Você está na página 1de 54

Dynamic model identification for industrial robots

An integrated experiment design and parameter estimation approach

Walter Verdonck, Jan Swevers, Joris De Schutter


Division PMA,
Dept. Mechanical Engineering
Katholieke Universiteit Leuven
Celestijnenlaan 300 B
B 3001 Heverlee, Belgium

I NTRODUCTION

Product differentiation, customization, and just-in-time logistics require a new approach


to manufacturing. Production lines have to be flexible, and therefore machine setup times must
be short, to be able to handle small batch sizes efficiently and economically. In addition, ever
higher quality standards and ever increasing international competition impose higher accuracy,
speed, and reliability requirements on manufacturing processes.

Industrial robot manipulators have become indispensable in view of increasing productivity and flexibility in fully automated production lines. Industrial robot manipulators are flexible
production machines. They can be used for a large variety of tasks, ranging from material
handling and assembly to cutting, welding, gluing and painting. The robot setup times can
be kept short if the robot task is programmed without interrupting the production process. This
1

approach is known as offline programming. In offline programming the robot task is programmed
in a software environment. The complete production process is simulated and optimized in
this software environment before transferring the obtained program to the robot controller. In
industrial practice, however, robots are still programmed manually by leading the actual robot
through the desired sequence of position points. This approach is known as online programming.
It has the advantage that the taught positions are run through more accurately during the task
execution than a path specified in a software environment. On the other hand, online programming
leads to idle times that are unacceptable in the case of small product batch sizes.

An important objective of process optimization during offline programming is minimization of the cycle time of a robot task. Evidently, smaller cycle times lead to higher productivity.
However, physical constraints, such as maximum motor and gear box torques and maximum
motor speeds, have to be accounted for in this optimization, to avoid mechanical overload of
these components. Mechanical overload causes accelerated wear and tear of the actuators, gears
and bearings, which in turn may cause loss of accuracy and even premature failure. In summary,
overload cuts back the reliability and accuracy of the process significantly, and must be avoided at
all times. Realistic simulation and optimization of the motion of a robot manipulator that accounts
for its physical constraints, requires an accurate dynamic model of the robot manipulator.

Accurate dynamic models are also needed in advanced model based robot control. A
standard industrial robot controller does not rely on a dynamic robot model. Usually such a
controller consists of independent PID-like position controllers, one for each robot joint. Although
these controllers yield sufficiently accurate path tracking for most industrial applications, some

recent applications like laser welding and laser cutting require a more advanced control approach.
These modern applications are characterized by complex six degree-of-freedom trajectories, fast
motion and stringent path tracking accuracy requirements. Fast motions result in high dynamic
coupling between the different robot links, that cannot be compensated for by a standard robot
controller. Advanced model based robot controllers take dynamic coupling into account, and yield
significantly better path tracking accuracy. Several robot manufacturers have already implemented
this type of controllers, or are developing them for their next generation of robots. It goes without
saying that the accuracy of the dynamic robot model is crucial for the performance of these
advanced controllers.

A dynamic robot model describes the rigid body dynamics of the robot and includes
Coulomb and viscous friction in the joints. The structure of such a model, that is the structure of
the model equations, is derived according to the recursive Newton-Euler or Lagrange formulation.
This structure depends only on the kinematic structure of the robot, which is well known for all
types of industrial robots. Consequently, the accuracy of a dynamic robot model depends mainly
on the accuracy of its parameters. Often, however, the values of these model parameters are not
known with sufficient accuracy. CAD data of robot parts are, if available, mostly not sufficiently
accurate. Dismantling the robot to measure the mass and inertia properties of the links is not a
realistic option. Moreover, estimates of friction parameters cannot be obtained using either of
these approaches.

Compared to the above approaches, experimental identification is a much more efficient


approach to obtain accurate estimates of all robot model parameters. This approach uses motion

and actuator torque data that are measured during well-designed experiments. This paper presents
such an experimental robot identification procedure based on a statistical framework. Its key
feature is the use of periodic bandlimited excitation [1]. Periodic excitation simplifies the
processing of measured signals as well as the estimation of model parameters using the maximum
likelihood method. The obtained models satisfy the requirements imposed by industry. This
method is therefore recognized by industry as an effective means of robot model identification.
A particular industrial application is robot payload identification.

The excitation trajectory for each joint is a finite Fourier series with optimized coefficients.
The optimization criterion is the expected uncertainty on the estimated model parameters. This
uncertainty depends on an estimate of the noise level on the actuator torque measurements.
The periodicity of the excitation trajectories and of the resulting response signals allows us
to improve the signal-to-noise ratio of the measured data by simply averaging over multiple
periods, and to estimate the measurement noise level. The estimated noise level is valuable in
view of optimizing the accuracy of the estimated model parameters within a maximum likelihood
estimation framework. In addition, precise frequency domain data filtering is possible, while the
measured joint positions can be accurately differentiated yielding consistent noise-free joint
velocity and acceleration estimates.

This robot identification approach is also the basis for a payload identification method,
that accurately estimates all inertial parameters of the payload attached to the robot end effector.
These estimates can be used to update the robot model each time another payload or tool is
attached to the robot.

The paper is organized as follows. First, two potential applications of dynamic robot
models, namely offline programming, or robot task optimization, and advanced model based robot
control are introduced. Next, the different steps of an experimental robot identification procedure
are described. The integrated approach and its main properties are discussed in detail. Then
follows the extension to robot payload identification. Finally, two case studies are considered:
experimental identification of an industrial robot manipulator, and estimation of the inertial
parameters of a reference payload.

T HE IMPORTANCE OF ACCURATE DYNAMIC

MODELS OF ROBOT MANIPULATORS

The outcome of an experimental robot identification procedure is a set of model parameter


estimates. In combination with the selected model structure, this parameter set yields a model
of the robot behavior. This model describes the dynamic relation between the actuator torques
(actuator data) and the joint positions (motion data) of the robot. Experimental robot identification
aims at providing the best possible model accuracy. Model accuracy can however be expressed
in different ways. One way is to focus on the accuracy of the model parameters themselves. In
this case model accuracy is expressed as the bias, or systematic error, and the uncertainty on
the model parameters. Another measure is the accuracy with which the model is able to predict
the actuator torques (model outputs) for a given set of motion data (model input). In this case
model accuracy is expressed as the root-mean-square or maximum value of the prediction error.
The application of the robot model determines which measure is more appropriate. This section
discusses the two main applications of dynamic robot models and indicates which accuracy
measure is most relevant.
5

Offline programming and task optimization

Nowadays, 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 (Figure 1). Furthermore, three-dimensional graphics support
inspection of the cooperation and coordination between the different elements of the production
process, as well as detection of possible collisions or other errors.

