Control Engineering Practice 11 (2003) 291300

Closed-loop identication of an industrial robot containing exibilities

. ( . Mans Ostring, Svante Gunnarsson*, Mikael Norrlof
Department of Electrical Engineering, Linkopings universitet, SE-58183 Linkoping, Sweden . .

Abstract Closed-loop identication of an industrial robot of the type ABB IRB 1400 is considered. Data are collected when the robot is subject to feedback control and moving around axis one. Both black-box and physically parameterized models are identied. A main purpose is to model the mechanical exibilities. It is found that a model consisting of three-masses connected by springs and dampers gives a good description of the dynamics of the robot. r 2002 Elsevier Science Ltd. All rights reserved.
Keywords: Identication; Robotics; Flexible arms

1. Introduction System identication is an established modeling tool in engineering and numerous successful applications have been reported. The theory is well developed, see . . e.g., Ljung (1999) or Soderstrom and Stoica (1989), and there are powerful software tools available, e.g., the System Identication Toolbox for Matlab (Ljung, 2000). Industrial robots represent an interesting challenge for system identication methods, and an overview of identication in robotics can be found in Kozlowski (1998). One application area for system identication within robotics is identication of the parameters in the kinematic description of the robot, while a second area deals with the problem of identifying the parameters in the dynamical model of the robot. A third area is to determine the parameters on the joint level including, for example, friction, motor characteristics, etc. Recent results from the latter two areas may be found in, e.g., Wang, Bi, and Zou (1996), Grotjahn, Daemi, and Heimann (2001) and Gautier and Poignet (2001). In these papers it is assumed that the robot is rigid. In the work presented here, the results of a study of the identication of robots, including exibilities are presented. This topic has been addressed in, e.g., ! ! Depince (1998), ElMaraghy, ElMaraghy, Zaki, and
*Corresponding author. Tel.: +46-13-281747; fax: +46-13-282622. E-mail address: svante@isy.liu.se (S. Gunnarsson).

Massoud (1994), Nissing and Polzer (2000), AlbuSch. ffer and Hirzinger (2001), Berglund and Hovland a (2000), Johansson, Robertsson, Nilsson, and Verhaegen . (2000), Pham, Gautier, and Poignet (2001) and Ostring, . Gunnarsson, and Norrlof (2001). The problem considered in Berglund and Hovland (2000) is closely related to the work reported below, but the proposed solution is based on frequency domain identication in combination with the solution of an eigenvalue problem. In Johansson et al. (2000), time domain identication methods of the black-box type are applied, which means that no physical parameters are obtained directly from the identication. The work presented in Albu-Sch. ffer a and Hirzinger (2001) deals with the identication of a lightweight robot with seven degrees of freedom, and the results involve identication of joint elasticity and damping parameters. These are found by applying external excitation on one axis at a time. Also, in Pham et al. (2001) the identication experiments are carried out by moving one axis at a time. The physical parameters sought are obtained as nonlinear functions of the estimates obtained using a model structure that is linear in the parameters. In the results of the work presented below, which is an . extension of the work reported in Ostring et al. (2001), a method is proposed in which inertial parameters and parameters describing exibility, can be identied directly in the time domain. This is done by utilizing a user-dened model structure in the System Identication Toolbox. As far as the authors know, this approach has

0967-0661/02/$ - see front matter r 2002 Elsevier Science Ltd. All rights reserved. PII: S 0 9 6 7 - 0 6 6 1 ( 0 2 ) 0 0 1 1 4 - 4

