resistance sensors. The transfer function r21 or k21 of the bridge including differential sensors of four or two increments ± å and transfer function r21 of two equal å in opposite arms may be linear but their measures depend on å by different way each other. From Table 3 it is possible to compare the accuracy measure formulas of the current or voltage supplied 4R bridges of few element and single element sensors. For example formulas of accuracy measures of similarly variable two opposite arm resistances are simpler then for the single variable arm. Form of the limited error | är21r | is similar for four linear A-D bridges and |äk21k| only for two of them: A and B.

6. Tolerances of industrial Pt sensors The good example of the broadly variable resistance sensors is platinum sensors Pt 100 of A and B classes commonly used in industrial temperature measurements. Tolerated differences from their nominal characteristic are given in standard EN 60751÷A2 1997. They are expressed in °C or as permissible resistance values in [ohm] - see |D| of classes A and B in Fig. 2. Characteristic of class A sensors is determined up to 650°C and for less accurate class Articles

11


Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 4,

N째 2

Table 3. Accuracy measures of the most common open-circuit 4R bridges of similar all arm initial resistances 4R10.

12

Articles

2010


Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 4,

N° 2

2010

B - up to 850°C. Initial limited errors |ä10| of both classes are 0,06% and 0,12%, respectively. On the base of nominal characteristic of Pt 100 sensors the maximum limited error |äR1|max º |ä|=|ä10|+|äåi| for 宥 of both classes is calculated as ratio of tolerances |D| and increments of sensor resistance [7], i.e. as

Obtained values are given in Fig. 2. They are only slightly changing and could be approximated by the single value and related to the maximum or mean value of the temperature range of each sensor. In the full range of positive Celsius temperatures the limited error | ä | doesn't exceed 0,2% of å for class A and | ä | £ 0,5% for class B.

7. Limited errors of the 4R bridge with single industrial Pt 100 sensors

Limited errors |är21|, |är21å| = |är21 - ä210|, |är21r| of the 4R bridge transfer function r21 with the single industrial sensor of A or B class has be calculated from formulas of Table 3. It was assumed that limited errors |äi0| of

Fig. 2. Tolerances | D | and maximum limited relative errors | ä | of temperature sensors Pt100 type A and B evaluated from their standard [7]. constant bridge arms are equal and not higher that the sensor initial error |ä10|, balance is at 0°C, and current of supply source is stable enough or ratio of output signal and this current is measured. Maximum temperature range (0 - 600)°C is taken for calculations and for it the relative

Table 4. Limited errors of few cases of the current supplied 4R bridge with the single resistor sensor, e.g. Pt 100 type A or B.

Articles

13


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

increment of sensor resistance is: åmax=2,137. As example numerical formulas of limited errors |är21| or |äk21| of the class A are also estimated. Limited errors of the class B sensor bridge have been similarly estimate. For clarifying considerations the lead resistances are taken as negligible. All results are presented in Table 4. In Table 4 five different cases of measuring circuit are considered, i.e.: bridge without any adjustments, outer and internal zero setting, negligible initial errors only of constant arms or of the sensor arm as well. Ratio of limited errors of the bridge without adjustments and Pt sensor is 1,7 for class A and 2,9 for class B. If errors of the bridge resistances are negligible (line 5) limited error is only slightly higher then for the sensor, but if also the initial sensors error is adjusted, then the bridge transfer function r21 error is even smaller then of the sensor itself ! (line 1). Results for examples 2, 3 and 4 are between 1 and 6. For comparison: relative limited error of the output voltage of the bridge including two similar Pt 100 sensors of the class A in opposite arms, calculated from line 3 of Table 3 for the same temperature range doesn't exceed 0,51% - without null correction, and 0,39% - if is corrected. These errors are calculated for the twice higher signal that for the single sensor bridge. They are slightly higher but signal linearly depends on resistance increment equal for both sensors. For lower temperature ranges relative limited errors and uncertainties type B become higher.

14

Articles

2010

impedance sensor circuits as DC and AC bridges of single and double supply, active bridges linearized by feedback or multipliers, Anderson loop (developed in NASA) and impedance converters with virtual DSP processing. Given in this paper methods of the simplification of their accuracy description could be also applied in many industrial measurements. Accuracy of current and voltage supplied strain bridges has been analyzed by M. Kreuzer [9], [10], but such a unified approach as given above to the accuracy description of unbalanced bridges and other circuits of broadly variable parameters, developed in [1], [4], [6] [8], is not found so far in literature. The presented method is also valuable for accuracy evaluation in testing any circuit from its terminals as twoport, which is commonly used in diagnostics and in impedance tomography. It was also used to describe the accuracy of two-parameter bridge measurements - see [3]-[5].

AUTHOR Zygmunt Lech Warsza* - Polish Metrological Society, Warsaw, Poland. E-mail: zlw@op.pl.

References [1]

8. General conclusions Two methods of describing the accuracy measures of the arbitrary imbalanced sensor bridges are presented together and compared., i.e.: - one component accuracy measure related to initial sensitivity of the bridge transfer functions, given before in [1], [6], [7] and - the new double component one of separately defined measures for zero and transfer function increment [8]. The second one is similar as used for the broad range instruments, e.g. digital voltmeters. Accuracy measures of bridge arms are defined for initial resistances and for their increments. Then these methods are independent from the sensor characteristic to the measured quantity. These methods are discussed using on few examples of 4R bridges of equal initial resistances, supplied by current or voltage source and with single, double and four element sensors. Given formulas allow to finding accuracy of the 4R bridge or uncertainty of measurements with bridge circuits if actual or limited values of errors or standard statistical measure of their resistances and sensors are known. Formulas of general and particular cases of the bridge may be used for computer simulation of the accuracy of various sensor bridges and measured objects of the X twoport equivalent circuit in different circumstances. Systematic errors could be calculated as random ones for set of sensor bridges in production or in exploitation and if correlation coefficients are small obtained values should be smaller than of the worst-case limited errors. Similar formulas as presented in this and other papers, e.g. [1], [4], [6], [7] could be formulated for any types of

N° 2

[2]

[3]

[4]

[5]

[6]

[7]

Warsza Z. L., “Four-Terminal (4T) Immittance Circuits in Multivariable Measurements” (Immitancyjne układy czterobiegunowe (4T) w pomiarach wieloparametrowych), Monograph of Industrial Institute of Control and Measurement PIAP, Warszawa 2004 (in Polish). Sydenham P.H., Thorn R. (eds.), Handbook of Measuring System Design, Chapters: 126: “Electrical Bridge Circuits Basic Information” and 127 “Unbalanced DC bridges” by Warsza Z. L., John Wiley & Sons, Ltd. New York, 2005, pp. 867-889. Warsza Z. L., “Two Parameter (2D) Measurements in Four Terminal (4T) Impedance Bridges as the New Tool for Signal Conditioning”. Part 1 and 2. In: Proc. of the 14th International Symposium on New Technologies in Measurement and Instrumentation and 10th Workshop on ADC Modeling and Testing IMEKO TC-4, Gdynia/Jurata, 2005, pp. 31-42. Warsza Z. L., “Two Parameter (2D) Measurements in Double-current Supply Four-terminal Resistance Circuits”, Metrology and Measurement Systems, vol. 13, no 1, 2006, pp. 49-65. Warsza Z. L., “Backgrounds of two variable (2D) measurements of resistance increments by bridge cascade circuit”. In: Proc. of SPIE Photonic Applications..., Wilga Symposium, Poland, vol. 6347, ed. Romaniuk R. S., 2006, pp. 634722R-1-10. Warsza Z.L., “Accuracy Measures of the Four Arm Bridge of Broadly Variable Resistances, Parts 1 and 2”. In: Proceedings of 15th IMEKO TC4 International Symposium on Novelties in Electrical Measurement and Instrumentations, TU Iasi , Romania, vol. 1, Sept. 2007, pp. 17-28. Warsza Z.L., “Miary dokładności transmitancji mostka rezystancyjnego w przypadkach szczególnych” (Accuracy measures of transmittance of variable resistance bridge in particular cases), Pomiary Automatyka Kontrola, no. 10, 2007, pp. 17 -24 (in Polish).


Journal of Automation, Mobile Robotics & Intelligent Systems

[8]

[9]

[10]

VOLUME 4,

N° 2

2010

Warsza Z.L., „Nowe ujęcie opisu dokładności mostka z przemysłowymi czujnikami Pt” (New Approach of the Accuracy of Industrial Pt Sensor's Bridge), Podstawowe Problemy Metrologii PPM'08, Prace Komisji Metrologii Oddziału PAN w Katowicach, Konferencje, no. 8, pp. 155-164 (in Polish). Kreuzer M., “Linearity and Sensitivity Error in the Use of Single Strain Gages with Voltage-Fed and Current-Fed Circuits”, Experimental Techniques, vol. 8, October 1984, pp. 30-35. Kreuzer M., “Whestone Bridge Circuits shows almost no Nonlinearity and Sensitivity Errors when used for Single Strain Gage Measurements”, Technical literature of Hottinger Baldwin Messtechnik, in Internet: ta_sg_kreuzer_06 2004.

Articles

15


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

Nツー 2

2010

DATA FUSION IN MEASUREMENTS OF ANGULAR POSITION Received 24th September 2009; accepted 3rd December 2009.

Sergiusz ナ「czak

Abstract: The paper describes a possibility of increasing accuracy of measurements of angular position with application of data fusion. Two cases of determining the angular position are considered: accelerometer-based measurements of tilt and measurements of angular position with application of incremental rotation-to-pulse sensor (coupled with an original measuring system described in the paper). Application of the proposed data fusion ensures in both cases decrease of the uncertainty of the related measurement of ca. 40%.

gx, gy, gz - components of the gravitational acceleration indicated by accelerometers with sensitive axes arranged as the Cartesian axes x, y, z, gxz, gyz - geometric sums of pairs of respective component accelerations, g - gravitational acceleration, j - arbitrarily oriented tilt angle, a - pitch, b - roll.

Keywords: data fusion, rotation-to-pulse angle sensor, tilt.

1. Introduction The paper refers to two specific cases of measuring angular position. The first is tilt sensing realized by means of an accelerometer (what recently has become very popular due to application of inexpensive MEMS acceleration sensors), while the second is an angle measurement by means of an incremental rotation-to-pulse sensor. In both cases we usually encounter a redundancy of the measurement data, which can be used for increasing accuracy of the considered measurements employing appropriate data fusion. According to one of many definitions, data fusion is a way of integrating into one whole data generated by various sources, or merging data generated by a single source, yet related to different features of an object or a phenomenon, and separated from a signal generated by a single sensor [1]. In the considered cases, we deal with a single sensor generating two or three output signals that can be merged in various ways. As for the dynamics of the considered measurements, accelerometers sensing tilt operate under static or quasistatic conditions, as far as the existing accelerations are concerned [2]. This is connected with low frequencies of performing such measurements, whereas in the case of incremental rotation-to-pulse sensors the conditions can be dynamic. However, because of the foreseen application, the proposed measuring system based on an incremental sensor is designed for operation at a very low rotational speed, what does not result in rigorous requirements with regard to operation speed of the electronic circuits as well as the software processing the output signals from the sensor.

2. Determining the tilt The case of determining a tilt angle with application of accelerometers is presented in Fig. 1, where: 16

Articles

Fig. 1. Components of the gravitational acceleration against the tilt angles. The basic dependencies used for calculating the angu-lar position represented by two component angles a and b can be found e.g. in [2]-[4]. However, application of these dependencies results in a low accuracy of deter-mining the tilt angles. An original method that makes it possible to determine with a higher accuracy an angular position (represented by the component angles a and b), employing appropriate data fusion, has been proposed in [5]. The idea of the measurement is based on computing these angles as a weighted average (having variable weight coefficients) of three different signals generated by the applied accelerometers: (1)

(2) where: w1, w2, w3, w4 - weight coefficients. In order to determine the weight coefficients, the following dependencies have been accepted that guarantee to obtain a minimal value of the combined standard uncertainty of the component tilt angles [5]:


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

w1 » cos2a w2 » sin2a w3 » cos2b w4 » sin2b

(3) (4) (5) (6)

The initial values of the component angles a and b that occur in the formulae (3) - (6) can be calculated in few ways: using the basic arc sine or arc cosine equations, which appear in formulae (1) - (2) (choice of appropriate equation is determined by value of angles a and b [2], [5], [6]) or using an arc tangent equation [4]. Additionally, an iterative method can be applied here, where in a successive step of computations the aforementioned weight coefficients are calculated once more, yet using values of angles a1 and b1 determined in the previous step according to formulae (1) - (2). In such case we employ recurrent formulae (7) - (8), and owing to that, the final values of angles a and b can be determined with a higher accuracy [6]:

N° 2

2010

the graduated wheel (same as in Fig. 2). For instance, a measuring system by Jenoptik Carl Zeiss with an incremental sensor IDW 2/16384 with 16,384 markers (what corresponds to a physical resolution of ca. 1.3 minute arc) that cooperates with a standard AE 101 counter unit, employs an 8-bit interpolation of the amplitude of the incremental signal, what makes it possible to obtain the accuracy and resolution of 1 or even 0.5 second arc (i.e. the measurement resolution is increased almost 160times) [8].

(7) Fig. 2. The measuring signals of the incremental sensor.

(8) However, in the case of the last method, when formulae (3) - (6) contain values of angles a and b determined in a previous iterative step according to (1) and (2), experimental studies indicated that no significant increase of the accuracy related to determination of tilt angles has been obtained in the following iterations. Additionally, application of the recurrent formulae may considerably limit dynamics of the sensor operation.

3. Measuring the angular position with incremental sensor Various types of incremental sensors are used for measuring all sorts of physical quantities. A principle of their operation is described in numerous works, e.g. in [7]. In order to determine an angular displacement of a graduated wheel, there are usually used two separate signals in space quadrature, generated by two detectors. The phase shift of p/2 results from the geometric configuration of the detectors and corresponds to a geometric angle of q/4 (where q is the pitch angle between adjacent markers on the wheel). Such solution makes it possible to distinguish the sense of rotation and to double the resolution of the sensor [7] (in the case when the two analogue output signals are directly converted into logical ones), which normally equals the number of markers created on the graduated wheel. Course of one period of the signals is depicted in Fig. 2 (assuming that their offsets are of 0, and their amplitudes of 1; electric phase angle g is expressed in degrees arc). High accuracy systems of this type, manufactured by such companies as Renishaw or Heidenhain, feature a subdivision resolution owing to interpolation of the analogue signals generated between adjacent markers on

While realizing the subdivision, one can alternately use the sine signal generated by the first detector within the range of the phase angle of -p/4¸p/4, and then the cosine signal generated by the second detector within the range of 1/4 p¸3/4 p (and respectively within the whole period), as it has been proposed in an analogous case of measuring tilt angles [2]. Such approach is also a kind of data fusion. Owing to this procedure, accuracy of determining the angular position within one period significantly increases, because the input signals are not used within the angular range where they feature a considerable nonlinearity, which results in a small sensitivity of the measurement [2]. The same result can be obtained while measuring a difference between instantaneous amplitude of the sine and cosine signal [7].

4. Data fusion in incremental measurements In order to obtain a further increase of the accuracy related to measurements of angular position, it has been decided to employ the same principle as in tilt measurements, described in section 2. It consists in a simultaneous, instead of an alternate as in the previous case, processing of both signals. Just as in the case of determining tilt angles, fusion of the information contained in both output signals is realized by calculating value of the phase angle g on the basis of a weighted average with variable weight coefficients. Analogously to formulae (1) - (2), this angle can be determined as follows: (9) where: w1, w2 - weight coefficients, U1 - analogue voltage signal from the first detector, U2 - analogue voltage signal from the second detector, U3 - amplitude of the signal from the first detector, U4 - amplitude of the signal from the second detector. Articles

17


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

Formula (7) is true when the weight coefficients meet the following self-evident condition: w1 + w2 = 1

(10)

Just as before, it is crucial to find such dependencies for determining the weight coefficients, being a function of phase angle g, that ensure obtaining a minimal value of the combined standard uncertainty of angle g, which can be expressed as [9]:

N° 2

2010

(16) Function (16) reaches its minimum, when its derivative with respect to w1 equals zero, i.e.:

(17) what boils down to the following equation: 2w1 - 2cos2 g = 0

(18)

Formula (18) yields dependencies identical compare to their counterparts (3) - (6) related to measurements of tilt angles: (11) where: u(U1) - uncertainty of determining the voltage signal from the first detector, u(U2) - uncertainty of determining the voltage signal from the second detector, u(w1) - uncertainty of determining the first weight coefficient, u(w2) - uncertainty of determining the second weight coefficient.

w1 Âť cos2 g w2 Âť sin2 g

(19) (20)

Courses of weight coefficients w1 (black curve) and w2 (gray curve) are presented in Fig. 3. The graphs indicate that for three characteristic values of phase angle g: 0, 1/4p and 1/2p, values of the coefficients will be respectively: (1, 0); (0.5, 0.5); (0, 1).

As it has been supported by an analysis of experimental results related to tilt measurements, in order to simplify formula (11) a certain assumption may be accepted while determining the respective partial derivatives of the weight coefficients. It is that the coefficients can be regarded as having a constant value, even though in fact they are functions of the measured angle. Under such assumption, combination of formula (9) with the simplified dependency (11) yields the following equation: Fig. 3. Graphs of formulae (19) and (20). (12) Still another simplification may be accepted here: uncertainties of determining the voltage signals from both detectors as well as their amplitudes are equal. It is fully justified in the cases when both detectors have an identical structure. So, let us assume: u(U1) = u(U2) = u(U)

(13)

U3 = U4 = U0

(14)

While expressing the voltages from the detectors as functions of phase angle g, formula (12) can be transformed as follows:

When the values of coefficients w1 and w2 are already known, we obtain a formula for determining the phase angle g: (21) Value of angle g determined in a previous step is to be calculated analogously as in the case of tilt angles (see section 2). Neglecting insignificant uncertainties of determining the weight coefficients, we can state that uncertainty of determining the phase angle g is constant and equals: (22)

5. Measuring System with Incremental Sensor (15) Taking into account formula (10), equation (15) can be rearranged as:

18

Articles

As it has been mentioned above, the proposed idea of data fusion has been already applied in the case of tilt measurements realized by means of systems based on MEMS accelerometers [6]. It is also planned to implement the data fusion in the case of using the aforementioned incremental sensor IDW 2/16384 manufactured by Zeiss.


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

Due to numerous shortcomings, the standard counter AE101 has been replaced with a custom built softwarecontrolled electronic system. Interpolation of the sine and cosine signal with subdivision accuracy is realized digitally with application of a commercial data acquisition card equipped with 16-bit A/D converters manufactured by Advantech [10] (cards with 12-bit A/D converters can be applied here as well). Theoretically, that would make it possible to increase accuracy of measuring the angular position few thousand times. It is obviously impossible because of various sources of errors that occur here. The most significant are the following: - noises of the output signals, - inaccuracies of the A/D converters of the data acquisition card, - inaccuracies of the markers on the graduated wheel, - error of the phase shift between the detectors, - variability of the phase shift during measurements, - different amplitudes of the output signals generated by the detectors, - variability of amplitudes of the output signals during measurements, - incorrect values of the offsets of the output signals due to alignment errors, - variability of the offsets of the output signals during measurements. As can be easily noted, it is a very significant matter to take into account the offsets, as well as amplitude magnitudes of the sine and cosine signal generated by the detectors in the incremental sensor (these values must be regarded while using formula (21)). Each sensor features an individual value of these parameters (in the case of the original counters AE 101, each of them was tuned to a single incremental sensor [8]). Therefore, it is necessary to determine the offsets and amplitude magnitudes during initialisation of the system or beforehand, for a specific sensor. Still another option, accepted in the considered measuring system, is measuring these parameters in real time and introducing possible corrections systematically. Such solution ensures obtaining the highest accuracy of the system, as it compensates for most of the errors listed above. Schematic of the described measuring system is presented in Fig. 4. As mentioned above, it features a subdivision accuracy due to a software interpolation of the measuring signals. A shortcoming of such system architecture is a significantly limited dynamics of its operation. In the case of measurements performed at high rotational speeds of the graduated wheel another kind of system should be built. It should be an autonomous microprocessor system processing the output signals generated by the rotation-topulse sensor in a real time, and transferring the computed results to a computer memory only from time to time in an off-line mode. It is also possible to eliminate the electronic interface while applying a data acquisition module featuring a high gain for the analogue inputs (such modules are designed for measurements of temperature by means of

N째 2

2010

thermocouples). However, such solution significantly limits the range of dynamic operation of the system.

Fig. 4. Schematic of the incremental measuring system.

6. Experimental Results As far as application of the considered data fusion is concerned, the experimental studies have been carried out so far only on a tilt sensor employing MEMS accelerometers. The following graph illustrates the obtained increase of the measurement accuracy.

Fig. 5. Decrease of error values owing to data fusion. The indication error expressed over axis y has been defined as an absolute value of the difference between the value of the tilt angle applied by means of the used test station and the value determined with respect to the average of respective indications of the tested sensor. The black curve is related to application of the weighted average, whereas the gray curve to application of the sine and cosine signal alternatively (as described in section 3). Increase of the measurement accuracy due to application of data fusion can be clearly observed.

7. Summary The author demonstrated in the paper an analogy between the case of measuring component tilt angles (pitch and roll) by means of accelerometers, and quite a different case of measuring angular position by means of incremental sensors with subdivision accuracy. Owing to application of the proposed data fusion, in both cases a significant decrease of the uncertainties of determiArticles

19


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

ning the angular values can be obtained (at least of 40 percent with respect to the basic measuring systems where the output signals are processed individually, using arc sine or arc cosine type of formulas). The same result can be obtained (yet in an easier way) in the case of determining tilt angles with application of an arc tangent function. A respective algorithm has been presented in [11]. The same applies to incremental sensors with subdivision accuracy, where a similar dependency can be used for calculating the phase angle realizing the subdivision [12]:

[2]

[3] [4]

[5]

(23) [6]

Experimental works carried out at the Faculty of Mechatronics, Warsaw University of Technology, proved that interpolation makes it practically possible to increase the sensor accuracy even 300 times (with respect to its physical resolution resulting from the number of the markers). However, it should be noted that some manufactures realize in their incremental measuring systems a 1024-fold division [7], or even a 4096-fold one [13]. Nevertheless, it should be emphasized that application of data fusion based on a weighted average, even though more complicated than application of the arc tangent function, has this advantage over the later that it is possible to regard different uncertainties related to both output signals. Then, formulae (7) - (8) as well as (21) change, and thus relation between the phase angle and the weight coefficients takes on a form other than (3) - (6) and (19) - (20). A case when values of uncertainties related to the output signals are different, concerns first of all triaxial MEMS accelerometers (often used in tilt measurements), since the signal related to their vertical sensitive axis is usually few times less accurate compare to the signals related to their horizontal sensitive axes. The author plans in the nearest future to carry out experimental works that are to determine how the accuracy of measuring tilt angle increases in the case of regarding the aforementioned difference and compensating for it. Application of the measurement principle based on the weighted average has yet another advantage over the arc tangent function. It yields a unique result in the case when the measured angle is of ±p/2, whereas while using the arc tangent function a respective angle value must be determined according to an arc cosine function, i.e. yet another data fusion must be applied.