Modern offline programming software does not consider uncertainties on 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 absolute position accuracy. That
is, inaccuracies caused by kinematic errors as well as robot dynamics are not considered. As a
result, the trajectories executed by the real robot manipulator may differ significantly from the
programmed and simulated trajectories. Often, the programmed trajectories have to be corrected
manually on the shop floor to compensate for these differences. These corrections are very time
consuming and therefore costly.

We now analyze the causes of these differences in more detail. Commercial offline
programming systems use kinematic models that take into account only the nominal robot
geometry and nominal constraints on position and velocities. Inaccuracies caused by kinematic
errors, which are mainly due to offsets in the joint position measurements as well as geometric
errors in the joints and links, can be compensated for by using a calibrated kinematic model of
the robot manipulator [2]. This approach results in a path following accuracy that is sufficient for
6

applications involving slow robot motions. However, for high speed motions, or in case of a heavy
payload, the tracking error remains significant. In those cases the tracking error is mainly caused
by the dynamic forces, for example centrifugal and Coriolis forces, dynamic coupling between
the joint axes, and actuator dynamics. These dynamic effects are not sufficiently compensated
for by the industrial robot joint velocity and position controllers. They can be compensated
however by switching to more advanced model based controllers, which are discussed in the
following section, or by adapting the programmed robot trajectory. In the latter case, the offline
programming environment has to be extended to include dynamic (robot) models, and a dynamic
simulation functionality. The dynamic robot model in combination with a model of the industrial
robot controller constitutes a filter, that is used to pre-filter the desired trajectory. The result is
a modified trajectory. Improved tracking accuracy is obtained when this modified trajectory is
applied to the robot with its industrial controller.

Offline programming offers several additional benefits once accurate path tracking is
guaranteed. In many industrial applications considerable gains in productivity can be achieved
by minimizing the cycle time. However, while minimizing the cycle time, the accuracy must
be kept constant. When reliable dynamic models are available, this problem can be solved with
efficient optimal motion planning algorithms. These algorithms take into account physical limits
of the robot such as workspace constraints, actuator power and torque constraints, and constraints
on the reaction forces and torques that act on the robot base plate. Alternative optimization
criteria, such as minimal energy consumption, are easily incorporated. These algorithms allow
us to distribute the load efficiently over the different actuators to prevent dynamic overload and
consequent wear and failure.

Advanced robot control

Although standard industrial controllers are still widely used, they yield insufficient path
tracking accuracy at high velocities and accelerations, especially when the robots are carrying
heavy payloads. There are two main reasons for their popularity. These controllers consist of
independent PID-like position and velocity control loops, one for each joint, and are therefore
simple with respect to implementation and tuning. These controllers are also very robust. This
robustness is due to the collocation of the actuators and sensors, that is, the sensors used for
feedback are mounted directly on the axle of the actuator.

These controllers, however, cannot appropriately cope with high velocities and accelerations as they lack any information about the system dynamics. Consequently these controllers
cannot sufficiently compensate for centrifugal and Coriolis forces, and for dynamic coupling
between the different robot links. For that reason, more advanced robot controllers have
been developed, for example the computed torque controller as well as feedforward dynamic
compensation control [3], [4], [5], [6]. These controllers are much more complex than standard
industrial robot controllers as they consider the complete robot dynamics and include a robot
model. Using this model, the above mentioned dynamic effects can be compensated for. Although
practical industrial conditions are far from those in laboratory environments, the results obtained
on an industrial setup have confirmed that it is worthwile using model based control [7]. Improved
tracking accuracy is obtained, provided that the robot model is sufficiently accurate.

For many years, these control algorithms have only been adopted in robotic research
labs as their complexity required advanced experimental digital controller hardware. Recently
8

however, such advanced controller hardware has become available for industrial use. This evolution creates a renewed interest in these control algorithms and stimulates robot manufacturers
to develop advanced model based industrial controllers.

The type of application of the dynamic robot model determines which measure of
accuracy is more appropriate. For both applications discussed above, the principal requirement
is that the model can accurately predict the actuator torques that correspond to a desired motion.
The requirement that the estimated parameters are close to the real parameter values is only
important in some specific applications, for example robot payload identification. Robot payload
identification is in most cases performed on the shop floor, by the robot operator. The operator
will only accept the outcome of the identification if the obtained parameters are close to his or
her expectations.

A N INTEGRATED EXPERIMENTAL ROBOT IDENTIFICATION

APPROACH

In experimental robot identification the parameters of dynamic robot models are estimated
using response data of the robot, measured during specially designed experiments. The response
data consist of sequences of joint positions and motor currents. From these data sequences of
joint velocities, accelerations, and motor torques are calculated. Experimental robot identification
is the only efficient way to obtain accurate robot models together with indications about their
accuracy, confidence interval and validity.

The following section gives an overview of a standard robot identification procedure. This
procedure consists of several steps, which are briefly described in the subsequent sections. The
9

advantages of using periodic excitation are discussed and illustrated. It enables the integration
of the main steps of the identification procedure: experiment design, signal processing, and
parameter estimation. Periodic excitation is the key feature of the identification procedure and
the main reason for its success.

A standard experimental robot identification procedure

Figure 2 shows the scheme of a standard robot identification procedure. The input to the
identification procedure is a priori kinematic and geometric information of the robot manipulator,
as well as specifications about the model accuracy. These inputs determine the choices to be
made in many steps of the procedure. For example, kinematic and geometric information of the
robot includes the number of joints, the orientation of the joint axes, the length of the robot
links, etc. 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 on the
different steps of the identification procedure as well.

The output of the identification procedure is a robot model that passes the model validation
tests. It consists of the selected model structure together with the parameter estimates. 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.

The different steps of the procedure are indicated in blue and red. Periodic excitation
contributes to the steps indicated in red. These steps are discussed in detail. The other steps are
only mentioned briefly.
10

Model generation

The generation of a dynamic robot model is the first step of an identification procedure.
Two conflicting effects have to be considered here: bias and uncertainty. Bias errors are avoided
by including all elements that determine the dynamic behavior of the robot manipulator in the
model structure. This often results in very complex models with many unknown parameters.
However, parameter uncertainty, and consequently model uncertainty, increases for a certain set
of measurement data, with the number of unknown model parameters. A compromise is required.
This compromise depends on the model accuracy specifications and on the intended application
of the model.

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 parts. For most robot manipulators, both link and joint flexibility
may be neglected. The robot system is represented by a number of rigid bodies. The rigid
body dynamic equations are the basis for the models considered in most robot identification
procedures. Depending on the system and on the specifications, these rigid body equations have
to be complemented with other effects such as friction and gravity compensation springs.

Rigid body dynamics

Many approaches exist to derive the dynamic equations of kinematic chains of rigid
bodies such as robots. The most common approaches are the Newton-Euler iterative calculation

11

scheme and the Lagrangian methods [8]. Both approaches yield the inverse dynamics equation
= M (q)
q + C(q, q)
q + g(q).

