Dynamic Model Identification For Industrial Robots: Integrated Experiment Design and Parameter Estimation
Dynamic Model Identification For Industrial Robots: Integrated Experiment Design and Parameter Estimation
Dynamic Model Identification For Industrial Robots: Integrated Experiment Design and Parameter Estimation
ndustrial robot manipulators are indispensable for achieving productivity and flexibility in fully automated production lines, where they are used for a wide variety of tasks, ranging from material handling and assembly to cutting, welding, gluing, and painting. To improve productivity and accuracy, robot manufacturers invest time and effort in developing advanced offline programming tools and controllers. Dynamic robot models are a key element in these developments.
OCTOBER 2007
1066-033X/07/$25.002007IEEE
environment. Many robot manufacturers, such as KUKA, ABB, and Fanuc, as well as third-party companies, offer software for offline programming. These interactive simulation environments allow the user to create virtual models of production environments and to program and simulate robot manipulator tasks before transferring the code to the robot controller (Figure 1). In particular, three-dimensional graphics are used to coordinate the elements of the production process and detect possible collisions and other system failures. Offline programming software does not consider uncertainty in the position and orientation of the objects in the production environment. In addition, it is assumed that the robot manipulator executes the programmed task with perfect position accuracy. However, the trajectories executed by the robot manipulator in practice differ significantly from the programmed and simulated trajectories due to kinematic modeling errors and robot dynamics. Inaccuracies due to kinematic errors can be compensated for by using a calibrated kinematic model of the robot manipulator [1]. However, for high-speed motions, or in case of a heavy payload, the tracking error is caused mainly by dynamic forces, for example, centrifugal and Coriolis forces, dynamic coupling between the joint axes, and actuator dynamics. These effects are not sufficiently accounted for by standard industrial robot controllers. Dynamic effects can be compensated, however, by switching to more advanced model-based controllers or by adapting the programmed robot trajectory. In the latter case, the offline programming environment must include dynamic robot models as well as dynamicsimulation functionality. Another objective of offline programming is to minimize the cycle time of a robot task, since shorter cycle times lead to higher productivity. However, physical constraints, such as maximum motor and gearbox torques as well as maximum motor speeds, must be accounted for to avoid overload, which causes accelerated wear and tear of the actuators, gears, and bearings, resulting in loss of accuracy and even premature failure. Realistic simulation and optimization of the motion of a robot manipulator, including its physical constraints, requires an accurate dynamic model.
controllers have been developed, for example, computedtorque control as well as feedforward dynamic compensation [2][5]. These controllers are more complex than standard industrial robot controllers since they are based on a model of the complete robot dynamics. Although practical industrial conditions are far from those in laboratory environments, results obtained on industrial setups confirm the benefits of model-based control [6]. Improved tracking accuracy is obtained, provided that the robot model is sufficiently accurate for torque prediction.
FIGURE 1 Illustrative graphical output of the KUKA offline programming environment KUKASim. A complete production process can be simulated in a virtual environment, collisions can be detected, and robot programs can be generated automatically. Cycle-time optimization is possible but is limited to geometric and kinematic information. (Courtesy of KUKA.)
OCTOBER 2007
sented by a kinematic chain of rigid bodies, and thus rigidbody dynamic equations are the basis for the model. Depending on the system and on the specifications, these rigid-body equations have to be complemented with models of other effects such as friction and gravity-compensating devices if present.
A priori Knowledge
Modeling Experiment Design Data Acquisition Signal Choose Parameter Processing Estimation Algorithm Parameter Estimation Validate Model Not Satisfied
Rigid-Body Dynamics
The Newton-Euler or Lagrangian method is used to derive the dynamic equations of kinematic chains of rigid bodies [8]. Both approaches yield the dynamics equation = M(q, ) + C(q, q, ) + g(q, ), q (1)
Satisfied
which expresses, for an n-degree-of-freedom robot, the n-vector of actuator torques as a function of the n-vectors of the joint positions q, velocities q, and accelerations q as well as the model parameters . In (1), M(q, ) is the n n inertia matrix, C(q, q, ) is the n-vector containing Coriolis and centrifugal forces, and g(q, ) represents gravitational torques. M(q, ), C(q, q, ), and g(q, ) are nonlinear functions of the model parameters , that is, the mass, centerof-gravity location, and moments and products of inertia of each link. Using the barycentric parameters [9] or the modified Newton-Euler parameters [8] yields a model of the form = (q, q, q) , (2)
FIGURE 2 Schematic representation of a standard experimental robot identification procedure. The kinematic and geometric information of the robot manipulator and model-accuracy specifications are the inputs to the identification procedure. This information is available prior to the identification and determines choices to be made in the procedure. The model validation step evaluates the accuracy of the model according to criteria that depend on the application of the model. If the identified model does not pass the validation tests, one or several steps of the procedure are repeated and choices are reconsidered.
