Robotica (1995) volume 13, pp 253-258.
© 1995 Cambridge University Press
A meta-study of PUMA 560 dynamics: A critical appraisal of
literature data
Peter I. Corke* and fBrian Armstrong-Helouvry
(Received in Final Form: July 25, 1994)
SUMMARY
The paper presents a meta-study of the kinematic,
dynamic and electrical parameters for the PUMA 560
robot. Parameter values which have been reported in the
literature are transformed into a single system of units
and coordinates, and differences in the data and
measurement techniques are discussed. New data have
been gathered and are presented where the record was
incomplete.
KEYWORDS: PUMA dynamics; Meta-study; Parameters.
1. INTRODUCTION
Research on visual servoing at CSIRO 1 led to a need to
implement dynamic-model-based control. The robot
platform was a PUMA 560, for which there is a
substantial literature. Rather than re-implementing
sophisticated model estimation experiments, the neces-
sary dynamic model parameters were sought in the
literature.2"10 Presented here are numerical comparisons
of data found in these reports.
In the absence of definitive data from the manufac-
turer, researchers have pursued many approaches to
estimating this data. The techniques used vary widely
and include estimation from geometric data,5'7 direct
disassembly and measurement 2 and unattributed in-
cidental data in tables. More recently there has been
considerable literature related to time-domain identifica-
tion techniques for manipulators such as the PUMA
260,11 MIT Direct Drive arm,12 the CMU DD-II arm,13-14
and a Manutec R3. 15 Apart from Mayeda,16 there have
been no reports on this approach for the PUMA 560.
The authors attribute this to the low velocity and
acceleration capabilities of the PUMA 560 which
preclude accurate determination of inertial parameters.
Online model-based controllers for the PUMA 560 have
been reported by a number of authors, see An et al}1 for
a discussion of reports through 1988, and Leahy et al.xs
for an experimental comparison of several model-based
control structures. Principally these controllers have been
based on model data from the sources cited here.
•Division of Manufacturing Technology CSIRO, Preston,
Victoria 3072 (Australia).
t Department of Electrical Engineering and Computer Science,
University of Wisconsin-Milwaukee, Milwaukee, Wisconsin
53201 (U.S.A.)
However, as shown below, there is considerable variation
in the model data available for the PUMA 560 robot.
The success of the model-based controllers may be
interpreted as a demonstration of the robustness of
model-based control approaches when applied to the
relatively slow and rigid PUMA 560 manipulator.
Bringing together results from the literature involves
more than might meet the eye. The values of dynamic
parameters depend upon the choice of coordinate frames
in which they are expressed, whether inertia is given in a
center-of-gravity or axis-of-rotation frame, and upon the
choice of physical units. Principally because there are
two Denavit-Hartenberg conventions in use19'20 and
within each convention there are user-defined degrees of
freedom arising from parallel axes and zero pose, no two
reports of PUMA dynamics present their results in
exactly comparable systems of coordinates.
Toward the goal of bringing together a complete and
consistent set of kinematic, dynamic and electrical
parameters of the PUMA 560, data presented in the
above cited papers, along with new data from the
original manufacturers and from CSIRO measurements,
have been translated into a single system of coordinates
and units. With the data in a directly comparable form,
several things may be accomplished:
1. Gaps which may exist in some reports may be filled
in using data available elsewhere.
2. It is possible to compare results. In some cases it is
possible to identify outliers in the data; and where
consensus can be established among reports, it will
give confidence that the data are reliable.
3. By looking at the spread in reported parameters, it
is possible to assess the challenge posed by accurate
dynamic parameter identification.
In medicine or the natural sciences it is common for
several research groups to repeat experiments and for
the data to be ultimately compared and contrasted in a
meta-study, that is a study of studies. In these fields a
measurement needs to be made several times before it is
fully trusted. In engineering we do less of this, but the
PUMA 560 arm presents something of a unique
opportunity to assess the challenges to accurate
parameter identification by observing the degree of
variation among reported values. This paper then is a
meta-study of PUMA 560 dynamic characteristics.
254 PUMA 560 Dynamics
246
In this paper, a table will be given for each group of
parameters showing data reported in the literature ZS Y46

translated into a single system of coordinates. Addition-

ally, new data, particularly electrical parameters, are
presented. Throughout the rest of the paper sources will
be referred to by the keys given in Table I. X4S6

1.1 A note on coordinate'frames

Two different systems for coordinate frame assignment
are now in use:
1. Denavit-Hartenberg (DH) assignments where
frame / has its origin along the axis of joint i + 1, as
described by Paul21 and Lee.3'4
2. Modified Denavit-Hartenberg (MDH) assignments
where frame i has its origin along the axis of joint i, X3
as described by Craig.20 This form will be used here
as the basis for comparison.
To permit direct comparison, PUMA 560 parameters
from the source papers have been translated to a single
system of coordinates. The modified Denavit-
Hartenberg representation20 is used, with frame assign-
ments and zero-angle pose as shown in Figure 1.

