International Journal of Solids and Structures 38 (2001) 1889±1902

www.elsevier.com/locate/ijsolstr

Friction and rigid body identi®cation of robot dynamics

M. Grotjahn *,1, M. Daemi, B. Heimann *,2
Institute of Mechanics, University of Hanover, Appelstrasse 11, D-30167 Hanover, Germany
Received 3 September 1999; in revised form 29 December 1999

Abstract
In this paper, an identi®cation method for industrial robots is described that does not require the a priori identi-
®cation of the friction model. First, the necessity for such a method is motivated by an overview on conventional
friction modelling and rigid body identi®cation. It is shown that the time variance of typical friction characteristics lead
to systematic identi®cation errors. They are avoided by the presented method, which is based on separating the base
parameters into three di€erent groups. Each group is identi®ed by simple measurements and a weighted least-squares
method. Measurements are carried out with simple motions in the neighbourhood of a number of especially selected
joint con®gurations. Further advantages of this method are its easy implementation for standard robot controls and the
possibility to ®nd modelling errors. The experimental implementation of this method to a typical industrial robot with
six rotational joints is carried out and yields remarkable results. Ó 2001 Elsevier Science Ltd. All rights reserved.

Keywords: Industrial robot; Modelling; Identi®cation; Dynamics; Friction

1. Introduction

Common identi®cation methods for industrial robots are based on the description of the equation of
motion in terms of a linear parameter vector of minimal order (base parameters) (Gautier and Khalil,
1988a; Khosla, 1988) and the use of least-squares (LS) error minimisation criteria (Armstrong, 1988, 1989;
Daemi and Heimann, 1996; Daemi and Grotjahn, 1996; Gautier and Khalil, 1988b; Gautier et al., 1995;
Lawson and Hanson, 1974; Pfei€er and H olzl, 1995; Presse and Gautier, 1993; Prufer et al., 1994). The
robot is typically moved along a trajectory, which is generated using an optimisation scheme to attain
maximum excitation of all base parameters. Joint motion and joint torque are measured and the torque due
to friction is compensated using a measured friction model. Then, a batch LS algorithm is used to calculate
the base parameter vector.
In practice, it can be observed that, even with good excitation of all parameters, the results still show
some systematic errors (Daemi and Heimann, 1996; Daemi and Grotjahn, 1996). One reason for these

*
Corresponding authors.
E-mail addresses: grotjahn@ifm.uni-hannover.de (M. Grotjahn), heimann@ifm.uni-hannover.de (B. Heimann).
1
Fax: +49-511-762-7010179.
2
Tel.: +49-511-762-4161; fax: +49-511-762-4164.

0020-7683/01/$ - see front matter Ó 2001 Elsevier Science Ltd. All rights reserved.
PII: S 0 0 2 0 - 7 6 8 3 ( 0 0 ) 0 0 1 4 1 - 4
1890 M. Grotjahn et al. / International Journal of Solids and Structures 38 (2001) 1889±1902

errors is the variation of friction characteristics during the measurements. Further diculties arise from the
restrictions of standard robot controls, since optimised trajectories usually require additional hardware for
generating arbitrary joint motions.
In this paper, a two-step approach for the identi®cation of the base parameter vector is described. It is an
extension of the algorithm introduced by Seeger (1992) which requires only trajectories with trapezoidal
velocity characteristic and no a priori compensation of the friction model. The idea is to measure directly
the moments of inertia and the gravitational torques for a number of di€erent joint con®gurations. The
resulting values are used to calculate the base parameter vector by a weighted LS method. The only as-
sumption that has to be veri®ed in advance is that the friction characteristics are symmetric.
The modelling and the time variance of friction are discussed in Sections 2 and 3. In Section 4, the rigid
body modelling is presented. Section 5 gives an overview on conventional identi®cation methods. The two-
step identi®cation scheme is explained in Section 6. And ®nally, in Sections 7 and 8, the experimental set-up
and results are depicted. The results hold for a wide range of industrial robots, since the tested manipulator
Siemens manutec-r15 is a typical industrial system with di€erent types of gears.

2. Modelling and identi®cation of gear friction

Normally, very sti€ gears with small backlash are used in industrial robotics. Therefore, gear elasticity
and backlash have small in¯uence on the controlled systemÕs dynamics. This in¯uence can be avoided by a
proper choice of identi®cation measurements.
Besides rigid body dynamics, only the losses in gears and bearings are taken into account by friction
models. For their identi®cation, friction has to be separated from other dynamic e€ects. By using trajec-
tories, where only one axis is moved, and selecting parts of the measurement with constant velocity, the
in¯uences of acceleration, centrifugal and Coriolis forces are eliminated. If gravitation has an in¯uence on
the torque of the regarded axis, it has to be compensated by a model identi®ed in advance. Thus, the
following measurements can be assumed to solely re¯ect the in¯uence of friction in gears and bearings.
The identi®cation of gravitation model can be done without the knowledge about friction behaviour by the
method proposed in Section 6.
Friction losses of a single robotic joint i are usually modelled as a torque Qi;f1 which is a function of its
own rotational joint speed q_ i . This non-linear function is mostly described by the sum of terms for viscous
damping and dry friction (Armstrong-Helouvry, 1991; Armstrong-Helouvry et al., 1994; Canudas de Wit
et al., 1991; Pfei€er and Holzl, 1995; Seeger, 1992):
Qi;f1 ˆ ai;1 q_ i ‡ ai;2 sign…q_ i †: …1†
This simple model needs to be re®ned for more precise modelling (Daemi and Heimann, 1996; Pr ufer
and Wahl, 1994). Fig. 1a shows measured friction torques of some axes of a robot over their full speed
range, normalised to their maximum speed and maximum torque. It can be seen that all axes show sig-
ni®cant degressive characteristics, not covered by the simple model given in Eq. (1). A better description of
the measured friction characteristics can be found by using one of the following equations:
1=3
Qi;f2 ˆ ai;1 q_ i ‡ ai;2 sign…q_ i † ‡ ai;3 q_ i ; …2†