Result
Figure 2. These inputs determine the choices to be made in the procedure. For example, kinematic and geometric information about the robot includes the number of joints, the orientations of the joint axes, and the lengths of the robot links. Model accuracy specifications determine the type of model to be used and the level of detail of the dynamics to be included in the model. These choices have implications for the various steps of the identification procedure. The last step of the identification procedure is model validation, where the user verifies that the model satisfies the accuracy specifications. If the obtained model does not pass the validation tests, one or several steps of the procedure are repeated and some of the choices are reconsidered.
which is linear in the unknown parameters. In (2), is the barycentric parameter vector, and is the observation or identification matrix, which depends only on the motion data. This property simplifies the parameter estimation considerably [10]. The barycentric parameters of a link are combinations of the inertial parameters of the link and its descendants in the kinematic chain [11]. For example, the barycentric mass of a link is defined as the mass of that link augmented by the total mass of all descendant links. The modified Newton-Euler parameters express the inertial parameters as first- and second-order moments with respect to a link frame located at the joint axis.
MODELING
Dynamic robot models define the relationship between the motion of the robot manipulator and the actuator torques. The motion of the robot is described by the position, velocity, and acceleration of all of its links. The robot is repre-
OCTOBER 2007
Periodic excitation allows us to integrate the experiment design, signal processing, and parameter estimation.
expressions that are linear in the parameters. Consequently, these effects fit conveniently within the linear-in-parameters model structure (2). Although friction is a complex nonlinear phenomenon, especially during motion reversal [14], [15], a friction model consisting of only Coulomb and viscous friction, that is, f ric = fC sign( ) + fv q, q (3)
is an acceptable simplification for many robotics applications [16]. Note that (3) is linear in the unknown Coulomb and viscous friction parameters fC and fv and thus is consistent with the linear model structure (2).
EXPERIMENT DESIGN
During the design of an identification experiment, it is necessary to ensure that the excitation is sufficient to provide accurate and fast parameter estimation in the presence of disturbances, and that the processing of the resulting data is simple and yields accurate results. The experiment design consists of two steps. First, a trajectory parameterization is selected, and second the trajectory parameters are calculated, usually by means of optimization.
tion. The coefficients ai,k and bi,k are the amplitudes of the sine and cosine functions along with the offset qi,0 of the position trajectory. In choosing the frequency range [ f , N f ] of the excitation trajectories, the following trade-off has to be considered. By selecting a low fundamental frequency f , that is, a long excitation period, a larger part of the robot workspace can be covered for given maximum joint velocities, however, at the cost of longer measurement time. Good coverage of the robot workspace improves the information content of the measurements as well as the accuracy of the parameter estimates [19]. On the other hand, including high frequencies provides high accelerations, which are required to accurately estimate the moments and products of inertia. The highest frequency of the commanded trajectory, however, is limited by the lowest resonance frequency of the robot structure. The structural flexibilities of the robot are excited if the highest frequencies of the trajectory are too close to the lowest resonance frequency. Excitation of flexible modes is disadvantageous because they are not accounted for in the rigid-body robot model. As discussed below, the commanded trajectory need not be followed exactly since the actual motion of the robot is measured.
(4)
where t represents time, and f is the same for all joints. This Fourier series is periodic with period T f = 2/ f , which is chosen to be an integer multiple of the sampling period. Each Fourier series contains 2N + 1 parameters, which are the degrees of freedom for trajectory optimiza-
Appropriate values for the trajectory parameters can be selected either by trial and error or by solving a nonlinear optimization problem with constraints imposed on the robot motion. Several objective functions exist for trajectory optimization. A popular optimization criterion is the logarithm of the determinant of the covariance matrix of the model parameter estimates, known as the d-optimality criterion [10]. This criterion measures the size of the uncertainty region of the model parameter estimates. Its calculation does not depend on the model parameters if the joint position, velocity, and acceleration data are free of noise but rather depends only on the robot trajectory through the identification matrix in (2) as well as on the covariance of the noise on the actuator torque measurements [20]. This property is useful in practice since the optimization of the robot excitation can be performed without any prior knowledge of the model parameters. The motion constraints impose limitations on the joint positions, velocities, and accelerations as well as on the robot end-effector position in Cartesian space. These limitations avoid collisions between the robot and objects in its environment as well as collisions between robot links.