292

. M. Ostring et al. / Control Engineering Practice 11 (2003) 291300

not been reported previously within the area of robot identication. The work presented is carried out under simplifying conditions. First, only movements around axis one are considered. Second, all experiments are carried out with the other axes in one position. Third, only a linear model structure is considered, which means that, for example, only viscous friction is included in the model. It should be noted that although the model is linear in the states it is not linear in the parameters. The simplifying assumptions can be motivated in different ways. First, the work presented can be seen as a feasibility study carried out in order to see whether or not this is a possible approach, and to try to go as far as possible with linear models. In some of the references cited above, the identication experiments are carried out on one axis at a time. The importance of the restriction that the identication is carried out on only one operating point is related to the intended use of the model. From the authors viewpoint there are at least two possible uses of the identied model. The rst is to use the model for control design on joint level, and the second is to use the model for diagnostic purposes. In both cases, two ways to extend the approach can be considered. One way is to identify linear time-invariant models in a number of operating points and use gain scheduling in the control or diagnostic functions. A second alternative is to move to a nonlinear model, where nonlinearity due to variations in operating point is captured. These extensions are left for future work. The paper is organized as follows. In Section 2, the robot system under consideration is described briey, and in Section 3 generation and pre-processing of data before identication, are discussed. In Section 4, a linear three-mass model is derived, and in Section 5 the results from identication using physically parameterized models are presented. For comparison, the system is also identied using black-box model structures, and the results of which are presented in Section 6. Finally, Section 7 contains some conclusions.

Fig. 1. ABB IRB 1400.

ref

u Controller Robot

Fig. 2. Block diagram of the robot control system.

and record signals in the robot control system is . described in Norrlof (2000).

3. Data collection and pre-processing 3.1. Implementation issues The data are collected while the control system is in operation, which means that the identication will be carried out using closed-loop data. The situation is schematically depicted in Fig. 2. During conventional operation, the desired path is specied using a high level programming language RAPID. In the control system, the desired path and the actual position are used to compute a desired torque, which is sent to the motor drive system. The torque results in a movement of the robot arm via the gear box and the links. Measurements of the actual torque are not available, but, instead, the torque reference of an inner loop, controlling the motor current, is used. Since the inner control loop has high bandwidth, the relationship between torque reference and generated torque can be approximated by a constant. Hence, the signals available for measurement for each joint are the torque (reference) and the motor angle. The data are collected and sent to a PC using the interface between S4C and Matlab.

2. The robot system The industrial robot studied is of the type ABB IRB 1400, as shown in Fig. 1. As discussed above the study in this paper is restricted to movements around axis one, since this movement is to a large extent decoupled from the other axes. The IRB 1400, which is one of the smaller members in the ABB family, is a six-degrees-of-freedom robot and carries a load of approximately 5 kg: The maximum speed for the TCP (tool center point) is 2:1 m=s and the maximum acceleration is 15 m=s2 : The robot is equipped with the control system S4C. In addition to the conventional control system, an interface between S4C and Matlab has been used. The interface used to inject

. M. Ostring et al. / Control Engineering Practice 11 (2003) 291300

293

To obtain a suitable excitation of the system an external signal is applied. More specically the excitation signal is superimposed on the position reference signal, which is kept constant during the experiment. An alternative would have been to superimpose the input signal on a linearly growing reference signal in order to avoid situations when the velocity changes sign, since the effects of static friction can be substantial in such cases. 3.2. Experiment design The determination of the excitation signal involves several choices. These choices are: number of data N; sampling interval T; and properties of the input signal. An upper bound on the sampling frequency is given by the internal sampling rate in the control system S4C, which is 2 kHz: There is a limitation on the duration of the data collection experiment, but the possible record length of approximately 24 103 is judged to be sufcient. The properties of the input signal depend on signal amplitude, frequency content and the shape of the signal. The choice of input signal is crucial in all types of identication for the quality of the identied model. In the area of robot identication, this problem has been studied in, for example, Swevers, Ganseman, Tukel, . Schutter, and Brussel (1997), where identication of rigid robots is considered. In Pintelon and Schoukens (2001) it is proposed to use the so-called special odd multi-sine signals. This choice has been investigated by . . . Norrlof, Tj. rnstrom, Ostring, and Aberger (2002). a The results of the identication experiments reported here are based on three data sets, where the excitation signal has different character:
*