Qi;f3 ˆ ai;1 q_ i ‡ ai;2 sign…q_ i † ‡ ai;3 atan…ai;4 q_ i †: …3†

For a given measured friction characteristic, an analytical description with a minimal square model error
can easily be calculated for Eqs. (1) and (2), since they have only linear dependencies with respect to their
parameters. Whereas for model (3), a non-linear optimisation procedure is applied. As depicted in Fig. 1b,
both models lead to much better results than the classic equation (1). The mean quadratic error between the
 M. Grotjahn et al. / International Journal of Solids and Structures 38 (2001) 1889±1902 1891

Fig. 1. (a) Normalised friction characteristics for di€erent axes of the manutec-r15. (b) Adaption of friction models to measured friction
characteristics of axis 1.

measured and modelled torque is used to decide whether model (2) or (3) gives a better description of the
measured characteristics. It has to be kept in mind that modelled parameters aj found in this way no longer
represent mechanical models (such as dry friction or viscous damping) but merely force the sum of all
e€ects of the multistage gears into a mathematical description.

3. Time variance of friction characteristics

The major problem in friction modelling is the time variance of friction (Daemi and Heimann, 1996,
1998; Pr ufer and Wahl, 1994). Commonly, relatively large time constants are assumed that arise from
temperature variation in gears and bearings. Thus, the typical approach for the friction modelling in ro-
botics is to use some Ôwarm-upÕ time, where the robot is supposed to reach stationary conditions (Arm-
strong, 1991; Daemi and Heimann, 1996; Seeger, 1992). The measurements in Fig. 2a shows that this does
not generally hold for geared robots.
The ®gure shows friction with respect to time for a multistage robot gear which is continuously moved
back and forth. The torque is measured every 12 s in periods, with constant velocity. The motion is in-
terrupted at di€erent times for short periods (5, 1 and 2 min). It shows that after these short interruptions,
the friction becomes signi®cantly larger. This e€ect implies that not only temperature but also the distri-
bution of lubricants determines friction conditions. Therefore, it is very dicult to ensure equal conditions
for the friction characteristics at two separate measurements. Actually, even during a precise friction
measurement (which takes a few minutes) operating conditions might change.

Fig. 2. Time variance of friction characteristics: (a) variations during steadily back and forth motion with short brakes, and (b) change
of friction characteristic of joint 2 during continuous operation.
1892 M. Grotjahn et al. / International Journal of Solids and Structures 38 (2001) 1889±1902

The changing of measured friction characteristics is depicted in Fig. 2b. Even with changing friction
characteristic, the graphs stay almost symmetric. This behaviour was observed for a number of di€erent
types of robot gears, although di€erent experiences are described in the literature (Armstrong, 1988; Pfei€er
and H olzl, 1995). But since a symmetric friction characteristic is essential for the proposed identi®cation
method, it should be veri®ed if necessary.

4. Parameter-linear modelling of rigid body dynamics

For identi®cation, a representation of robot dynamics has to be determined which is linear with respect
to the unknown inertial and gravitational parameters. In this section, the basic idea for this determination is
described. It is derived using the modi®ed Denavit±Hartenberg (MDH) notation for robot kinematics
(Khalil and Klein®nger, 1986). The resulting base parameter vector is grouped according to the require-
ments of the identi®cation scheme shown in Section 6.

4.1. Dynamic parameters

T
The dynamic parameters of each link i consist of its mass mi , its ®rst moment si :ˆ ‰ sxi syi szi Š ˆ mi riCi
i
(rCi is the vector from the linkÕs coordinate frame to its centre of mass) and its inertia tensor about the
…i† …i†
corresponding coordinate frame I i . Collecting the independent scalar components of si and I i results in
the parameter vector,
plink;i ˆ ‰ Ixxi Ixyi Ixzi Iyyi Iyzi Izzi sxi syi szi m i ŠT ; …4†
which leads to the over-all parameter vector,
 T
p ˆ plink;1 plink;2 . . . plink;n : …5†
In practical robotic applications, a number of parameters are zero because of the orientation and po-
sition of the coordinate frames as well as symmetric shapes of the bodies. Furthermore, a large number of
parameters can be replaced by linear combinations (Section 4.3).