OCTOBER 2007
Online programming has the advantage that positions defined by task execution are more accurate than a path specified in a software environment.
Measured Joint Position Sequence (One Period) Frequency Spectrum of Measured Joint Position Filtered Frequency Spectrum of Measured Joint Position Spectrum of Joint Velocity Noise-Free Joint Velocity
q(t)
DFT
commanded trajectories. Although these measurements may contain more or different harmonics than the commanded trajectories, the motion is periodic, with the same period as that of the commanded trajectories, and bandlimited, because the robot is controlled by a bandlimited position-feedback controller. The actuator torque data are obtained through actuator current measurements without additional sensors. The relationship between current and torque is modeled as linear or as a higher order polynomial [21], of which the parameters are provided by the motor manufacturer or identified in a separate experiment [22]. The aim of the signal processing step is to clean up the measured data. This step improves the signal-to-noise ratio of torque and joint position measurements, estimates the variance of the measurement noise, and calculates the joint velocity and acceleration estimates based on the measured joint positions.
Q(k)
Multiply by Rectangular Qf (k) Window
x j (k) j (k)Qf(k)
k
IDFT
q(t)
FIGURE 3 Exact frequency-domain differentiation of a measured periodic joint position signal. The left column explains the various steps of the procedure. The right column illustrates these steps for a periodic signal containing only one frequency, that is, a sinusoidal signal, perturbed with additive normally distributed noise. First, the signal is transformed to the frequency domain using the discrete Fourier transform (DFT). Next, the spectrum is multiplied by a rectangular window, which selects the frequencies that contain signal information. For this example, the multiplication corresponds to selecting one frequency. The spectrum is set to zero at all other frequencies. The resulting spectrum is then multiplied by the continuous-time frequency-domain representation of a differentiator at the selected frequency. That is, the spectrum is multiplied by j(k) = j2kfs /P , with fs the sampling frequency, P the number of time-domain samples, k = {1, 1} is the index of the selected frequency in the discrete spectrum. A transformation back into the time domain using the inverse DFT yields an estimate of the first time derivative of the original signal, that is, the velocity. The velocity signal is almost free of noise, that is, noise is removed from all frequencies except the selected ones. For simplicity, only the amplitude spectra are shown.
1 (MK 1)
(5)
k=1 m=1
where xm (k) indicates the kth sample within the mth peri od, and x(k) denotes the average of x, that is, x(k) = 1 M xm (k). M m=1 (6)
OCTOBER 2007
For robot identification the signal x corresponds to the measured actuator torques and joint positions q.
FIGURE 4 A KUKA IR 361 industrial robot in the robotics laboratory at KU Leuven. The position controller of the robot is provided by Orocos software [24], [25]. This software facilitates the application of periodic commanded trajectories, while synchronizing encoder and actuator current measurements.
The maximum likelihood estimation sits at the other side of the spectrum since it provides unbiased estimates with minimal uncertainty regardless of the spectrum of the measurement noise. The maximum likelihood estimate of the parameter vector is the value of that maximizes the likelihood of the measurements. This criterion, which is a nonconvex function of the unknown model parameters, depends on the covariance of the noise on all measurements and derived variables such as joint velocities and accelerations. Solving this nonlinear least squares optimization problem is often cumbersome because it requires an initial guess of the parameters and because it might converge to a local optimum, yielding a suboptimal solution that may be biased. When the joint position, velocity, and acceleration data are free of noise, which is for practical purposes realized by applying the above-mentioned signal processing, the identification matrix in (2) is free of noise, and the maximum likelihood estimation simplifies to weighted linear least squares estimation (WLSE). The complexity of WLSE is comparable to that of LLSE, in fact, the only difference is that WLSE weighs the data with the inverse of the covariance of the actuator torque measurement noise and there-
OCTOBER 2007
An objective of offline programming is to minimize the cycle time of a robot task, since shorter cycle times lead to higher productivity.
and TABLE 1 Comparison of actuator torque prediction errors with the measurement noise level. The root mean square of the actuator torque prediction errors is comparable to the standard deviation of the noise on the averaged actuator torque measurements.