2. A COMPARISON OF KINEMATIC
PARAMETERS
The kinematic models of the 11 sources must be
considered in order to transform the inertial parameters
into a single system of coordinates. Five sets of kinematic
parameters are compared in Table II. Each set of
parameters must be taken in the context of the axis and
angle conventions in the cited paper. However there is
clearly some variation in the link lengths and offsets cited
by various authors. These could conceivably reflect
changes to the design and manufacture of the robot with
time. Some comments on these parameters are:
• The modified Denavit-Hartenberg representation
has 5 length parameters compared to four in the
Table I. Keys to principal sources used in the text.
Key Source
Armstrong Armstrong et al}
Lee Fu, Gonzalez and Lee,3 and also Lee.4
Paul81 Paul, Rong and Zhang.5
Paul86 Paul and Zhang.6
Tarn Tarn et al?
BreakingAway "Breaking Away from Val" a memo
describing operations of the Unimation servo
system in some detail.8
Unimation A data sheet of unknown origin but
purporting to be from Unimation, listing
dynamic parameters of motors and links.
Kawasaki A data sheet of motor specifications obtained
from the local (CSIR) Kawasaki robot
distributor.
RCCL Source code of RCCL software.'-25
MU Measurements taken on University of
Melbourne's Kawasaki 560.10
CSIRO Measurements taken on CSIRO Division of
Manufacturing Technology's Unimate 560.
Fig. 1. Frame assignments made according to the modified
Denavit-Hartenberg convention.
standard representation. By inspection we can see
that D2MDH + D3MDHm D3OH, and Armstrong's
data thus agrees closely with Tarn and Lee.
• Paul81 apparently gives an incorrect sign for A3.
The sign of A3 in Paul86 is correct due to the
definition of the zero-angle pose.
• The a, values given by Paul86 are negative
compared to PaulSl and Lee. This is due to the
definition of a right-handed configuration for the
zero-angle pose in Paul86.
• The value of D3 from Paul86 is given as 125.4,
which is significantly lower than the other reported
values.
• Lee alone gives a value for D6, which is the distance
from the wrist center to the flange plate. Thus Lee
places the T6 coordinate frame on the flange, while
all others consider it as the center of the wrist.
3. A COMPARISON OF INERTIAL
PARAMETERS
3.1 Link mass
Reported values for link mass are presented in Table III.
Armstrong's data were determined by dismantling the
robot and weighing the components. Paul81's paper
provides only "normalized mass" figures with link 6
being assigned a relative mass of 1. The figures are
simply normalized versions of the "Plato areas" given in
the same table. In the table we have equated the mass of
link 3 with Armstrong's value. Tarn's data are from
PUMA 560 Dynamics 255

Table II. Comparison of kinematic constants.

Paul81 Paul86 Lee Tarn Armstrong

A, A a, A, A a, A, A <*i A, A a,_, A,-x A

-90 0 0 90 0 0 -90 0 0 -90 0 0 0 0 0
0 432 0 0 431.8 0 0 431.8 149.09 0 431.8 0 -90 0 243.5
90 20 149 -90 19.1 125.4 90 -20.32 0 90 -19.1 150.05 0 431.8 -93.4
-90 0 432 90 0 431.8 -90 0 433.07 -90 0 431.1 90 -20.3 433.1
90 0 0 -90 0 0 90 0 0 90 0 0 -90 0 0
0 0 0 0 0 0 0 0 56.25 0 0 0 90 0 0
Each is specified in terms of the coordinate frames of the cited paper (a in degrees, A and D are in mm).

estimation and measurement of the components of each have been translated to the center of gravity
link and are consistently higher than Armstrong's. representation for Table VI. Radii of gyration are
reported in Paul81. Using the values for link mass
3.2 Link center of gravity reported in Table III, these have been translated to
Link-center-of-gravity values are given in Table IV. inertia. Since the inertia and location of center of gravity
Paul81's values are given without explanation, but of link 1 are not separately identifiable in the
examination seems to indicate that uniform distribution manipulator dynamics, the value presented by Armstrong
of mass within the links is assumed. This, however, is as / j Z l is the combined influence in link coordinates, that
unlikely given the monocoque construction technique is, / zsl [Link Coordinates] = 7ZIl[Center of Gravity] +
and the heavy motors at one end of each of links two and mi{szxl + s2yi).
three. Tarn used a combination of measurement and Parameter values vary by 200%-450% throughout
estimation for each component within the link to Table VI. This may be taken as an indication of the
determine the overall value for the link. Armstrong used difficulty of obtaining accurate measurements of inertia.
a knife edge balance to determine the center-of-gravity On the other hand, the range between the largest and
of the disassembled links directly. smallest parameters is 10s, and so the 10OJ (or 300%)
variability in published values of inertia is an order of
3.2.1 Cross-check via gravity loading. Since gravity is a
magnitude less than the spread in these parameters.
dominant dynamic effect, a comparison of the gravity
Control based upon a model with 10°-S uncertainty in
loading coefficients may be more meaningful than link dynamic parameters is perhaps much better than control
mass and center of gravity values alone. The equations of based upon no dynamic model at all.
the gravity loading terms are given in Table A7 of
Armstrong. The coefficients of the gravity loading terms 4. A COMPARISON OF MOTOR PARAMETERS
are evaluated and compared in Table V. There is close There are two types of PUMA 560, one built by
agreement between the magnitudes of the coefficients Unimation (Danbury, CT) and the other by Kawasaki
from Armstrong and those used in RCCL. The sign (Japan). The two types are similar in most respects, but
difference on joints 1 and 3 is likely to be due to the sign
of the gear ratios,8 which are negative for those joints. Table IV. Link center of gravity (mm).
The values used within RCCL were obtained by an
experimental procedure as described by Lloyd for a Parameter Armstrong Paul81 Tarn
PUMA 2609 (Private communication, J. Lloyd, McGill
0 0
University, Canada).
80 4
0 -309
3.3 Link moments of inertia 68 216 103
Table VI gives the moments of inertia about the center 6 0 5
of gravity for each link. Tarn's inertial values are 16 -26 -40
reported relative to the joint axes. Using Tarn's center of 0 0 20
gravity parameters from Table IV, the inertial values 70 -216 -4
14 0 14