(1)

This equation 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. M (q) is the
n n inertia matrix, C(q, q)
is the n-vector containing Coriolis and centrifugal forces, and g(q)
represents the gravitational torques. Equation (1) is a nonlinear function of the classical inertia
parameters of each link: the mass, the position of the center of gravity of the link, the moments
and products of inertia.

Fortunately, other parametrizations as well as nonlinear parameter transformations exist


that transform equation (1) into a model that is linear in a new set of unknown parameters ,
= (q, q,
q) .

(2)

is the observation or identification matrix, which depends only on the motion data. is
the vector of transformed parameters. The barycentric parameters are an example of such an
alternative parametrization [9].

The linear property is interesting because it simplifies the parameter estimation considerably. This issue is discussed below. Almost all existing robot identification methods use this
type of linear models.

12

Additional dynamic effects

Rigid body equations (1) and (2) include only the effects of link masses and inertias.
Friction, dynamic coupling due to the inertia of the rotor of the actuators, as well as the effects
of gravitation compensation springs contribute significantly to the dynamic behavior of the robot
manipulator, and have to be taken into account in the robot model. Gravity compensating springs
reduce the static load on the base actuators, and are often used in robot systems. Dynamic
coupling and gravity compensating springs can be described by mathematical expressions that
are linear in the parameters. Consequently, they fit perfectly in the linear model structure (see
equation (2)).

Although friction is a complex nonlinear phenomenon, especially at motion reversal [10],


[11], a friction model consisting of only Coulomb and viscous friction, that is
f ric = fC sign(q)
+ fv q,

(3)

has proved to be an acceptable simplification for robotics applications. This friction model is
linear in the unknown Coulomb and viscous friction parameters fC and fv respectively. Hence,
it also fits in the linear model structure of equation (2).

Experiment design

Accurate robot identification requires specially designed experiments. During the design
of an identification experiment, it is essential to consider (1) whether the excitation is sufficient
to provide accurate and fast parameter estimation in the presence of disturbances, and (2)
13

whether the processing of the resulting data is simple and yields consistent and accurate results.
The experiment design consists of two steps. First, an appropriate trajectory parametrization is
selected, and secondly appropriate trajectory parameters are calculated, preferably through some
kind of optimization.

Trajectory parametrization

Several approaches exist to parametrize robot excitation trajectories, for example, finite
sequences of joint accelerations [12], or fifth-order polynomials interpolating between finite sets
of joint positions and velocities which are separated in time [13]. Although these trajectories
provide a good excitation of the robot dynamics after optimization, the resulting measurement
data are not periodic, nor bandlimited. These two properties are important.

The signal processing that precedes the parameter estimation simplifies considerably and
is consistent as well as accurate if the robot excitation is periodic and bandlimited. Both are
key features of the presented integrated identification method. The excitation trajectory for each
joint is periodic. Periodic excitation yields a periodic response, that is, all measured signals are
periodic after the transient robot response has died out. The angular position qi for each joint i
is written as a a finite Fourier series, that is, a finite sum of sine and cosine functions
qi (t) = qi,0 +

N
X

(ai,k sin(kf t) + bi,k cos(kf t)) ,

(4)

k=1

where t represents the time, and f is the fundamental pulsation of the finite Fourier series.
This fundamental pulsation is the same for all joints. This Fourier series is periodic, with period
Tf =

2
.
f

Each Fourier series contains 2N +1 parameters, which are the degrees-of-freedom for
14

the trajectory optimization: the amplitudes of the sine and cosine functions ai,k , and bi,k , for
k = 1 to N , and the offset qi,0 on the position trajectory.

The advantages of using bandlimited periodic excitation are obvious. Time domain data
averaging is possible. This reduces the amount of data that is used for the parameter estimation
and improves the signal-to-noise ratio of the data. This property is very important for the motor
current (torque) measurements which are usually very noisy. In addition, periodicity allows us to
estimate the variance of the measurement noise by simply calculating the sample variance. This
information is valuable for the maximum likelihood parameter estimation. Furthermore, joint
velocities and accelerations can be calculated through analytical differentiation of the measured
joint angles. While analytical differentiation yields accurate noise-free velocity and accelerations
estimates, it simplifies the model parameter estimation and it improves the accuracy of the
parameter estimates. Finally, the bandwidth of the excitation trajectories can be specified, such
that excitation of the robot flexibilities can be either completely avoided or intentionally brought
about. All these advantages are discussed further below.

None of the other existing robot excitation methods possesses the above mentioned
important features.

Trajectory optimization

Appropriate values for the trajectory parameters can either be selected by means of trial
and error, as is often done in industry, or by solving a complex nonlinear optimization problem
with constraints imposed on the robot motion.
15

Several objective functions exist for this optimization problem. A popular optimization
criterion is the logarithm of the determinant (log det() ) of the covariance matrix of the parameter
estimates, known as the d-optimality criterion [14]. This criterion is a measure for the size of the
overall uncertainty region of the parameter estimates. It does not depend on the model parameters
if the joint position, velocity and acceleration data are free of noise. It depends only on the robot
trajectory via the identification matrix [15], and on the variance of the noise on the actuator
torque measurements. This property is important as 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 the Cartesian space. These limitations
avoid collision between the robot and objects in its environment as well as collision between
robot links. This last type of constraint involves straightforward forward kinematics calculations.

Robot excitation and data acquisition

The optimized robot excitation trajectory is programmed in the robot controller. The
robot executes the trajectory repeatedly, while data is being collected at a constant, user specified
sampling frequency. Data collection starts after the transient robot response has died out.

Joint positions are measured using the joint encoders that are mounted on the axle of
the actuators. The measured joint trajectories can deviate significantly from the desired ones,
which is due to the limitations of the robot controller. In other words, the robot executes a
trajectory that does not correspond perfectly to the designed excitation trajectory. Therefore the
16

measured joint positions are used for the parameter estimation instead of the designed excitation
trajectories.

The actuator torque data are obtained through actuator current measurements, such that
no additional sensors are required. The relation between current and torque is assumed to be
linear.

Signal processing

The aim of the signal processing step is to clean up the measured data. This includes
improvement of the signal-to-noise ratio, estimation of the variance of the measurement noise,
and calculation of the joint velocity and acceleration estimates based on the measured joint
positions. In general these operations are not straightforward. Noise filtering often requires adhoc decisions and introduces noise coloration that yields biased parameter estimates. Numerical
differentiation blows up the noise which is inherently present in the measurements. Noisy
joint velocity and acceleration estimates result, which deteriorate the accuracy of the parameter
estimates.

This section illustrates that these problems can be avoided if all measured signals are
periodic.