Joint 1 6.1 N-m 3.3 N-m Joint 2 6.3 N-m 4.6 N-m Joint 3 2.9 N-m 1.5 N-m
t1 . = . . . tK
(10)
fore discriminates between accurate and inaccurate data. WLSE has the same favorable properties as the maximum likelihood estimation. The weighted least squares estimate of the model parameters is WLS = (FT =(
1
F)1 FT
1/2
(7) (8)
1/2
F)+
where ()+ denotes the pseudoinverse, (qt1 , qt1 , qt1 ) . . F= , . (qtK , qtK , qtK )
Here, (qtk , qtk , qtk ) and tk , for k = 1, . . . K, are the identification matrix and torque vector evaluated using the joint position, velocity, and acceleration data, as well as the actuator torque data at the discrete-time instant tk . Furthermore, is the covariance matrix of the actuator torque data, which is a dense matrix if the noise is correlated. Here, the assumption is made that the noise is white, which is made for convenience. Since the matrix F in (9) is free of noise, the parameter estimates are unbiased even if the actual noise is colored, although the estimates may have nonminimal uncertainty [23, pp. 92106]. The covariance matrix of the estimated parameter vec tor WLS is C = (FT
1
F)1 ,
(11)
(9)
2 1.5 Joint Position (rad) 1 0.5 0 0.5 1 1.5 2 0 1 Joint 1 Joint 2 Joint 3 2 3 4 5 6 Time (s) 7 8 9 10
which, along with the related uncertainty bounds on the parameters, can be used for model validation. The aim of the model-validation step is to obtain confidence in the estimated robot model in view of its intended application. Obviously, the most appropriate validation test is to use the model in the application and evaluate its success. This validation method, however, can lead to undesirable and even dangerous situations. Therefore, validation must be addressed prior to the actual application. Given this requirement, a validation test must evaluate the properties of the estimated model and the parameter estimates that are relevant to the intended application. Unsatisfactory validation leads to a reconsideration of some previous steps of the identification procedure, for instance, a new experiment design or a more elaborate dynamic model. We consider two model-validation measures, the actuator torque-prediction accuracy and the parameter-estimation accuracy.
FIGURE 5 Optimized robot-excitation trajectory. One period of the optimized joint trajectories is shown. These trajectories consist of a five-term Fourier series with a base frequency of 0.1 Hz. The trajectory parameters are optimized according to the d-optimality criterion, taking into account workspace limitations, constraints on joint velocities, and accelerations. The optimization criterion is a measure of the size of the overall uncertainty region of the parameter estimates.
OCTOBER 2007
yielding actuator torque predictions . The same motion is executed by the robot. The actuator torques are measured and compared with the predicted actuator torques. The difference between predicted and measured torque, that is, the prediction error, is expected to be small compared to the torque signal and preferably comparable to the noise level on the actuator torque measurements. Good validation practice requires that the validation trajectories be different from the excitation trajectory, while being representative of the intended application. For example, robot trajectories for painting and laser cutting applications are smooth and continuous in time, while applications such as pick and place and spot welding are characterized by trajectories with many starts and stops.
Parameter Accuracy
An alternative method for validating the model is to check the accuracy of the estimated parameters. Two approaches are possible. The estimated parameters are compared with an estimate of their confidence interval, or with estimates available from other sources, for example, CAD data of robot parts. Based on the parameter covariance matrix (11), it is possible to derive confidence intervals for each parameter. Comparison of the parameter values with their confidence interval gives an indication of the accuracy of the individual parameters. If this parameter covariance matrix is not available, an estimate of the parameter uncertainty can still be obtained by repeating the parameter estimation for different excitation trajectories. Subsequently, the mean value of the model parameters and their sample variance can be calculated. This approach is more time consuming than the confidence interval approach but is more reliable.
FIGURE 6 Three-dimensional visualization of the optimized excitation trajectory. The curve shows the trajectory of the robot tool center point of the KUKA IR 361 robot. Only the first, second, and third links of the robot are used in this experiment. The robot model identification is based on joint-angle and actuator-torque data measured during successive repetitions of the trajectory.
Frequency (Hz) FIGURE 7 Amplitude of the frequency spectrum of one period of the measured joint trajectories. These signals have a base frequency of 0.1 Hz and contain five dominant harmonics, in accordance with the commanded trajectories. The amplitudes of the higher frequency components are significantly smaller than the five low-frequency components, which shows that the measurement noise and tracking errors caused by the limited bandwidth of the robot controller are small. During the signal processing step, only the first five frequencies are selected. By ignoring the higher frequencies, for practical purposes noiseless joint position signals are obtained.