AUTHOR Sergiusz Łuczak - Division of Design of Precision Devices, Institute of Micromechanics and Photonics, Faculty of Mechatronics, Warsaw University of Technology, ul. Boboli 8, 02-525 Warsaw, Poland; tel. (+48)(22) 2348315, fax (+48)(22) 2348601. E-mail: s.luczak@mchtr.pw.edu.pl.

References [1]

20

Sroka R., Metody fuzji danych w pomiarach parametrów ruchu drogowego, (Rozprawy Monografie, no. 182),

Articles

[7] [8]

[9] [10]

[11]

[12] [13]

N° 2

2010

Uczelniane Wydawnictwa Naukowo-Dydaktyczne AGH: Cracow, 2008 (in Polish). Łuczak S., Oleksiuk W., Bodnicki M., “Sensing Tilt with MEMS Accelerometers”, IEEE Sensors J., vol. 6, no. 6, 2006, pp. 1669-1675. “Tilt Sensing with Kionix MEMS Accelerometers”, AN 005, Kionix, Ithaca, NY 2005. Łuczak S., „Pomiary odchylenia od pionu przy użyciu akcelerometrów MEMS”, Pomiary Automatyka Robotyka, no. 7-8/2008, 2008, pp. 14-16 (in Polish). Łuczak S., Oleksiuk W., “Increasing Accuracy of Tilt Measurements”, Engineering Mechanics, vol. 14, no. 3, 2007, pp. 143-154. Łuczak S., „Algorytm wyznaczania odchylenia od pionu przy użyciu akcelerometrów MEMS”, [in] Proc. I Kongres Mechaniki Polskiej, Warsaw, 2007, p. 119 & CD-ROM (in Polish). Zakrzewski J., Czujniki i przetworniki pomiarowe, Wyd. Politechniki Śląskiej: Gliwice, 2004 (in Polish). IDW Incremental Translumination Angle-Measuring System, Jenoptik Carl Zeiss JENA GmbH, Berlin Leipzig, 1989. Wyrażanie niepewności pomiaru. Przewodnik, Główny Urząd Miar, Warsaw, 1999 (in Polish). “PCI-1716/1716L 16-bit, 250kS/s, High-Resolution Multifunction Card. Startup Manual”, Advantech Co., Ltd., Taipei, 2007. Łuczak S., “Advanced Algorithm for Measuring Tilt with MEMS Accelerometers” [in] R. Jabłoński et al., Recent Advances in Mechatronics, Springer-Verlag: Berlin Heidelberg, 2007, pp. 511-515. Chapman M., “Heterodyne and homodyne interferometry”, Renishaw, New Mills, 2002. “Angle Encoders”, Heidenhain, 2005.


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

SPECIAL ISSUE SECTION

Brain-like Computing and Applications Guest Editors: Tsutomu Miki Takeshi Yamakawa

N째 2

2010



Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 4,

N° 2

2010

Editorial Oscar Castillo*, Patricia Melin

Special issue section on Brain-like Computing and Applications It is our great pleasure to present our readers with this special issue of the JAMRIS (Journal of Automation, Mobile Robotics & Intelligent Systems) on Brain-like Computing and Applications. Today’s advanced information technology based on high-performance and inexpensive digital computers make our daily life convenient. Moreover it facilitates a high-speed and complicated calculation and handling of a huge amount of data. On the other hand, it can not perform tasks easy for a human being, such as, recognizing facial expressions, distinguishing an animal from a human, grasping a soft material, etc. Researchers are now facing to the limitations of system intellectualization. To cope with this limitation, engineering applications of crucial functions of the biological brain have drawn great attention of researchers and developers. Collaborations among different fields are necessary for accomplishing the break through. We promoted the brain-inspired information technology (Brain-IT) under the Kyutech 21st century COE program (2003-2008) entitled, "World of brain computing interwoven out of animals and robots", (Project leader: Takeshi Yamakawa), which has been driven by the department of Brain Science and Engineering, Graduate School of Life Science and Systems Engineering, Kyushu Institute of Technology. We think that developing the brain-inspired system demands for facilitating cross-disciplinary researches. Five research fields, namely Physiology, Psychology, Theory and Models, Devices, and Robotics, are indispensable to realize a true brain-inspired information system. The Kyutech 21st century COE program produced the several interdisciplinary technology fusions which becomes a base of brain-inspired information technology. Each core technology is new approach brought by conjunction of research results in plural different fields. As one of the results of the Brain-IT, we organized the special session on “Brain-like Computing: Theory and Applications” in the SCIS & ISIS 2008, Joint 4th International Conference on Soft Computing and Intelligent Systems and 9th International Symposium on advanced Intelligent Systems, held on Nagoya, Japan, during 17th-21st September 2008 and had deep and fruitful discussions on this impressive topics. The papers in this issue are mainly the extended and revised version of papers that have been selected from submissions to the special session and the related research results of the 21st Kyutech COE program. The first paper by G. Ogata, K. Natsume, S. Ishizuka, and H. Hayashi is focused on the spike-timingdependent plasticity (STDP) in Hippocampus. Readers of this journal are familiar with computational models but not neuroscience ones. The editors venture to put the neuroscience paper at the top of this special section. We believe that the research results of neuroscience will give suggestive information to researchers in engineering field and promote researches on brain-inspired systems. The STDP function, described in this paper, is known as relating to sequential learning in brain, which is shown by computational studies but not yet neurophysiological experiments. In this paper, experiments on STDP phenomena were performed in rat. The dentate gyrus (DG) in the hippocampus binds or integrates spatial and non-spatial information in parallel. We hope this topic give a new impression to readers. In the second paper, T. Miki, H. Hayashi, Y. Goto, M. Watanabe, and T. Inoue present a practical simple local navigation system inspired Hippocampal function. The authors develop a human-like local navigation using a sequence of landmarks with a simple model easy to implement. The proposed method adapts to a slight change of landmark in the real world. Mechanisms for storing, recalling, and updating the landmark sequence are described. The validity of the proposed system is confirmed using an autonomous mobile robot with the proposed navigation mechanism. The third paper, by K. Tokunaga and T. Furukawa, deals with a new effective method for building an environmental map in a self-organizing manner which is based on a Higher Rank of Self-Organizing Map (SOM2). This is an important and great interest topic in mobile robot researches. The proposed method creates an environmental map in a self-organizing manner from visual information obtained with a camera on a mobile robot. The effectiveness is shown by simulation results. Editorial

23


Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 4,

N째 2

The fourth paper by S. Sonoh, S. Aou, K. Horio, H. Tamukoh, T. Koga, and T. Yamakawa is focused on emotional expressions on robots. Usage of emotions in robotics is one of the attractive themes. The authors proposed an Emotional expression Model of the Amygdala (EMA), which realizes both recognition of sensory inputs and a classical conditioning of emotional inputs. The EMA was developed with a massively parallel architecture by using an FPGA, and the effectiveness is confirmed by demonstrations of a human-robot interaction with the emotions. The fifth paper by T. Matsuo, T. Yokoyama, D. Ueno, and K. Ishii deals with a robot motion control system using Central Pattern Generator (CPG) which exists in nervous system of animals and generates rhythmical motion pattern. The authors propose a robot motion control system using CPG and applied it to an amphibious multi-link mobile robot. The proposed system adapts dynamically to an environmental change by switching a controller due to environment and robot of state. The effectiveness is confirmed by segmental results. In the last paper, T. Sonoda, Y. Nishida, A. Nassiraei, and K. Ishii present a unique actuator which is a new antagonistic joint mechanism using kinematic transmission mechanism (KTM). The proposed model gives a solution for problems that are difficulties in control caused by complex and nonlinear properties, downsizing of actuator, and response time of articular compliance. The performance of KTM is evaluated through stiffness and position control simulations and experiments.

Guest Editors: Tsutomu Miki Kyushu Institute of Technology, Kitakyushu, Japan E-mail: miki@brain.kyutech.ac.jp Takeshi Yamakawa Kyushu Institute of Technology, Kitakyushu, Japan E-mail: yamakawa@brain.kyutech.ac.jp

24

Editorial

2010


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N째 2

2010

THE SPIKE-TIMING-DEPENDENT PLASTICITY FUNCTION BASED ON A BRAIN SEQUENTIAL LEARNING SYSTEM USING A RECURRENT NEURONAL NETWORK Genki Ogata, Kiyohisa Natsume, Satoru Ishizuka, Hatsuo Hayashi

Abstract: This paper examines the spike-timing-dependent plasticity (STDP) at the synapses of the medial entorhinal cortex (EC) and the dentate gyrus (DG) in the hippocampus. The medial and lateral ECs respectively convey spatial and non-spatial information to the hippocampus, and the DG of the hippocampus integrates or binds them. There is a recurrent neuronal network between the EC and the hippocampus called the EC-hippocampus loop. A computational study has shown that using this loop and STDP phenomena at the recurrent EC synapse, sequential learning can be accomplished. But the STDP functions at the synapses of the EC and DG have not yet been studied by neurophysiological experiments. Experiments on STDP phenomena were performed in rats. The STDP function was asymmetrical in the EC synapse and symmetrical in the DG. The medial EC mainly processes the time-series signals for spatial information about visual landmarks when a rat is running in an environment, the lateral EC processes their features, and the DG binds or integrates the information on the positions and features of the landmarks. Thus, the EC-hippocampus loop processes sequential learning of spatial and non-spatial information in parallel, and the DG binds or integrates the two kinds of signals. A system based on this biological phenomenon could have similar characteristics of parallel processing of object features and positions, and their binding.

(MEC) and flows into the hippocampus via the medial perforant path (mPP). Non-spatial information and information on color or shape, which are processed in the occipital and temporal cortices, respectively, enter into layer II of the lateral EC (LEC) and flow to the hippocampus via the lateral perforant path (lPP) (Fig. 1). Then the output signal of the hippocampus returns to layer II of the EC (ECII) through layer V of the EC (ECV). Thus, the connection between the EC and hippocampus is recurrent [4]. This recurrent neuronal network (EC-hippocampus-EC, etc.) is called the EC-hippocampus loop.

1. Introduction

Fig. 1. Spatial information and non-spatial information flow in a brain. The signals on the spatial information first enter the medial entorhinal cortex (MEC) and are conveyed to the dentate gyrus (DG) of the hippocampus. Signals on non-spatial information enter the lateral entorhinal cortex (LEC) and are also conveyed to the dentate gyrus.

Animals live in a temporal world. They can experience several events and store them as episodic memories using their brains. They can also sense a sequence of the events and memorize the sequence. In the engineering field, memorization, i.e., sequence learning, is processed as a recurrent neural network [1]. Michael Jordan developed a recurrent network and applied it to word recognition and the production of speech, etc. [2]. But how does the brain itself process such sequence learning? The entorhinal cortex and hippocampus in the brain are thought to be involved in sequence learning. The hippocampus processes episodic memory. In episodic memory, many sensory signals flow into the hippocampus one by one. The sensory signals are processed in the cortex first. They then flow into the hippocampus via the entorhinal cortex (EC) [3]. For example, when a rat is running in an environment where some visual landmarks are located, spatial information, which is processed in the parietal cortex, first enters into layer II of the medial EC

Fig. 2. The neuronal network between the entorhinal cortex and hippocampus is recurrent. A cross section including hippocampus and EC in the lower figure is shown in the upper picture. You can see the size of the rat brain compared

Keywords: hippocampus, entorhinal cortex, STDP function, brain science.

Articles

25


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

26

N° 2

2010

with a lighter. mECII, layer II of the MEC; SUB, subiculum. The subiculum is the gateway from hippocampus to the EC in the EC-hippocampus loop. mECV stands for layer V of the mEC. The arrows indicate the signal flow in the EC-hippocampus loop in the lower figure.

pe. But the STDP function at the synapse between the mPP and granule cells has not yet been clarified. In the present study, we explored the STDP rules at the synapse between ECII and ECV neurons and at the mPP synapse of the DG in the hippocampus.

In memory processes in the brain, the synapses, i.e., the junctions between two neurons, undergo two kinds of change in the neuronal network [5]. One is a long-term potentiation, LTP, and the other is a long-term depression, LTD. In the LTP, the signal easily passes through the synapse, while in the LTD, it is difficult for the signal to pass through. In the learning process, both LTP and LTD occur, and it is thought that their occurrence forms a spatial pattern. These synaptic changes are controlled by the precise timing of the pre- and the postsynaptic spikes [6]. The resulting long-lasting synaptic change in spike-timing-dependent plasticity (STDP) is expressed as a function of the spike timing Dt between pre- and postsynaptic firing. In most systems studied to date, a presynaptic neuronal spike before the postsynaptic neuronal spike leads to LTP while a postsynaptic neuronal spike before the presynaptic neuronal spike leads to LTD [6]. Cases in which a presynaptic spike before a postsynaptic spike actually leads to LTD while the opposite timing leads to LTP, have also been observed [7]. These STDP rules have an asymmetrical function. There is also a symmetrical STDP function in which the synchronous firing of pre- and postsynaptic neurons leads to LTP while a shift of the timing results in LTD. The STDP rules are used in sequence learning [8] and the control of synchronous activity of neuronal oscillation [9]. In the research of the neural networks, Hebb’s rule is ordinarily used as a learning rule. This rule has no temporal information, while the STDP rule does. When the neural network adopts the STDP rule as a learning rule, the network can use the temporal information of the neuronal spike more effectively. Igarashi and his collaborators including one of the coauthors of the present paper have proposed a model in which the EC-hippocampus loop and the STDP phenomena at the recurrent synapse of ECII are used for sequence learning in the brain [10]. In the model it takes a few tens of msec for the signal to propagate along the loop. When the first signal comes into the loop, the signal can be associated with the next signal, which has a delay around a few tens of msec according to the STDP rule at the recurrent synapses in ECII. Their model adopted a symmetrical STDP function. But the function at the synapse has not yet been clarified by a neurophysiological experiment. The dentate gyrus (DG) of the hippocampus receives a signal from the EC via the medial perforant path (mPP) and lateral perforant path (lPP) in the EC-hippocampus loop. When a rat runs through a course of objects, the mPP conveys the spatial information of the objects, and the lPP conveys their non-spatial features (Fig. 1). The DG integrates the two kinds of information. To clarify how the DG integrates the information, the characteristics of the STDP function at the synapses of mPP and lPP must be studied. The STDP function at the synapse between the lPP and granule cells in the DG has already been measured experimentally [11]. It is symmetrical in sha-

2. Materials and Methods

Articles

Experiments were carried out in compliance with the Guide for the Care and Use of Laboratory Animals at the Graduate School of Life Science and Systems Engineering of Kyushu Institute of Technology. The STDP rule at the ECII synapse was recorded in an EC slice cut from a rat brain as shown in Fig. 2. The STDP rule at the DG synapse was recorded in a hippocampal slice cut from a brain as shown in Fig. 2. Rats were anaesthetized and decapitated, and the brains were removed. Then the slices were cut from the brains using a microslicer. Fifty-nine slices (450 μm thick) of the EC and hippocampus were prepared from twenty-six 3- to 4-week-old Wistar rats. They were transferred to each recording chamber, and perfused with oxygenated artificial nutrition solution. In the STDP experiment at the synapse of the ECII, a recording electrode was placed in the ECII cell layer to record the field excitatory postsynaptic potential (fEPSP) (Fig. 3). The fEPSP indicates the synaptic transmission at the synapse. One of the two stimulation electrodes was placed in the axon layer to stimulate the axons of the presynaptic neurons, and the other electrode was placed in the cell layer to stimulate the postsynaptic neurons of mEC (Fig. 3a). The electrode did not stimulate just one axon or neuron but several. In the experiment at the DG synapse in the hippocampus (Fig. 4), the recording electrode was put into the cell layer of the DG, and the two stimulation electrodes were located to stimulate the presynaptic and postsynaptic neurons. The stimulation protocols were the same in the EC and hippocampus (Fig. 3). The baseline response induced by the baseline stimulus was first recorded; then the paired stimulus at both the presynaptic and postsynaptic neurons was given to induce STDP at the synapse, and the baseline stimulus was given again to check whether the STDP was induced or not. The degree of stimulation at the baseline was adjusted so that the amplitude of fEPSP was 50 % of the maximum so that changes in fEPSP could be easily observed. The stimulus in the paired stimulus was adjusted to the minimum strength at which the postsynaptic neuron induced the spike. Twenty baseline stimuli were fed at intervals of 30 sec and the baseline responses of fEPSP were recorded. After the baseline stimuli, 120 paired stimuli were fed at intervals of 5 sec for the pairing process. In the pairing, the timing of the stimulation to the presynaptic and postsynaptic neurons shifted. After the pairing the baseline responses were again recorded for an hour at the same interval as before (Fig. 3b). The regression line of fEPSP at the latencies between 4 – 8 msec after the stimulation (Fig. 3b arrow) was extrapolated, and the slope of the line was calculated. It was used as an indication of the synaptic strength. To check whether LTP or LTD was induced, the averaged slope of fEPSP 10 min before the pairing was compared with that between 50 and 60 min after the pairing. The statistical test (student's t-test) was performed and the data with a significant difference


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N째 2

2010

Fig. 3. The recording protocol for the STDP function at the ECII synapse. A. Sites where the recording and the stimulation electrodes were located. In the right inset, an example of the electrical signal from the recording electrode is shown. The x and y axes indicate the time and the field potential, respectively. B. The stimulation protocol to record STDP phenomena is shown. The vertical bars indicate the stimulations. The x-axis indicates the time. The inset shows the typical fEPSP. The arrow indicates the time when the slope of fEPSP is calculated. An extrapolated line is also shown.

Fig. 4. Schematic diagram of the recording of the STDP function in the DG of the hippocampus. (p < 0.05) were adopted. At the STDP rule, a positive spike timing (Dt > 0) indicates that the presynaptic neurons were stimulated first, and then the postsynaptic neurons were stimulated. Negative timing (Dt < 0) indicates the opposite. The synaptic change was recorded at spike timings from -60 to 60 msec at the ECII synapse, and from -40 to 20 msec at the DG synapse of the hippocampus.

3. Results 3.1. STDP function at the synapse of ECII In EC slices, after the pairing of the positive timing of Dt = 20 msec, fEPSP was suppressed and the suppression

lasted for at least 60 min (Fig. 5). The fEPSP slope was significantly decreased compared with the baseline slope before the pairing (significance probability p < 0.01). Therefore, LTD was induced (Fig. 5). On the other hand, the pairing at Dt = 0 msec induced the potentiation of fEPSP and lasted for an hour. The pairing induced LTP (Fig. 6). STDP function at the synapse of ECII (Fig. 7) shows that at the spike timing Dt between the presynaptic and postsynaptic cells around 0-msec LTP was induced, while at the shifted timing between them around 20-msec LTD was induced. The shape of the STDP function was asymmetrical. This type of STDP function was first found at the distal synapse of the pyramidal cells in the cortex [12].

Articles

27


Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 4,

N째 2

2010

while the pairings at the negative timing of Dt = -10 msec induced LTD (Fig. 9).

Fig. 5. LTD induced by the pairing of the positive spike timing Dt = 20 msec at the ECII recurrent synapse. The x-axis indicates the time, and the y axis indicates the relative fEPSP slope. The relative fEPSP slope is defined by the ratio of the fEPSP slope to the average for ten minutes before the pairing. The time zero indicates the end of the pairing. The horizontal thick bar indicates the pairing.

Fig. 8. LTP induced by the pairing of the positive timing Dt = 5 msec at the synapse in the DG of the hippocampus.

Fig. 6. LTP induced by the pairing. Fig. 9. LTD induced by the pairing of the negative timing Dt = -10 msec at the synapse in the DG of the hippocampus. The pairing suppressed the transmission.

Fig. 7. STDP function at the ECII recurrent synapse. The filled circles and bars indicate the relative fEPSP slope and the standard errors of the means (SEM). The x-axis shows the spike timing, and the y-axis indicates the relative fEPSP slope. These data are obtained from forty slices of twenty male rats. 3.2.

STDP function in dentate gyrus (DG) of hippocampus In hippocampal slices, after pairing at the positive timing of Dt = 5 msec, LTP was slightly induced (Fig. 8), 28

Articles

Fig. 10. STDP function at the synapse in the DG of the hippocampus. Circles and bars indicate the relative fEPSP slope and SEM. These data are obtained from nineteen slices of six male rats. The STDP function at the DG synapse of hippocampus (Fig. 10) shows that at the spike timing between the pre-


Journal of Automation, Mobile Robotics & Intelligent Systems

synaptic and postsynaptic neurons around 5 msec LTP was induced, while when Dt is between 10 and 18 msec, and between -20 and -1 msec, LTD was induced. The shape of the STDP function is almost symmetrical, unlike the function at the ECII synapse.