17

Data averaging and noise variance estimation

Measurement noise is a stochastic disturbance. It causes uncertainty and bias errors on


the estimated parameters. For a given set of data, the bias error can be avoided and uncertainty
can be minimized by using an efficient estimator, for example a maximum likelihood estimator.
This estimator yields unbiased parameter estimates with minimal uncertainty when correct values
for the noise variance are used. The model uncertainty can further be reduced by using longer
data records. Unfortunately, the time required to calculate the parameter estimates often increases
dramatically with the length of the data record. A better solution is to improve the signal-to-noise
ratio of the data.

When the data is periodic, the signal-to-noise ratio can be improved by data averaging,
without using a lowpass noise filter. Especially the quality of the noisy actuator current (torque)
measurements improves by performing this operation.

The sample variance formula estimates the variance of a stochastic signal. This formula
can be used to estimate the noise on measured time-varying signals when they are periodic. For
a signal x consisting of M periods of K samples, the sample variance equals
K

x2

XX
1
=
(xm (k) x(k))2 ,
(M K 1) k=1 m=1

(5)

where index k indicates the k-th sample within a period of the signal, and subscript m indicates
the m-th period. x denotes the average of x, that is,
M
1 X
xm (k).
x(k) =
M m=1

18

(6)

Knowledge of the noise variance is essential for the maximum likelihood parameter
estimation discussed below. The variance of the noise on the averaged signals is obtained by
dividing the sample variance formula by M .

Accurate estimation of the joint velocities and accelerations

Calculation of the identification matrix requires estimates of the joint velocities and
accelerations. Numerical differentiation is only accurate in a small frequency band, far below
the Nyquist frequency. In addition, numerical differentiation amplifies measurement noise. These
problems can be avoided in case of periodic measurement signals. In that case, joint velocities and
accelerations can be calculated through analytical differentiation. This operation is performed in
the frequency domain and combined with frequency domain data windowing. Figure 3 illustrates
this operation for the estimation of the velocity. The example presented in this figure is a
measured position signal that consists of a pure harmonic signal and additive noise.

First, the averaged joint position measurements are transformed to the frequency domain
using the discrete Fourier transform. This transformation does not introduce leakage errors
because of the periodicity of the signals. Next, the relevant frequency lines are selected by
frequency domain windowing, using a rectangular window. This corresponds to frequency domain
data filtering. The other frequency lines contain only noise, and are removed. The selected
frequency lines are then multiplied by the frequency response of a pure single and double
differentiator. That is, the frequency lines are multiplied by j and 2 respectively, with the
frequency in radians per second. The resulting frequency spectra are then transformed back to

19

the time domain using the inverse discrete Fourier transform. This procedure yields noise-free
estimates of the joint velocities and accelerations.

Parameter estimation

The parameter estimation discussed in this section, is based on model equation (2), which
is linear the parameters.

The selection of an appropriate parameter estimation method is a compromise between


accuracy and complexity of implementation. Linear least squares parameter estimation (LLSE)
sits at one side of this spectrum. LLSE is a simple non-iterative method that finds the parameter
estimates in a single step using singular value decomposition. LLSE does not discriminate
between accurate and inaccurate data, and consequently, it often yields biased estimates with a
nonminimal uncertainty.

The maximum likelihood estimation sits at the other side of the spectrum as it provides
consistent estimates with minimal uncertainty. The maximum likelihood estimate of the parameter
vector is the value of that maximizes the likelihood of the measurements. This criterion in
its most general formulation is a nonconvex function of the unknown model parameters while
it depends on the variance of the noise on all measurements and derived variables such as joint
velocities and accelerations. The minimization of this likelihood function is a nonlinear least
squares optimization problem. This optimization problem is usually solved with iterative search
algorithms, for example the Gauss-Newton or the Levenberg-Marquardt algorithms. This search
is often cumbersome because it requires an initial guess of the parameters and because it might
20

converge to a local optimum yielding a suboptimal solution.

When however, the joint position, velocity and acceleration data are free of noise, the
identification matrix in equation (2) is obviously free of noise. In this case the maximum
likelihood estimation simplifies to the Markov estimation, that is, a weighted linear least squares
estimation (WLSE). The complexity of WLSE is comparable to that of LLSE. The only difference
is that WLSE weighs the data with the inverse of the variance of the actuator torque measurement
noise, and therefore discriminates between accurate and inaccurate data. WLSE has the same
favorable properties as the maximum likelihood estimation. This simplification to a WLSE is a
direct consequence of the periodic excitation of the robot dynamics and the analytic calculation
of the velocities and accelerations, as described above.

The weighted least squares estimate of the model parameters is


W LS = (FT 1 F)1 FT 1 ,
= (0.5 F)+ 0.5 ,

(7)
(8)

where ()+ denotes the pseudo-inverse,

(qt1 , qt1 , qt1 )

(qt2 , qt2 , qt2 )

F=
,

..

(qtK , qtK , qtK )

21

(9)

and

t1


t2

= .
..
.


tK

(10)

(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
discrete time instance tk . is the diagonal covariance matrix of the actuator torque data. It
contains the variance estimates of the noise on the measured torque data, obtained using the
sample variance formula (5).

The covariance matrix of the estimated parameter vector W LS is


C = (FT 1 F)1 .

(11)

This covariance matrix, and the related uncertainty bounds on the parameters, are valuable
measures for the model validation.

Model validation

Model validation is an essential step in the identification procedure. The accuracy of


the model is evaluated using one or several criteria. An unsatisfactory validation leads to
a reconsideration of some previous steps of the identification procedure, for instance a new
experiment design or a new, more elaborate dynamic model.

22

The aim of the model validation step is to gain confidence in the obtained robot model
in view of its intended application. The validation must provide an answer to the question
whether the model is sufficiently accurate for the application to be successful. Obviously, the
most appropriate validation test would be to actually use the model in the application, and
to evaluate its success. This can however lead to undesirable and even dangerous situations.
Therefore, validation must be addressed prior to the actual application. This requires tests that
evaluate the properties of the obtained model and its estimated parameters that are important for
the intended application.

This section discusses two measures to validate the accuracy of the obtained model: (1)
the torque prediction accuracy of the model, and (2) the accuracy of the parameter estimates.

Actuator torque prediction accuracy

The actuator torque prediction accuracy of a robot model is important in view of offline
programming, task optimization, as well as advanced robot control. The robot model (2) is
evaluated for some desired motion, described by a set of joint positons q, velocities q,
and
accelerations q, yielding actuator torque predictions . The same motion is executed by the
robot. The actuator torques are measured and compared with the predicted actuator torques.
Good validation practice requires that the validation trajectories are different from the excitation
trajectory, while being representative for the intended application. 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.

23

To validate the model, measured and predicted actuator torques are compared. The
difference between predicted and measured torque, that is, the prediction error, should be small
in comparison to the torque signal and preferably comparable to the noise level on the actuator
torque measurements. A graphical analysis of the prediction error allows us to detect the parts
of the model that are (in)sufficiently accurate. The root mean square (RMS) of the prediction
errors gives a more global impression of the quality of the model.