OCTOBER 2007
TABLE 2 Model validation based on actuator torque prediction for the validation trajectory. The RMS of the actuator torque prediction errors is small compared to the actuator torques, which indicates that the identified model can accurately predict the actuator torques for point-to-point motions. These motions are typical of pick-and-place and spot-welding operations.
Joint 1 8.9 N-m Joint 2 11.8 N-m Joint 3 4.4 N-m
Parameter Estimation
The robot model for the first three links contains 15 barycentric parameters, six friction parameters (viscous and Coulomb friction parameters for each joint), and two additional parameters that model the gravity-compensating spring. The rotor inertia of the actuator for the third axis is added as an extra parameter to account for the coupling effect introduced by the mounting of the third actuator on the second link [12]. These model parameters are estimated using the robot identification procedure discussed above. Figure 8 compares the averaged measured torques for the excitation trajectory with the predicted actuator torques based on the identified model parameters and the available motion data. The third column shows that the prediction error is small, except during velocity reversal due to the limited accuracy of the friction model at low velocities. Table 1 compares the RMS of the prediction errors with the standard deviation of the noise on the averaged actuator torque measurements.
trajectories are implemented on the Orocos [24], [25] control platform, which replaces the standard industrial position controller for this robot. After the transients die out, the joint positions and motor currents of the first three robot joints are measured. The total measurement time is 160 s, corresponding to 16 periods of the excitation trajectory. The data sampling rate is 150 Hz. Figure 7 shows the frequency spectrum of one period of the joint trajectories measured during the robot-excitation experiment.
Measured Torque Torque Joint 1 (Nm) 100 0 100 200 100 0 100 200
10
10
10
200 100 0 100 200 0 100 50 0 50 0 5 Time (s) (a) 10 100 0 5 Time (s) (b) 10 5 10
100 50 0 50 100
FIGURE 8 Model validation based on actuator torque prediction for the excitation trajectory. The predicted actuator torques of the first three joints of the KUKA IR 361 robot are compared with the averaged measured torques for the excitation trajectory. (c) shows that the prediction errors are small. The remaining peaks occur during velocity reversal, indicating that the friction model cannot capture the complex dynamic friction behavior at low velocities.
OCTOBER 2007
Validation Experiment
For validation, we choose a trajectory that is representative for pick-and-place as well as spot-welding operations. This trajectory (see Figure 9) consists of point-to-point motions between 20 randomly chosen points in the robot workspace. The robot moves with maximum acceleration and deceleration between these points, resulting in a trapezoidal velocity profile between two successive points, and comes to rest with zero dwell time at each point. Figure 10 shows the measured and predicted actuator torques and the corresponding prediction errors. Comparison of the prediction errors with the measured torques shows that the model we obtain is capable of accurately predicting the actuator torque data. The peaks in the prediction error indicate again that the assumed friction model, which includes viscous and Coulomb friction only, is unable to capture the complex dynamic friction behavior. Table 2 shows the RMS of the actuator torque prediction errors for this validation experiment. Extending the robot model to include more advanced friction models, as well as models of the mechanical losses in the actuators and the efficiency of the transmissions, further improves the prediction accuracy [26]. These model extensions are, however, nonlinear in the parameters, and thus considerably complicate parameter estimation.
2.5 2 1.5 Joint Position (rad) 1 0.5 0 0.5 1 1.5 2 2.5 0 10 20 30 Time (s) 40 50 Joint 1 Joint 2 Joint 3
FIGURE 9 Robot model validation trajectory. The trajectory consists of point-to-point motions between 20 randomly chosen points in the robot workspace. The robot moves with maximum acceleration and deceleration between these points, and comes to rest with zero dwell time at each point. This validation trajectory is representative of pick-and-place and spot-welding operations.
is usually not feasible inside a production line, yielding inaccurate parameter estimates.
where L is the payload identification matrix, and L is a vector that contains the barycentric parameters of the payload. These barycentric parameters contain the payload mass as well as the first- and second-order moments with respect to the last robot rotation axis. Since the robot payload is the last body in the serial robot structure, an unambiguous transformation exists between the barycentric parameters of the payload and its classical inertia parameters. The first term in (13), which corresponds to (2), describes the robot dynamics. This term represents the actuator torque contributions that result from the inertial parameters, possibly gravity-compensating springs, and joint friction. The first two contributions do not change by adding a payload to the robot. Therefore, the corresponding parameter set is considered as a priori information for the robot payload identification. The joint friction changes with the payload, and, consequently, the parameters are reestimated. Altogether, the number of unknown parameters
OCTOBER 2007
remains small, which is beneficial for the robustness and accuracy of the payload parameter estimation.
tion, because the former is most often performed inside a production line.