4. Discussion At the synapse of ECII, the STDP function was asymmetrical (Fig. 7). It had an LTP region around Dt = 0 msec, and had an LTD region at Dt > 0 msec. On the other hand, at the synapse of DG in the hippocampus, it was symmetrical (Fig. 10). It had an LTP region around Dt=5 msec, and had two LTD regions at Dt < 0 and Dt = -15 msec on both sides of the LTP region. These results suggest that the STDP function is different in different brain regions [13]. It is thought that induction of LTP or LTD depends on the rise in intracellular Ca2+ concentration at the postsynaptic neuron ([Ca2+]i) by the pairing [14]. When [Ca2+]i increases so little by the pairing, the synapse does not change. When it increases moderately, LTD is induced. [Ca2+]i is increased by the neuronal spike. When the increase in [Ca2+]i is great, LTP is induced. Therefore, when the presynaptic and postsynaptic neurons fire simultaneously (Figs. 7 and 10), [Ca2+]i increases greatly in the postsynaptic neurons, and LTP is induced at both the EC and hippocampus synapse (Figs. 7 and 10). The spike timing between the postsynaptic and presynaptic neurons shifts, [Ca2+]i does not increase so much by the stimulation, and LTD is induced at the DG synapse of the hippocampus (Fig. 10). Only this mechanism does not explain the finding that there was no LTD region in the STDP function of the ECII synapse. The present asymmetrical STDP function found at the ECII synapse is a novel type. A similar asymmetrical STDP function was found at the remote synapse from the soma of the pyramidal neurons located in the neocortex of the brain [12], while the function mainly had an LTP region at the negative timing different from the present function. The mechanism of the function has not yet been clarified. There are feedback inhibitory neurons in the ECII network. The neurons may contribute to the STDP function. In an open field with landmarks, a rat runs watching the landmarks. The rat memorizes the time-series of the position of the landmarks and associates their positions and features. From this information the rat recognizes its position. As a result, place cells emerge in the hippocampus [15]. The place cell in the hippocampus fires in the respective place field whenever the rat runs into the field from any direction. As a result, the hippocampus of the rat memorizes its position in the environment. The granule cells in the DG of the hippocampus receive information from the EC via the mPP and lPP. The mPP conveys the spatial (positional) information of the landmarks and the lPP conveys information about their features. The mPP and lPP come from the mECII and lECII, respectively. The two signals converge in granule cells of the DG in the hippocampus. The granule cells in the DG of the hippocampus integrate these two signals (Fig. 1). The STDP function at the DG synapse with mPP (Fig. 10) and lPP in the hippocampus was asymmetrical [11]. It has an LTP area near Dt = 0 msec, and there are two LTD areas with negative and positive timing. When the granule cells

VOLUME 4,

N째 2

2010

fire simultaneously with the firing of mPP or lPP, the synapse of mPP and lPP on the granule cells is long-term potentiated. Therefore, there is a possibility that the granule cells will associate the non-spatial information of the landmark brought by the lPP with their positional information conveyed by the mPP when the two inputs arrive at the granule cells simultaneously. The hippocampus associates the features of the landmark objects with their positional information. The DG neurons, which associate the position and the features of the landmarks, can be regarded as place cells. Igarashi et al. [10] has proposed a model of the brain by which the EC-hippocampus loop processes sequence learning of sensory inputs. In rats, sensory information to the EC is coded by the spike train. It is assumed that the frequency of the spike train is dependent on the distance between the rat and the landmark. When the rat is far from the landmark, the frequency of the spike train is low, for example, 30 Hz, and when the rat comes closer to the landmark, its frequency increases, say, to 40 Hz. The stellate cells in the ECII connect with each other via the recurrent neuronal network through the hippocampus (Fig. 2). Thus, the output of the stellate cells in the EC returns to the other stellate cells in the mEC with some delay. Igarashi's model has found that when two stellate cells acquire the inputs of spike trains of 40 Hz and 30 Hz, respectively, LTP is induced at the synapse from a 40 Hzfiring stellate cell to a 30 Hz-firing stellate cell according to the STDP rule [10]. Thus, when the rat is running through the landmarks A, B and C, the EC-hippocampus loop memorizes the sequence of the landmarks A, B, and C. Igarashi et al. [10] used the STDP function with the symmetrical Mexican hat type shape. One of the present coauthors, Hayashi, has shown that when EC stellate cells receive sensory inputs which contain the background noise, the asymmetrical STDP function with an LTD area for the negative timing and LTP area for the positive timing induces irrelevant synaptic enhancement at the ECII synapse [16]. Thus, they suggest that the LTD area at the positive timing of the symmetrical STDP function prevents the irrelevant synaptic change and enables robust sequence learning for sensory inputs containing the background noise. The STDP function found in the present paper has an LTD area at the positive timing. Using the function, the EC-hippocampus loop could process the sensory input signal with background noise more robustly. Actually there are two loops, a mEC-mPP-hippocampus loop, which processes the time-series of the position, and a lEC-lPP-hippocampus loop, which processes the time-series of the features separately and in parallel. The time-series of the position of the landmark and the features of the landmarks are processed in the EC-hippocampus loop in parallel, and the position and features of the landmark are integrated and bound in the granule cells of the DG in the hippocampus. The parallel processing of the feature of the objects and the integration or binding of them may be the characteristics of the information processing of a brain system based on EC-hippocampus loop. In results, the system can interpret the environment around it freely, and it can adjust to its environment. Articles

29


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

In brain-inspired systems, a new concept for studies linking the fields of brain science and engineering, it is thought that the results from brain science can be applied to engineering technologies [17]. The present work suggests the EC-hippocampus loop processes sequence learning based on features and positioning of landmarks, and the hippocampus integrates or binds the features and the positions in parallel. Using these processes a brain may navigate a path of movement. The proposed brain sequence-learning model using the EC-hippocampus loop may be applicable to the sequential learning of landmarks by a mobile robot. A navigational system for a robot based on the present results can be developed. Our results suggest that there is the possibility for a brain to use a recurrent neuronal network and asymmetrical STDP function to achieve sequential learning. This processing has a feed-forward character. On the other hand, recurrent neural networks in the engineering field usually adopt an algorithm of back propagation through time (BPTT) to learn the sequence of the time-series signal [18]. If the algorithm were used in a brain, some neurons could fire retrospectively. Actually, the hippocampal neurons replay in reverse [19]. Which algorithm the brain adopts must be determined in the future studies. In addition, which algorithm is useful for a robot to navigate a path of moving will be clarified by the studies of the brain-inspired systems.

[7]

[8]

[9]

[10]

[11]

[12]

[13]

ACKNOWLEDGMENTS This work was partially supported by the 21st Century COE (Center of Excellence) Program at the Kyushu Institute of Technology, entitled "World of brain computing interwoven out of integrating animals and robots".

[14]

[15]

AUTHORS Genki Ogata, Kiyohisa Natsume*, Satoru Ishizuka, Hatsuo Hayashi - Kyushu Institute of Technology, 2-4 Hibikino, Wakamatsu-ku, Kita-kyushu, Fukuoka, Japan 808-0196. Kiyohisa Natsume, tel and fax: +81-93-6956094. Email: natume@brain.kyutech.ac.jp. * Corresponding author

[16]

References

[18]

[1] [2] [3]

[4]

[5]

[6]

30

Elman J.L., "Finding structure in time", Cognitive Science, vol. 14, 1990, pp. 179–211. Jordan M.I., "Serial Order: A parallel distributed processing approach" , ICS report, vol. 8604, 1986, pp. 1-40. Kloosterman F., Van Haeften T., Witter M.P., Lopes Da Silva F.H., "Electrophysiological characterization of interlaminar entorhinal connections: an essential link for re-entrance in the hippocampal-entorhinal system", Eur. J . Neurosci., vol. 18, 2003, pp. 3037-52. Witter M.P., Moser E.I., "Spatial representation and the architecture of the entorhinal cortex", Trends Neurosci., vol. 29, 2006, pp. 671-8. Bliss T.V., Collingridge G.L., "A synaptic model of memory: long-term potentiation in the hippocampus", Nature, vol. 361, 1993, pp. 31-9. Bi G.Q., Poo M.M., "Synaptic modifications in cultured hippocampal neurons: dependence on spike timing,

Articles

[17]

[19]

N° 2

2010

synaptic strength, and postsynaptic cell type", J. Neurosci., vol. 18, 1998, pp. 10464-72. Markram H., Lubke J., Frotscher M., Sakmann B., "Regulation of synaptic efficacy by coincidence of postsynaptic APs and EPSPs", Science, vol. 275, 1997, pp. 213-5. Abbott L.F., Blum K.I., "Functional significance of longterm potentiation for sequence learning and prediction", Cereb Cortex, vol. 6, 1996, pp. 406-16. Lubenov E.V., Siapas A.G., "Decoupling through synchrony in neuronal circuits with propagation delays", Neuron, vol. 58, 2008, pp. 118-31. Igarashi J., Hayashi H., Tateno K., "Theta phase coding in a network model of the entorhinal cortex layer II with entorhinal-hippocampal loop connections", Cogn. Neuro-dyn., vol. 1, 2007, pp. 169-84. Lin Y. W., Yang H.W., Wang H.J., Gong C.L., Chiu T.H., Min M.Y., "Spike-timing-dependent plasticity at resting and conditioned lateral perforant path synapses on granule cells in the dentate gyrus: different roles of N-methyl-D-aspartate and group I metabotropic glutamate receptors", Eur. J. Neurosci., vol. 23, 2006, pp. 2362-2374. Kampa B.M., Letzkus J.J., Stuart G.J., "Dendritic mechanisms controlling spike-timing-dependent synaptic plasticity", Trends Neurosci., vol. 30, 2007, pp. 456-63. Sjostrom P.J., Rancz E.A., Roth A., Hausser M., "Dendritic excitability and synaptic plasticity", Physiol. Rev., vol. 88, 2008, pp. 769-840. Bienenstock E.L., Cooper L.N., Munro P.W., "Theory for the development of neuron selectivity: orientation specificity and binocular interaction in visual cortex", J. Neurosci., vol. 2, 1982, pp. 32-48. O'Keefe J., "Place units in the hippocampus of the freely moving rat", Exp. Neurol, vol. 51, 1976, pp. 78-109. Hayashi H., Igarashi J., "LTD windows of the STDP learning rule and synaptic connections having a large transmission delay enable robust sequence learning amid background noise", Cogn. Neurodyn., no. 2(3), June 2009, pp. 119-130. Natsume K., Furukawa T., "Introduction of Brain-inspired systems". In: 2009 International Symposium on Nonlinear Theory and its Applications, Sapporo, Japan, 2009, pp. 195-197. Rumelhart D.E., Hinton G.E., Williams R.J., Learning internal representations by error propagation, vol. 1, MIT Press, 1986. Foster D.J., Wilson M.A., "Reverse replay of behavioural sequences in hippocampal place cells during the awake state", Nature, vol. 440, 2006, pp. 680-683.


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N째 2

2010

A SIMPLE LOCAL NAVIGATION SYSTEM INSPIRED BY HIPPOCAMPAL FUNCTION AND ITS AUTONOMOUS MOBILE ROBOT IMPLEMENTATION Tsutomu Miki, Hatsuo Hayashi, Yuta Goto, Masahiko Watanabe, Takuya Inoue

Abstract: We propose a practical simple local navigation system inspired by the sequence learning mechanism of the entorhino-hippocampal system. The proposed system memorizes a route as sequences of landmarks in the same way humans do. The proposed local navigation system includes a local route memory unit, landmark extraction unit, and learning-type matching unit. In the local route memory unit, the concept of the sequence learning mechanism of the entorhino-hippocampal system is implemented using a fully connected network, while a sequence of landmarks is embedded in the connection matrix as the local route memory. This system has two operation modes: learning and recall modes. In learning mode, a sequence of landmarks, i.e. a local route, is represented by enhanced loop connections in the connection matrix. In recall mode, the system traces the stored route comparing current landmarks with the stored landmarks using the landmark extraction and learning-type matching units. The system uses a prospective sequence to match the current landmark sequence with the recalled one. Using a prospective sequence in the route comparison allows confirmation of the correct route and deals with any slight change in the current sequence of landmarks. A certainty index is also introduced for judging the validity of the route selection. We describe a basic update mechanism for the stored landmark sequence in the case of a small change in the local route memory. The validity of the proposed system is confirmed using an autonomous mobile robot with the proposed navigation system. Keywords: human-like local navigation, sequence learning, entorhino-hippocampal system, autonomous mobile robot.

1. Introduction Recently, due to the rapid growth in digital technology, there has been accelerated development of highly intelligent machines. Intelligent machines have made our daily lives far more convenient. Huge quantities of data are easily handled electronically at high speed and many electrical devices have been provided with powerful functionality. In particular, the latest vehicles come equipped with very sophisticated information devices [1]. One of the most remarkable technologies is the navigation system, which guides us to our destinations in unfamiliar territory. The guidance is supported by a GPS system and an accurate digital map. In other words, the system is highly dependent on data and therefore, has a weakness with respect to recent changes and mistakes in the data. Humans on the other hand, can handle such

changes flexibly. Introducing such human-like information processing would be vital in compensating for the weakness in digital equipment. Our aim is to develop an effective human-like system, which compensates for this weakness and is able to suggest a plausible route even when conventional navigation systems fail due to insufficient information. Recently, many researchers have focused on the function of spatio-temporal representation in the hippocampus [2], [3]. Yoshida and Hayashi proposed a computational model of the sequence learning in the hippocampus; that is, neurons that learn a sequence of signals can be characterized, in the hippocampal CA1, by using propagation of neuronal activity in the hippocampal CA3 [5]. A network model of the entorhinal cortex layer II (ECII) with entorhino-hippocampal loop circuitry was proposed by Igarashi et al. [6]. Loop connections between stellate cells in the ECII are selectively potentiated by afferent signals to the ECII, and consequently stellate cells connected by potentiated loop connections fire successively in each theta cycle. The mechanism has also been investigated from a neurobiology viewpoint [7]. We focus on the attractive abilities of sequential learning and apply them to a local navigation system. Several navigation mechanisms, inspired by crucial brain functions especially in the hippocampus and its surroundings, have been proposed [4], [8], [9]. Most of these mechanisms, however, tend to become very intricate as a result of faithfully mimicking the brain mechanism. On the other hand, simplicity of the model is an important factor for practical embedded systems. We aim to develop a simple local navigation system introducing the essence of the remarkable brain functions. We have proposed a practical simple local navigation system inspired by the sequence learning mechanism of the ECII with entorhino-hippocampal loop circuitry [10]. In the system, a route is represented as a sequence of landmarks as is the case in humans. The proposed local navigation system consists of a simple local route memory unit, a landmark extraction unit, and a learning-type matching unit. In the local route memory unit, the sequence learning mechanism of the entorhino-hippocampal system is implemented using a fully connected network, while a sequence of landmarks is embedded in the connection matrix as the local route memory. This system has two operation modes: learning and recall modes. In the learning mode, a sequence of landmarks is represented by enhanced loop connections in the connection matrix. In the system, a certainty index is introduced to evaluate the validity of the route selection. We have realized a flexible local navigation system in a simple architecture Articles

31


Journal of Automation, Mobile Robotics & Intelligent Systems

using the prospective landmark sequence and certainty index. In this paper, we describe the mechanisms for storing and recall the landmark sequence and present a basic update mechanism for the stored landmark sequence. We confirm the validity of the proposed system and investigate its adaptability to changes in circumstance through experiments using an autonomous mobile robot with the proposed mechanism.

2. Sequence learning mechanism of the entorhinal-hippocampal system Igarashi et al. proposed a computational model of the ECII network with entorhino-hippocampal loop connections [6]. In their model, a pair of afferent pulse trains to the ECII, with clearly different frequencies, is selected by virtue of loop connections that are selectively potentiated by the pairs of afferent signals. The frequency depends on the strength of sensory input. The signal transmission delay through the loop connections produces the order of places. Here a “place” means a place in the real world corresponding to a landmark. We assume that the observer moves at a constant speed. Here, let us assume that a route is coordinated by a sequence of places, A, B, C, D, and E. The signal for each place is represented by a frequency depending on the distance between the observer and the place, where a high frequency corresponds to a shorter distance. A higher frequency signal makes a so-called “place cell” fire in a shorter period of time. When a signal observed at place A is fed into the ECII, the signal of the place cell A is transmitted from the ECII to the ECV through the DG-CA3CA1 in the entorhino-hippocampal loop circuit as shown in Fig. 1. If signal B observed at place B fires another place cell B in the ECII at the time the transmitted signal A arrives at the ECV cell A, the connection between the ECV cell A and ECII cell B is enhanced by learning mechanisms in the brain. This means that a relationship between place A and place B is established. The relationships for the following signals are established in the same manner. As a result, the order of places is embedded in the loop connection weights from the ECV to ECII neurons.

VOLUME 4,

N° 2

2010

chanism in the hippocampus and the entorhinal cortex. The system uses signals obtained from images of landmarks specifying places and the route is stored as a chain connection of landmarks.

3. Simple local navigation system inspired by the sequence learning mechanism in the entorhino-hippocampal loop circuit We propose a dedicated navigation system inspired by the structure of the entorhino-hippocampal loop circuitry. In the proposed system, the sequence learning mechanism is implemented using a fully connected network (as shown in Fig. 2), while the order of landmarks is embedded in the matrix of loop connections as a local route memory unit. Here each landmark corresponds to a place in the real world. The entorhino-hippocampal loop circuitry illustrated in Fig. 1 is represented by neurons corresponding to place cells in the ECII and connecting loops with connection weights. A connection weight represents an established connection between the corresponding place cells in the ECV and ECII. The navigation system also includes a landmark extraction unit and a learning-type matching unit as shown in Fig. 3. The landmark extraction unit extracts landmarks from an image obtained with a camera, and the learning-type matching unit evaluates the degree of matching with the current tracing route.

Fig. 2. Local route memory unit: the route is coordinated by a sequence of observed landmarks. Landmark sequence: A®B®C®D®E®F®. Each landmark corresponds to a place in the real world. The loop circuit illustrated in Fig. 1 is represented by neurons corresponding to place cells in ECII and a loop connected with a connection weight. The connection weight represents an established connection between place cells in ECV and ECII.

Fig. 1. Entorhino-hippocampal loop circuit. Here “place” means a place in the real world, and represents “a place cell” which is a neuron in EC corresponding to a place in the real world. In this paper, we develop a practical simple local navigation system inspired by the sequence learning me32

Articles

Fig. 3. Proposed navigation system includes a landmark extraction unit, a learning-type matching unit, and a local route memory unit.


Journal of Automation, Mobile Robotics & Intelligent Systems

The system has two operation modes, learning mode and recall mode. In learning mode, the order of the obtained landmarks is stored in the local route memory unit. In recall mode, the local route memory unit recalls the prospective landmark sequence and the learning-type matching unit evaluates the selected current route by comparing the observed landmark sequence with the recalled prospective landmark sequence. Here, we assume the following procedure. 1) First, in learning mode, the system memorizes a route by moving along the route or storing the data of the route. 2) Thereafter, in recall mode, the system traces the route automatically according to the stored one. Fig. 4 shows the basic elements of the local route memory unit. Fig. 4a) depicts a neuron unit corresponding to a place cell in the ECII. Here Ij, aj, uj and prgj are the landmark input, activation input, output of the neuron, and a program signal, respectively. A LRN / RCL is a mode select signal equal to “1” in learning mode and “0” in recall mode. Once the neuron has been activated, output uj retains the activated state until another neuron is activated. Fig. 4b) depicts the connection weight unit, where wij is a connection weight. The connection weight unit stores the ordering relationship between the i-th and j-th neurons. When Ij is fed into the j-th neuron, uj is activated in learning mode, prgj becomes “1” and the connection weight wij is set to “1”. (a) j-th neuron unit

VOLUME 4,

N° 2

2010

3.1. Learning mode In learning mode (Fig. 5), the system coordinates the route as a sequence of observed landmarks. A signal corresponding to the landmark is fed into the local route memory unit and a connection is made between the current landmark and the one immediately before. Here, let us assume that a signal of the j-th landmark Ij is observed after a signal of the i-th landmark Ii. The relationship between the landmarks is represented by enhanced loop connections in the connection matrix as follows: 1) ui is activated by signal Ii and the activation state is retained until the next neuron is activated. 2) Then, the landmark signal Ij is fed into the j-th neuron, and a program signal prgj is set to “1”. 3) The connection weight wij is assigned by the following equation. (1) 4) Moreover, an activation signal aj is assigned by the following equation. (2) 5) aj activates uj in preparation of the next landmark signal. By repeating the above steps, the local route is stored in the local route memory unit.

(b) Connection weight unit

Fig. 4. (a) Neuron unit and (b) connection weight unit. The neuron corresponds to a place cell in ECII. Here Ij, aj, uj and prgj are the landmark input, activation input, output of the neuron, and a program signal, respectively. A LRN / RCL is a mode select signal equal to “1” in learning mode and “0” in recall mode. Once the neuron has been activated, uj retains an activated state until another neuron is activated. The connection weight unit memorizes the ordering relationship between the i-th and j-th neurons. When Ij is fed into the j-th neuron, uj is activated in learning mode, prgj changes to “1” and the connection weight wij is set to “1”. In the neurobiological computational model, the loop delay corresponds to the sampling period of capturing the landmark. Conversely, in the proposed system, the landmark is captured every time it is observed and the activation state of the neuron is kept with a latching mechanism until the next landmark appears.

Fig. 5. Assignment of the connection weight in learning mode. 3.2. Recall mode In recall mode (Fig. 6), LRN / RCL = “0”, the system traces the stored route in the local route memory unit as follows: 1) When the landmark signal Ii is observed and fed into the i-th neuron, Ii activates ui. 2) If the connection weight wij on the activated line ui is “1”, then ai is set to “1”. 3) The neuron output signal uj is activated by aj. 4) One recall leads to another, until the (i+k)-th landmark signal Ii+k is recalled in the same manner. Here, the observed and prospective sequences consist of K landmarks. Articles

33


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

5) The recalled sequence consists of K landmarks is fed into the matching unit and compared with the current sequence of observed landmarks.

N° 2

2010

tigates the subsequent sequence after a change point. If the subsequent landmarks match the stored landmarks, the system accepts that the selected route is correct and that an environmental change has occurred in the stored landmark sequence. If an update of memory is required, the connections are updated as described below. (a) Stored route

(b) Adaptability to a missing landmark in the real world: the case in which landmark “B” disappears in the real world.

Fig. 6. Landmark sequence activation in recall mode. Certainty index CI for judging the validity of the selected current route In recall mode, the matching unit matches the current sequence of landmarks and the corresponding recalled sequence (where both sequences consist of K landmarks). The certainty index CI is introduced to verify the correctness of the current selected route. The certainty index CI represents the validity of the route currently selected and is defined as 3.3.

(3)

(c) Adaptability to an additional landmark in the real world: the case in which landmark “b” appears between landmarks “A” and “B”.

(4)

where K is the length of the recalled sequence used for R O route matching, and Ii+k and Ii+k represent the (i+k)-th recalled landmark and currently observed landmark, respectively. A slight change in the circumstances can be represented by a change in the CI. For example, a crossing can be detected by a change in the certainty index CIi. As the observer reaches the crossing, the certainty index CIi decreases to 1/K. It is because that the number of the matched landmarks decreases as closing to a crossing. 3.4.