Parameter accuracy

An alternative method to validate the model is to check the accuracy of the estimated
parameters. Two approaches are possible. The estimated parameters can be compared with an
estimate of their confidence interval, or with estimates available from other sources, for example
CAD data of robot parts. CAD data are however rarely sufficiently accurate.

Based on the parameter covariance matrix (equation (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 first approach but is in general more reliable.

24

Advantages of periodic excitation

Periodic excitation is the key feature of the presented identification approach. Periodic
excitation is strongly integrated in the experiment design, signal processing and parameter
estimation steps. The advantages of periodic excitation in robot identification have briefly been
mentioned in [16], but have never been demonstrated experimentally. The discussion in this
section is based on experimental data, presented further on in this paper.

Analytic differentiation of joint position measurements according to the procedure


described above, yields unbiased and noise-free estimates of the joint velocities and accelerations.
However, this procedure can only be applied for periodic signals. In the case of nonperiodic
signals, the discrete Fourier transform that is used in this differentiation introduces leakage
errors, which are systematic errors. The mentioned frequency domain windowing corresponds
to ideal noise filtering, providing noise-free velocity and acceleration data. Figure 4 compares
the joint accelerations obtained by analytical and numerical differentiation of measured joint
position data. The data is taken from the first case study presented in the following section.
Tustins bilinear differentiation rule is used for the numerical differentiation. The difference is
obvious.

Numerical differentiation techniques can be combined with lowpass filtering to improve


the signal-to-noise ratio. However, these noise filters introduce amplitude and phase distortions.
The selection of the cutoff frequency is crucial. This selection is quite difficult when the
bandwidth of the excitation signal is not clearly specified, as it is the case for nonperiodic
excitation trajectories, for example presented in [12], [13]. The selection of the filter cutoff
25

frequency is then a compromise between eliminating noise and eliminating valuable information.
The phase distortion of the filter can be compensated for if forward and backward filtering are
combined.

The noise on the velocity and acceleration data affects the accuracy of the model
parameter estimation. Noise on the elements of the identification matrix introduces systematic
estimation errors when it is not considered properly in the parameter estimation. Linear least
squares and Markov estimation suffer from this drawback. On the other hand, the maximum
likelihood estimation method presented in [17] can deal with this problem properly. This
parameter estimation approach is however much more complex. It involves an iterative search
algorithm that requires good initial parameter estimates to avoid suboptimal solutions.

The following comparison illustrates how noise on the joint velocity and acceleration
data affects the accuracy of the Markov estimation method. The Markov estimation method is
applied twice with the same set of measurement data. For the first estimation, the joint velocity
and acceleration data are calculated through analytical differentiation. The second estimation
is based on noisy velocity and acceleration data that are obtained through Tustins numerical
differentiation method. No data filtering is applied. Table I compares the accuracy of both models
using the RMS of the actuator torque prediction errors for the validation trajectory presented in
Figure 9. A loss of accuracy up to 75% is observed.

26

E XTENSION TO

ROBOT PAYLOAD IDENTIFICATION

Since industrial robot manipulators evolve toward more lightweight structures, the relative
contribution of their payloads and end effector tools to the actuator torques increases. As a
consequence, knowledge of the inertial properties of these payloads becomes more important
for offline programming, task optimization, and advanced robot control. Identification methods
that can efficiently and effectively estimate these payload parameters are required.

For many years already experimental robot identification techniques have been applied
successfully. The literature on robot identification is very extensive, but in most cases no
distinction is made between the robot links and the robot payload. The payload is considered as
a part of the last link. If the payload changes, the robot dynamics change, and a re-identification
of the complete robot is required. This approach is very time consuming and not efficient since
most elements of the robot model do not depend on the payload. A more dedicated identification
approach is most welcome.

Identification procedure

This section presents a robot payload identification method based on the integrated robot
identification approach presented above. The main goal of a dedicated payload identification is
to avoid a full dynamic identification each time another payload is attached to the end effector
of the robot. Only a minimal set of relevant parameters is re-estimated. More precisely, we are
mainly interested in an update of the payload inertial parameters, while all other robot inertial
parameters are assumed to be known a priori.
27

The derivation of the payload identification method starts from the observation that the
robot links and the payload each contribute separately to the joint torques, that is
= robot links + payload
= (q, q,
q) + L (q, q,
q)L ,

(12)
(13)

where L is the payload identification matrix, and L the 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 equation (13) corresponds to equation (2), and describes the robot dynamics. This term represents the actuator torque contributions that result from the inertial parameters,
possibly gravitation compensation devices, and joint friction. The first two contributions do not
change by adding a payload to the robot. Therefore the corresponding parameters are considered
as a priori information for the robot payload identification. The joint friction changes with the
payload, and consequently the parameters are re-estimated. Altogether, the number of unknown
parameters remains small, which improves the robustness and accuracy of the payload parameter
estimation.

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
28

for the base joints. On the other hand, accurate estimation of the inertial parameters of the payload
requires fast payload motions. These fast payload motions are most easily accomplished using
the robot wrist joints and one additional base joint.

The other steps of the payload identification procedure are almost the same as for the
general robot identification procedure. Hence the payload identification procedure benefits from
the advantages associated with periodic excitation.

Additional accuracy requirements for payload identification

Robot payload identification must fulfill two requirements. For simulation purposes and
controller design, the total robot model, consisting of both the robot parameters and the identified
payload parameters, must be able to accurately predict the actuator torques that correspond to
any desired robot motion. This requirement corresponds to that of a general robot identification
procedure. As an additional requirement the estimated parameters of the robot payload must be
close to the real parameter values as robot payload identification is in most cases performed on
the shop floor. The robot operator will only accept the outcome of the identification when the
estimated parameters are close to his or her expectations.

E XPERIMENTAL RESULTS

ON INDUSTRIAL ROBOT MANIPULATORS

This section discusses experimental results of the robot and payload identification procedures using two industrial robot manipulators. The first case study consists of the identification

29

of the first three links of an industrial KUKA IR 361 robot. The second case study consists of
the identification of the inertial parameters of a calibrated reference payload attached to the end
effector of a KUKA KR 15.

Identification of the first three links of a KUKA IR 361

Description of the experimental setup

The KUKA IR 361, which is shown in Figure 5, is a six-degree-of-freedom industrial


robot manipulator with a payload capacity of 8 kg. Only the first three links are considered here.
The robot is equipped with a gravitation compensation spring that is mounted between the first
and second link. This spring reduces the load due to gravitation on the actuator of the second
link.