Excitation Trajectory
An excitation trajectory that is optimal for payload identification is quite different from the trajectories used for robot identification. This difference is explained as follows. The inertial parameters of the base links, which are in most cases the first three links, dominate the dynamics of the robot. Therefore, trajectory optimization for robot identification typically yields fast motions for the base joints. On the other hand, accurate estimation of the inertial parameters of the payload requires fast payload motions, which are easily accomplished using the robot wrist joints and one additional base joint. Moreover, the workspace available for performing the experiments is often more limited for payload than for robot identifica-
100 0 200 20 40
20
40
20
40
20
40
50
50
50
FIGURE 10 Model validation based on actuator torque prediction for the validation trajectory. The predicted actuator torques of the first three joints of the KUKA IR 361 robot are compared with the measured torques for the validation trajectory. Part (c) shows that the prediction errors are small. These small prediction errors show that the estimated robot model can accurately predict the actuator torques, and is thus well suited for model-based robot control and dynamic task optimization.
OCTOBER 2007
controller and a payload capacity of 15 kg. To validate the accuracy of the payload identification method, a calibrated reference payload is used, shown in Figure 12. A detailed CAD model provides estimates of all ten inertial parameters of this payload. These values are used as reference values for the robot payload identification experiment presented here and are shown in the second column of Table 3. The robot excitation is limited to the last four joints of the manipulator. A fundamental frequency of 0.033 Hz is selected for the excitation trajectories, yielding a period of 30 s. The trajectories are two-term Fourier series, consisting of the fundamental frequency and its 20th or 25th harmonic. These choices for the robot excitation are imposed by the payload identification plug-in of the robot controller. The data are sampled with a sampling period of 12 ms.
TABLE 3 Validation of the accuracy of the inertial parameter estimates for the reference payload shown in Figure 12. The reference values of the inertial parameters result from a CAD model of the payload. Ten experiments with the same payload, each using a different excitation trajectory, are performed. From the resulting ten sets of parameter estimates, the mean value and standard deviation are calculated, given in the third and fourth columns, respectively. All reference values lie within the 2 confidence interval of the mean estimated values.
Reference Value 9.579 0.024 0.090 0.202 0.612 0.063 0.637 0.158 0.002 0.008 Mean Value 9.618 0.0246 0.0930 0.2065 0.677 0.121 0.621 0.164 0.004 0.010 Standard Deviation 0.174 0.0010 0.0016 0.0064 0.077 0.077 0.020 0.011 0.010 0.016
Parameter m (kg) cx (m) cy (m) cz (m) Ixx (kg-m2 ) Iyy (kg-m2 ) Izz (kg-m2 ) Ixy (kg-m2 ) Ixz (kg-m2 ) Iyz (kg-m2 )
with the payload of the robot. A re-identification of the friction parameters is therefore necessary to obtain accurate parameter estimates. Ten identification experiments are performed. Each experiment uses a different excitation trajectory and yields a set of parameter estimates based on data measured during ten periods. Based on the resulting ten sets of parameter estimates, mean parameter values and their standard deviation are calculated. The results are presented in Table 3 and compared with the reference values. The reference values lie within the 2 confidence interval around the average parameter estimates.
FIGURE 11 The KUKA KR 15 industrial robot in the robotics laboratory at KU Leuven with a reference payload. The KUKA KR 15 is a sixdegree-of-freedom industrial robot with a PC-based controller and a payload capacity of 15 kg. A software plug-in makes it possible to perform robot payload identification experiments that use periodic trajectories for the last four robot axes, that is, the elbow and wrist axes.
FIGURE 12 The reference payload used to validate the payload identification procedure. This payload is reconfigurable. For example, its mass and principal moments of inertia can be changed between 1.4 and 15 kg and from nearly zero to 3 kg-m2 , respectively. The reference values for the inertia parameters are calculated from a CAD model. The second column of Table 3 shows these reference values for the configuration of the payload shown here.
OCTOBER 2007
The use of periodic excitation is the key feature of the robot identification method.
Figure 13 evaluates the model accuracy based on its actuator torque prediction accuracy for a validation trajectory. The torque prediction error is small except during velocity reversals, as expected. The accuracy of the parameter estimates and the actuator torque prediction satisfy the requirements imposed by industry. The mass of a payload can be estimated within 5% and its center of mass within 1 cm.