4.2. Equations of motions in parameter-linear form

The equations of motion of tree-structured robots can be derived in parameter-linear form by the
Newton±Euler approach (Khosla, 1988) or the Lagrangian equations of second kind (Gautier and Khalil,
1988a):
_ ‡ g…q† () Q ˆ A…q; q;
Q ˆ M…q†q ‡ c…q; q† _ q†p: …6†
The left-hand side of Eq. (6) shows the conventional form of the dynamic equations, whereas on the
right side, the parameter-linear form of the equation of motion is given. The mass matrix M…q† is a function
_ describes centrifugal and Coriolis e€ects and g…q† resembles the in¯uence of
of the joint angles q; c…q; q†
_ q† contains the known kinematic quantities
gravitation. Q is the vector of joint torques. The matrix A…q; q;
of the robot, which is multiplied by the corresponding parameter vector p.
The moments of inertia of the motors Imi are usually given by the manufacturer so that they are not
included in the identi®cation model. Nevertheless, they have to be compensated by means of an additional
model
_
Qm ˆ H…q†q ‡ cm …q; q† …7†
as described in Daemi and Heimann (1998).
 M. Grotjahn et al. / International Journal of Solids and Structures 38 (2001) 1889±1902 1893

4.3. Determination of base parameters

For application of LS-estimation techniques (Lawson and Hanson, 1974), the parameter vector p must
have minimal dimension (referred to as base parameter vector). This means that all parameters without
in¯uence must be removed at ®rst. For example, a revolute base axis can be regarded. If the base is ®xed,
…1†
obviously only one element (Izz1 ) in I 1 has an in¯uence on the dynamics. Second, the parameters must be
regrouped in order to eliminate all linear dependencies from A…q; q;_ q† so that it becomes a full rank matrix.
The determination of the base parameter vector is demanding, especially for robots with more than three
degrees of freedom (d.o.f.). One possibility is the use of numerical methods, like singular-value-decompo-
sition (Gautier, 1990; Pfei€er and H olzl, 1995). The main disadvantage of this approach is that the de-
termination of the new parameter combinations is dicult and very time consuming.
Gautier and Khalil presented an analytical method in 1988. Their algorithm is based on a recursive
Lagrange algorithm. General rules were formulated for calculating the energy contributions of the pa-
rameters of link i with respect to those of link i ÿ 1. These rules were further exploited to ®nd a recursive
algorithm to calculate the base parameters.

4.4. Grouping of the base parameters and the dynamic equations

The two-step identi®cation scheme presented in Section 6 is based on partitioning the base parameters
into three vectors:

1. The gravitational parameters pg occur in g…q†.

2. The diagonal parameters pMd occur on the diagonal of M…q† but not in g…q†.
3. The o€-diagonal parameters pMod occur only in the o€-diagonal elements of M…q†.

The grouping property holds for any open kinematic chain, i.e. for typical serial industrial robots.
Furthermore, typically the number of elements of pMod is small due the symmetric structure of the links, e.g.
for the manutec-r15, pMod has only one element. The grouping leads to a separation of matrix A…q; q; _ q† in
Eq. (6) into di€erent components:

Q ˆ AMd …q; q†pMd ‡ AMod …q;  q†pMod ‡ AMg …q; q†pg ‡ Ac …q; q†p _ ‡ Ag …q†pg : …8†
|‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚{z‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚} |‚‚‚‚‚{z‚‚‚‚‚} |‚‚‚‚{z‚‚‚‚}
M…q†
q _
c…q;q† g…q†

_ Nevertheless, they can be separately

Of course, the parameter vectors are still coupled by c…q; q†.
identi®ed since this coupling is eliminated by a proper choice of identi®cation measurements (Section
6.3).

5. Conventional identi®cation scheme

There exists a vast amount of literature on the identi®cation of the rigid body model (Armstrong, 1988,
1989; Daemi and Heimann, 1996; Daemi and Grotjahn, 1996; Gautier and Khalil, 1988b; Gautier et al.,
1995; Pfei€er and H olzl, 1995; Khosla, 1988; Khosla and Kanade, 1987; Presse and Gautier, 1993; Pr ufer
et al., 1994). However, most of the methods are variations of the same identi®cation scheme. The robot is
moved along a trajectory, where joint motion and torque are measured. Finally, the parameters are esti-
mated by the use of the LS technique.
1894 M. Grotjahn et al. / International Journal of Solids and Structures 38 (2001) 1889±1902

5.1. Least-squares estimation

The LS identi®cation is based on the representation of robot dynamics shown on the right-hand side of
Eq. (6). For a certain point in time tk , measurements of m 6 n di€erent axes c1 ; . . . ; cm are combined as
 T h iT
Qtk ˆ Qc1 ;tk . . . Qcm ;tk and Wtk ˆ ATc1 ;tk . . . ATcm ;tk …9†

with Qcj ;tk , the torque of joint cj and Acj ;tk , the corresponding row of A…q…tk †; q_ …tk †; q…tk ††. Further combi-
nation of measurements at r (r
m  n) di€erent time steps leads to the over-determined vector equation,
Q ˆ Wp ‡ e; …10†
 T T
with the measurement vector Q ˆ Qt1 . . . Qtr , the observation matrix W ˆ ‰WTt1
and the unknown . . . WTtr Š
error e which has to be introduced due to measurement noise. The general solution for Eq. (10) can be
found by a pseudo-inverse of the observation matrix:

^p ˆ … W e T Q:
e T W†ÿ1 W …11†
This shows why p has to have minimal dimension. Otherwise, W would not have full rank and the
e is W
pseudo-inverse would not exist. The most common de®nition of W e ˆ W which leads to a minimum of
the error vector e in the LS sense:
ÿ
ÿ
ÿ1
^pLS ˆ min eT e ) ^pLS ˆ WT W WT Q: …12†
p

Relation (12) is the most general case that does not take into account any a priori information of the
system or disturbances. In the literature, a lot of re®nements can be found. Seeger (1992) proposed
weighting with maximal joint torque to account for the di€erent torque ranges of the drives. Gautier and
Khalil (1992) suggested another weighting method using expected values of the parameters. Other well-
known re®nements are the instrumental variable method (Seeger, 1992) or the total least-squares estimation
(Gautier et al., 1994).

5.2. Trajectory optimisation

A sucient excitation of all parameters is necessary to ensure their identi®ability. If one or more pa-
rameters are not excited, W loses its full rank and Eq. (11) cannot be solved. However, already a poor
excitation leads to a large in¯uence of disturbances on the estimation result. An upper bound for the
relative error of the estimated parameter vector is given by (Armstrong, 1988; Lawson and Hanson, 1974)


p ÿ ^p ke k rmax …W†
6 cond…W† with cond…W† ˆ …13†
kpk kQ k rmin …W†
and rmax …W† and rmin …W† is the largest and smallest singular value of W, respectively. By minimising the
condition of the observation matrix, the standard criterion for trajectory optimisation is de®ned (Arm-
strong 1988, 1989; Gautier and Khalil, 1988b; Pfei€er and H olzl, 1995). Other optimisation criteria, like
maximisation of the smallest singular value, can be found in Armstrong (1988, 1989), Presse and Gautier
(1993), Shaefers et al. (1993) and Daemi and Grotjahn (1996).
One disadvantage of the optimisation is the computational burden (Armstrong, 1988, 1989; Daemi and
Grotjahn, 1996; Gautier and Khalil, 1992). The main problem, however, is that the robot control must
provide the possibility of driving arbitrary trajectories. With standard industrial controls, however, only
very simple trajectories (line, circles, ptp) can be generated.
 M. Grotjahn et al. / International Journal of Solids and Structures 38 (2001) 1889±1902 1895

5.3. Friction compensation

The most common way of compensating the in¯uence of friction on rigid body identi®cation is to
perform Ôwarm-upÕ motions to reach stationary temperature and lubrication conditions. Then, the rigid
body model is identi®ed together with the friction model (Gautier and Khalil, 1988b; Pr ufer et al., 1994;
Pfei€er and H olzl, 1995). The disadvantage of this approach is the large number of parameters which
results from the more complex friction models (2) and (3). For example, for a six d.o.f. robot, it would lead
to 18 additional parameters which would make the identi®cation much more complex or even impossible.
The other possibility is to compensate friction by a model identi®ed in advance (Daemi and Heimann,
1996). As shown in Section 3, however, the operating conditions change so fast that an exact friction
compensation is impossible. This leads to systematic errors in rigid body identi®cation.

6. Two-step identi®cation

In Section 5, the conventional identi®cation scheme and its main disadvantages ± the need of optimised
trajectories and model-based friction compensation ± are discussed. In this section, a di€erent approach is
introduced which eliminates these disadvantages.

6.1. Identi®cation scheme

The identi®cation scheme is based on the grouping of the base parameters presented in Section 4.4. In
the ®rst step, the elements of g…q† and M…q† are ÔmeasuredÕ for a number of di€erent joint con®gurations
(Sections 6.2 and 6.3). In the second step, parameters are estimated by a combined evaluation of the
separate measurements.
The measuring of the gravitational torque at m di€erent joint con®gurations and the combination of all
these measurements leads to
2 …1† 3 2 ÿ
3 2 3
Q c1 Ag;c1 q…1† e…1†
6 . 7 6 .. 7 6 . 7
4 .. 5 ˆ 4 . 5 pg ‡ 4 .. 5 ; …14†
ÿ

Q…cmm † Ag;cm q…m† e…m†
|‚‚‚‚{z‚‚‚‚} |‚‚‚‚‚‚‚‚‚‚‚{z‚‚‚‚‚‚‚‚‚‚‚} |‚‚‚‚{z‚‚‚‚}
Cg Wg eg

ÿ …i†

where Q…n†ci are the measured gravitational torques, Ag;ci q are the corresponding rows of Ag …q† for the
given joint con®guration, q…i† and e…i† are the unknown measurement errors. Suitable joint con®gurations
are selected by analysis of the gravitational vector g…q† (Section 6.2). The application of a weighted LS
estimation leads to
   ÿ1
^pg ˆ min eTg Weg ) ^ pg ˆ WTg WWg WTg WWg : …15†
pg

The diagonal weighting matrix takes into account the di€erent ranges of the torque measurements by
weighting them with the maximum torque of each axis
ÿ