PRBS-signal as position reference would result in vary large amplitudes in the velocity reference. By letting the transfer function of the controller and the system in Fig. 2 be denoted by F s and Gs; respectively, the spectrum of the torque signal is given by  2   F io   Fy o: Ft o  1 F ioGio r Hence the spectrum of the torque depends not only on the spectrum of the applied reference signal but also on the regulator and the system itself. Since the system in this case contains an integrator, and it is likely that the regulator contains integral action, even though the regulator is not known, the energy of the input torque will be small in the low-frequency range. To illustrate this phenomenon the periodogram of the input and output signals for the three data sets are presented in Fig. 3. The gure shows that the input energy decreases rapidly below 10 Hz: 3.3. Data preprocessing It has turned out to be practical, e.g., for model validation purposes, to use the motor angle velocity as output instead of the motor angle directly. Therefore, the measured motor angle is differentiated numerically. In the next step of preprocessing, the sample means are removed, and nally the input and output signals are low-pass ltered, using a fourth order Butterworth lter with cut-off frequency 100 Hz, to reduce the inuence of high-frequency disturbances. 3.4. Closed-loop identication The assumption in this paper is that data are collected while the control system is active, i.e., in closed loop, while the aim is to obtain a model of the open loop system from input torque to motor position. On the other hand, in e.g., Johansson et al. (2000), experiments are presented where the controller is disconnected and the input torques are applied directly. For identication of systems operating in closed loop there are three main approaches, see Ljung (1999) or . . Soderstrom and Stoica (1989): The indirect approach, the joint inputoutput approach, and the direct approach. In the indirect approach it is required that the controller is known, and, since this is not the case here, this approach is not applicable. In the joint input output approach, the controller is estimated, which is a difcult task since the structure of the controller is not known, and it is likely that the controller contains nonlinear elements. The direct approach can, however, be applied and this is the method that will be exploited here.

Set 1: A chirp signal where the frequency changes linearly from 50 to 0:5 Hz in 12 s: The resulting input P torque has a mean power measured as 1=N u2 t; of 0.56. Set 2: A sum of 10 sinusoids with equidistant spacing in frequency, between 5 and 50 Hz: The resulting mean input power is 0.13. Set 3: Similar to Set 2, but with larger amplitude. The mean input power is 0.64.

These choices are by no means claimed to be optimal. The chirp signal, Set 1, is chosen since this is a common input signal in other publications, see e.g. Johansson et al. (2000). The sum of sinusoids, Set 3, is chosen in order to provide a signal for validation purposes with comparable frequency content and energy but different waveform. Set 2 is included in order to provide indications of the inuence of the signal amplitude. The reason for not choosing an excitation signal of PRBS-type is that the excitation signal is applied as both a position reference and a velocity reference. Using a

294

. M. Ostring et al. / Control Engineering Practice 11 (2003) 291300

Fig. 3. Lower: Periodogram of input signal (torque). Upper: Periodogram of output signal (angular velocity).

Therefore, this present work is restricted to three-mass models. The input is the torque t generated by the electrical motor, while the output is the motor angle ym : The angles of the other masses, yg and ya ; are not measurable. The physical parameters are dened as follows:
Fig. 4. Three-mass exible model.

Jm ; Ja ; Jg moments of inertia kg ; ka spring constants An important question in closed-loop identication is how the model quality in general, and the bias distribution in particular, is affected by the feedback. A thorough discussion of this topic is found in Ljung (1999) and Forssell and Ljung (1999). The main factors that affect the quality of the estimated model is the level of measurement noise, the properties of the feedback and the use of a noise model. Since the two rst factors cannot be controlled, an important task will be to study the inuence of noise model on the model quality. fm ; fa ; fg dg ; da r viscous friction coefficients dampings 1 gear box ratio r 118

4. Physical model The dynamics of the robot system, when moving around axis one, will be approximated by a model consisting of three masses connected via springs and . dampers as shown in Fig. 4. In Ostring et al. (2001) models with two and three masses were compared. It was found that models with three masses gave considerably better results than models with two masses.

The rst mass represents the rotating part of the electrical motor driving the robot. The motor is followed by a gear box. Flexibility in the gear box is represented here by the spring between the gear box and the second mass. Alternatively, the exibility could be placed before the gear box. The second spring between the second and third mass represents the exibility in the robot arm. Applying torque balances for the three masses and introducing the states 2 6 6 yg t ya t 6 6 ym t xt 6 6 6 yg t 4 ya t rym t yg t 3 7 7 7 7 7 7 7 5

. M. Ostring et al. / Control Engineering Practice 11 (2003) 291300

295

the input signal ut tt and the output signal yt ym t gives the state-space model xt Axt But; where 2 0 6 0 6 6 6 rkg A 6 Jm 6 6 kg 6 J 4 g 0 BT 0 0 0 0 0
a kg J

y Cxt; r 0 1 1
rdg Jm fg dg da Jg da Ja

2 3 0 1 7 7 7 7 0 7; 7 7 da 7 Jg 5 da fa Ja 1 0 0:

r2 dg fm Jm rdg Jg

ka Ja 1 Jm

0 0 0;