CONCLUSIONS
The use of periodic excitation is the key feature of the presented robot identification method. Periodic excitation
allows us to integrate the experiment design, signal processing, and parameter estimation. This integration simplifies the identification procedure and yields accurate models. Experimental results on an industrial robot manipulator show that the estimated dynamic robot model can accurately predict the actuator torques for a given robot motion. Accurate actuator torque prediction is a fundamental requirement for robot models that are used for offline programming, task optimization, and advanced model-based control. A payload identification approach is derived from the integrated robot identification method, and possesses the same favorable properties.
Measured Torque 200 Axis 3 (Nm) 0 200 400 600 200 Axis 4 (Nm) 100 0 100 200 200 Axis 5 (Nm) 100 0 100 200 100 Axis 6 (Nm) 50 0 50 100 0 5 10 15 Time (s) (a) 20 25 30 0 5 10 15 20 25 30 0 5 10 15 20 25 30 0 5 10 15 20 25 30 200 0 200 400 600 200 100 0 100 200 200 100 0 100 200 100 50 0 50 100 0 5 0 5 0 5 0 5
10
15
20
25
30
10
15
20
25
30
10
15
20
25
30
10
20
25
30
FIGURE 13 Model validation based on actuator torque prediction for the validation trajectory. The predicted actuator torques of joints three to six of the KUKA KR 15 robot with reference payload are plotted with the measured torques in (a). (b) shows that the prediction errors are small.
OCTOBER 2007
ACKNOWLEDGMENTS
This article presents research results of the K.U. Leuven Concerted Research Action GOA/05/10. The work is also sponsored by a specialization grant of the Institute for the Promotion of Innovation through Science and Technology in Flanders (IWT-Vlaanderen), by the Research Council K.U.Leuven, CoE EF/05/006 Optimization in Engineering (OPTEC), and by the Belgian Network Dynamical Systems, Control, and Optimization, funded by the Interuniversity Attraction Poles Programme, initiated by the Belgian State, Science Policy Office. The scientific responsibility rests with its authors.
[18] M. Gautier and W. Khalil, Exciting trajectories for the identification of base inertial parameters of robots, Int. J. Robot. Res., vol. 11, no. 4, pp. 362375, 1992. [19] R. Pintelon and J. Schoukens, System Identification: A Frequency Domain Approach. New York: Wiley, 2001. [20] M. Gautier, Identification of robots dynamics, in Proc. IFAC Symp. Theory Robots, Vienna, Austria, 1986, pp. 351356. [21] J. Swevers, B. Naumer, S. Pieters, E. Biber, W. Verdonck, and J. De Schutter, An experimental robot load identification method for industrial application, in Advanced Robotics: Experimental Robotics VIII, New York: Springer-Verlag, 2003, pp. 318327. [22] P. Corke, In situ measurement of robot motor electrical constants, Robotica, vol. 14, no. 4, pp. 433436, 1996. [23] J.P. Norton, An Introduction to Identification. London: Academic, 1986. [24] H. Bruyninckx, (2001) Open Robot Control Software, Leuven, Belgium. [Online]. Available: http://www.orocos.org/ [25] P. Soetens and H. Bruyninckx, Realtime hybrid task-based control for robots and machine tools, in Proc. IEEE Int. Conf. Robot. Automation, Barcelona, Spain, 2005, pp. 260265. [26] J. Swevers, W. Verdonck, B. Naumer, S. Pieters, and E. Biber, An experimental robot load identification method for industrial application, Int. J. Robot. Res., vol. 21, no. 8, pp. 701712, 2002. [27] G. Hirzinger, N. Sporer, A. Albu-Schffer, M. Hhnle, R. Krenn, A. Pascucci, and M. Schedl, DLRs torque-controlled light weight robot III Are we reaching the technological limits now? in Proc. IEEE Int. Conf. Robot. Automation, Washington DC, 2002, pp. 17101716.