A fundamental frequency of 0.1 Hz is selected for the excitation trajectories, which results
in a period of 10 seconds. The trajectories are five-term Fourier series, yielding 11 trajectory
parameters for each joint, and a 0.5 Hz bandwidth. Optimization of the trajectory parameters
is based on the d-optimality criterion. Figure 6 shows the optimized trajectory for the three
robot joints. Figure 7 shows a three-dimensional representation of this optimized trajectory in
the workspace of the robot. The trajectories are implemented on the Orocos [18], [19] control
platform that replaces the standard industrial position controller of this robot. After the transients
have died out, the joint angles and the motor currents of the first three robot joints are measured.
The total measurement time is 160 seconds. This time interval corresponds to 16 periods of the
excitation trajectory. The data sampling rate is 150 Hz.
30

Parameter estimation

The 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 compensation spring for the second link. The rotor inertia
of the actuator of 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 [20].

These model parameters are estimated using the Markov estimator or the WLSE, using
the joint angle and actuator torque measurements. These measurements are averaged over 16
periods. Velocity and acceleration data are derived from the averaged joint angle data through
analytical differentiation. The variance of the noise on the averaged torque measurements is
estimated using the approach presented above. Next, the accuracy of the model is evaluated.

Figure 8 compares the averaged measured torques for the excitation trajectory and the
predicted actuator torques based on the identified model parameters and the available motion
data. The third column shows the prediction error. This error is small, except at velocity reversal
due to the limited accuracy of the friction model at low velocities. Table II compares the RMS
of the prediction errors with the standard deviation of the noise on the averaged actuator torque
measurements.

31

Validation experiment

For the validation, we choose a trajectory that is representative for pick and place as well
as spot welding operations. Figure 9 shows the selected validation trajectory. This trajectory
goes through 20 points randomly chosen in the workspace of the robot. The robot moves with
maximum acceleration and deceleration between these points, and comes to a full stop at each
point. This results in a trapezoidal velocity profile between two successive points, which is
commonly used in industry.

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
obtained model is capable of accurately predicting the actuator torque data. The peaks in the
prediction error occur at low velocities. Again, this indicates that the assumed friction model,
which includes viscous and Coulomb friction only, is too simple. Extending the robot model with
more advanced friction models, coupling between the actuator torques, models of the mechanical
losses in the actuators and of the efficiency of the transmissions further improve the prediction
accuracy. These extensions are discussed in [21]. Most of these model extensions are however
nonlinear in the parameters, and therefore complicate the parameter estimation considerably.

The first row of table I shows the RMS of the actuator torque prediction errors for this
validation experiment.

32

Payload identification

Description of the experimental setup

The test setup consists of a six-degree-of-freedom KUKA KR 15 industrial manipulator


(Figure 11) equipped with a payload. The KUKA KR15 is a modern industrial robot with a
PC-based controller and a payload capacity of 15 kg. To validate the accuracy of the payload
identification method, a calibrated reference payload is used. This payload is shown in Figure 12.
A detailed CAD model provides estimates of all ten inertial parameters of this payload. The
estimates are sufficiently accurate in this case, because of the simplicity of the payload. Simple
experiments, like for example weighing, have confirmed the accuracy of these estimates. They
are used as the reference values for the robot payload identification experiment presented here,
and are shown in the second column of table III.

The robot excitation is limited to the last four joints of the manipulator. This limitation
is imposed by the payload identification plugin of the robot controller. A fundamental frequency
of 0.033 Hz is selected for the excitation trajectories, yielding a period of 30 seconds. The
trajectories are 2-term Fourier series. They consist of the fundamental frequency and its 20th or
25th harmonic. This constraint is again imposed by the software plugin. Data are sampled with
a sampling period of 12 ms, which corresponds to 2499 samples per excitation period.

33

Parameter estimation and experimental results

The payload identification model contains 18 identifiable parameters: the ten inertial
parameters of the payload and eight friction parameters. The friction parameters consist of one
Coulomb and one viscous friction parameter for each of the last four joints. From experimental
experience we learned that friction changes significantly with the payload of the robot. A reidentification of the friction parameters is therefore necessary to obtain accurate and consistent
parameter estimates.

Ten different 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 III and compared with the
reference values. The reference values lie within the 2 confidence interval around the average
parameter estimates.

Figure 13 evaluates the model accuracy based on its actuator torque prediction accuracy
for a validation trajectory. This validation trajectory is different from the excitation trajectories.
The total robot model consists of the available rigid body model of the robot without payload,
and is extended with the average payload parameter estimates as well as the joint Coulomb and
viscous friction parameter estimates. This robot model is able to accurately predict the measured
actuator torques. The largest prediction errors occur at the velocity reversals, as expected.

The accuracy of the parameter estimates and of the actuator torque prediction satisfy the
34

requirements imposed by industry. The mass of a payload can be estimated with an accuracy of
a few hundred grams, and its center of mass with an accuracy of ten millimeter.

An experimental analysis presented in [21] reveals that it is much easier to pass the
actuator torque prediction validation tests than to obtain sufficiently accurate estimates of the
payload parameters. The industrial requirements with respect to actuator prediction accuracy are
less strict than those for the payload parameter accuracy. If the model contains a sufficient number
of parameters, acceptable prediction accuracy can be obtained even with payload parameter
estimates that are insufficiently accurate.

C ONCLUSIONS

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 simplifies the identification procedure considerably and yields
accurate models. Experimental results on an industrial robot manipulator show that the obtained
dynamic robot model is able to 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. In addition, the approach is able to provide
estimates of the payload inertial parameters that satisfy the strict accuracy requirements imposed
by industry. This approach is therefore recognized by industry as an effective means of robot
35

payload identification.

ACKNOWLEDGMENTS

The work is sponsored by a specialization grant of the Institute for the Promotion of
Innovation through Science and Technology in Flanders (IWT-Vlaanderen).

R EFERENCES

[1] M. Olsen, J. Swevers, and W. Verdonck, Maximum likelihood identification of a dynamic


robot model: implementation issues, International Journal on Robotics Research, vol. 21,
no. 2, pp. 8996, February 2002.
[2] N. N., RRS robot controller simulation (RCS) interface, Berlin, 2001. [Online].
Available: http://www.realistic-robot-simulation.org
[3] J. Y. S. Luh, M. W. Walker, and R. P. Paul, On-line computational scheme for mechanical
manipulators, Journal of Dynamic Systems, Measurement, and Control, vol. 102, no. 2,
pp. 6976, June 1980.
[4] J.-J. E. Slotine, The robust control of robot manipulators, The International Journal of
Robotics Research, vol. 4, no. 2, pp. 4964, 1985.
[5] P. Khosla and T. Kanade, Experimental evaluation of nonlinear feedback and feedforward
control schemes for manipulators, The International Journal of Robotics Research, vol. 7,
no. 1, pp. 18 28, February 1988.
[6] C. Canudas de Wit, B. Siciliano, and G. Bastin, Theory of robot control. Springer, 1996.
36