C 0 0

In the robot control system, the motor angle ym is the only available output signal, but since the measurement noise is fairly small, a reasonable estimate of the motor velocity is easily obtained. Therefore, the motor velocity is used as the output signal in the model above. Describing the movement of the robot by a fth order linear model is of course a simplication from different viewpoints. First, the properties of the robot, e.g., the moments of inertia, depend on the actual position of the robot. In this work the same position is considered in all experiments. Second, it is a simplication to describe the mechanical exibilities by nite dimensional models, and third, several nonlinearities have been neglected. Some nonlinearities that can be expected to have inuence are backlash in the gear-box, nonlinear friction, and nonlinear stiffness in the springs. The approach in this paper is to stick to linear models and evaluate how well the system can be modeled. The investigation of nonlinear effects, which are left for future work, include methods to detect nonlinearities using the measured data, insight into how different types of nonlinearities affect the identied models, and nally methods to identify models including nonlinear effects. Some results on the use of so called special odd multisine signals for detection of nonlinearities are presented . in Norrlof et al. (2002).

denotes that a disturbance model is estimated. The unknown parameters in the vector y are estimated using the System Identication Toolbox, see Ljung (2000), and the model structure given by the problem is specied as a Matlab m-le. Using a priori knowledge the parameters are suitably scaled in order to estimate parameters having the same order of magnitude. Furthermore, the inverse of the moments of inertia are introduced as unknown parameters. The identication is carried out both without disturbance model, denoted K 0; and with estimation of disturbance model, denoted Ka0; respectively. This means that 10 physical parameters are identied in the rst case, and ve additional parameters for the noise description are estimated in the second case. In Table 1 the value of the loss function of the estimated models are shown for the three data sets and the two model structures. For all data sets the loss function is much smaller when the noise model is included. The loss function is dened as the sum of the square of the residuals obtained when the output is compared to the one step ahead prediction. Without disturbance model the prediction is formed entirely using the input signal, i.e., as a simulation of the model. With a model containing a disturbance description the prediction is based on both previous outputs and previous inputs. Since the sampling interval is short in this problem it is logical that the one step ahead predictions are improved by also using the previous outputs. Table 2 shows the results when models are validated using the three data sets. The measure of the t is given by 0 q1 PN # 2 t1 yt yt C B fit 100@1 q A; 7 PN yt y2 % t1 # where yt is the measured output, yt is the simulated output and y is the mean value of the measured output. % In the table, case i; j denotes the situation when the model is identied using data set i; and validated using data set j: Thus, i j means a case when the model is validated using the data set used for estimation. When the t is computed using the estimation set, i.e., cases 1; 1; 2; 2 and 3; 3; the t is rather high in all three cases. The t is also high when a model estimate using Set 1 is validated using Set 3 and vice versa. In crossvalidations including Set 2, e.g., validating the model

5. Identication of physically parameterized models Consider the state-space description on innovations form xt Ayxt Byut Kyet; yt Cxt et; 5 6

where the matrices Ay; By and C are given by Eqs. (3) and (4), and y denotes a vector containing the parameters to be identied. The signal et is a zero mean white disturbance, and Ky is the Kalman lter gain. Choosing Ky 0; the model structure corresponds to an output error (OE) structure, while Kya0

Table 1 Loss functions for the physically parameterized models Data set 1 2 3 K 0 0.83 0.50 1.03 Ka0 9 105 3:7 106 1:2 106

296

. M. Ostring et al. / Control Engineering Practice 11 (2003) 291300

obtained from Set 2 using Set 3 (case 2; 3; the t is essentially lower. This is an interesting observation since the only difference between Sets 2 and 3 is the lower amplitude in Set 2. This difference is probably caused by nonlinear effects, as will be discussed further below. The N-sign in the second column of Table 2, corresponding to the model estimated using Set 2, indicates that the identied model is unstable. Therefore, no simulation can be carried out using this model. The values of the loss function, Table 1, and the simulated t, Table 2, offer two possibilities to judge the quality of the identied models, and in particular to decide whether to include a disturbance model or not. Since simulation is closely related to control the simulated t is judged to be a more relevant measure in this problem. Since models without disturbance description give better t, also in cross-validation, they are favored in this case.
Table 2 Simulated t according to Eq. (7) in case i; j; i.e. when a model estimated using data set i is simulated using data set j: The N-sign denotes that the identied model is unstable and cannot be simulated Case 1; 1 1; 2 1; 3 2; 1 2; 2 2; 3 3; 1 3; 2 3; 3 K0 85% 66% 84% 62% 82% 62% 83% 65% 86% Ka0 70% 55% 70% N N N 69% 51% 70%