Update mechanism of the stored route in the local route memory In the proposed method, the order of landmarks is embedded in a matrix of loop connections as the local route memory. The stored route in the local route memory unit can easily be updated by rearranging the loop connections. In this section, we describe the procedure for updating the connections in two different cases: a missing or added landmark observed in the selected route. When the mismatched landmark is detected, the system inves34

Articles

Fig. 7. Proposed system can adapt to stored route changes caused by a slight change in landmarks. The stored route in the local route memory can be updated by rearranging the loop connections. Let us assume that the route depicted in Fig. 7a) is stored in learning mode. Fig. 7b) illustrates the case where landmark “B” disappears in the real world. In this case, first, the landmark of place A is observed. The system recalls the prospective landmark sequence and expects the landmark of place B as shown on the left in Fig. 7b). If “C” is observed instead of “B”, the system confirms that landmark “B” is missing and generates a new connection-node between “A” and “C” as shown on the right in Fig. 7b). As a result, the stored route is updated. On the other hand, Fig. 7c) illustrates the case where landmark “b” appears


Journal of Automation, Mobile Robotics & Intelligent Systems

between landmarks “A” and “B”. In this case, a new route is created by generating two new connections instead of the old one as shown in Fig. 7c).

4. Experimental results To confirm the validity of the proposed local navigation system for practical engineering applications, we made use of an autonomous mobile robot “WITH” [11], which is a basic omni-directional mobile robot platform developed as a result of the Kyutech 21st COE program. We confirm that the proposed system can extract landmarks and store a route in the form of a sequence of landmarks and that the robot with the proposed mechanism can trace the route corresponding to the stored landmark sequence in the local route memory unit. Moreover, we show that the robot behaves like a human when circumstances change slightly. Fig. 8 shows the autonomous mobile robot consisting of the robot base WITH, a USB camera, and palmtop computer VAIO type-U in which the proposed navigation system is embedded. The setting signal of the operation modes and the cruise control signal in learning mode are given wirelessly by an external computer. The experiments were designed to investigate the following behavior: 1) route tracing according to the stored route memorized in learning mode; and 2) selecting a plausible route when a slightly altered sequence of landmarks is encountered as a result of either a missing landmark or the addition of an unknown landmark in the stored route sequence.

VOLUME 4,

N° 2

2010

lighting has an effect on the appearance of landmarks obtained with a camera. Different color images can therefore be obtained for the same landmark. Here, we avoid the problem by using a standard three-layer-perceptron trained in advance with all the landmarks used in the experiments. In this experiment, the MLP is trained using the hue data of an HSV image set for each landmark. The trained MLP works as a classifier of the landmark obtained with a camera.

Fig. 9. Experimental field: field size is 5.5m x 5.5m, path is drawn using two black lines, and width of the path is 24cm. Each landmark is about 10cm x 10cm and labeled with a color: red, blue, light green, yellow, or orange.

Fig. 10. Arrangement of landmarks and the stored route in the experiments.

Fig. 8. Autonomous mobile robot consisting of the mobile robot base WITH, a USB camera, and a palmtop computer VAIO type-U in which the proposed navigation system is embedded. 4.1.

Route tracing according to the route stored in the local route memory unit Landmarks are set along the path on an experimental field as shown in Fig. 9. A path, of width 24 cm, is drawn using two black lines. Each landmark is about 10 cm x 10 cm and labeled with one of the colors, red (R), blue (B), light-green (G), yellow (Y), or orange (O). Fig. 10 shows the arrangement of the landmarks and the route stored in learning mode in this experiment. In general, a change in

First, in learning mode, an operator controls the robot navigating through the planned route. The system obtains landmarks beside the path along the route traversed and stores the route as a sequence of these landmarks. Once this has been completed, we make the robot trace the stored route autonomously. Fig. 11a) shows the camera view in which landmarks obtained by the landmark-extracting unit are displayed. The robot traces the correct route as shown in Fig. 11b). Fig 11c) shows the prospective sequence recalled by the local route memory unit and a change in the certainty index as defined in Eq. (3). As shown in Fig.11c), by monitoring the change in the CI, it is known when the robot reaches a crossing. The CI returns to "1" after the robot turns at the crossing. This means that the correct route has been selected. 4.2. Route tracing with a slight route change Two situations are assumed: a missing landmark and the addition of an unknown one in the current route. The Articles

35


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

(a) Camera view: landmarks are extracted from an image obtained with a single camera

N째 2

2010

(b) Autonomous mobile robot traces the stored route automatically

(c) Prospective sequence recalled by the local route memory unit and a change in the certainty index

Fig. 11. Result of route tracing according to the route memorized in learning mode. (a) Missing landmark in the current route

(b) Change in the certainty index CI

Fig. 12. Behavior of route tracing in the case of a slight change in the route stored in the local route memory. (a) At a crossing, the system evaluates all possible routes and chooses the route with the highest matching degree as the correct direction. (b) The robot recognizes a missing landmark from a change in the CI and by comparing the current route with a shifted memory route. 36

Articles


Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 4,

N° 2

2010

(a) Additional landmark

(b) Change in the certainty index CI

Fig. 13. Behavior of route tracing in the case of a slight change in the route stored in the local route memory. (a) At a crossing, the system evaluates all possible routes and chooses the route with the highest matching degree as the correct direction. (b) The robot recognizes an additional landmark from a change in the CI and by comparing the current route with a shifted memory route. behavior in the case of a missing landmark or an additional one is shown in Figs. 12 and 13, respectively. Fig. 12. shows the case that landmark is missing. The robot recognizes a missing landmark from a change in the CI and by comparing the current route with a shifted memory route. Then the system generates a new route by update mechanism described in 3.4. Conversely, in the case that an additional landmark is appeared in the real world (Fig. 13), the robot recognizes the additional landmark from a change in the CI and by comparing the current route with a shifted memory route. Consequently, the system generates new routes by update mechanism described in 3.4. At a crossing, the robot decides which way to move by comparing the matching degree (CI) of the local landmark sequence for all possible paths. Despite checking all possible paths at the crossing, perfectly matching paths are not detected in these cases. Therefore, the robot selects the path with the highest matching degree (CI) as the plausible direction. When a local route stored in memory is found in the selected route, the system concludes that its decision was correct and that the route had changed slightly.

5. Conclusions A practical simple human-like navigation system inspired by the entorhino-hippocampal loop mechanism was proposed and the validity confirmed through experiments using an autonomous mobile robot. The system navigates using landmarks stored in the memory unit. By using a prospective landmark sequence, the system is able

to adapt to slight changes in the local route stored in the route memory unit. In this paper, we confirmed that the correct action is taken when there is either a missing or an additional landmark in a local route. This ability makes the navigation system flexible. In addition, we introduced the certainty index as a measure of recognizing the present situation in a route tracing. We also presented the basic idea for updating the route stored in the local route memory unit in the case of a slight change in circumstances by changing the connection weight of “the fullyconnected network�. The authors emphasize that the proposed system implementing the sequence learning mechanism can be completed even by a small palmtop computer. This feature is very important from an engineering viewpoint. The system compensates for the latest digital navigation systems requiring a GPS and up-to-date map for accurate navigation. In particular, the system is promising as a low-cost and effective navigation system. It can be used in places where GPS signals are unavailable and in shopping malls where costly dedicated equipment such as RFID-tag is not employed. In our future work, we aim to develop a decision mechanism for updating the timing and to investigate the adaptability of the proposed method. Improvement of the representation ability of the certainty index is necessary for recognizing more complex situations. A robust landmark extraction technique that can operate in real scenes is also important to use the proposed method in practical applications. Articles

37


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

ACKNOWLEDGMENTS This work was supported in part by a 21st Century Center of Excellence Program <Center#J19>, "World of Brain Computing Interwoven out of Animals and Robots (PL: T. Yamakawa)" granted in 2003 to the Department of Brain Science and Engineering, Graduate School of Life Science and Systems Engineering, Kyushu Institute of Technology by the Japan Ministry of Education, Culture, Sports, Science and Technology.

AUTHORS Tsutomu Miki*, Hatsuo Hayashi, Yuta Goto, Masahiko Watanabe, Takuya Inoue - Graduate School of Life Science and Systems Engineering, Kyushu Institute of Technology, 2-4, Hibikino, Wakamatsu-ku, Kitakyushu 808-0196, Japan. Tel/Fax: +81-93-695-6125. E-mail: miki@brain.kyutech.ac.jp. * Corresponding author

References [1]

[2]

[3]

[4]

[5]

[6]

[7]

[8]

[9]

[10]

38

Makoto M., "Recent Topics of Car Electronics”, Technical report of IEICE. ICD (in Japanese), vol. 94, no. 243, 1994, pp. 69-75. Yamaguchi Y., “A Theory of hippocampal memory based on theta phase precession”, Biological Cybernetics, no. 89, 2003, pp. 1-9. Hafting T., et al., “Hippocampus-independent phase precession in entorhinal grid cells”, Nature, no. 453, 2008, pp.1248-1252. Wagatsuma H.,Yamaguchi Y., “Content-Dependent Adaptive Behavior Generated in the Theta Phase Coding Network”, ICONIP 2007, Part II, LNCS 4985, 2008, pp. 177-184. Yoshida M., Hayashi H., "Emergence of sequence sensitivity in a hippocampal CA3-CA1 model”, Neural Networks, vol. 20, 2007, pp. 653-667. Igarashi J., Hayashi H., Tateno K., "Theta phase coding in a network model of the entorhinal cortex layer II with entorhinal-hippocampal loop connections”, Cognitive Neurodynamics, vol. 1, no. 2, 2007, pp. 169-184. Natsume K., et al., “The Possibility of Brain-Inspired Time-Series Memory System using the Recurrent Neuronal Network between Entorhinal Cortex and Hippocampus”, Joint 4th Int. Conf. on Soft Computing and Intelligent Systems and 9th Int. Sympo. on advanced Intelligent Systems (SCIS&ISIS2008), Nagoya, 2008, pp. 1778-1782. Gionannangeli C., Gaussier P., “Autonomous visionbased navigation: Goal-oriented action planning by transient states prediction, cognitive map building, and sensory-motor learning”, International Conference on Intelligent Robots and Systems (IROS 2008), 2008, pp. 676-683. Barrera A., Weitzenfeld A., “Biologically-inspired robot spatial cognition based on rat neurophysiological studies ”, Autonomous Robots, vol. 25, no. 1-2, 2008, pp. 147-169. Miki T., et al., “Practical Local Navigation System Based in Entorhino-hippocampal Functions”, Joint 4th Int. Conf. on Soft Computing and Intelligent Systems and 9th

Articles

[11]

N° 2

2010

Int. Sympo. on advanced Intelligent Systems (SCIS& ISIS2008), 2008, Nagoya, pp. 1783-1787. Ishii K., Miki T., "Mobile robot platforms for artificial and swarm intelligence researches ", Brain-Inspired IT, vol. 3, 2007, pp. 39-42.


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 2

2010

BUILDING A COGNITIVE MAP USING AN SOM2

Kazuhiro Tokunaga, Tetsuo Furukawa

Abstract:

1.1. Aim of this study The ability to build an environmental map based on sensor information is necessary for an autonomous robot to perform self-localization, identification of direction and self-navigation. In animals, a map building capability is very important to accomplish crucial behavior such as predation, nest homing, path planning, and so on. With regards to research on map building in animals, O’Keefe and Dostrovsky identified “place cells”, which respond preferentially to specific spatial locations in the hippocampus of a rat [1]. The place cells encode the

observed sensory information as the animal explores its environment. Moreover, O’Keefe and Nadel propounded the theory that animals build a “cognitive map” within the brain, based on research of the place cells [2]. It is thus evident that animals build a cognitive map, which plays a role in path planning using landmarks (i.e., particular information of local environments) coded by the place cells. Furthermore, Taube et al. identified “head direction cells”, which respond preferentially according to the direction of the head [3]. The head direction cells are seen to be involved in the map building, since it is important to know one’s own direction before moving to a destination. Therefore, a robot is expected to be able to perform navigation automatically using a cognitive map model that incorporates the mechanism of place cells and head direction cells in its implementation. Moreover, it may be possible to identify a mechanism from the model akin to the map building of animals. With regards to a technical model for map building, we propose using a Higher Rank Self-Organizing Map (SOM2) in this study. The SOM2 proposed by Furukawa [4] is generally an extended model of Kohonen’s SOM [5]. The SOM2 has a SOM-type modular network structure but with nesting SOMs (Fig. 1(a)). It is the task of each SOM module in the SOM2 to identify a manifold, which approximates a data vector set, thereby enabling the entire SOM2 to find the formation of the fiber bundle in a self-organizing manner (Fig. 1b)). It has been suggested in [4] that by using this feature, the SOM2 can, with unsupervised learning, build a map, which is able to estimate the

(a)

(b)

In this paper, we propose a new method for building an environmental map in a self-organizing manner using visual information from a mobile robot. This method is based on a Higher Rank of Self-Organizing Map (SOM2), in which Kohonen’s SOM is extended to create a map of data distributions (set of manifolds). It is expected that the “SOM” is capable of creating an environmental map in a self-organizing manner from visual information, since the set of visual information obtained from each position in the environment forms a manifold at every position. We also show the effectiveness of the proposed method. Keywords: self-Organizing map, map building, place cells, head direction cells, autonomous robot.

1. Introduction

Fig. 1. Architecture of SOM2. (a) the SOM2 is a nest structure of SOMs. The position of each child SOM is kept with connecting between neighborhood child SOMs by path. (b) Each child SOM approximates each episode with a graph map (i.e., a manifold) through training of the episodes. The connecting the correspondence point of each map represents the fiber. Articles

39


Journal of Automation, Mobile Robotics & Intelligent Systems

location and azimuth direction independently, based only on sets of the image data vectors observed from omni-directional vision sensors. It has also been suggested that the SOM2 can, with unsupervised learning, build a cognitive map with the features given below, using only visual information; each module of the SOM2 represents a place cell which codes the specific location, while each reference vector unit in the SOM module represents a head direction cell. However, a detailed verification has not yet been done. Hence, this study aims to confirm through various computer simulations, whether or not the position and the azimuth direction can be estimated from a map acquired by unsupervised learning of the SOM2. 1.2. Related works Map building is an important theme in studies involving autonomous robots. Consequently, in recent years, various methods for building an environmental map have been proposed. A classical map building method, dead reckoning, can estimate the location and pose (or direction) of a robot by calculating the displacement from an internal sensor, such as a rotary encoder, acceleration sensor, and so on. The proposal method builds the map from only the arrangement of the memory of sensor information, while the dead reckoning build the map by using odometry information. On the other hand, several methods for map building using external sensors such as a laser range finder, vision sensor, and so on, have also been proposed. The most popular method, known as SLAM [6], is often used in map building using external sensors [7], [8], [9]. SLAM can perform self-localization and estimate the structure of the environment around the robot simultaneously, making it a technologically excellent method for map building. Nevertheless, SLAM requires a highly accurate observation model and locomotion model, a priori, since it is necessary to understand the correct structure of the environment [10]. The observation and locomotion models provide the physical measurements of the environment and the physical location of the robot using external sensors, respectively. The observation and locomotion models provide the physical measurements of the environment and the physical location of the robot using external sensors, respectively. The SLAM builds the map based on measurements provided from these models. However, it is difficult to develop the models, which flexibly build the map, since the conditions of the environment, and the sensors are nonstationary. Besides, the correspondence between SLAM and map building in an animal’s brain has not yet been identified. In contrast, a method has been proposed, called “topological map building”, that builds the map abstracted by a graph [11]. Nodes and edges in the graph represent specific locations (areas) and pathways between areas, respectively. Typically, sensor information for landmarks is memorized to nodes, while pathways between landmarks are stored as edges. Then self-localization and path planning can be performed by matching the sensor information to the map. Since each node memorizes the information that represents a local area, it is not necessary to comprehend the correct structure of the environment. In addition, the method requires no physics 40

Articles

VOLUME 4,

N° 2

2010

models beforehand. Moreover, it is very interesting because this method resembles the cognitive map based on place cells in the hippocampus of an animal’s brain. However, this method has two important issues: how are the allocation of nodes and the connection of paths decided in a self-organizing manner. As a solution to these issues, applying a self-organizing neural network (SONN), such as the Self-Organizing Map (SOM), Topology Representing Network (TRN), and so on, has been suggested [12]. Reference vectors of the SONN are the nodes that memorize the information of specific areas. Moreover, paths that connect reference vectors represent the pathway between nodes. The SONN can perform the allocation of nodes and the connection of paths in a self-organizing manner with unsupervised learning. Tanaka et al. proposed an implementation model incorporating the place cell in the hippocampus using a TRN [13]. This method does not, however, build a cognitive map in the same way as an animal, because GPS information is included as training input. In addition, K. Chokshi et al. proposed a method for self-localization using the categorization of vision information by an SOM [14]. Their method is also an implementation model of the place cell. Nevertheless, since these methods do not include the functionality of the head direction cell, a robot cannot identify its own direction. In contrast, our proposal method builds the map that can independently estimate the position and direction from only the visual information with the unsupervised learning. Each module of SOM2 is a node, which memorizes the visual information that represents the local environment in a self-organizing manner. In addition, the memorized visual information is ordered corresponding to direction with the unsupervised learning. The features are similar to ones of the place cells and head direction cells in the hippocampus. Thus, the method incorporates not only the functionality of the place cells, but also that of the head direction cells. Moreover, the topology of the map acquired with SOM2 and the topology of the environment’s geography are nearly equal. Therefore, the selfnavigation of the robot can be very easily performed by using the map.

2. Map building using an SOM2 In this study, we aim to show that the SOM2 can create a self-organizing map in which the position and orientation of a robot can be estimated using only vision sensor information. First, we explain how to acquire the map using an SOM2. When a mobile robot equipped with a vision sensor gets a birds-eye view of the surrounding area at place A in the environment as shown in Fig. 2a), the episode of vision sensor information is distributed as a manifold in a multi-dimensional vector space (sensor space). If the mobile robot observes vision sensor information by rotating 360 degrees at place A, then the episode of vision sensor information is distributed as a one-dimensional toroidal manifold (Fig. 2b)). In addition, if the mobile robot moves from place A to place B, the episode of sensor information at place B forms a manifold near place A (Fig. 2b)). Thus, the episodes observed at consecutive places in the environment form continuous manifolds in


Journal of Automation, Mobile Robotics & Intelligent Systems

sensor space (Fig. 2b)). Moreover, the correspondence point between the manifolds corresponds to the shooting angle of the vision sensor (that is, the azimuth direction of the environment). Therefore, it is expected that the position and azimuth direction can be estimated using a map created based on the distance and correspondence point between manifolds. (a)

(b)

Fig. 2. The episode of vision sensor information is distributed as a manifold in a sensor space. If the mobile robot moves from place A to place B, the episodes of sensor information at place B from a manifold near place A. Moreover, the correspondence point between the manifolds corresponds to the shooting angle of the vision sensor. For this method, we employ the SOM2 proposed by Furukawa. The SOM2 is an extension of the SOM in which each reference vector unit in the conventional SOM is replaced by an SOM module. In other words, the SOM2 is a nest structure of SOMs (Fig. 1a)). In this paper, the SOM module (child level) is called the “child SOM”, while the whole SOM (parent level) is called the “parent SOM”. In the SOM2, sets (episodes) of vector data are given to the SOM2 as training data. The vector data for each episode are distributed on each of the subspaces in vector space

VOLUME 4,

N° 2

2010

(Fig. 1b)). Each child SOM approximates each episode with a graph map (i.e., a manifold) through training of the episodes. Here, training of each child SOM is performed in such a way that the correspondence points on the map in each child SOM are uniform. Thus, connecting the correspondence point of each map represents the fiber. Besides, the parent SOM orders the maps formed by child SOMs. Therefore, an SOM2 can create a map of manifolds. In building an environmental map using an SOM2, a Ring SOM (RSOM), which approximates the distribution of data vectors by a one-dimensional toroidal manifold, is em-ployed for each child SOM (Fig. 3). Hereafter, this SOM2 is referred to as the RSOM×SOM. The episode sets of vision sensor information observed at various places are given to the SOM2 as training episodes. After training, each child SOM forms a manifold of each place’s episodes. Here, the correspondence points in the map of each child SOM are constant, that is, the environmental azimuth direction is constant. Each reference vector unit of the RSOM represents a head direction cell. Moreover, the parent SOM creates a map with the topology (i.e., topology at the positions of the manifolds in sensor space) of the positions in the environment preserved. As a result, the map of the parent SOM itself represents a geometrical map of the environment. Each module of parent SOM represents the place cells. In addition, the azimuth directions of the environment in each RSOM are ordered in a self-organizing manner. Restrictions on the method, however, include that the working environment is open without obstacles, in which robot cannot pass through and robot’s view is interrupted, and that similar visual information does not exist in the environment. It is certainly possible to apply the method in a non-limited environment by enhancing the SOM2. (This is addressed in subsection 5.1.) Nonetheless, the aim of this study is to verify that an SOM2 can create a self-organizing map in which the position and orientation of a robot can be estimated using only vision sensor information. The robot’s working environment for this study is set as follows. (A) The working environment is open without obstacles. Moreover, the robot can see faraway buildings and mountains, etc. as shown in Figs. 4 and 5. (B) The robot has an omni-directional camera as vision sensor. (C) Only visual information is assumed to be observed by the robot sensors. In (A), under normal circumstances, it is preferable that the robot can build the map while looking around with a single directional camera. Building the preferable map from partial information is difficult without enhancing the algorithm for the SOM2. Thus, in this study, the episodes are acquired from an omni-directional camera.

3. Algorithm for the SOM2 (RSOM×SOM) In this section, the algorithm for the RSOM×SOM is explained. The RSOM×SOM is an SOM2 in which each child SOM is replaced by a RSOM. The difference between an RSOM and SOM is the definition of the distance measure between reference vectors on the map, since the reference vector is allocated on a one-dimensional toroid in Articles

41


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 2

2010

Fig. 3. Architecture of RSOM×SOM that is engineering model of the place cells and the head direction cells. The map of the parent SOM itself represents a geometrical map of the environment. Each reference vector unit of RSOM represents a head direction cell. Moreover, Each RSOM represents the place cells.

Fig. 4. Type 1 environment.

Fig. 5. Type 2 environment.

the RSOM, but on a lattice in the SOM. Otherwise they are the same. First, we define certain variables. Suppose there are I training episodes where each episode is composed of J data vectors. The i-th episode is defined as Di = {xi1, xi2,..., xij,..., xiJ}, where xij is the vector data. Furthermore, the parent SOM is composed of K RSOM modules. Each RSOM module (i.e., child SOM) has L reference vectors. Now, the set of reference vectors in the k-th RSOM module is defined as wk = {xk1, xk2,..., xkl,..., xkL}. In the training of a RSOM×SOM, the following three processes are repeated: (1) evaluative process, (2) cooperative process, and (3) adaptive process. These processes are explained below.