[7] F. Caccavale and P. Chiacchio, Identification of dynamic parameters and feedforward


control for a conventional industrial manipulator, Control Engineering Practice, vol. 2,
no. 6, pp. 10391050, 1994.
[8] C. Atkeson, C. An, and J. Hollerbach, Estimation of inertial parameters of manipulator
loads and links, The International Journal of Robotics Research, 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, pp. 165184, 1996.
[10] J. Swevers, F. Al-Bender, C. Ganseman, and T. Prajogo, An integrated friction model
structure with improved presliding behaviour for accurate friction compensation, IEEE
Transactions on Automatic Control, vol. 45, no. 4, pp. 675686, 2000.
[11] F. Al-Bender, V. Lampaert, and J. Swevers, The generalized Maxwell-slip model: a novel
model for friction simulation and compensation, IEEE Transactions on Automatic Control,
vol. 50, no. 11, pp. 1883 1887, November 2005.
[12] B. Armstrong, On finding exciting trajectories for identification experiments involving
systems with nonlinear dynamics, The International Journal of Robotics Research, vol. 8,
no. 6, pp. 2848, 1989.
[13] M. Gautier and W. Khalil, Exciting trajectories for the identification of base inertial
parameters of robots, The International Journal of Robotics Research, vol. 11, no. 4,
pp. 362375, 1992.
[14] L. Ljung, System identification: theory for the user.

Englewood Cliffs, New Jersey:

Prentice-Hall Inc., 1987.


[15] M. Gautier, Identification of robots dynamics, in Proceedings of the IFAC Symposium on

37

Theory of Robots, Vienna, Austria, 1986, pp. 351356.


[16] J. Swevers, C. Ganseman, D. Bilgin, J. De Schutter, and H. Van Brussel, Optimal robot
excitation and identification, IEEE Transactions on Robotics and Automation, vol. 13,
no. 5, pp. 730740, 1997.
[17] M. M. Olsen and H. G. Petersen, A new method for estimating parameters of a dynamic
robot model, IEEE Transactions on Robotics and Automation, vol. 17, no. 1, pp. 95100,
2001.
[18] H. Bruyninckx, Open RObot COntrol Software, http://www.orocos.org/, 2001.
[19] P. Soetens and H. Bruyninckx, Realtime hybrid task-based control for robots and machine
tools, in Proceedings of the IEEE International Conference on Robotics and Automation,
Barcelona, Spain, 2005, pp. 260265.
[20] 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.
[21] J. Swevers, W. Verdonck, B. Naumer, S. Pieters, and E. Biber, An experimental robot
load identification method for industrial application, International Journal on Robotics
Research, vol. 21, no. 8, pp. 701712, 2002.

38

Figure 1.

Screen shot of the 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. Some environments include
cycle time optimization, however based on geometric and kinematic information only. (Courtesy
of KUKA)

39

A priori
knowledge

Kinematic &
geometric
information

Model
accuracy
specification

Robot identification procedure

Model
generation
Experiment
design
Data
acquisition
Signal
processing

Choose parameter
estimation algorithm

Parameter
estimation
Validate
model

Not satisfied

Result

Satisfied

Figure 2.

Accurate model
parameters

Accurate torque
prediction

Schematic representation of a standard experimental robot identification procedure.

The kinematic and geometric information of the considered robot manipulator and model
accuracy specification is the input to the identification procedure. This information is available
prior to the identification, and determines the choices that have to be made in most of the steps
of the procedure. The identification procedure is a sequence of several steps, of which the last
one is the model validation step. This step evaluates the accuracy of the model according to
one or several criteria depending on the application of the model. If the obtained model does
not pass the validation tests, one or several steps of the procedure are repeated and choices
are reconsidered. Periodic robot excitation is the key feature of the developed identification
40
procedure, and has significant impact on the steps indicated in red.

Measured joint position


sequence (one period)

q(t)

DFT
Frequency spectrum of
measured joint position

Q(w)
w
Multiply by
rectangular window

Filtered frequency
spectrum of measured
joint position

Qf(w)
w
x jw

Spectrum of joint
velocity

jwQf(w)
w
IDFT

Noise-free
joint velocity

Figure 3.

q(t)

Analytical differentiation of a measured periodic joint angle signal. The left column

explains the different steps of the calculations. The right column illustrates these steps for one
period of sinusoidal signal perturbed with additive normally distributed noise. First, this signal
is transformed to the frequency domain using the discrete Fourier transform (DFT). Then the
spectrum is multiplied by a rectangular window that selects the frequency lines that contain
signal information. For this example, this corresponds to selecting one frequency line. All other
frequency lines are set to zero. The resulting spectrum is then multiplied by j, that is the
frequency domain representation of a differentiator. A backtransformation into time domain,
using the inverse DFT, yields a noise-free estimate of first time derivative of the original signal,
that is the velocity. Remark that the spectra consist of complex numbers. For reasons of simplicity,
only the amplitude spectra are shown.
41

Analytical calculation of acceleration

Acceleration (rad/s2)

10

10

Numerical calculation of acceleration

Acceleration (rad/s2)

Figure 4.

5
Time (s)

Comparison of the accuracy of analytical and numerical differentiation. The top

and bottom figures show the accelerations that are obtained through respectively analytical
and numerical differentiation of joint position (joint angle) data measured on a KUKA IR 361
industrial robot. The blue, green, and red lines indicate joints 1, 2, and 3 respectively. Analytical
differentiation is possible because of the periodicity of the signals, and yields noise-free estimates
of the acceleration. The numerical differentiation using Tustins bilinear differentiation rule yields
noisy estimates of the acceleration.

42

differentiation method

joint 1

joint 2

joint 3

analytical

8.9 Nm

11.8 Nm 4.4 Nm

numerical

10.6 Nm

17.5 Nm 7.8 Nm

TABLE I
I NFLUENCE OF THE APPLIED DIFFERENTIATION METHOD ON THE ROBOT MODEL ACCURACY.
T HE ACCURACY OF THE ROBOT MODEL DEPENDS ON THE DIFFERENTIATION

METHOD THAT

IS USED IN THE DATA PROCESSING STEP PROCEEDING THE PARAMETER ESTIMATION .


TABLE COMPARES THE
MODELS .

RMS

T HE

OF THE ACTUATOR TORQUE PREDICTION ERRORS FOR TWO

B OTH MODELS ARE OBTAINED WITH THE M ARKOV METHOD USING THE SAME SET

OF MEASUREMENT DATA .

F OR THE CALCULATION OF THE FIRST MODEL , THE JOINT

VELOCITIES AND ACCELERATIONS ARE DERIVED WITH THE ANALYTICAL DIFFERENTIATION


METHOD .

F OR THE CALCULATION OF SECOND MODEL , T USTIN S NUMERICAL

DIFFERENTIATION METHOD IS USED .

T HE FIRST MODEL IS SIGNIFICANTLY MORE ACCURATE

THAN THE SECOND MODEL .

43

Figure 5.