Figs. 5 and 6 show the amplitude and phase curves of the estimated models. The models obtained using Sets 1 and 3 agree very well in these gures. There is a small deviation for low frequencies and around the second peak. The model from Set 2 agrees in the mid frequency range while the deviation is larger in the low-frequency range and around the second peak. This difference might be caused by nonlinear effects. For control purposes it is important to determine the resonance and anti-resonance frequencies of the system. For simplicity, only the rst anti-resonance (notch) frequency on will be considered here. In Table 3, the value of on of the identied models are presented. The relationship between the zeros of a continuous time system and the zeros of the corresponding discrete time system is complicated, and simple relationships can only be obtained as the sampling interval tends to zero. See ( . e.g., Astrom and Wittenmark (1984). To make a correct comparison between the anti-resonance of the continuous time models and the discrete time model identied below the model is transformed to their discrete time counterparts before the location of the notch frequency is determined. In Table 3, it is seen that when K 0 the notch frequency on is the same for Sets 1 and 3, while it is lower for Set 2. A possible explanation to the reduction of on for Set 2 is the presence of backlash in the gear box. In Aberger (2000) it is found using simulations that the presence of backlash reduces the rst notch frequency. The reduction depends on the width of the backlash and the magnitude of the input signal. When including a noise model Ka0 the variation in estimated notch frequency is much larger.

10

From u1 to y1

Amplitude

10

10

10 Frequency (rad/s)

Fig. 5. Amplitude curves of physically parameterized models when K 0: SolidSet 1, DashedSet 3, Dash-dottedSet 2.

. M. Ostring et al. / Control Engineering Practice 11 (2003) 291300

297

100 80 60 Phase (degrees) 40 20 0 20 40 60 80 100

From u1 to y1

10 Frequency (rad/s)
Fig. 6. Phase curves of physically parameterized models when K 0: SolidSet 1, DashedSet 3, Dash-dottedSet 2.

Table 3 Estimate of the rst notch frequency on in the Bode diagram for the physically parameterized models Data set 1 2 3 K 0 93.0 91.5 93.0 Ka0 92.2 88.9 91.1

Table 4 Estimates of physical parameters for different data sets using K 0 Par. Jm kg Jg fm ka Ja dg da fg fa Set 1 3:9 10 1:9 105 6.3 2:6 102 1:5 105 12.2 168 12 232 11
4

Set 2 4:0 10 2:0 105 7.2 3:1 102 1:9 105 10.3 18 6 392 17
4

Set 3 4:4 104 1:9 105 6.2 2:3 102 1:6 105 11.6 91 21 201 8

The estimates of the physical parameters using the three data sets are shown in Table 4, where some observations can be made. The rst six parameters show a fairly stable behavior when estimated using different data sets. Considering Sets 1 and 3 the difference is approximately 10% for Jm and fm while the difference is lower for the spring constants and moments of inertia. For Set 2 the variations are somewhat larger. For a corresponding two-mass exible model the p notch frequency on is approximately given by on E k=Ja where Ja is the moment of inertia of the second mass. For the three-mass model the zeros of the transfer function from input torque to motor angle velocity are approximately (when da dg fa fg 0) given by the roots of the polynomial   kg ka 2 ka kg s4 2 0: 8 s Jg Ja Ja Jg By comparing the results for Sets 1 and 3 it is found that even though the estimates of the individual physical parameters differ between the data sets the notch frequency on is the same. The estimate of the notch

frequency is hence less sensitive for the character of the excitation signal than the individual physical parameters. Another measure of the accuracy of the estimated models is obtained by plotting the estimated frequency functions together with estimate condence regions. Figs. 7 and 8 show the amplitude curves of the models (using K 0) obtained using Sets 1 and 2, respectively. The curves for Set 3 are similar to those of Set 1. The gures clearly show that the accuracy is lower when using Set 2, and this is natural effect due to the lower signal-to-noise ratio and nonlinear effects. The variations in fm can probably be explained by the fact that the friction in reality also consists of static friction, which is not considered here. The variations in Jm are probably explained by the fact

298

. M. Ostring et al. / Control Engineering Practice 11 (2003) 291300

10

From u1 to y1

Amplitude

10

10

10 Frequency (rad/s)

Fig. 7. Amplitude curve and condence region for the mode obtained using Set 1.