Next, error Eik in each child SOM module for each episode is calculated as

(1) Evaluative process First, error eijkl between each data vector and each reference vector in all child SOMs is calculated as follows: 2

eijkl = w kl - x ij .

Here, the index of the best matching unit (BMU) l is defined as l

42

Articles

(2)

1 J k* å eij J j =1

(3)

where eijk* is the error of BMU lijk* for xij. Thus, error Eik in each child SOM module is the mean of eijk* for all data vectors in one episode. Moreover, the best matching module (BMM) li* for the i-th episode is defined as

ki* = arg min Eik .

(4)

k

(2) Cooperative process In the cooperative process, the learning rates aik and bijl are calculated to decide the update values of all reference vectors. aik is defined as follows:

a ik =

(1) k* ij

lijk * = arg min eijkl .

Eik =

fik åi ' fik'

(5)

( )

æ d k , k * 2ö i ÷ f = expç ç 2s P2 (t ) ÷ ø è k i

(6)


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

æ t ö s P () t = s P min + (s P max - s P min )× expçç - ÷÷ , (7) è tP ø where aik is the learning rate at the parent level, f is a neighborhood function, and d(k,ki* ) represents the distance between the k-th module and the BMM on the map. In addition, sP(t) represents a neighborhood radius, which decreases exponentially with learning step t. In this study, sP(t) is defined as Eq. (7). sPmax and sPmin are the maximum and minimum radii of the neighborhood function, respectively. Moreover, tP is a time constant for decreasing the speed of the neighborhood radius. Next, bijl is defined as follows: l ij

b =

fijl

å

l j ' ij '

f

æ d (l , l **)2 æ ij ÷ f = expç ç 2s C2 (t) ÷ è è l ij

æ t æ s C (t)= s C min + (s C max - s C min )× expçç - çç . è tC è

(9)

(10)

(3) Adaptive process All reference vectors are updated by I

2010

Fig. 6. Upper image is example of panoramic image observed from vision sensor. The episode given to RSOM×SOM is created from the panoramic image.

(8)

bijl is the learning rate at the child level. Note that the distance between the l-th reference vector and the BMU of the BMM on the map is calculated as Eq. (9). This encourages the preservation of homogeneity in the map of each child SOM. sCmax, sCmin, and tC are the maximum radius, minimum radius, and time constant at the child level.

J

w kl = åå a ik b ijl x ij .

N° 2

Next, we explain the episodes given to the RSOM×SOM. An episode Di = {xij} is created from the observed panoramic image. Data vector xij is a color histogram vector extracted from an image (size 64x64) clipped from the panoramic image. For all data vectors, the color histograms are extracted evenly from the entire panorama image. Next, the simulation flow is explained. First, the robot moves around randomly in the environment, while simultaneously, observing the panoramic images at various positions in the working environment. Next, the set of episodes extracted from the panoramic images are given to the RSOM×SOM as training episodes. Here, note that X-Y coordinate is not included in training episodes. The training of the RSOM×SOM is performed offline. After the training process, two places (A) and (B) are verified to confirm whether the RSOM has built an environmental map that can estimate the position (brief X-Y coordinate) and head direction using only vision sensor information from the two types of environments.

(11)

i =1 j =1

In training an RSOM×SOM, the above three processes are repeated.

4. Simulation This section presents the verification results for two types of simulations. The purpose of the simulations is to confirm whether the RSOM can build an environmental map to estimate the position and head direction using only vision sensor information. 4.1. Framework for simulations Using the “Webots” robotics simulation software developed by Cyberbotics Ltd., we created two types of working environments for the simulations. Type 1 is an environment in which four walls are painted red, blue, green, and yellow (illustrated in Fig. 4). Type 2 is a park-like working environment shown in Fig. 5. The area in which the robot is able to move is a meter long by a meter wide (Figs. 4 and 5). Furthermore, the robot has an omnidirectional vision sensor. Fig. 6 is an example of a panoramic image taken from the omni-directional vision sensor. The size of the panoramic image is 512 x 64 pixels. In addition, the colors of the panoramic image are converted to 64 colors (in other words, red, green and blue are converted to 4 colors).

(A)Confirmation of estimating the position The BMM on the RSOM×SOM is monitored when the robot moves to an arbitrary place in the working environments. If the RSOM×SOM can build a map in which the topology of the geography is preserved, then the topology of the robot’s places is almost the same as that of the BMM’s positions. (B)Confirmation of estimating the head direction First, after the robot is put in an arbitrary place, the episode observed from this place is given to the RSOM× SOM. The robot is turned to face north and then, a BMM corresponding to the episode is decided. In addition, a BMU in the BMM is decided after the color histogram vector of the front image (64x64 pixels) is given to the BMM. If the robot is rotated on the spot, the BMU will change continuously on the map of the RSOM. Thus, the map of the RSOM preserves the topology of the direction. In addition, it is expected consistency be maintained in the reference vectors of every module. 4.2. Simulation results in Type 1 environment First, before the training of the RSOMxSOM, the robot moved randomly in the environment, and simultaneously, observed the set of panoramic images. In Fig. 7 the trajectory of the robot is depicted by “-“ , while “ ” denotes the positions at which the panoramic images were obserArticles

43


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

ved. Note that the direction of the robot was not fixed. Panoramic images were observed from 200 positions; in other words, there were 200 training episodes. In addition, there were 32 data vectors per episode.

Fig. 7. The trajectory of the robot at observation of visual information in type 1 environment. The trajectory of the robot is depicted by “-“, while “ ” denotes the positions at which the panoramic images were observed.

N° 2

2010

The results of “(A) Confirmation of estimating the position” are shown in Figs. 8 and 9. Fig. 8 shows the trajectory of the robot. In addition, Fig. 9 shows the RSOM× SOM’s map in which each lattice corresponds to an RSOM module. The episodes observed at positions “A” to “T” in Fig. 8 were given to the RSOM×SOM as test data. Moreover, the lattices denoted by letters of the alphabet in Fig.9 are the BMMs at the positions shown in Fig. 8. Thus, it is possible for an RSOM×SOM to build a map that preserves the topology of the geography. The same result was obtained consistently despite training being repeated several times. The results of “(B) Confirmation of estimating the head direction” are shown in Figs. 10 a), (b), and (c). Each figure is a result at putting the robot on A, F, and J places in Figure 8, respectively. Having been placed at each position, the robot was rotated 360 degrees in intervals of 5 degrees. In Figs. 10a), (b), and (c), the relationship between the head direction of the robot and the BMU is shown. These results confirm that the head direction of the robot and the BMUs change continuously. Moreover, the head direction was able to be estimated by BMU easily. (a)

(b)

Fig. 8. the trajectory of the robot in “(A) Confirmation of estimating the position” in type 1 environment.

(c)

Fig. 9. The result of map building using RSOM×SOM in type 1 environment. The lattices denoted by letters of the alphabet are the BMMs at the positions shown in Fig. 8. 44

Articles

Fig. 10. The result of “(B) Confirmation of estimating the direction” in type 1 environment. (a), (b), and (c) are the relationships between head direction of robot and RSOM’s unit at position A,F, and J in Fig. 8 respectively.


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 2

2010

200 positions; in other words, there were 200 training episodes. In addition, there were 32 data vectors per episode. The results of “(A) Confirmation of estimating the position” are shown in Figs. 12 and 13. These results suggest that an RSOM×SOM is able to build a map that preserves the topology of the geography even if the visual information varies. The results were consistent despite training being done several times. The results of “(B) Confirmation of estimating the head direction” are shown in Figs. 14. Each figure is a result at putting the robot on A, G, and K places in Figure 14, respectively. In the results, BMU corresponding to the head direction of the robot has not been continuously changed. These results suggest that estimation of the head direction was difficult because of the existence of a similar color histogram. (a) Fig. 11. The trajectory of the robot at observation of visual information in type 2 environment.

(b)

Fig.12. the trajectory of the robot in “(A) Confirmation of estimating the position” in type 1 environment.

(c)

Fig.14. The result of “(B) Confirmation of estimating the direction” in type 2 environment. (a), (b), and (c) are the relationships between head direction of robot and RSOM’s unit at position A,G, and K in Fig. 12, respectively.

Fig.13. The result of map building using RSOM×SOM in type 2 environment. The lattices denoted by letters of the alphabet are the BMMs at the positions shown in Fig. 12.

5. Discussion

4.3. Simulation results in Type 2 environment Fig. 11 shows the trajectory of the robot in environment of Fig. 5. The panoramic images were observed at

5.1. Map building in complex environments We have described map building using an SOM2 in a complex environment containing obstacles and similar Articles

45


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 2

2010

vision information. In an adaptation of the SOM2, child SOMs (i.e., place cells) are allocated to unreachable places, since the parent SOM is fixed in the lattice topology. A solution to this problem is to use a self-organizing neural network as the parent SOM, such as the NG and TRN, where the network topology is not fixed. It is shown in [4] that the parent and child of the SOM2 can be designed using any SONN, besides the SOM. It is expected that the place cells are allocated only to the subspace in which input episodes are distributed, by replacing parent SOMs with NGs. Besides, when similar vision information exists, then it is considered that it is necessary to introduce the method of the map building including a time transition to the algorithm of SOM2.

Even if some of the environment change, it is possible to estimate the position and direction with map that was built by training of the SOM2. We experimented as follows to verify the issue. First, the map was built in the type 2 environment. Next, a human’s object was put on the environment (Fig.15). Namely, some of the environment was changed with the human’s object. Final, (A) and (B) of simulation were verified. The results are shown to Fig.16. BMMs on the map changed as shown in Fig.16 when the robot was moved as shown in Fig.12 on the environment. There is little difference between result (Fig.13) of the environment where the human’s object is not put and this result (Fig.16). Thus, it is suggested that a part of change not influence the position estimation.

5.2. Feature extraction from vision information In this study, the color histogram was used as a simple feature extraction, since the study aims to verify the map building by SOM2. It was shown that the map building from only color information was possible by the simulations. However, it is difficult in the real environment to distinguish the local environment from only color information. Therefore, the technique for recognizing the environment such as SIFT [15] is requested to be used as a feature extraction. 5.3. RSOMxSOM’s responses for changing environment.

6. Conclusion In this paper, we confirmed that the SOM2 could build a cognitive map that includes features of the place cells and head direction cells. It was shown that both the position and the azimuth direction could be estimated from the map acquired by unsupervised learning of the SOM2. The SOM2 model is not based on the neurological function of the hippocampus, but is modeled technologically in a topological way. A model that imitates the function of the cognitive map in animals more closely can be developed by creating an algorithm that introduces a time transition of information into the SOM2.

AUTHORS Kazuhiro Tokunaga*, Tetsuo Furukawa - Department of Brain Science and Engineering, Kyushu Institute of Technology, 2-4 Hibikino, Wakamatsu-ku, Kitakyushu 8080196, Japan. E-mails: tokunaga@brain.kyutech.ac.jp, furukawa@brain.kyutech.ac.jp. * Corresponding author

References [1]

Fig. 15. Type 2 environment to which the human is added. [2] [3]

[4] [5] [6]

[7]

Fig.16. The result in type 2 environment to which the human is added. 46

Articles

O’keefe J., Dostrovsky J., “The hippocampus as a spatial map: preliminary evidence from unit activity in the freely moving rat “, Brain. Res., vol. 34, 1971, pp. 171175. O’keefe J. , Nadal L., The Hippocampus as a Cognitive Map, Oxford: Oxford. University Press, 1978. Taube J.S., Muller R.U., Ranck J.B. Jr., “Head-direction Cells Recorded from the Postsubiculum in Freely Moving Rats”, Journal of Neuroscience, vol. 10, 1990, pp. 436447. Furukawa T., “SOM of SOMs”, Neural Networks, vol. 22, issue 4, 2009, pp. 463-478. Kohone T., Self-Organizing Maps, Springer-Verlag: New York, 2001. Gamini Dissanayake M.W., Newman P., Clark S., Durrantwhyte H.F., Csorba M., “A solution to the simultaneous localization and map building (SLAM) problem”, IEEE Transactions on Robotics and Automation, vol. 17, 2001, pp. 229-241. Durrant-Whyte H.F., Bailey T., “Simultaneous Localisation and Mapping (SLAM): Part I The Essential Algorithms”, Robotics and Automation Magazine, vol. 13, 2006, pp. 99–110.


Journal of Automation, Mobile Robotics & Intelligent Systems

[8]

[9]

[10]

[11]

[12]

[13]

[14]

[15]

VOLUME 4,

N° 2

2010

Sim R., Elinas P., Griffin M., “Vision-based SLAM using the rao-blackwellised particle filter”.” In: IJCAI Workshop on Reasoning with Uncertainty in Robotics, 2005. Se S., Lowe D.G., Little J.J., “Vision-based global localization and mapping for mobile robots,” IEEE Transactions on Robotics, vol. 21, Issue 3, 2005, pp. 364375. Yairi T., “Map Building without Localization by Dimensionality Reduction Techniques”. In: The 24th International Conference on Machine Learning (ICML-2007), 2007. Trullier O., Wiener S.I., Berthoz A., Meyer J., Berthelot P.M., “Biologically-based Artificial Navigation Systems: Review and prospects”, Progress in Neurobiology, vol. 51, 1997, pp. 483-544. Martinetz T., Schulten K., “Topology Representing Networks”, Neural Networks, vol. 7, issue 3, 1994, pp. 507522. Takahashi T., Tanaka T., Nishida K., Kurita T., “Self-organization of place cells and reward-based navigation for a mobile robot”. In: Proceedings of ICONIP’01, 2001, pp. 1164-1169. Chokshi K., Wermter S., Panchev C., Burn K., “Image Invariant Robot Navigation Based on Self Organizing Neural Place Codes”, Lecture notes in computer science, vol. 3575, 2005, pp. 88-106. Loww D.G., “Distinctive Image Features from ScaleInvariant Keypoints”, International Journal of Computer Vision, vol. 60, no. 3, 2004, pp. 91-110.

Articles

47


Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 4,

N째 2

2010

A HUMAN ROBOT INTERACTION BY A MODEL OF THE EMOTIONAL LEARNING IN THE BRAIN Satoshi Sonoh, Shuji Aou, Keiichi Horio, Hakaru Tamukoh, Takanori Koga, Takeshi Yamakawa

Abstract: In this paper, we proposed an emotional expression system as a brain-inspired system. The emotional expression was achieved by an Emotional expression Model of the Amygdala (EMA), which was an engineering model inspired by an emotional learning in the brain. EMA can realize both recognition of sensory inputs and a classical conditioning of emotional inputs. Furthermore, a specific hardware of EMA was developed with a massively parallel architecture by using an FPGA, and achieved a calculation speed that is over 20 times faster than an embedded general-purpose computer. Finally, we confirmed an effectiveness of a human-robot interaction with the emotions, which were generated by the proposed emotional expression system. Keywords: emotional learning, amygdala, classical conditioning, human-robot interaction.

1. Introduction Recently, emotions are introduced into intelligence robots [1]. An expression of the emotion has advantageous effects on a communication between a human and a robot, as well as behaviors of the robot. The communication through the expression of the emotions is a nonverbal and intuitive for the human. For example, internal states of the robot, which are not usually visualized, can be communicated to the human by using facial expressions based on the emotions that the robot generates. As conventional methods for the expression of the emotion, artificial intelligence models have emulated some emotions of the robots [2], [3]. However, the expression of the emotion must be programmed with defined inputs and certain situations in advance, and it is not an emergence as a result of interactions between unknown environments. Therefore, a novel model, which can generate wide variety of emotions by learning in the unknown environments, has growing requirements to improve the human robot interaction. Emotional-expression Model of the Amygdala (EMA) has been proposed as an artificial neural network of the amygdala from a viewpoint of an engineering approach [4]. EMA has been established based on neuroscience findings of the amygdala that is an emotional learning system in the brain [5]. The learning of the emotions by EMA is interactively achieved by both recognition and a classical conditioning of inputs from environments. Furthermore, EMA is suitable for applications to the robot, because it has superior recognition abilities compared to other models of the emotion [6], [7]. In this paper, we apply EMA to the expression of the 48

Articles

emotion for an autonomous mobile robot as a brain-inspired system. First, we demonstrate an effectiveness of EMA using a robot simulator. Furthermore, we develop an accelerator of EMA, which allows real-time interactions, using FPGA. Finally, we implement EMA in the robot as an emotional expression system including sensors, expression-devices and the accelerator of EMA. An effectiveness of the developed system is confirmed by an interactive training of the expression of the emotion between the human and the robot.

2. Emotional-expression model of the amygdala 2.1. Amygdala A limbic system of the brain is an information processing system involved in an emotion and a memory. The amygdala, which is a part of the limbic system, involves in the emotional learning. The amygdala receives various sensory stimuli from an inside and an outside of the body via a sensory thalamus [5]. The sensory stimuli are integrated in a lateral nucleus of the amygdala (LA) and are localized and recognized based on the characteristics. Furthermore, a value of the stimulus is evaluated for corresponding emotions in a central nucleus of the amygdala (CE). As a consequence, emotional reactions, such as freezing and stress-hormone release, arise in whole of the body as emotional responses. A relationship between the sensory stimulus and the emotional responses is acquired by a classical conditioning by using the sensory stimulus and the emotional stimulus [8]. The sensory stimulus, such as an auditory stimulus, acts as a conditioned stimulus (CS), and is naturally irrelevant to the emotional responses. On the other hand, the emotional stimulus, such as an electrical shock, is called an unconditioned stimulus (US) since the stimulus potentially generates the emotional responses. The classical conditioning is achieved by simultaneously presenting CS and US. After the conditioning, the emotional responses are induced by observing CS only. The amygdala is related to the conditioning with a fear emotion in particular. 2.2. Architecture of EMA EMA emulates two essential functions in complex and diverse functions of the amygdala, recognition of the sensory stimulus and conditioning to the emotional response. Two functions are absolutely imperative in the emotional human robot interaction. It is preferable that the recognition system is self-organized thorough the interaction. Furthermore, the recognition should be


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 2

2010

Fig. 1. Architecture of EMA. EMA is inspired by anatomical findings of the amygdala and consists of three layers, the sensory input layer, the LA layer and the CE layer. adaptive to environmental changes. In order to realize interactive and adaptive recognition system, we adopted Self-Organizing Map (SOM) [9] and its adaptive learning rule in EMA. The EMA architecture to satisfy two functions is inspired by the anatomical findings of the amygdala. The architecture is three-layers to integrate the sensory stimuli shown in Fig. 1. The sensory input layer has several input units, and corresponds to an entrance area of the amygdala including the sensory thalamus. The LA layer has a number of competitive units, and receives the sensory stimulus. The competitive units are arranged in a two-dimensional array, and can extract characteristics of the sensory stimulus as a feature map in a self-organizing manner. Finally, emotional values, which represent strength of the emotional responses, are evaluated in the CE layer. Several types of emotions, such as fear, pleasure and surprise, are available in EMA although the amygdala is specifically related to the fear emotion. 2.3. Algorithm of EMA Let x(t)=(x1(t),…,xm(t),…,xM(t)) be an input vector that represents the sensory stimulus at time step t, and wi(t)=(w1,i(t),…,wm,i(t),…,wM,i(t)) be a reference vector of the i-th unit on the LA layer. The best matching unit (BMU) for the sensory stimulus is selected by the following equation

BMU (t ) = arg min (errori (t )) i

2

= arg min ( x(t ) - w i (t ) ). i

(1)

EMA regards the BMU as a classified CS of the sensory stimulus. The reference vectors are updated by the following equations.

w i (t + 1) = w i (t ) + a (t )hi , BMU (t ) (x(t ) - w i (t )) hi , BMU

(t )

æ D (i , BMU (t )) 2 = exp çç 2 2ó ( t ) è

ö ÷ ÷ ø

(2)

(3)

Here, hi,BMU(t) means a neighborhood function and D(i,BMU(t)) means a distance function between the ith unit and BMU on the LA layer. á(t) and â(t) are the learning ratio and the neighboring width at time step t, respectively. These parameters in EMA are determined based on an adaptation degree to the sensory stimulus, because an adaptive learning is significant for an interactive learning between the human and the robot. The adaptive learning is achieved by the following equations [11].

á (t ) =

errori (t ) errormax

(4)

ó (t ) = max(á (t ) ó max , ó min )

(5)

Here, errormax is the maximum error until time step t from the initial state, á(t) means a normalized error as the adaptive degree. ómax and ómin are the maximum and the minimum neighboring widths, respectively. These equations mean that the learning ratio and the neighboring width increase when the normalized error is large, but these decrease when the normalized error is small. Furthermore, a relationship between the sensory stimulus and the emotional stimulus is acquired by using a learning model of the classical conditioning [10]. Let E(t)=(E1(t),…,Ek(t),…,EK(t)) be an input vector that represents strength of the emotional stimulus, V(t)=(V1(t),…,Vk(t),…,VK(t)) be an output vector of the emotional value that represents strength of the generated emotional responses at the time step t, where the suffix k means an index corresponding to a kind of the emotion. The emotional value is calculated by the following equations, when the sensory stimulus is presented.

V (t ) = å acti (t )u(t ),

(6)

t

acti (t ) =

hi , BMU (t )

åh j

.

(7)

i , BMU ( t )

Here, acti(t) is a normalized activity of the i-th unit to Articles

49


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

the sensory stimulus and ui(t)=(u1,i(t),…,uk,i(t),…, uK,i(t)) is an emotional weight of the i-th unit. The emotional weight is updated by the following equation.

u i (t + 1) = u i (t ) + äacti (t )(E(t ) - V (t )) .

3. Computational simulations 3.1. Simulation environment An experiment of the expression of the emotion for a dog-like robot is performed by computational simulations. A simulation environment is created by a robotics simulator “Webots” [12], and is shown in Fig. 2. The simulated robot (SONY AIBO ERS210) in the environment has a vision sensor and a touch sensor. The robot can detect a color intensity of the front ball as the sensory stimulus by using the vision sensor. Furthermore the robot sometimes receives tactual stimulus such as hitting and gentle stroking from the environment, when the sensory stimulus is presented. Here, we assume that the tactual stimulus induces a corresponding emotional response to the robot as the emotional stimulus. EMA implemented in the robot performs the recognition and the conditioning for the emotional learning like the amygdala.