W ˆ diag wc1 . . . wcm with wci ˆ Qÿ1
ci ;max : …16†
^ ii in Section 6.3) at k
Next, pMd is identi®ed by measuring diagonal elements of the mass matrix Mii (see M
di€erent operating points which are chosen by analysing M…q† (Section 6.2). These measurements are
1896 M. Grotjahn et al. / International Journal of Solids and Structures 38 (2001) 1889±1902

combined and the in¯uence of the already known ^pg is compensated as well as the diagonal elements Hii of
the mass matrix H…q†
2 3 2 ÿ
3 2 …1† 3 2 ÿ
3 2 3
M^ …1† AMg;c1 q…1† ; uc1 Hc1 AMd;c1 q…1† ; uc1 e…1†
c1
6 . 7 6 7 6 7 6 7 6 . 7
6 . 7ÿ4
4 . 5
..
. 5^pg ÿ 4 ... 5 ˆ 4 ..
. 5 pMd ‡ 4 .. 5 : …17†
ÿ
ÿ

^
M ck…k †
A Mg;ck q …k †
; u ck H …k †
ck A Md;ck q …k †
; u ck e …k †

|‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚{z‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚} |‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚{z‚‚‚‚‚‚‚‚‚‚‚‚‚‚‚} |‚‚‚{z‚‚‚}

WMd eMd
CMd

The compensation of p ^g does not signi®cantly worsen the estimation result. The measurements of the
elements of g…q† are very simple and a relatively high precision of ^pg can be achieved. The vector uci is a
vector of same size as 
q with a 1 in the component ci of the measured inertia and zeroes elsewhere. This
means that Eq. (17) has the dimension of a moment of inertia for rotational joints and the dimension of a
mass for translational joints. An estimation for pMd can then be found by
ÿ
ÿ
ÿ1
^pMd ˆ min eTMd WeMd ) ^ pMd ˆ WTMd WWMd WTMd WWMd : …18†
pMd

The identi®cation of pMod is completely analogous to Eqs. (17) and (18). The di€erence lies in the exe-
cution of the measurement, where acceleration of one axis is related to the torque of another axis (Daemi
and Heimann, 1998). Finally, ^ pg , ^
pMd and ^
pMod are combined to ^p.

6.2. Choice of joint con®gurations

It must be ensured that all parameters are suciently excited. The in¯uence of one special parameter in
g…q† and M…q† depends on the joint angles. This dependence is excited by varying the joint con®guration.
As an example, one can regard the fourth diagonal element of M…q† of the manutec-r15: M44 ˆ
p10 ‡ p11 sin2 …q5 † ‡ p14 cos2 …q5 †. The parameters p10 , p11 and p14 can be excited and estimated by varying q5
for di€erent measurements of M44 . In our experiment, for example, we changed q5 from ÿ100° to 100° in
steps of 25°. An analogous analysis and variation of all angular dependencies in g…q† and M…q† yields
relatively good results for the characteristic values of the estimation problem, like condition number or
smallest singular value. Therefore, an optimisation is not necessary to ensure sucient excitation.

6.3. Measurement trajectories

A major advantage of the measurements at operating points is the fact that they can be implemented
easily in typical industrial robot controls. Each measurement is carried out by moving the axis Ôback and
forthÕ along some trapezoidal velocity pro®le in the neighbourhood of the operating point. No optimisation
strategy is necessary as only one quantity is interesting and the trajectories have a special characteristic to
ensure excitation of this quantity.
For measurements of gravitational torques, long periods with constant velocity have to be included (Fig.
3a) in order to excite gravitational e€ects and to eliminate inertial in¯uences. The mean value between an
averaged torque at forward and backward motion gives the desired gravitational torque.
Motions with a higher share of acceleration are used to identify the moments of inertia (Fig. 3b). The
gravitational component is compensated and a joint model
1=3
Qi ˆ Mii qi ‡ a1 q_ i ‡ a2 sign…q_ i † ‡ a3 q_ i …19†

is adapted to the measurement again by use of a LS estimation

 M. Grotjahn et al. / International Journal of Solids and Structures 38 (2001) 1889±1902 1897

Fig. 3. Typical trajectories for identi®cation of (a) gravitational and (b) inertial parameters.

2   3
1=3 2 3
q …t † q_ i …t1 † sign q_ i …t1 † q_ i …t1 † Qi …t1 †
6 i 1 7 ÿ
ÿ1
6 . 7 6 7
Wˆ6 . .. . .. 7; C ˆ 4 ... 5 ) ^pr ˆ WT W WT C …20†
4 . .  ..  . 5
qi …tr † q_ i …tr † sign q_ i …tr †
1=3
q_ i …tr † Qi …tr †

 T
with ^pr ˆ M^ ii ; a^1 ; a^2 ; a^3 .
The measurement of the o€-diagonal elements of the inertia matrix is similar to the gravitational
measurements. Axis i is moved at a very low constant velocity to avoid sticking and axis j is accelerated in
the neighbourhood of the operating point. Since speed of axis i is constant, a very simple linear friction
model is applied

Qj ˆ Mji qi ‡ a1 q_ j : …21†

Finally, the coupling inertia Mji is identi®ed similarly to Eq. (20).