10

From u1 to y1

Amplitude

10

10

10 Frequency (rad/s)

Fig. 8. Amplitude curve and condence region for the mode obtained using Set 2.

that this estimate is affected by the high frequency components of the signals, and that these differ between the data sets. For two data sets some of the last four damping and friction parameters are negative. This is obviously unrealistic for a physically parameterized linear model, but a possible explanation here is the presence of nonlinearities. This phenomenon need further investiga-

tion. The dampings and friction coefcients show large variations between the data sets.

6. Black-box identication Even though a main point of the paper is the use of physically parameterized models black-box models will

. M. Ostring et al. / Control Engineering Practice 11 (2003) 291300

299

also be identied for comparison. The starting point in the black-box identication is the general linear model structure, see, e.g., Ljung (1999), yt Gq; yut Hq; yet; 9 where Gq; y and Hq; y are transfer operators, q denotes the shift operator qyt yt 1; and y is the vector of parameters to be determined. ut and yt denote the input and output signals, respectively, and et is white noise. Two different model structures will be studied, and these are the OE structure, dened by Bq; y Gq; y ; Hq; y 1 10 F q; y and the BoxJenkins (BJ) structure, where Bq; y Cq; y ; Hq; y : Gq; y F q; y Dq; y

Table 5 Model t according to Eq. (7) in case i; j; i.e. when a model estimated using data set i is simulated using data set j Case 1; 1 1; 2 1; 3 2; 1 2; 2 2; 3 3; 1 3; 2 3; 3 OE(%) 85 66 83 62 82 62 83 65 86 BJ(%) 82 67 82 45 55 45 53 50 55

11

Table 6 Estimate of the rst notch frequency (rad/s) in the Bode diagram Data set 1 2 3 OE 93.3 91.5 93.0 BJ 94.5 92.8 96.2

For the model structure in Eq. (9) the corresponding one-step-ahead predictor is # yt; y H 1 q; yGq; yut 1 Hq; yyt: 12 The aim is to nd the parameter vector that minimizes the prediction error # et; y yt yt; y by minimizing a criterion of the type N 1 X 2 e t; y: VN y N t1 13

14

obtained for the physically parameterized models. The similarities are also supported by the values of the t. The physically parameterized and black-box models, e.g. for Set 1, can therefore be considered to be the same.

For the two model structures mentioned above the predictor is a nonlinear function of the parameter vector, which implies that the parameter estimate has to be computed using an iterative search method. The identication experiments are carried out using the System Identication Toolbox, Ljung (2000). In order to compare with the results using physically parameterized models above the order of the input output dynamics is ve in all experiments. Increasing the model order does not give any improvement of the t, and this indicates that order ve is adequate for this problem. For the BJ models second-order noise dynamics was used. Table 5 shows the simulation t when calculated similarly to how the t was calculated for the physically parameterized models. The results for the OE-models are almost identical to the results for the physically parameterized models. The results for the BJstructure are not directly comparable with the results for physically parameterized models with noise model since the noise descriptions are different. Table 6 shows the rst notch frequency of the identied models, and also here the results for the OEmodel are very close to the results for physically parameterized models. When plotting the Bode diagrams of the OEmodels these curves are almost identical to the curves

7. Conclusions and future work A physically parameterized linear three-mass model of the open loop dynamics of an industrial robot during one axis movement has been identied. Three data sets, with different input properties, have been used. Cross validations show that a model of OE type, i.e., without disturbance model, is adequate, even though data have been collected in closed loop. The estimates of moments of inertia and spring stiffness, obtained using different data sets, are fairly stable, while the estimates of friction coefcients and dampings uctuate considerably. A possible explanation is the presence of nonlinearities in the real robot. There are a number of aspects of the presented results that are subjects for future work. One important problem is to apply the same technique for identication of several axes simultaneously. Some initial work dealing with axes two and three have started. The choice of exciting signal can also be studied further, and one step in this direction is taken in . Norrlof et al. (2002). A further topic is to continue the work in Aberger (2000) and investigate the effects of nonlinearities, like e.g., nonlinear spring stiffness, on the model properties.