A KUKA IR 361 industrial robot in our robotics laboratory. This six-degree-of-

freedom robot is shown here with a reference payload like the one discussed in the next case
study. This payload is however not considered here. This robot is equipped with brushless dcmotors. It was manufactured in 1987, and is present in our lab since 1990. The standard industrial
position controller of this robot is replaced by the Orocos software [18], [19]. This software
allows us to apply periodic excitation trajectories, and to synchronize encoder and actuator
current measurements.

44

1.5

Joint position (rad)

0.5

0.5

1
joint 1
joint 2
joint 3

1.5

10

Time (s)

Figure 6. Optimized robot excitation trajectory: joint 1 (full blue line), joint 2 (dashed green line)
and joint 3 (dash-dotted red line). One period of the optimized joint trajectories is shown. These
trajectories consist of five-term Fourier series with a base frequency of 0.1 Hz. This frequency
corresponds to a period of 10 second. The trajectory parameters are optimized according to the
d-optimality criterion, taking into account workspace limitations, constraints on joint velocities
and accelerations. This optimization criterion is a measure of the size of the overall uncertainty
region of the parameter estimates.

45

Figure 7. Three-dimensional visualization of the optimized excitation trajectory. The line shows
the trajectory of the robot tool center point of the KUKA IR 361 robot. Only the first three links
of the robot are considered in this experiment. The robot model identification is based on joint
angle and actuator torque data measured during 16 successive repetitions of this trajectory.

46

Torque joint 1 (Nm)

Measured torque
100

100

100

100

100

Torque joint 2 (Nm)

10

200

10

200

200

200

200

100

100

100

100

100

100

200

200
0

Torque joint 3 (Nm)

Torque prediction error

100

200

10

10

100

100

50

50

50

50

50

50

Time (s)

10

100

10

10

10

200
0

100

100

Figure 8.

Predicted torque

Time (s)

10

100

Time (s)

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. The third column shows that the
prediction errors are small. A small prediction error is not unusual because the model parameters
are estimated by minimizing these errors for this trajectory. The remaining peeks occur at velocity
reversal. This indicates that the friction model is too simple to model the complex dynamic
friction behavior at low velocities.

47

RMS value

joint 1

joint 2

joint 3

prediction error 6.1 Nm

6.3 Nm

2.9 Nm

4.6 Nm

1.5 Nm

noise level

3.3 Nm
TABLE II

C OMPARISON OF ACTUATOR TORQUE PREDICTION ERRORS WITH


LEVEL .

T HE ROOT MEAN

THE MEASUREMENT NOISE

SQUARE OF ACTUATOR TORQUE PREDICTION ERRORS ARE

COMPARABLE TO THE STANDARD DEVIATION OF THE NOISE ON THE AVERAGED ACTUATOR


TORQUE MEASUREMENTS .

48

2.5
2
1.5

Joint position (rad)

1
0.5
0
0.5
1
1.5
2
2.5

Figure 9.

10

20

30
Time (s)

40

50

Robot model validation trajectory: joint angles for joint 1 (full blue line), joint 2

(dashed green line) and joint 3 (dash-dotted red line). The trajectory goes through 20 poses
randomly chosen in the workspace of the robot. The robot moves with maximum acceleration
and deceleration between these points, and comes to a full stop at each point. This validation
trajectory is representative for pick and place as well as spot welding operations.

49

Torque joint 1 (Nm)

Measured torque

Torque joint 2 (Nm)

Torque prediction error

100

100

100

100

100

100

Torque joint 3 (Nm)

Predicted torque

20

40

20

40

200

200

200

100

100

100

100

100

100

200

20

40

200

20

40

200

50

50

50

50

50

50

20
40
Time (s)

20
40
Time (s)

20

40

20

40

20
40
Time (s)

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. The third column shows that the prediction
errors are small. This shows that the estimated robot model is capable of accurately predicting the
actuator torques, and therefore well suited for advanced model based robot control and dynamic
task optimization.

50

Figure 11.

The KUKA KR 15 industrial robot in our robotics laboratory with a reference

payload. This robot is a six-degree-of-freedom modern industrial robot with a payload capacity
of 15 kg. It has a PC-based controller and is equipped with ac-motors. A software plugin allows
us to perform robot payload identification experiments that use periodic trajectories for the last
four robot axes.

51

Figure 12. The reference payload that has been developed to validate the payload identification
procedure. This payload is reconfigurable. For example, its mass and principal inertias can be
changed between 1.4 and 15 kg and from nearly zero to over 3 kgm2 , respectively. A detailed
CAD model of this payload is available. This model provides accurate estimates of the inertial
parameters. These parameter estimates have been used as reference values to validate the robot
payload identification procedure. The second column of table III shows these reference values
for the presented configuration of the payload.

52

Parameter

Reference

Mean

Standard

value

value

deviation

m [kg]

9.579

9.618

0.174

cx [m]

0.024

0.0246

0.0010

cy [m]

0.090

0.0930

0.0016

cz [m]

0.202

0.2065

0.0064

Ixx [kg m2 ]

0.612

0.677

0.077

Iyy [kg m2 ]

0.063

0.121

0.077

Izz [kg m2 ]

0.637

0.621

0.020

Ixy [kg m2 ]

0.158

0.164

0.011

Ixz [kg m2 ]

0.002

0.004

0.010

Iyz [kg m2 ]

0.008

0.010

0.016

TABLE III
VALIDATION OF THE ACCURACY
REFERENCE PAYLOAD SHOWN IN

OF THE INERTIAL PARAMETER ESTIMATES FOR THE

F IGURE 12. T HE REFERENCE VALUES OF THE INERTIAL

PARAMETERS RESULT FROM A DETAILED

CAD MODEL OF THE PAYLOAD . T HE MEAN

PARAMETER VALUES ARE OBTAINED FROM TEN SETS OF PARAMETER ESTIMATES .


FOUR PRESENTS THE CORRESPONDING STANDARD DEVIATION .
LIE WITHIN THE TWO

C OLUMN

A LL REFERENCE VALUES

CONFIDENCE INTERVAL OF THE MEAN ESTIMATED VALUES .

53

Measured and predicted actuator torque

Actuator torque prediction error

Joint 3 (Nm)

200
0

200

200

400

400

Joint 4 (Nm)

600

Joint 5 (Nm)

10

15

20

25

600

30

200

100

100

100

100
0

10

15

20

25

200

30

200

200

100

100

100

100

200

Joint 6 (Nm)

200

200

10

15

20

25

200

30

100

100

50

50

50

50

100

Figure 13.

200

10

15
Time (s)

20

25

100

30

10

15

20

25

30

10

15

20

25

30

10

15

20

25

30

10

15
Time (s)

20

25

30

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 on top of the measured torques. The second column shows that the prediction
errors are small. The remaining peaks in the prediction errors occur at the velocity reversals,
and indicate that the considered Coulomb and viscous friction models are too simple.

54

Você também pode gostar