Fig. 2. A simulation environment. The robot has a vision sensor to detect the sensory stimulus and has a touch sensor to detect the emotional stimulus. The sensory stimulus is a color of ball objects and the emotional stimulus is a tactual sense of the robot. The sensory stimulus is represented as a three-dimensional vector x(t)=(IR , IG , IB), where IR, IG and IB are color intensities of red, green and blue, respectively. Articles

2010

Strength of the emotional stimuli is represented as E(t)=(Ep , Ef). In this simulation, the parameters of EMA are as follows; the number of the LA unit is 64 (8 x 8), ómax=1.0, ómin=0.1, ó=0.3.

(8)

Here, ä is a conditioning ratio. The algorithm of EMA is summarized two computational processes; (1) the stimulus recognition process and (2) the conditioning process. Recognition of the sensory stimulus can be achieved in the stimulus recognition process. In parallel, a prediction and an update of the emotional value can be achieved in the conditioning process. In application such as human robot interaction, an advantage of EMA over other classical conditioning models, for example TD model [11], is the self-organizing and adaptive recognition of the sensory stimulus (See [4]). We confirm an effectiveness of EMA by software simulations and experiments with developed hardware of EMA in the following section.

50

N° 2

3.2. Recognition of the sensory stimulus At the beginning, a performance of the stimulus recognition was confirmed. We randomly presented the color ball to the robot. Fig. 3(a) shows the feature map of EMA after presenting 1000 times. In the feature map, each color patch shows features of the sensory stimulus that are represented by the reference vector of the LA unit. The neighboring LA units have similar feature, and the distant LA units have different features, for For BMUR, BMUG and BMUB, R, G and B are subfix. Although uniformed feature map was obtained by the randomly sensory stimuli, we presented red-biased stimuli in order to confirm effectiveness of the adaptive learning rule, additionally 500 times. The feature map was changed and specialized in red feature by additional stimulus, as shown in Fig. 3(b). The feature map is updated depending on the adaptive degree to the sensory stimulus. Thus, the recognition of EMA works well even if the environment is dynamically changed, although the conventional model [11] cannot accommodate.

Fig. 3. A feature map in EMA obtained by the stimulus recognition process. Each color patch represents the reference vector of unit on the LA layer. (a) The obtained map by presenting 1000 random stimuli. (b) The map adapted additional stimuli (red color).

Fig. 4. Emotional values in the basic classical conditioning experiment. (a) The emotional value to fear emotion. (b) The emotional value to pleasure emotion. Emotional values represent strength of emotional responses to the sensory stimulus. EMA can achieve the acquisition and the extinction for more than one emotion. 3.3. Emotional conditioning Next, the classical conditioning experiment was


Journal of Automation, Mobile Robotics & Intelligent Systems

performed by simultaneously presenting both the sensory stimulus and the emotional stimulus. The emotional stimulus, E=(1,0), was associated with the sensory stimulus, x=(1,0,0), every 20 steps in the first 500 steps. In the next 500 steps, the emotional stimulus, E=(0,1), was associated with the same sensory stimulus every 20 steps. Fig. 4 shows the emotional values in the learning steps, where V1 and V2 correspond to fear and pleasure emotions, respectively. Emotional values were acquired by the classical conditioning. As a result, the emotional responses to the sensory stimulus were generated without the emotional stimulus. In the last of the conditioning, the emotional value (V1) was eventually lost because the corresponding emotional stimulus was not presented. Thus, the “acquisition” and the “extinction”, which are a basic principle of the classical conditioning, can be achieved in EMA. In this simulation, it was confirmed that the robot with EMA was able to generate the emotions by a combination of the recognition and the conditioning. This con-

VOLUME 4,

N° 2

2010

tribution by EMA makes behaviors of robots more intelligent, and the human robot interaction becomes natural and interactive.

4. FPGA implementation of EMA 4.1. Architecture of EMA hardware To realize real-time processing of EMA in robotics applications, we proposed specific hardware architecture of EMA in which a hardware-oriented algorithm is employed. The proposed EMA hardware (EMAHW) was developed based on a massively parallel architecture as well as conventional SOM hardware [14], [15] because the algorithm of EMA was based on SOM. Fig. 5 shows the massively parallel architecture of EMAHW including 81 units. The architecture is achieved by several local circuits and one global circuit. Each local circuit corresponds to one LA unit, and has a memory of the reference vector and the emotional vector. The global circuit calculates following processes; finding BMU, adapting the learning parame-

Fig. 5. A massively parallel architecture of the EMA hard-ware. The EMA hardware consists of several local circuits and one global circuit.

Fig. 6. A block diagram of the emotional expression system including the digital hardware of EMA, the emotional sensors and the emotional expression devices. Articles

51


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

ters, calculating the emotional value and conditioning. Furthermore, the learning parameters (Eq. (3) and Eq. (4)) were modified in EMAHW as follows;

hi , BMU (t ) Î 2 - n

á (t ) Î 2 - n

(n = 0,1,..., N ) ,

(n = 0,1,..., N ) .

(9) (10)

N° 2

2010

most famous pet, and their emotional expressions have been investigated in detail. The robot’s emotions, which are generated by EMAHW, are expressed by simple movements of the emotional-expression devices. For example, the ear is laid back and the tail is wagged in small motions, when the robot is feeling a fear emotion. In addition, pleasure, confusion and attention emotions have been elaborated.

These learning parameters allow to EMAHW use shift register as a substitute for multiplier. Thus, it drastically reduces a circuit area and calculation cost. 4.2. Performance of EMA hardware EMAHW was implemented on FPGA (Xilinx Spartan-3E, sc3s1600E) with 81 LA units and 8-bit accuracy. EMAHW calculated at clock frequency up to 50 MHz. A performance of EMAHW was estimated at 632.8 MCUPS (Million Connection Update Per Second), when the reference vectors were three dimensions and the emotional weights were two dimensions. For comparison, a performance of a general-purpose PC (Intel Core 2 Duo, 2.66GHz) is 43.2 MCUPS in the original algorithm. Furthermore, a performance of EMAHW was confirmed by comparison with embedded processors that were available an autonomous mobile robot. A benchmark test using three-dimensional 1000 sensory stimuli was performed by each processor. Table 1 shows a comparative result. The calculation speed of the EMAHW was twentytimes or more as fast as the portable PC in spite of the lowest clock frequency.

Fig. 7. The robot “WITH” equipped the emotional expression system.

5. Human robot interaction with EMA 5.1. Robot with emotional expression system To verify an effectiveness of EMA in the human robot interaction, an emotional expression system was implemented in an autonomous mobile robot “WITH” [16]. A block diagram of the emotional expression system is shown in Fig. 6, and the robot including the sensors and the devices is shown in Fig. 7. The emotional sensors include a CMOS camera to capture front images of the robot, and a capacitance sensor array to detect tactual senses from the human to the robot. The main controller receives the captured image from the CMOS camera, and sends an averaged the color value as the sensory stimulus to EMAHW. Furthermore, the main controller estimates tactile information, hitting or gentle stroking, by using the number of responded sensors, and sends the information as the emotional stimulus to EMAHW. EMAHW calculates the emotional value by using the sensory and emotional stimulus provided. The emotional-expression devices were developed based on an ear and a tail of a dog in order to communicate the robot's emotions to human. The dog is the

Fig. 8. The experimental results of the human robot interaction. Here, the colored marker is the sensory stimulus as CS and the touch to the robot is the emotional stimulus as US. (a) A scene that the human is training the robot with the red maker and gentle stroking. (b) A scene that the robot is expressing the pleasure emotion to the only red marker. (c) The feature map obtained by the interaction. (d) Emotional values acquired by the interaction.

Table 1. A comparison of calculation time to benchmark test between EMAHW and other embedded processor.

52

Articles


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

5.2. Experiment in human robot interaction The emotional expression system including EMAHW was implemented in the robot as a robot’s amygdala system. As a result, the robot got to recognize the sensory stimulus from the environment, and to generate the emotional value and the expression of the emotion. The human robot interaction was perfumed in the real environment. In the interaction, the human as a trainer presented a colored-marker as the sensory stimulus, and touched with one’s hands as the emotional stimulus to the robot. Here, the sensory stimulus as CS is a presentation of colored-marker, and the emotional stimulus as US is a touch of the robot with gentle stroking or hitting. At the begging, the robot unconditionally responded using the specific emotion to the touch, but the colored-marker did not induce any emotions. Fig. 8 (a) shows a scene that the human was training the robot with the red marker and the gentle stroking. As the interaction was repeated, the robot became to express the emotion to the red marker only. Fig. 8 (b) shows a scene that the robot was expressing pleasure emotion using the tail wagging to the red marker without the touches after the training. Fig. 8 (c) shows the obtained feature map in EMAHW. The feature map preserves the topology of the sensory stimulus as well as the result of the computational simulation. Fig. 8 (d) shows the emotional value corresponding to the interaction step. The red and blue lines represent the emotional value of fear and pleasure emotion, respectively. In the real environment, the acquisition and the extinction can be successfully achieved. In the human robot interaction experiment, the robot recognized the sensory stimulus, and predicted the emotional value as well as animals. The emotional expression of the robot makes the interaction to the human more intelligent and human-friendly.

AUTHORS Shuji Aou, Keiichi Horio, Takeshi Yamakawa - Department of Brain Science and Engineering, Kyushu Institute of Technology, 2-4 Hibikino, Wakamatsu-ku, Kitakyushushi, 808-0196 Japan. E-mails: aou@brain.kyutech.ac.jp, horio@brain.kyutech.ac.jp, yamakawa@brain.kyutech.ac.jp. Satoshi Sonoh*- Corporate Research & Development Center, Toshiba Corporation, 1 Komukai-Toshiba-cho, Saiwai-ku, Kawasaki-shi, JAPAN. E-mail: satoshi.sonoo@toshiba.co.jp. Hakaru Tamukoh - Institute of Symbiotic Science and Technology, Tokyo University of Agliculture and Technology, 2-24-16 Nakamachi, Koganei-shi, 183-8509 Japan. Email: tamukoh@cc.tuat.ac.jp. Takanori Koga - Fuzzy Logic Systems Institute, 680-41 Kawazu, Iizuka-shi, 820-0067, Japan. E-mail: koga@flsi.or.jp. * Corresponding author

References [1] [2]

[2]

[4]

[5] [6]

[7]

[8]

ACKNOWLEDGMENTS This work was supported by a 21st Century Center of Excellence Program, ”World of Brain Computing Interwoven out of Animals and Robots (PI: T. Yamakawa),” granted in 2003 to the Department of Brain Science and Engineering (Graduate School of Life Science and Systems Engineering), Kyushu Institute of

2010

Technology by Japan Ministry of Education, Culture, Sports, Science and Technology. This work is also supported by Grantin-Aid for Scientific Research (A), 2006, 18200015.

6. Conclusion In this paper, we implemented EMA in the simulated and real robot in order to realize the expression of the emotions from the interaction. The expression of the emotions can be achieved by the recognition of the sensory stimulus and the classical conditioning in EMA. Furthermore, we proposed the emotional expression system including the accelerator of EMA. The EMA hardware has the computational ability, which is 20 times or more fast than other embedded processors. The human robot interaction with the expression of the emotion can be achieved in simulated and real environment. In the future work, a system that estimates a stimulus involving the expression of the emotion by consideration of internal and contextual status of the robot is needed as an extension of EMA. Then, we believe that the brain inspired system will achieve a breakthrough in the human robot interaction.

N° 2

[9] [10]

Breazeal C.L., Designing Sociable Robots, Cambridge, MA: The MIT Press, 2002. Fujita M., Kuroki Y., Ishida T., Doi T., “Autonomous Behavior Control Architecture of Entertainment Humanoid Robot SDR-4X”. In: Proc. of 2003 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2003, pp. 960-967. Sawada T., Takagi T., Fujita M., “Behavior Selection and Motion Modulation in Emotionally Grounded Architecture for QRIO SDR-4X II”. In: Proc. of 2004 IEEE/RSJ International Conference on Intelligent Robots and Systems, 2004, pp. 2514-2519. Sonoh S., Horio K., Aou S., Yamakawa T., “An Emotional Expression Model Inspired by the Amygdala”, International Journal of Innovative Computing, Information and Control, vol. 5, 2009, no. 5, pp. 1147-1160. Aggleton J.P., The Amygdala: A Functional Analysis, New York: Oxford University Press, 2000. Armony J.L., Servan-Schreiber D., Cohen J.D., LeDoux J.E., “An anatomically constrained neural network model of fear conditioning”, Behavioral Neuroscience, vol. 109, 1995, no. 2, pp. 246-257. Mor'en J., Balkenius C., “A computational model of emotional learning in the amygdala”. In: Proc. of the 6th International Conference on the Simulation of Adaptive Behavior, 2000, pp. 383-391. Phelps E.A., LeDoux J.E., “Contributions of the amygdala to emotion processing: from animal models to human behavior”, Neuron, vol. 48, 2005, pp. 175-187. Kohonen T., Self-Organizating Maps, Berlin: SpringerVerlag, 1997. Rescorla R.A., Wagner A.R., “A theory of Pavlovian conditioning: variations in the effectiveness of reinforcement and nonreinforcement”, Classical conditioning II: Articles

53


Journal of Automation, Mobile Robotics & Intelligent Systems

[11]

[12] [13]

[14]

[15]

[16]

54

current research and theory, New York: Appleton-Century-Crofts, 1972, pp. 64-99. Sutton R.S., “Learning to predict by the methods of temporal difference learning”, Machine Learning, vol. 3, 1988, no 1, pp. 9-44. Cyberbotics Webots “http://www.cyberbotics.com/” Berglund E., Sitte J., “The parameterless self-organizing map algorithm”, IEEE Transaction on Neural Networks, vol. 17, 2006, pp. 305-316. Tamukoh H., Horio K., Yamakawa T., “Fast Learning Algorithm for Self-Organizing Map Employing Rough Comparison WTA and its Digital Hardware Implementation”, IEICE Trans. on Electronics, vol. E87- C, 2004, no. 11, pp. 1787-1794. Hikawa H., “FPGA implementation of self organizing map with digital phase locked loops”, Neural Networks, vol. 18, 2005, pp. 514-522. Takemura Y., Sato M., Ishii K., “Toward Realiation of Swarm Interlligene Mobile Robots”, Brain-Inspired IT II, Amsterdam: Elsevier B.V., 2006, pp. 273-276.

Articles

VOLUME 4,

N° 2

2010


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 2

2010

THE STUDY OF BIO-INSPIRED ROBOT MOTION CONTROL SYSTEM

Takayuki Matsuo, Takeshi Yokoyama, Daishi Ueno, Kazuo Ishii

Abstract: Robots and robotics technologies are expected as new tools for inspection and manipulation. The dynamics of robot always are changed by environment and robot of state in mission. Therefore, an adaptation system, which is able to switch controller due to environment and robot of state, is needed. Meanwhile, animals are able to go through several environments and adapt several own states. The adaptation system is realized Central Pattern Generator (CPG). CPG exists in nervous system of animals and generates rhythmical motion pattern. In this paper, a robot motion control system using CPG is proposed and applied to an amphibious multi-link mobile robot. Keywords: CPG, snake-like robot, biomimetics.

1. Introduction Robot and robotics technologies are expected to provide new tools for inspection and manipulation. The dynamics of Robot is changed by environment and state of robots: weight, attitude and so on. Operations in several environments and conditions of robot need switching system, which is able to change due to a variety of circumstances. The system, which is worked, is called “Hybrid Dynamical System (HDS)”. HDS is able to manage many-controlled object. Therefore, one is very complex theory. But, in the nature, Animals are able to move and adapt between different environments. Adaptation system of animal is realized by Central Pattern Generator (CPG). CPG exists in nervous system of animal and generates rhythmical pattern. Several motion patterns of animals such as swimming, waking, flapping and so on are generated by CPG. CPG has “entrainment feature” which synchronizes a wave that matches the resonance frequency. The wave of CPG is able to be adjusted using sensory feedback. Taga [1] realized simulation of biped walking robot based on the idea that motion pattern is generated by interaction between CPG, body-dynamics and environment. Williamson [2] applied CPG to a humanoid robot that performs cranking, sawing, and hitting drum. Matsuoka et al. [3] developed control system of a giant swing robot which is switched among swing mode and rotation mode using CPG. The goal of our study is development of bio-inspired robot control system using CPG and application into HDS. In this paper, we report about development of an amphibious multi-link mobile robot for test bed of HDS and dynamics property analysis. Additionally, distributed motion control system using entrainment feature of CPG without sensory feedback was developed.

2. Development of Amphibious Multi-Link Mobile Robot “AMMR” 2.1. Motion of snake-like animals In previous works, Hirose et al. developed multi-link mobile robot [4] and underwater multi-link mobile robot [5]. These robots move forward using reaction forces from frictional force on ground or fluid drag force in underwater. Robot dynamics are different between ground motion and underwater motion. Especially, an amphibious multi-link mobile robot[6] is very interesting for HDS to apply CPG. In this research, we applied motion control system using CPG to amphibious multi-link mobile robot. The motion mechanisms of snake-like animals have already been studied by Hirose [7] and Azuma [8]. A snake’s body is covered with special scales that have low friction in the tangential direction and high friction in the normal direction. This feature enables thrust to be produced from a wriggle motion. An eel swims under water by generating an impellent force from a hydrodynamic force. Snake and eel generate impellent force actuating each joint with a certain phase difference. And, at turning motion, joint trajectories of snake and eel have bias, which is balance of oscillation of joint. 2.2. Mechanism In our previous work [9], we developed the multi-link mobile robot, as a test bed for the evaluation of a motion control system using a CPG. We realized wriggle motion for forward and turning motions using periodical output signals of the CPG control system. However, the multilink mobile robot was not able to evaluate HDS in which the dynamics of a robot transfer from one mode to another because previous one was not able to move in underwater. In previous works, electrical circuit did not have feedback connection because servomotors, which were not able to extract angle data, were used. Thus we developed an amphibious multi-link mobile robot (see Fig. 1) as a new test bed for motion control and adaptation in two different environments that require different dynamics: land and underwater environments. Table 1 gives the specifications of an amphibious multi-link mobile robot. An amphibious multi-link mobile robot moves both over land and under water. Therefore, waterproofness is an important design consideration. O-rings are employed on the shaft of each joint and under the cylinder lids to ensure the waterproofness of the robot. The robot comprises eight cylinders that are joined so that each cylinder can rotate around a yaw axis via DC motors, a gearbox and a control circuit, as shown in Fig. 2. A range of joint movement is ±ð/3[rad]. Hydrodynamic forces produced Articles

55


Journal of Automation, Mobile Robotics & Intelligent Systems

by fins and the body produce thrust forces under water and passive wheels are used on ground (Fig. 3).

Fig. 1. Overview of AMMR.

VOLUME 4,

N° 2

2010

2.3. Electrical system An amphibious multi-link mobile robot consists of two kinds of cylinders: a cylinder for the head (head module) and seven cylinders for the body (motor modules). Main voltage for each module is 8 V and decrease voltage into 5 V for MPU and sensors using DC/DC converter. However, connection method of power line is a cascade connection. Therefore, voltage decreases from tail to head module. One cannot evaluate torque of each joint in this condition. As main power, 24 V are used and voltage is stepped down to 8 V in each module using switching regulator, which is able to apply high electrical current. A motor module has a motor to control the joint angle and a circuit board. The circuit has a microprocessor unit (MPU, PIC18F452), a potentiometer to measure the joint angle, an RS485 transceiver (MAX1487) for communication and a current sensor to measure joint torque. The MPU calculates the target trajectory using the neuron potential of the CPG or sinusoidal function, controls the motor using a PID control, manages sensor information (e.g., current and angle data) and communicates with circuits of other modules using the RS485. The head module transfers the target behavior to other modules and control of communication.

3. The experimental analysis of dynamics property for “AMMR” Fig. 2. Internal architecture of a motor module.

Fig. 3. Passive Wheels and Fins. Table 1. Specification of robot.

56

Articles

3.1. Method of dynamics property analysis In this paper, dynamics property analyses are carried out in 120 motion patterns which are expressed by combination of amplitude a (between 10 deg and 60 deg with a step of 10 deg), frequency f (between 0.1 Hz and 0.4 Hz with a step of 0.1Hz) and phase difference ö (between 20 deg and 100 deg with a step of 20 deg) on ground and in underwater. For dynamics property analyses, travel velocity v m/s and dissipation power per a meter ÄW J/m of robot analyses were evaluated by average value of 3 trials in each motion pattern. And a motion capture system was used for measurements of robot motion. 8 markers are put on head, tail and joints of robot and these markers were tracked by a motion capture system. Travel distance of head and tail markers in a straight line dh, dt m were measured and average of dh and dt was defined as travel distance d m of robot. Travel velocity v m/s is evaluated from travel distance d and operation time t=20 s.


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

Dissipation powers per travel distance of a meter ÄW J/m were evaluated from total dissipation power W J in operation time t=20 s and travel distance d m as shown in Eq. (2) and total dissipation power were defined as Eq. (1). V V and I(n) A shows impressed voltage and carried electrical current of motor per sampling time Ät=0.05 s.

N° 2

2010

3.2. Velocity analysis on ground and underwater Fig. 4a)-d) and Fig. 5a)-(d) show velocities of robot on ground and in underwater by adjustment of a deg and ö deg at four fixed frequencies of f=0.1, 0.2, 0.3 Hz and 0.4 Hz. Velocities were expressed by brightness and the brighter grids are the faster velocity. In ground motions, velocities were fast in white grids and fastest velocity was v=0.54 m/s at f=0.4 Hz, a=30 deg, ö=40 deg]. Travel velocities v m/s were fast at t a=30, 40 deg by adjusting a deg and by adjusting phase difference ö deg, travel velocities v m/s were fast at ö=40 deg.

Meanwhile, in underwater motion, fastest velocity was v=0.10m/s in white grids. Travel velocities v m/s are fast at a=50 deg by adjusting a deg and by adjusting phase difference ö deg, travel velocities v m/s were fast at ö=60 deg. The parameters, which generate maximum velocity, were difference between ground motion and underwater motion. Robot was not able to move in parameter which include a=10 deg and ö=20 deg. Next, we focused on variations of velocities. Fig. 6a)c) shows examples of variations of velocities in ground motions. Fig. 6a) shows variations of velocities by adjustment of ö deg and a deg at fixed parameter of f=0.4 Hz, Fig. 6b) shows variations of velocities by adjustment of f Hz and a deg at fixed parameter of ö=40 deg and Fig. 6c) show variations of velocities by adjustment of a deg and f Hz at fixed parameter of ö=40 deg. Fig. 7a)-c) shows examples of variations of velocities in underwater motion at same parameters as ground motions. On ground, velocities are changed by adjustment ö and f, however velocities were not changed or changed a few by adjusting a between a=40 deg and a=60 deg. Velocities became max speed at ö=40 deg and f=0.4 Hz. These results show that adjustment of ö and f are more effective than adjustment of a in control of velocity on ground. Robot snakes its way, therefore, definition of velocity is different whether to define the distance in straight line which robot moves

Fig. 4. Variations of travel velocity in ground motions (a) f=0.1[Hz] (b)f=0.2[Hz] (c)f=0.3[Hz] (d)f=0.4[Hz].