The coupling expressed by c…q; q† _ (Section 4.4) has no in¯uence on the measurements of gravitational
torques and the diagonal elements of the mass matrix. The element of c…q; q† _ that corresponds to the
measured torque becomes zero since only this single axis is moved. For the measurements of the o€-di-
agonal elements of the mass matrix, however, the in¯uence of c…q; q† _ is not zero in general. But it can
usually be neglected as both axes are moved at low velocity.

7. Experimental set-up

The test stand consists of the commercial robot Siemens manutec-r15 and its standard control ACR-20.
The measurements are carried out by a dSPACE real-time computer which is connected to a WinNT-PC
with the program package M A T L A B / S I M U L I N K .

7.1. Measurement and signal processing

For the identi®cation joint positions and driving torques are obtained from the motor encoder signals
and motor currents. Velocities and accelerations are calculated by numerical di€erentiation. All mea-
surements are ®ltered with a non-causal, phase-neutral eighth-order Butterworth low-pass ®lter. Although
all signals are measured on the motor side, the in¯uence of gear elasticity and backlash can be neglected
since only shock- and jerk-less trajectories are used.
1898 M. Grotjahn et al. / International Journal of Solids and Structures 38 (2001) 1889±1902

7.2. The industrial robot Siemens manutec-r15

The Siemens manutec-r15 is a six d.o.f. industrial robot with revolute joints. In Fig. 4, the de®nition of
the MDH coordinate frames and parameters are given. As mentioned in Section 4.1, a lot of parameters can
be set to zero since they have no in¯uence on the dynamics. Table 1 shows the remaining elements of the
plink;i for all links. The application of Gautier and KhalilÕs algorithm (1988) to these de®nitions leads to the
base parameter vector
2 3 2 3
p1 Izz1 ‡ Iyy2 ‡ Iyy3 ‡ l21 …m3 ‡ m4 ‡ m5 ‡ m6 †
6 p2 7 6 Ixx2 ÿ Iyy2 ÿ l21 …m3 ‡ m4 ‡ m5 ‡ m6 † 7
6 7 6 7
6 p3 7 6 Ixz2 7
6 7 6 7
6 p4 7 6 2
Izz2 ‡ l1 …m3 ‡ m4 ‡ m5 ‡ m6 † 7
6 7 6 7
6 p5 7 6 sx2 ‡ l1 …m3 ‡ m4 ‡ m5 ‡ m6 † 7
6 7 6 7
6 p6 7 6 Ixx3 ÿ Iyy3 ‡ Iyy4 ‡ 2l2 sz4 ‡ l2 …m4 ‡ m5 ‡ m6 † 7
6 7 6 2 7
6 p7 7 6 Izz3 ‡ Iyy4 ‡ 2l2 sz4 ‡ l2 …m4 ‡ m5 ‡ m6 † 7
pˆ6 6 7 6 7:
7ˆ6 …22†
2
7
6 p8 7 6 sy3 ÿ sz4 ÿ l2 …m4 ‡ m5 ‡ m6 † 7
6 p9 7 6 Ixx4 ‡ Iyy5 ÿ Iyy4 7
6 7 6 7
6 p10 7 6 Izz4 ‡ Iyy5 7
6 7 6 7
6 p11 7 6 Ixx5 ‡ Ixx6 ÿ Iyy5 7
6 7 6 7
6 p12 7 6 Izz5 ‡ Ixx6 7
6 7 6 7
4 p13 5 4 sy5 ÿ sz6 5
p14 Izz6

The in¯uence of the base parameters on g…q† and M…q† are given in Table 2. It shows that the base pa-
rameters can be divided according to Eq. (8) into three groups:

1. p5 ; p8 ; p13 are the only parameters contained in g…q†. They are combined in pg .
2. All remaining parameters with the exception of p3 can be found on the diagonal elements of M…q†. They
are combined in pMd .

Fig. 4. MDH coordinate frames and parameters for the manutec-r15.

Table 1
Non-negligible parameters of the manutec-r15
plink;1 plink;2 plink;3 plink;4 plink;5 plink;6
Izz1 Ixx2 ; Ixz2 ; Iyy2 ; Izz2 ; sx2 Ixx3 ; Iyy3 ; Izz3 ; sy3 ; m3 Ixx4 ; Iyy4 ; Izz4 ; sz4 ; m4 Ixx5 ; Iyy5 ; Izz5 ; sy5 ; m5 Ixx6 ; Iyy6 ; Izz6 ; sz6 ; m6
 M. Grotjahn et al. / International Journal of Solids and Structures 38 (2001) 1889±1902 1899