300

. M. Ostring et al. / Control Engineering Practice 11 (2003) 291300 Johansson, R., Robertsson, A., Nilsson, K., & Verhaegen, M. (2000). State-space system identication of robot manipulator dynamics. Mechatronics, 10, 403418. Kozlowski, K. (1998). Modelling and identication in robotics. Berlin: Springer. Ljung, L. (1999). System identication: Theory for the user (2nd ed.). Upper Saddle River, NJ, USA: Prentice-Hall. Ljung, L. (2000). System identication toolboxusers guide. Sherborn, MA, USA: The MathWorks Inc. Nissing, D., & Polzer, J. (2000). Parameter identication of a substitution model for a exible link. System Identication Symposium SYSID 2000. . Norrlof, M. (2000). Iterative learning control analysis, design and . . experiments. Ph.D. thesis, Linkoping University, Linkoping, Sweden. . . Norrlof, M., Tj. rnstr. a om, F., Ostring, M., & Aberger, M. (2002). Modeling and identication of a mechanical industrial manipulator. In: IFAC World congress. Barcelona, Spain. . . Ostring, M., Gunnarsson, S., & Norrlof, M., 2001. Closed loop identication of physical parameters of an industrial robot. In: Proceedings of the 32nd international symposium on robotics. Seoul, Korea. Pham, M., Gautier, M., & Poignet, P. (2001). Identication of joint stiffness with bandpass ltering. In: Proceedings of the 2001 IEEE international conference on robotics & automation. Seoul, Korea. Pintelon, R., & Schoukens, J. (2001). System identication: A frequency domain approach. Somerset, NJ: IEEE Press and John Wiley & Sons. . . Soderstrom, T., & Stoica, P. (1989). System identication. PrenticeHall International, Hemel Hemptstead, Hertfordshire. Swevers, J., Ganseman, C., Tukel, D., Schutter, J. D., & Brussel, H. V. . (1997). Optimal robot excitation and identication. IEEE Transactions on Robotics and Automation, 13, 730740. Wang, Q.-G., Bi, Q., & Zou, B. (1996). Parameter identication of continuous-time mechanical systems without sensing acceleration. Computers in Industry, 28, 207217.

Acknowledgements This work was supported by CENIIT and ISIS at . Linkopings universitet.

References
Aberger, M. (2000). Effects of nonlinearities in black box identication of an industrial robot, Tech. Rep. LiTH-ISY-R-2322, Department . of Electrical Engineering, Linkoping University, SE-581 83. Link. oping, Sweden. Albu-Sch. ffer, A., & Hirzinger, G. (2001). Parameter identication a and passivity based joint control for a 7 dof torque controlled lightweight robot. In: Proceedings of the 2001 IEEE conference on robotics & automation. Seoul, Korea. ( . Astrom, K. J., & Wittenmark, B. (1984). Computer controlled systems: Theory and design. Englewood Cliffs, NJ: Prentice-Hall. Berglund, E., & Hovland, G. (2000). Automatic elasticity tuning of industrial robot manipulators. In: Proceedings of the 39th IEEE conference on decision and control. Sydney, Australia. D! pinc! , P. (1998). Parameters identication of exible robots. IEEE e e Proceedings on Robotics and Automation, 11161121. ElMaraghy, W.H., ElMaraghy, H.A., Zaki, A., & Massoud, A. (1994). A study on the design and control of robot manipulators with exibilities. In: Fourth IFAC symposium on robot control. Capri, Italy. Forssell, U., & Ljung, L. (1999). Closed-loop identication revisited. Automatica, 35, 12151241. Gautier, M., & Poignet, P. (2001). Extended Kalman ltering and weighted least squares dynamic identication of robot. Control Engineering Practice, 9, 13611372. Grotjahn, M., Daemi, M., & Heimann, B. (2001). Friction and rigid body identication of robot dynamic. International Journal of Solids and Structures, 38, 18891902.