Fig. 5. Variations of travel velocity in underwater motions (a) f=0.1[Hz] (b)f=0.2[Hz] (c)f=0.3[Hz] (d)f=0.4[Hz].

(1)

(2)

Fig. 6. Variations of travel velocity on Ground. Articles

57


Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 4,

N° 2

2010

Fig. 7. Variations of travel velocity in Underwater. as travel distance or define the distance which robot snakes its way as one. In this paper, the distance in straight line which robot moves is defined as travel distance. Therefore, velocities were not changed so much by adjustment of a. In underwater, velocities are changed by adjustment ö, f and a. Velocity became max speed at ö=60 deg and f=0.3 Hz. In underwater, robot is affected from fluid drag force. Fluid drag force becomes big value if velocity and acceleration of robot are big value. Therefore, robot velocity decreased in high frequency (f=0.4 Hz).

4. Motion control system using CPG

Dissipation power analysis on ground and underwater Fig. 8a)-d) and Fig. 9a)-d) shows results that dissipation power per travel distance of a meter ÄW J/m was analyzed. In the analysis, ÄW is shown by brightness like velocity analysis. The brighter grid is the lower value. In the result of analysis, on ground, the lowest value was 23.49 J/m at a=30 deg, f=0.4 deg, ö=80 deg. Meanwhile, in underwater, the lowest value was 103.74 J/m at a=60 deg, f=0.2 deg, ö=60 deg. Fig. 10a)-c) and Fig. 11a)-c) show examples of variation of dissipation power per travel distance of a meter. Fig. 10a)-c) and Fig. 11a)-c) don’t include data at a=10 deg because robot was not able to go forward so much. Overall, underwater motions need more power than ground motions. As shown in Fig. 10a) and

4.1. CPG Models Several CPG models are proposed such as Von der Pol [10] model, Matsuoka model [11],[12], Terman-Wang model [13], Willson-Cowan model [14], and so on. We employed Matsuoka model. The Matsuoka model is expressed in Eqs. (3)–(5), where ui is the membrane potential of the i-th neuron, vi represents the degree of adaptation, u0 is the external input with a constant rate, fi is the feedback signal from a sensory input, ôu, ôv and â are parameters that specify the time constant for the adaptation, wij is the neuron weighting, yi is output of neuron and n is the number of neurons. The CPG wave in the Matsuoka model can entrain external oscillatory input using sensor input fi. The neural oscillator wave is able to entrain a wave that matches the resonance frequency. Additionally, the wave in the Matsuoka model is able to change amplitude and bias by adjusting u0.

Fig. 8. Variations of dissipation power in ground motions (a)f=0.1[Hz] (b)f=0.2[Hz] (c)f=0.3[Hz] (d)f=0.4[Hz].

Fig. 9. Variations of dissipation power in underwater motions (a) f=0.1[Hz] (b)f=0.2[Hz] (c)f=0.3[Hz] (d)f=0.4[Hz].

3.3

58

Fig. 11a), if phase difference ö is adjusted from 20 to 100 deg, ÄW becomes minimum value at =60 deg on ground and =80 deg in underwater. As shown in Fig. 10b) and Fig. 11b), if frequency f is adjusted from 0.1 to 0.4 Hz, ÄW becomes minimum value at f=0.3 Hz on ground and f=0.4 Hz in underwater.

Articles


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 2

2010

Fig. 10. Variation of Dissipation Power on Ground.

Fig. 11. Variation of Dissipation Power in underwater. In this paper, we realized wave generation, which has phase difference and shift neutral position of wave oscillation adjusting CPG parameters such asôu, ôv and u0 are used for motion control. (3) (4) (5) 4.2. Design of Motion Control System Using CPG The CPG for the multi-link mobile robot is shown in Fig. 12. A neural oscillator consists of an extensor neuron (ENk) and a flexor neuron (FNk), which connects each other through the weights of wef and wfe. k means neural oscillator number. Extensor neurons are connected to flexor neurons of the neighboring neural oscillator, and flexor neurons are connected to extensor neurons of the neighboring neural oscillator. We defined weights between neural oscillators as w1 and w2. And, an external input of extensor neuron and flexor neuron are defined as u0_e and u0_f. A set of neural oscillator is assigned to each of our robot’s seven joints. Taga[15],[16] simulated biped walking robot connecting neural oscillator to musculoskeletal model. In this paper, we used same method as neural oscillator model. The output of neural oscillator Ok follows Eq. (6). The network architecture is designed as a closed loop to generate periodically successive signals with a certain phase. After CPG simulations, we employed sets of CPG parameters for generating waves with a certain phase difference and CPG waves are converted into target joint angles following Eq. (7). Table 2 gives

the parameters for forward motion and Fig. 13 shows target joint angles, which are made by CPG wave. The amplitude a of target joint angles are 40 deg and frequency f is 0.1 Hz. We determined the parameters through trial and error using Matlab. We simulated and calculated equation of Matsuoka model adjusting parameters for finding parameter sets, which are able to generate waves. And we determined parameter sets which are able to realize a certain amplitude, phase difference and frequency. Amplitude and frequency of target joint angles are linked to robot’s speed. If amplitude and frequency are set large value, robot’s speed becomes fast. On the contrary, if their values are set to be small, robot’s speed becomes slow. The motion control system changes frequency adjusting ôu and ôv and changes amplitude adjusting u0_e and u0_f. Table 3 shows the relationship between ôu, ôv and frequency. Other parameters are same with Table 2. Table 4 shows the relationship between u0_e, u0_f and amplitude. Other parameters are same with Table 2. The amphibious multi-link mobile robot can change direction with a change in parameter u0_e and u0_f, which shifts the neutral position. If parameters are set u0_e =0.99 and u0_f =0.94, the robot turns and moves toward the right. Figure 14 shows the output of the CPG network, which has bias 15 deg for a right turn motion. If u0_e and u0_f are adjusted, bias is changed. Table 5 shows the variations in bias. (6) (7)

Articles

59


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N째 2

Fig. 12. A CPG architecture of robot.

Fig. 13. Output of CPG for forward motion.

Fig. 14. Output of CPG for right turn.

Table 2. Parameters of CPG in forward motion.

Table 3. Parameters for adjustment of frequency.

Table 4. Parameters for adjustment of amplitude.

Table 5. Parameters for adjustment of bias.

60

Articles

2010


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

5. Conclusions In this paper, an amphibious multi-link mobile robot that has seven joints and moves on ground and under water was developed. And dynamics property analysis was carried out on ground and in underwater. Velocity was max speed at a=30 deg f=0.4 Hz =40 deg on ground and at a=40 deg f=0.3 Hz =60 deg in underwater. Additionally, a distributed robot motion control system using a CPG was developed and realized forward and turning motion. In future work, we will include a sensor feedback in the bio-inspired robot motion control system.

[12]

[13]

[14]

[15]

ACKNOWLEDGMENTS This work was supported by a 21st Century Center of Excellence Program, “World of Brain Computing Interwoven out of Animals and Robots (Pl: T. Yamakawa)” (center#J19) granted to the Kyushu Institute of Technology by the Ministry of Education, Culture, Sports, Science and Technology of Japan.

[16]

N° 2

2010

tually inhibiting neurons with adaptation”, Biological Cybernetics, vol. 52, 1985, pp. 367-376. Matsuoka K., “Mechanisms of frequency and pattern control in the neural rhythm generators”, Biological Cybernetics, vol. 56, 1987, pp. 345-353. Terman D., Wang D.L. ,”Global competition and local cooperation in a network of neural oscillators”, Physica D81, 1995, pp. 148-176. Wilson H.R., Cowan J.D., “Excitatory and inhibitory interactions in localized populations of model neurons”, Biological Journal, vol. 12, 1972, pp. 1-24. Taga G., Yamaguchi Y., Shimizu H., “Self-organized control of bipedal locomotion by neural oscillators in unpredictable environment”, Biological Cybernetics, vol. 65, 1991, pp. 147-159. Taga G., “A model of the neuro-musculo-skeletal system for human locomotion II. Real-time adaptability under various constraints”, Biological Cybernetics, vol. 73, 1995, pp. 113-121.

AUTHORS Takayuki Matsuo*, Daishi Ueno, Kazuo Ishii - Department of Brain Science and Engineering, Kyushu Institute of Technology, 2-4 Hibikino, Wakamatsu, Kitakyushu, Fukuoka, 808-0196, Japan. Tel&Fax: +81-93-695-6102. E-mail: t-matsuo@brain.kyutech.ac.jp. Takeshi Yokoyama - YASUKAWA Electric Co. 2-1, Kurosaki-shiroishi, Kitakyushu, Fukuoka, 806-0004, Japan. * Corresponding author

References [1] [2]

[3]

[4]

[5]

[6]

[7] [8] [9]

[10] [11]

Taga G., “Emergence of Locomotion”, J. of R.S.J., vol. 15, no. 5, 1997, pp. 680-683 (in Japanese). Williamson M.M., Robot Arm Control Exploiting Natural Dynamics, PhD thesis, Massachusetts Institute of Technology, 1999 Matsuoka K., Ohyama N., Watanabe A., Ooshima M., “Control of a giant swing robot using a neural oscillator”, Advances in Natural Computation (Proc. ICNC 2005, Part II), 2005, pp.274-282. Mori M., Yamada H., Hirose S., “Design and Development of Active Cord Mechanism ”ACM-R3” and its 3dimentional Locomotion Control”, J. of R.S.J., vol. 23, no. 7, 2005 , pp. 886-897 (in Japanese). Takayama T., Hirose S., ”Study on 3D Active Cord Mechanism with Helical Rotational Motion”, J. of R.S.J., vol. 22, no.5, 2004, pp.625-635 (in Japanese). Chigasaki S., Mori M., Yamada H., Hirose S., “Design and Control of Amphibious Snake-Like Robot “ACM-R5””. In: Proceedings of the 2005 JSME Conference on Robotics and Mechatronics, ALL-N-020(1)-(3). Hirose S., “Bionic machine engineering”, Kougyo Chosakai, 1987 (in Japanese). Azuma A., The subject-book of Animal’s Motion, Asakura, 1997. Matsuo T., Yokoyama T., Ishii K., “Development of Neural Oscillator Based Motion Control System and Applied to Snake-like Robot”. In: IROS’07, 2007, pp. 3697-3702. Val der Pol B., “On relaxation oscillations”, Phil. Mag., vol. 2, no.11, 1926, pp. 978-992. Matsuoka K., “Sustained oscillations generated by muArticles

61


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 2

2010

DEVELOPMENT OF ANTAGONISTIC WIRE-DRIVEN JOINT EMPLOYING KINEMATIC TRANSMISSION MECHANISM Takashi Sonoda, Yuya Nishida, Amir Ali Forough Nassiraei, Kazuo Ishii

Abstract: Antagonistic mechanisms attract attentions as joint actuators of linkage mechanisms, which control output torque, joint stiffness and position simultaneously. As the actuators or components of antagonistic driven joints, special devices with nonlinear elasticity property such as pneumatic actuators, nonlinear springs are often utilized to satisfy the requirements of antagonistic mechanisms. However, these devices have difficulties in control caused by complex and nonlinear properties, downsizing of actuator, and response time of articular compliance. In order to solve these problems, we propose a new antagonistic joint mechanism using kinematic transmission mechanism (KTM), which is composed of links and cams with dedicated design. The performance of KTM is evaluated through stiffness and position control simulations and experiments. Keywords: antagonistic-driven joints, equilibrium point hypothesis, controllable stiffness mechanisms.

1. Introduction Main applications of robotics research in the 1960s and 1970s target the automation of industrial processes using industrial robots/manipulators [1], and manipulators resembling human arms are deployed for various automation tasks of the factories. In the 1980s, robots began to spread to the manufacturing floors in the form of wheeled or legged mobile mechatronic systems called AGV and extreme environments such as space, underwater, nuclear plants. The roles of robots are no longer limited to tasks in automated factories and expand into exploration of hazardous, human-unfriendly, extreme environments. Recently, various kinds of robots are provided to surveillance, security, and cleaning tasks [2]. Ingenious autonomous robotic systems emerging in late 1990s have artificial intelligence, e.g. Sony’s Aibo robotic dog [3]-[5], and Hondas humanoid robots from P2, P3 to Asimo [6], [7], and make robots familiar to our surrounding and daily life. One of recent research topics in robotics is the co-habit of human and robots. It is expected that robots with different degrees of autonomy and mobility will play an increasingly important role in all aspects of human life. To enable coexistence of human and robots, robots will be much more complex in their hardware, software and mechanical structures, and biomimetic technology and brain inspired processing would be a breakthrough of future robotics. One of the important technical issues in robotics is motion control problem. In the most of application, robots are required to work quickly, precisely and safety, 62

Articles

however, these requirements sometimes make conflicts and trade-off. In order to realize precise, stable and less vibration motion control, the robot joints should be rigid. On the other hand, the more rigid robot joints become, the less safe and flexible to unexpected things. To overcome this problem, the compliance control of robot joint should be considered from the viewpoints of hardware and software developments. Hardware approaches are; for example, to use soft materials as the mechanical structure components, cover the body with shock absorbers, or implement soft actuators such as pneumatic actuators. In the software approaches, force control and impedance control are often used, where virtual springs and dampers model are supposed to exist in the working space. Morita et al. showed the possibility to absorb shocks not only by protections using shockabsorbing materials but also by softness of the joints. In the tasks which need skillful motion control such as peg-in-hole insertion, inaccuracy of position control of robot arm should be compensated by compliance methods. As one of the methods, the remote center compliance (RCC) device with elastic elements is proposed. However, the position and orientation control are not enough for the peg-in-hole problem, because the information about the forces around the joints in the control aren't employed. Although the industrial robots are preferred to be made rigid as much as possible, the service robots should have soft joints to avoid damages to their environments and to humans. As a solution of this problem, bio-inspired approaches attract attentions. Animals including human can control articular impedances during motion [8], and human can realize the quick motion and the supple behavior. The system, which can change articular impedances of robots, is one of research trends in robot joints and software impedance controls have been developed [9]. One of the most important technical problems in the compliance control is the time delay caused by electric factors and computing time. Regarding collision, the time delay in the response becomes very critic problem. So, the robot joints with mechanical softness are researched and proposed, for examples, programmable passive impedance (PPI) which controls its compliance by using nonlinear springs slotting directly between the actuator and the wire [10], nonlinear spring tensioner (NST) which controls its compliance by using the movable pulleys and the linear springs [11], the mechanism with Ltype link added to NST [12], mechanical impedance adjuster (MIA) which controls its compliance by changing the moment arm by using a flat spring and a slider connected to the spring [13], actuator with nonlinear elastic


Journal of Automation, Mobile Robotics & Intelligent Systems

system (ANLES) which produces nonlinear elasticity by winding a linear spring around a guide shaft [14], and the mechanism using pneumatic rubber muscles [15]. The above mechanisms have quick response to sudden external forces because they can set the stiffness of joints beforehand, and guarantee delays mechanically. Koganezawa et al. assert that above mechanisms have two problems as following: A) Design of the nonlinear elastic elements is difficult. B) The mechanisms employing nonlinear elastic elements become large size. Including these above problems, we also consider a new item: C) Slow responsiveness in the dynamic compliance control. In the conventional mechanisms, elastic materials must be deformed to obtain target stiffness, and the delays occur by deforming. Animals change the articular impedance during the sequence of motion from one moment to the next. For example, the joints are well adjusted to high or low stiffness at an instant in the hopping motion. That is, in a series of motion, the joints of robots should have good responsiveness regarding articular compliance. The conventional joint mechanisms with stiffness adjustment function show better performance than software-based methods in the static compliance control responsiveness, on the other hand, have difficulty in the dynamic compliance control responsiveness. In our research, we especially address controlling compliance by variable mechanical impedance. We propose an antagonistically wire-driven mechanism, which has good control performance regarding dynamic compliance. We describe the viscoelasticity property of musculoskeletal system, and propose the model describing the compliance control system of musculoskeletal system. Then we show the antagonistically wire-driven mechanism realizing the model mechanically.

2. Impedance control mechanisms of musculoskeletal system 2.1. Kinetic features of muscles In general, regarding the mechanical impedance, three elements: mass, viscosity and elasticity, are con-

VOLUME 4,

N° 2

2010

trollable parameters, and musculoskeletal systems control only the viscosity and the elasticity because the change of mass is not so easy. Although the mechanism of impedance control system in the musculoskeletal systems is not fully understood yet, a few researches report on kinetic features of muscles as follows [16]: 1) A muscle can work in the contracting direction only. 2) A tension of the muscles is bigger when the frequency of the input pulses to the muscle is higher. 3) A tension force of the muscle depends on the muscle length. A mechanism which rolls round the wire using actuator is generally used to realize the kinetic feature regarding the above 1). However, this mechanism alone cannot produce a similar feature of muscles. The input pulses of the 2) are neural pulses, which transmit contraction signals to muscles. The contracting force of muscles becomes large when the pulses arrive more frequently. This pulse input is a manipulating variable in the control. The item 3) means that a muscle force is the function of the muscle length (or joint’s position; see Fig. 1). Muscles change their outputs depending on their length (that is joints’ angles), however, conventional actuators don't change the output according to their position, which is the main difference between muscles and general actuators. 2.2. Neural system of muscles and servo system In muscles, there is an organ called muscle spindle, which sends the afferent signals by sensing the change of position and velocity of the muscle, and neurotendinous spindle in tendons responding to the tension of the muscle, which also sends the afferent signals. Those signals are transmitted to the central nerve system via alpha motor neuron. Then, the alpha motor neuron sends the contraction signal to the muscles, so that the muscle spindles, the neurotendinous spindles and the alpha motor neuron comprise a structure that has a servo system characteristic with feedback loop regarding the motion (see Fig. 2). There is a response called stretch reflex generating in this servo system. In stretch reflex the muscle contracts to the resisting direction to stretch of the muscle. However, it is still disputable at the point of whether the impedance control of musculoskeletal system is also realized by structured servo system mentioned above.

Fig. 1. Length vs. tensions of muscles. Articles

63


Journal of Automation, Mobile Robotics & Intelligent Systems

Fig. 2. Neural system of the muscle. Regarding to position control of the joints in the motion, such as reaching, even if the neural pathways of muscle spindles and neurotendinous spindles are interrupted, the position control is possible [16], however, the accuracy of position control decreases. This fact means that the position control is possible in open-loop and, in a musculoskeletal system, the result suggests that the position control is realized by balancing the muscle’s tensions, which is called equilibrium point (EP) hypothesis in reference [17]. An articular EP is a point of balanced forces between extensors and flexors (see Fig. 3). At EP external forces are not generated. The muscles generate only the returning forces to EP, if articular position moves away from EP. The articular position therefore changes when EP changes. This is the principle of thinking in the position control by EP hypothesis. As the external forces are not generated around EP, the joints don't move. Important point is that the internal forces are generated, so that the forces realize the articular

VOLUME 4,

N° 2

2010

stiffness. For example, in isometric contractions, which change the muscle tensions and don't change the position of the joint, the articular stiffness is proportional to the muscle tensions [8]. Moreover, in closed-link mechanisms and antagonistically wire-driven mechanisms, it is known that the internal forces define the mechanical stiffness [18], [19]. Summarizing the above: Controlling the position without a neural servo system is possible. Internal forces are proportional to the muscle tensions. Internal forces relate to articular stiffness.

3. Modeling of musculoskeletal system 3.1. Antagonistic wire-driven models Two kinds of wire-driven mechanisms models with 2input (torques ô1 and ô2 in Fig. 4) and 2-output (joint angle è and stiffness) are shown in Fig. 4. Figure 4a) shows a conventional antagonistically wire-driven mechanism, and 4b) is the EP hypothesis-based antagonistically wire-driven mechanism. The main difference between Fig. 4 a) and Fig. 4b) is the torque around the equilibrium point. In the wire-driven system of Fig. 4a), the reaction torque will not occur and not go back to the original position against to external forces, because the joint can take any angles if the input torques ô1 and ô2 are equal. However, in the case of Fig. 4b), the reaction force will be generated to go back to the equilibrium point caused by the non-linear property of muscle-like actuators, when the external force is given to the joint. That is, the

Fig. 3. Equilibrium point hypothesis. (a) The mechanism using general actuators.

Fig. 4. Wire-driven mechanisms. 64

Articles

(b) The mechanism using actuators like muscles under the equilibrium point hypothesis.


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

wire-driven mechanism in Fig. 4b) can keep the position of the joint without feedback control regarding position. Another important aspect of the joint is the stiffness. The gradient of output torque, ô/è, means the stiffness of the joint, which is proportional to the output of musclelike actuator in the mechanism in Fig. 4b). The higher the gradient, ô/è, becomes, the larger the stiffness. This fact shows that the actuator mechanism, which changes the output according to its position, is needed to realize musculoskeletal system. In the following sections, we describe a mathematical model considering these characteristics. 3.2. Proposed mathematical model We use sigmoid function as the output model of muscle-like actuator for the joint as shown in Fig. 3. The musculoskeletal system is the antagonistic system, so this system is expressed by using two sigmoid functions as:

k11 k 21 ì ì ü ü + k10 ýô c1 + í + k 20 ýô c 2 ôo = -í - k12 (è o -å 1 ) - k 22 (è o -å 2 ) î1 + e î1 + e þ þ

N° 2

2010

4. Kinematic transmission mechanism (KTM) In order to realize the mathematical model in eq. (1), we propose a wire-driven mechanism shown in Fig. 5. In the mechanism, the cam (the input axis) rotation moves the passive link, where the wire is connected, and the passive link pulls the pulley (the output axis). Figure 5a) shows the geometric design of the proposed model. In the whole mechanism, because it is an antagonistic joint mechanism, the mechanism is symmetric. The cam shape can be determined based on the kinematic input-output relation of mechanism. We call this mechanism Kinematic Transmission Mechanism (KTM), as the kinematic elements such as the cam and the link are used instead of elastic materials. Although in case of employing elastic materials, analysis of its feature is not easy, the design of KTM is clear because we can design the cam to satisfy the kinematic constraints of the mechanism. Moreover, regarding to miniaturization of KTM, the similarity shape keeps the same input-output relation because its kinematic relationship depends on only the ratio of lengths, so that it will be possible to realize small size KTM.