Table 2
Dependencies of g…q† and M…q† on p for the manutec-r15
Mij …q† jˆ1 jˆ2 jˆ3 jˆ4 jˆ5 jˆ6 gi …q†
iˆ1 p1 ; p2 ; p6 ; p8 ; p9 ; p10 ; p3 ; p9 ; p11 ; p12 ; p13 ; p14 p9 ; p11 ; p12 ; p13 ; p14 p10 ; p11 ; p13 ; p14 p12 ; p13 p14 0
p11 ; p12 ; p13 ; p14
iˆ2 Sym. p4 ; p7 ; p8 ; p9 ; p11 ; p7 ; p8 ; p9 ; p11 ; p12 ; p11 ; p13 ; p14 p12 ; p13 p14 p5 ; p8 ; p13
p12 ; p13 ; p14 p13 ; p14
iˆ3 Sym. Sym. p7 ; p9 ; p11 ; p12 ; p13 ; p14 p11 ; p13 ; p14 p12 ; p13 p14 p8 ; p13
iˆ4 Sym. Sym. Sym. p10 ; p11 ; p14 0 p14 p13
iˆ5 Sym. Sym. Sym. Sym. p12 0 p13
iˆ6 Sym. Sym. Sym. Sym. Sym. p14 0

3. For this robot p3 is the only parameter that can only be found on the o€-diagonal elements of M…q† and
thus, has to be identi®ed by measuring coupling torque.

A more complete description of the robot is given in Daemi and Heimann (1998) and Daemi (1998).

8. Results

The algorithm described in this paper is applied to the manutec-r15. For the identi®cation of pg , mea-
surements at 64 joint con®gurations are used, 92 con®gurations for pMd and nine con®gurations for pMod .
The duration for all measurements is about 45 min. A second series of measurements is carried out with an
additional load of 10.86 kg mounted on the end e€ector. The distance of its centre of mass to the hand wrist
is 140 mm. Figs. 5 and 6 give a comparison of the measured gravitational torques and moments of inertia
with those modelled by the identi®ed parameter vectors. The dependencies on the joint angles are accurately

Fig. 5. Gravitational torques with (´) and without (s) additional mass compared to their model representation (Ð and - - -).

Fig. 6. Moments of inertia with (´) and without (s) additional mass compared to their model representation (Ð and - - -).
1900 M. Grotjahn et al. / International Journal of Solids and Structures 38 (2001) 1889±1902

Table 3
Identi®cation results for manutec-r15 with (A) and without additional (B) payload
i 1 2 3 4 5 6 7
pi (A) 32.015 ÿ28.058 ÿ2.187 19.905 48.491 9.602 13.313
pi (B) 36.907 ÿ33.782 ÿ2.435 22.565 51.896 16.331 18.562
Dpi 4.892 ÿ5.724 ÿ0.247 2.6597 3.405 6.7296 5.2493
Dpi (exp.) 2.713 ÿ2.713 0 2.713 5.409 5.786 5.786

i 8 9 10 11 12 13 14
pi (A) ÿ15.836 0.0000 0.0443 0.2883 0.1554 ÿ0.1885 0.3760
pi (B) ÿ23.244 ÿ0.0895 0.2198 0.2274 0.4381 ÿ1.9983 0.3980
Dpi ÿ7.408 ÿ0.0895 0.1755 ÿ0.0609 0.2826 ÿ1.8097 0.0220
Dpi (exp.) ÿ7.885 0.0 0.0 0.0 0.0 ÿ1.518 0.0470

Fig. 7. Clipping from Fig. 6 without additional mass (con®guration number 10±24).

re¯ected by the model and the in¯uence of the additional mass is obvious. Many of the measured moments
of inertia are constant with respect to changing joint con®guration. But the main reason for this behaviour
is the chosen representation in which the in¯uence of the gravitational parameters is compensated. They
contribute signi®cantly to the angular dependency since they re¯ect contributions due to Steiner's theorem.
Numerical results for the base parameters with and without additional mass are shown in Table 3. The
di€erences between the two identi®cation results are also given, which re¯ect the in¯uence of the additional
mass. They are compared to their expected values resulting from the base parameter de®nition in Eq. (22).
It has to be considered that due to the compensation of the motor inertia and the friction, only a small
fraction of the measured torques contain information for the identi®cation process. Especially, for pa-
rameters with higher indices results become more inaccurate because the in¯uence of friction in the hand
axes increases. Still, the increase of parameters is well re¯ected in the measurements indicating a successful
identi®cation of the complex dynamics of the robotic system.
Besides simplicity of the measurements and no need of model-based friction compensation, another
advantage of this method becomes visible. By analysing the measurements, an evaluation of the model
structure is possible to ®nd unmodelled e€ects as well as unnecessary parameters. An example is given in
Fig. 7, where the gravitational torque of joint 3 is plotted as a function joint angle four. An o€set of about
30° of the measured sinusoidal torque characteristic is observable which cannot be re¯ected by the model.
This is due to the unmodelled distance of the centre of mass of joint four to its rotational axis.

9. Conclusions

The conventional model consisting of dry friction and viscous damping is usually not sucient for the
modelling of friction losses in industrial robotsÕ gears. But even by more complex models, the change of
 M. Grotjahn et al. / International Journal of Solids and Structures 38 (2001) 1889±1902 1901

friction with respect to temperature and lubrication conditions cannot be re¯ected with maintainable e€ort.
The lack of predictability can lead to a deterioration of rigid body identi®cation. The presented identi®-
cation method eliminates this disadvantage. Furthermore, it is much easier to apply to standard industrial
robots since only simple trajectories with trapezoidal velocity pro®le are used. The third main advantage of
the presented method is the possibility of evaluating the model by investigation of the measurements. The
eciency of the method is illustrated by experimental application to the standard industrial robot manutec-
r15.