REFERENCES
[1] RRS robot controller simulation (RCS) interface, Berlin, 2001. [Online]. Available: http://www.realistic-robot-simulation.org [2] J.Y.S. Luh, M.W. Walker, and R.P. Paul, On-line computational scheme for mechanical manipulators, J. Dynamic Syst., Measure. Control, vol. 102, no. 2, pp. 6976, June 1980. [3] J.-J.E. Slotine, The robust control of robot manipulators, Int. J. Robot. Res., vol. 4, no. 2, pp. 4964, 1985. [4] P. Khosla and T. Kanade, Experimental evaluation of nonlinear feedback and feedforward control schemes for manipulators, Int. J. Robot. Res., vol. 7, no. 1, pp. 1828, Feb. 1988. [5] C. Canudas de Wit, B. Siciliano, and G. Bastin, Theory of Robot Control. New York: Springer-Verlag, 1996. [6] F. Caccavale and P. Chiacchio, Identification of dynamic parameters and feedforward control for a conventional industrial manipulator, Control Eng. Practice, vol. 2, no. 6, pp. 10391050, 1994. [7] M. Olsen, J. Swevers, and W. Verdonck, Maximum likelihood identification of a dynamic robot model: Implementation issues, Int. J. Robot. Res., vol. 21, no. 2, pp. 8996, Feb. 2002. [8] C. Atkeson, C. An, and J. Hollerbach, Estimation of inertial parameters of manipulator loads and links, Int. J. Robot. Res., vol. 5, no. 3, pp. 101119, 1986. [9] P. Fisette, B. Raucent, and J.-C. Samin, Minimal dynamic characterization of tree-like multibody systems, Nonlinear Dynamics, vol. 9, no. 12, pp. 165184, 1996. [10] L. Ljung, System Identification: Theory for the User. Englewood Cliffs, New Jersey: Prentice-Hall, 1987. [11] M. Renaud, Quasi-minimal computation of the dynamic model of a robot manipulator utilizing the Newton-Euler formalism and the notion of augmented body, in Proc. IEEE Int. Conf. Robot. Automation, Raleigh, North Carolina, 1987, pp. 16771682. [12] L. Sciavicco, B. Siciliano, and L. Villani, Lagrange and Newton-Euler dynamic modelling of a gear-driven rigid robot manipulator with inclusion of motor inertia effects, Advanced Robotics, vol. 10, no. 3, pp. 317334, 1996. [13] W. Khalil and E. Dombre, Modeling, Identification and Control of Robots. London: Hermes Penton Ltd, 2002. [14] J. Swevers, F. Al-Bender, C. Ganseman, and T. Prajogo, An integrated friction model structure with improved presliding behavior for accurate friction compensation, IEEE Trans. Automat. Contr., vol. 45, no. 4, pp. 675686, 2000. [15] F. Al-Bender, V. Lampaert, and J. Swevers, The generalized Maxwellslip model: A novel model for friction simulation and compensation, IEEE Trans. Automat. Contr., vol. 50, no. 11, pp. 18831887, Nov. 2005. [16] M. Daemi and B. Heimann, Identification and compensation of gear friction for modeling of robots, in Proc. 11th CISM-IFToMM Symp. Theory and Practice Robots and Manipulators, Udine, Italy, 1996, pp. 8996. [17] B. Armstrong, On finding exciting trajectories for identification experiments involving systems with nonlinear dynamics, Int. J. Robot. Res., vol. 8, no. 6, pp. 2848, 1989.
AUTHOR INFORMATION
Jan Swevers (jan.swevers@mech.kuleuven.be) received the M.Sc. degree in electrical engineering and the Ph.D. degree in mechanical engineering from the Katholieke Universiteit Leuven, Belgium, in 1986 and 1992, respectively, where he is currently a professor in the Department of Mechanical Engineering, Division of Production Engineering, Machine Design and Automation (PMA). His research interests include modeling, identification, and control of mechatronic systems. He can be contacted at the Department of Mechanical Engineering, Division PMA, Katholieke Universiteit Leuven, Celestijnenlaan 300 B, B3001 Heverlee, Belgium. Walter Verdonck received the M.Sc. degree in mechanical engineering and the Ph.D. degree in mechanical engineering from the Katholieke Universiteit Leuven, Belgium, in 1999 and 2004, respectively. He is currently employed as a research engineer at the Flanders Mechatronics Technology Centre. His research interests include modeling, identification, and control of robots and mechatronic systems. Joris De Schutter received the M.Sc. degree in mechanical engineering and the Ph.D. degree in mechanical engineering from the Katholieke Universiteit Leuven, Belgium, in 1980 and 1986, respectively, as well as an M.Sc. degree from M.I.T. in 1981. He is currently full professor in the Department of Mechanical Engineering, Division of Production Engineering, Machine Design and Automation at K.U. Leuven. His research interests include programming and control of sensor-based robots, as well as optimal design and control of high performance mechanical and mechatronic drive systems.
OCTOBER 2007