(1) where the constant parameters, kij(i= 1,2; j = 0,1,2), are arbitrary design parameters, constant parameters åi are the offset values to decide the initial position, variable èo is the joint angle, variable ôci are the torques of the actuators, and variable ôo is the torque around rotation axis of the joint. Equation (1) expresses the curve shown in Fig. 1. Next, the elastic coefficient is obtained by partial derivative of ôo with respect to èo as: .

(2)

In the right side of the above equation, all values are constants except for ôc1 and ôc2. So, we can set the elastic coefficient by changing only the values, ôc1 and ôc2. This means that the joint stiffness of the mechanism which has the relation of eq. (1) is controllable without elastic materials by the principle of eq. (2). And, at ôo = 0, the joint converges to the EP, which is decided by the ratio of ôc1 and ôc2. Within the range to satisfy ôo = 0, the position of èo can be changed. (a) Structure of KTM.

4.1. Analysis of KTM The relation between the output torque ôoi (i =1,2,...) (around the output axis of the joint mechanism) and the input torque ôci (around the rotation axis of the cam) is expressed by eqs. (1) and (2): (3) .

(4)

Equation (4) means the kinematic transmission property of the cam, which corresponds to the gear ratio between output and input of KTM. Here, the relation between ôci and ôoi can be obtained by the principle of virtual work as: ,

(5)

where äèo and äèci are the virtual displacements. Solving the above equation with regard to äèo, eq. (6) is obtained.

(b) Analysis of kinematics.

Fig. 5. Kinematic transmission mechanism (KTM):(This image shows only one side. Practically, a symmetrical mechanism is added). Articles

65


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

,

(6)

Here, äèo and äèci are nearly zero, so the above equation is transformed into the following form: ,

2010

The wire is rolled round via point P when the input displacement increases. And èo is displaced by Äèa. Then displaced wire length Älqp is:

ro

.

(9)

(7)

where, Äèo and Äèci is small displacement. Integrate the above equation with regard to , and then we obtain:

From law of cosines we obtain the increment Äèa of èa when the side length of the triangle AQP is increased by Älqp. From the Äèa, we obtain lab after rotation, where

. (8)

(10)

The obtained èo and èci from eq. (8) are the displacements regarding the input and the output which satisfy eq. (3). From here on, we discuss only one-side KTM except when especially noted. So, the subscript i denoting the distinction number of KTM is abbreviated. Using obtained èo and èci we analyze the KTM. The final purpose is to obtain the cam shape satisfying the kinematic constraints on the KTM. Figure 5 (b) shows the vectors and the angles of the KTM. The meaning of each symbol is: A,...D,O,...,Q : Names of each point l : Position vectors from a point to a point Subscripts express the points e.g. lcb is the position vector from point C to B r : Vectors that express the diameter of a circle nb : Normal vector of a follower tb : Tangent vector of a follower ô : Torque around a joint è : Joint angle

The coordinate of cam shape vector lcd can be transformed to the coordinate frame of the cam as: (11) where, R(è) is rotation matrix. The cam shape is obtained by calculating C lcd by using eq. (11). 4.2. Design of KTM In this section, we discuss about the design of KTM. The parameters regarding dimensions such as the link length, the size of the bearing and the positioning of each rotary shaft are decided by adjusting to the designer’s purpose. The design steps are: 1. Decide the design parameters regarding the dimensions such as the lengths of the links. 2. Decide the design parameters regarding the sigmoid function from eq. (1) or (3).

Fig. 6. Joint angle èo vs. input-output torque ratio, and transitions of the equilibrium points. (a) Designed cam shape

Fig. 7. Cam for KTM. 66

N° 2

Articles

(b) CAD drawing of cam


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 2

2010

Table 1. Parameters. Parameters

Values

Units

3. Obtain èo and èc from eq. (8). 4. Derive the cam shape coordinates from eq. (9) to (11). Figure 6 shows the graphs of the output torque of joint vs. the joint angle when the ratio of the input torques of two actuators is changed. These graphs are obtained by eq. (1) and Table 1 show the corresponding parameters. In phase 2, we decide some parameters by checking the curves such as shown in Fig. 6. Table 2. Specifications of developed joint mechanism.

The designed cam is shown in Fig. 7-(a). Fig. 7b) shows the CAD drawing of the produced cam. Figures 8a) and 8b) are the graphs that show kinematic transmission feature expressed by the gear ratio C(èc) with respect to the output joint angle èo and the input cam angle èc. We verified the cam functioning whether it works according to design. The verifying way is check of the matching ratio about the curves of the theoretically designed cam and the made cam when the abscissa is the input angle and the ordinate is the output angle. The result is shown in Fig. 9. The actual values match to theoretical values well. So, we conclude that the cam made according to our design behaves as expected.

5. Experimental Prototype Using designed KTM, we developed the 1-DOF joint mechanism (see Fig. 10). The specifications of the joint mechanism are shown in Table 2. The EP moves by changing the ratio of input torques as shown in Fig. 6. Changing the ratio, we examined the po-

Parameters

Values

Units

sition control capability of the joint mechanism employing KTM without position feedback. Only the motor’s currents are controlled by the PI-controller in the equipment. The joint angle and the motor angles are measured by using a potentiometer and two rotary encoders. In the experiment, the responses of the current and the joint angle are measured; at first, the joint angle is set to a position of 40[deg], then the current values are changed and the joint angle target is set to 0[deg]. The currents are controlled to be a constant value while measuring. The same values of the currents are set to the two motors in the joint mechanism while targeting the joint angle to 0[deg]. In order to investigate the difference of the joint stiffness, we compared the responses with respect to the joint angle in cases of 0.04[A] and 0.08[A]. The results are shown in Fig. 11, where (a) is the response of the joint angle and (b) is the response of the motor’s currents. The joint angle gets close to the target angle 0[deg], by controlling only the current inputs. However, there are the steady-state errors caused by the effect of friction and the loss of back drivability. The main factor of loss of back drivability is the counter electromotive force, which is caused by the motor. The size of the steady-state error in the case of 0.08[A] is smaller than in the case of 0.04[A]. And the responsiveness in the case of 0.08[A] is better than in the case of 0.04[A] as well. As the next step, we evaluate the responsiveness regarding dynamic stiffness control when the joint mechanism is loaded by the external force. The currents are controlled at 3 seconds and 6 seconds after the experiment starts, and the joint stiffness is changed. The loaded external force is cyclic and the value of amplitude is invariable. The result is shown in Fig. 12. As shown in the figure, the amplitude of the joint angle becomes large when the current values are set low. This occurs due to the decrease of the joint’s stiffness. The amplitude changes, immediately after the change of currents. So it is verified that the joint mechanism’s stiffness can be controlled dynamically.

6. Joint Mechanism and Biogenic Motion Control Here, we discuss about the relationship of motion controls and brains. Marr and Albus have been proposed a theory that cerebellums have the function of perceptron from their minute observation at the part [20],[21]. Therefore, cerebellums output a pattern when a pattern is inputted to them, such as perceptron. In biogenic motion controls, it has cerebral-cerebellum loop, and it has been shown that cerebellums participate in the motion control. The movement disorder occurs when cerebral-cerebellum Articles

67


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

(a) The input (cam) angle èo

N° 2

2010

(b) The joint angle èo

Fig. 8. Gear ratio of designed cam (the case of cam1).

Fig. 9. Cam curves of theoretical and actual value.

Fig. 10. Antagonized wire-driven mechanism using KTM (1 DOF).

(a) The joint angles èo

(b) The currents of the motors of the joint (M1 and M2 are the motors of left and right side)

Fig. 11. Experimental results of angle and stiffness control of the joint (The cases of low and high stiffness). loop has broken, but the intention of trying to move (voluntary motor control) is not inhibited. So, it has been suggested that cerebellums participate in the motion skill. If the theory of cerebellar perceptron is true, the motion should be converted to any patterns as the input or the output to the central nervous system. The proposed joint mechanism is able to control the position and the compliance of the joint by changing the motor output (this equals the tension of a muscle) . So, controlling the 68

Articles

joint position, it is able to control the joint directions. Therefore, the joint mechanism is possible to control the motion by the static inputs and not the dynamic inputs. This means that the position, the compliance, the direction and the outputs of the joint are regarded as the static patterns. We suggest that the mechanism changing the movement elements to the static patterns such as the proposed mechanism by us is able to connect the biogenic motion control to neural system models.


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 2

2010

(a) The joint angle

(b) The currents of the motors of the joint

Fig. 12. Experimental results of response of changing stiffness.

7. Conclusions In this paper, we proposed the robot joint “KTM” which is based to the EP hypothesis, which is a leading hypothesis of the position control system of musculoskeletal system. One of the important issues of the EP hypothesis is that the nonlinear kinetic features of muscles are the essential to control the articuler position and stiffness. In the conventional mechanical softness joints, elastic materials are employed for obtaining nonlinear performance. However, deformation of elastic materials occur delays. KTM has the advantage of the nonlinear kinematic transmission features such as linkage mechanisms have it. We proposed an EP hypothesis based kinetic model, and we discussed the kinematic transmission feature expressing by the model with the design steps. And we showed that the position and the stiffness control are possible from experiments. We proved that the joint angle moves to around the target angle but the steadystate errors remains. The steady-state errors become smaller by making the articuler stiffness higher. However, the improvement of position controllability is expected by dealing with the counter electromotive force of the motors. We observe the responsiveness of the stiffness control by loading the constant cyclic external force to the joint. Then, we verified that oscillation of the joint angle just the changing of the articuler stiffness. A future work is to verify the response in loading high frequency external force to the joint. As other problems, there are the stretch of wire, optimization of KTM, multi-DOF and closed loop control. KTM is more simple mechanism than conventional mechanical softness joints. Moreover, freedom of design is higher, and minimizing is possible. We expect

the large applications of KTM such as an active shock absorber for vehicles and such as the application to manufacturing and household robot.

AUTHORS Takashi Sonoda* - Fukuoka Industry, Science & Technology Foundation, The University of Kitakyushu, 1-1 Hibikino, Wakamatsu-ku, Kitakyushu, 808-0135, Japan, +8193-695-6102 (ext. 2838). E-mail: sonoda@lab-ist.jp. Yuya Nishida, Amir Ali Forough Nassiraei, Kazuo Ishii Kyushu Institute of Technology, 2-4 Hibikino, Wakamatsu-ku, Kitakyushu 808-0196, Japan, +81-93-695-6102, E-mails: {nishida-yuya@, nassiraei@, ishii@}brain.kyutech.ac.jp. * Corresponding author

References [1]

[2] [3]

[4]

[5]

Crisman J. D., Bekey G., “The Grand challenge for robotics and automation”, IEEE Robotics & Automation Magazine, vol. 3, 1996, pp. 10-16. Engelberger J. F., Robotics in Service, The MIT Press, Cambridge, MA, 1989. Friedman B., et al., “Hardware companions?: What online AIBO discussion forums reveal about the humanrobotic relationship” . In: Proc. of CHI 2003, ACM Press, 2003, pp. 273-280. MelsonG. F., et al., “Robots as Dogs?: Children's Interactions with the Robotic Dog AIBO and a Live Australian Shepherd”. In: Proc. in conference on Human Factors in Computing Systems, 2005, pp. 1642-1659. Bartneck C., Forlizzi J., “Shaping human-robot interacArticles

69


Journal of Automation, Mobile Robotics & Intelligent Systems

[6]

[7] [8]

[9]

[10]

[11]

[12]

[13]

[14]

[15]

[16]

[17]

[18]

[19]

[20] [21]

70

tion: Understanding the social aspects of intelligent robot products”, Ext. Abstracts CHI 2004, ACM Press, 2004, pp. 1731-1732. Hirai K., Hirise M., Haikawa Y., Takenaka T., “The development of Honda Humanoid Robot”. In: Proc. of the 1998 IEEE International Conference on Robotics & Automation, Leuven, Beligium, 1998 , pp. 1321-1326. Honda Motor Co., Ltd., Asimo year 2000 model, http: //world.honda.com/ASIMO/technology/spec.html. Akazawa K., Aldridge J.W., Steeves J.D., Stein R.B., “Modulation of Stretch Reflexes During Locomotion in the Mesencephalic Cat”, Journal of Physiology, vol. 329, 1982, pp. 553-567. Siciliano B., Khatib O. (Eds.), Springer Handbook of Robotics, Springer Berlin Heidelberg, 2008, pp. 161185. Laurin-Kovitz K.F., Colkgate J.E., Carnes S.D.R, “Design of Components for Programmable Passive Impedance”. In: Proc. of the 1991 International Conference on Robotics & Automation, 1991, pp. 1476-1481. Noborisaka H., Kobayashi H., “Design of a Tendon-Driven Articulated Finger-Hand Mechanism and Its Stiffness Adjustability”, JSME International Journal. Series C: Mechanical Systems, Machine Elements and Manufacturing, vol. 43, 2000, no. 3, pp. 638-644. Yamaguchi J., Takanishi A., “Development of a Leg Part of a Humanoid Robot-Design of a Biped Walking Robot Having Antagonistic Driven Joints Using a Nonlinear Spring Mechanism”, Advanced robotics: the International Journal of the Robotics Society of Japan, vol. 11, 1997, no. 6, pp. 633-652. Morita T., Sugano S., “Development and Evaluation of Seven-D.O.F. MIA ARM”. In: Proc. of the 1997 IEEE International Conference on Robotics & Automation, USA, 1997, pp. 462-467. Koganezawa K., Nakazawa T., Inaba T., “Antagonistic Control of Multi-DOF Joint by Using the Actuator with Non-Linear Elasticity”. In: Proc. of the 2006 IEEE International Conference on Robotics & Automation, USA, 2006, pp. 2201-2207. Schulte H.F., “The Characteristics of the McKibben Artificial Muscle”. In: Application of External Power in Prosthetics and Orthotics, National Academy of Science, 1961, pp. 94-115. Akazawa, Biomechanism Library Biological Information Engineering, Tokyo Denki University Press (in Japanese), 2001, pp. 81-103. Feldman A.G., “Once more on the Equilibrium-Point hypothesis (ë model) for motor control” , Journal of Motor Behavior, vol. 18, 1986, no. 1, pp. 17-54. Hanafusa H., Adli M.A.. “Effect of Internal Forces on Stiffness of Closed Mechanisms”. In: Proc. 5th International Conference on Advanced Robotics, Italy, 1991, pp. 845-850. Li Z., Kubo K., Kawamura S., “Effect of internal force on rotational stiffness of a bicycle handle”. In: Proc. of the 1996 IEEE International Conference on Systems, Man, and Cybernetics, 1996, pp. 2839-2844. Marr D., “A theory of cerebellar cortex”, Journal of Physiology, vol. 202, 1969, pp. 437-470. Albus J.S., “A theory of cerebellar function”, Mathematical Bioscience, vol. 10, 1971, pp. 25-61.

Articles

VOLUME 4,

N° 2

2010


Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 4,

N째 2

2010

INFocus THE SPOTLIGHT on new n New chapter in the cooperation NASA with GE Engineers and scientists from NASA and General Motors with the help of engineers from Oceaneering Space Systems of Houston, developed and built the next generation of Robonaut 2 (R2) - a faster, more dexterous and more technologically advanced robot. R2 can use the same tools as humans, which allows them to work safely side-by-side humans both on Earth and in space. Going everywhere the risks are too great for people, Robonaut 2 will expand people's capability for construction and discovery. Using leading edge control, sensor and vision technologies, future robots could assist astronauts during hazardous space missions and help GM build safer cars and plants. The idea of using dexterous, human-like robots capable of using their hands to do intricate work is not new to the aerospace industry. NASA and GM cooperate since 1960s (through a Space Act Agreement at the agency's Johnson Space Center in Houston) with the development of the navigation systems for the Apollo missions. The original Robonaut, a humanoid robot designed for space travel, was built by the software, robotics and simulation division at Johnson in a collaborative effort with the Defense Advanced Research Project Agency 10 years ago. More information on Robonaut and video at: http://robonaut.jsc.nasa.gov Based on http://www.sciencedaily.com n New simulator for da Vinci Robotic Surgical System RoSS - the Robotic Surgical Simulator, developed by Thenkurussi ("Kesh") Kesavadas (University of Buffalo's School of Engineering and Applied Sciences) and Khurshid A. Guru (Roswell Park Cancer Institute) allows surgeons to practice skills needed to perform robot-assisted surgery without risk to patients. It is one of the world's first simulators that closely approximates the "touch and feel" of the da Vinci robotic surgical system. Khurshid A. Guru, MD, director of the Center for Robotic Surgery and attending surgeon in RPCI's Department of Urology, and Thenkurussi Kesavadas, PhD, professor of mechanical and aerospace engineering at UB and head of its Virtual Reality Lab, founded the Western New York-based spin-off company, Simulated Surgical Systems, LLC, to commercialize the simulators. Both stress how important is training for proficiency for surgeons, especially in robot-assisted operations, just like in the aircraft, they say. "Think of the RoSS as a flight simulator for surgeons", said Mr Kesavadas to RobotWorldNews' reporter. The RoSS will play an educational role for RPCI and other similar institutions involved in robot-assisted surgical systems. Already, at least 70 percent of all prostate surgeries in the U.S. are performed using robotic surgical systems; robotic surgeries are generally less invasive, cause less pain, require shorter hospital stays and allow faster recoveries than conventional surgery. Robotic surgical systems are increasingly being used for gynecologic, gastrointestinal, cardiothoracic, pediatric and other urologic surgeries. Is expected that Ross will be commercially available by the end of 2010. More information: http://www.buffalo.edu/news/10997 Based on: http://www.robotworldnews.com In the spotlight

71 ?


Journal of Automation, Mobile Robotics & Intelligent Systems

VOLUME 4,

N째 2

2010

n Precise security technology based on voice recognition According to Dr. Robert Rodman, professor of computer science at North Carolina State University, and his fellow researchers, their new research will help improve the speed of speech authentication, with keeping accuracy. There are no two identical voices, just like the fingerprints or faces. Current technology is still too slow to gain widespread acceptance person's voice recognition may take several seconds or more. The response time needs to improve without increasing the error rate. Scientists extended SGMM method (sorted GMM - a novel Gaussian selection method) by using 2-dimensional indexing, which leads to simultaneous frame and Gaussian selection. They modified existing computer models to streamline the authentication process so that it operates more efficiently. Potential users of this technology are governments, financial, insurance, health-care and telecommunications industries everywhere where high level of data protection is needed. The others co-authors of the research are: Rahim Saeidi, Tomi Kinnunen and Pasi Franti of the University of Joensuu in Finland; and Hamid Reza Sadegh Mohammadi of the Iranian Academic Center for Education, Culture & Research. The research, "Joint Frame and Gaussian Selection for Text Independent Speaker Verification", was presented in March at the International Conference on Acoustics, Speech and Signal Processing (ICASSP) in Dallas. Source: http://www.sciencedaily.com http://www.ncsu.edu n New feature of robots': a runny nose At Tsukuba University in Japan Yotaro, a robot, which emulates a real baby, has been constructed. It's full-cheeked face, made of soft translucent silicon with a rosy hue, looks a little weird with so luminous blue eyes. It also has sporting a pair of teddy-bear ears. Robot's face is backlit by a projector connected to a computer to simulate crying, sneezing, sleeping, smiling, while a speaker can let out bursts of baby giggles. Mood's changes are based on the frequency of touches. It moves its arms and legs when different parts of its face and body are touched. Yotaro also simulates a runny nose, with the help of a water pump that releases bodytemperature droplets of water through the nostrils. The inventors hope that cute Yotaro will induce young Japanese people to parenting through showing its pleasures as Japan faces a demographic crisis. Japan has the world's longest average life expectancy and one of the lowest birth rates. The fifth of the population is aged 65 or older. By 2050, that figure is expected to rise to 40 percent. Source: http://www.physorg.com

72

In the spotlight


VOLUME 4,

Journal of Automation, Mobile Robotics & Intelligent Systems

N° 2

2010

EVENTS SPRING-SUMMER 2010 April 16 – 18 23 – 25 23 – 25

ICMAE 2010 – IEEE International Conference on Mechanical and Aerospace Engineering, Chengdu, Sichuan, China. http://www.iacsit.org/cmae ICIII 2010 – International Conference on Industrial and Intelligent Information, Bangkok, Thailand. http://www.iciii.org ICTLE 2010 – International Conference on Traffic and Logistic Engineering, Bangkok, Thailand. http://www.ictle.org

June 11 – 13 13 – 17 16 – 18 22 – 24 23 – 24 24 – 26 28 – 30

ICNIT 2010 – International Conference on Networking and Information Technology. Manila, Philippines. http://www.icnit.org ICAISC 2010 – 10th International Conference on Artificial Intelligence and Soft Computing, Zakopane, Poland. http://icaisc.org SEDE 2010 – 19th International Conference on Software Engineering and Data Engineering. San Francisco Fisherman's Wharf, San Francisco, California, USA. http://www.users.csbsju.edu/~irahal/sede2010 ICINT 2010 – 2nd International Conference on Information and Network Technology. http://www.icint.org 3rd International Conference on Human-Robot Personal Relationships, Leiden, Netherlands. http://hrpr.liacs.nl RoEduNet 2010 – 9th RoEduNet International Conference, Sibiu, Romania. http://roedu2010.ulbsibiu.ro Design and Nature 2010 – 5th International Conference on Comparing Design in Nature with Science and Engineering, Pisa, Italy. http://www.wessex.ac.uk/10-conferences/design-and-nature-2010.html

July 9 – 11

9 – 11

ICIAE 2010 – IEEE International Conference on Information and Applied Electronic. Chengdu, Sichuan, China. ICIAE 2010 will be held in conjunction with the 3rd IEEE ICCSIT 2010. http://www.iciae.org ICCSIT 2010 – 3rd IEEE International Conference on Computer Science and Information Technology. Chengdu, Sichuan, China. http://www.iccsit.org

August 1–3 20 – 22

ICMEE 2010 – 2nd IEEE International Conference on Mechanical and Electronics Engineering, Tokyo, Japan. http://www.icmee.org ICACTE 2010 – 3rd IEEE International Conference on Advanced Computer Theory and Engineering. Chengdu, Sichuan, China. http://www.icacte.org

September 8 – 10

ADM 2010 – 3rd International Conference on Advanced Design and Manufacture, Nottingham, United Kingdom. http://www.admec.ntu.ac.uk/adm2010/home.html

Events

73 ?


Issuu converts static files into: digital portfolios, online yearbooks, online catalogs, digital photo albums and more. Sign up and create your flipbook.