References

Armstrong, B., 1988. Dynamics for robot control: Friction modeling and ensuring excitation during parameter identi®cation. Stanford
University, Stanford, CA, 1988.
Armstrong, B., 1989. On ®nding exiting trajectories for identi®cation experiments involving systems with nonlinear dynamics. Int. J.
Robotics Res. 8 (6), 28±48.
Armstrong-Helouvry, B., 1991. Control of Machines with Friction. Kluwer Academic Publishers, Boston.
Armstrong-Helouvry, B., Dupont, P., Canudas de Wit, C., 1994. A survey of models, analysis tools and compensation methods for the
control of machines with friction. Automatica 30 (7), 1083±1138.
Canudas de Wit, C., Noel, P., Aubin, A., Brogliato, B., 1991. Adaptive friction compensation in robot manipulators: low velocities.
The Int. J. Robotics Res. 10 (3), 189±199.
Daemi, M., 1998. Modellierung und Identi®kation der Dynamik von Industrierobotern f ur den Einsatz in Regelungen. Fortschritt-
Berichte VDI. VDI-Verlag, D usseldorf (in German).
Daemi, M., Grotjahn, M., 1996. Practical experiences with L.S. methods for the identi®cation of robot dynamics. In: Proc. of the 2nd
ECPD International Conference on Advanced Robotics, Vienna.
Daemi, M., Heimann, B., 1996. Identi®cation and compensation of gear friction for modeling of robots. In: Proc. of the 11th CISM-
IFToMM Symp. on the Theory and Practice of Robots and Manipulators, Udine, pp. 89±96.
Daemi, M., Heimann, B., 1998. Separation of friction and rigid body identi®cation for industrial robots. In: Proc. of the 11th CISM-
IFToMM Symp. on the Theory and Practice of Robots and Manipulators, Paris, pp. 35±42.
Gautier, M., 1990. Numerical calculation of the base inertial parameters. In: Proc. of the IEEE Int. Conf. on Robotics and
Automation, Cincinati, pp. 1020±1025.
Gautier, M., Khalil, W., 1988a. A direct determination of minimum inertial parameters of robots. In: Proc. of the IEEE Int. Conf. on
Robotics and Automation, pp. 1682±1687.
Gautier, M., Khalil, W., 1988b. On the identi®cation of the inertial parameters of robots. In: Proc. of the IEEE Conf. on Decision and
Control, pp. 2264±2269.
Gautier, M., Khalil, W., 1992. Exciting trajectories for the identi®cation of base inertial parameters of robots. Int. J. Robotics Res. 11,
362±375.
Gautier, M., Khalil, W., Restrepo, P.P., 1995. Identi®cation of the dynamic parameters of a closed loop robot. In: Proc. of the IEEE
Int. Conf. on Robotics and Automation, Nagoya, pp. 3045±3050.
Gautier, M., Vandanjon, P.O., Presse, C., 1994. Identi®cation of inertial and drive gain parameters of robots. In: Proc. of the IEEE
Conf. on Decision and Control, Lake Buena Vista, pp. 3764±3769.
Khalil, W., Klein®nger, J.F, 1986. A new geometric notation for open and closed-loop robots. In: Proc. of the IEEE Int. Conf. on
Robotics and Automation, pp. 1174±1179.
Khosla, P.K., 1988. Estimation of robot dynamics parameters: theory and application. Int. J. Robotics and Automation 3 (1), 35±40.
Khosla, P.K., Kanade, T., 1987. An algorithm to estimate manipulator dynamics parameters. Int. J. Robotics and Automation 2 (3),
127±135.
Lawson, C.L., Hanson, R.J., 1974. Solving Least Squares Problems. Prentice-Hall, Englewood Cli€s, NJ.
Pfei€er, F., Holzl, J., 1995. Parameter identi®cation for industrial robots. In: Proc. of the IEEE Int. Conf. on Robotics and
Automation, Nagoya, pp. 1468±1476.
Presse, C., Gautier, M., 1993. New criteria of exciting trajectories for robot identi®cation. In: Proc. of the IEEE Int. Conf. on Robotics
and Automation, Atlanta, pp. 907±912.
Pr
ufer, M., Schmidt, C., Wahl, F., 1994. Identi®cation of robot dynamics with di€erential and integral models: a comparison. In: Proc.
of the IEEE Int. Conf. on Robotics and Automation, San Diego, pp. 340±345.
Pr
ufer, M., Wahl, F., 1994. Friction analysis and modelling for geared robots. In: Proc. of the 4th IFAC Symposium on Robot
Control, Capri, pp. 551±556.
1902 M. Grotjahn et al. / International Journal of Solids and Structures 38 (2001) 1889±1902

Seeger, G., 1992. Selbsteinstellende, modellgest

utze Regelung eines Industrieroboters. In: Fortschritte der Robotik, vol. 13. Vieweg-
Verlag, Wiesbaden, (in German).
Shaefers, J., Xu, S.J., Darouach, M., 1993. Sensitivity and error analysis of parameter identi®cation for a class of industrial robots. In:
Proc. of the IEEE Int. Conference on Systems, Man and Cybernetics, Le Touquet, pp. 195±200.