Table III. Link mass values (kg). 0 0 0

0 0 -3
Parameter Armstrong Paul81 Tarn 19 -20 -86

__ 0 0 0
m, 4.43 13.0
0 0 -1
m2 17.40 10.2 22.4
0 0 -10
4.80 4.80 5.0
m4 0.82 1.18 1.2 0 0 0
m5 0.35 0.32 0.62 0 0 0
0.09 0.13 0.16 32 10 3
256 PUMA 560 Dynamics

Table V. Comparison of gravity coefficients (Nm). Table VII. Experimental motor torque constants—motor
referenced (Nm/A).
Parameter Armstrong Tarn RCCL
Parameter Armstrong CSIRO Paul81 MU
•37.2 -20.8 37.196 (CP21)
gz -8.44 -7.51 -8.44 (CP30) 0.189 0.223 0.255 0.202
gi 1.02 1.10 -1.023 (CP20) 0.219 0.226 0.220 0.258
g* 0.249 0.390 0.248 (CP22) K, 0.202 0.240 0.239 0.245
-0.0282 0.00455 -0.028 (CP50) 0.075 0.069 0.078 0.095
0.066 0.072 0.070 0.101
0.066 0.066 0.079 0.089

different servo motors are used. In each PUMA two sizes

of servo motor are used; a larger motor for joints 1-3
and a smaller motor for joints 4-6. The following
sections compare data for the two most significant motor
parameters; torque constant and armature inertia.
Information about other motor electrical parameters is
given by Corke and Armstrong-He'louvry.22
operation or that have experienced several abrupt power
shut-down events. Reduced motor magnetization, along
with friction, perhaps accounts for the experimental data
being consistently lower than the manufacturers'
specifications seen in Table VIII.

4.1 Motor torque constant 4.2 Armature inertia

Reported motor torque constants are presented in Table Motor armature inertias are tabulated in Table IX.
VII. These have been translated from the original reports Armstrong's values for motor plus transmission inertia
using gear ratio or knowledge of full motor current (load referenced) were determined by parameter
where necessary. Striking variation is seen, given that identification experiments. Here, they are divided by the
motors 1-3 are identical, as are motors 4-6. The data gear ratio squared to give the values of armature inertia
reported in Table VII are averaged and compared with presented in Table IX (motor referenced). From our
data from the motor manufacturers in Table VIII. The knowledge of motor similarity, Armstrong's value for
manufacturer's data indicate that a higher torque motor 2 seems anomalous. Armstrong's data for the wrist
constant in the wrist motors of the Kawasaki model is joint motor inertias is consistently higher than the
one difference between the robots. Considerable manufacturer's data. This may be explained by the
variation is also seen in Table VIII. Communication with inertia of the shaft and flexible coupling which cannot
a robotics firm supporting PUMAs, (Private communica- experimentally be separated from armature inertia.
tion, F. Pagano, AR2 Inc., Oxford, Connecticut, USA)
indicates that the Unimation data are specifications of a 5. FRICTION
lower limit, and that variation of up to 20% above the Machine friction will vary substantially between robots,
specification is normal. A remagnetizing process is and for a given robot with changes in temperature and
needed for motors with many hours of lubrication. Due to the gear type transmission, joint
friction also has a substantial position dependence.23 A
seven-parameter friction model has recently been
Table VI. Link moments of inertia. proposed,23-24 but in the literature one most often finds
Parameter Armstrong Tarn Paul81 Coulomb and viscous friction parameters reported.
Typically, each parameter is dependent upon direction of
1.100 0.195 motion, and has a component due to the motor and a
1.110 0.026 component due to the transmission. Writing friction on
0.350 0.177 0.195 the ith joint in the form:
0.130 0.403 0.588
0.524 0.969 1.886 •Bje, e<o
0.539 0.965 1.470 B?9, 0 > 0 (i)

'r 66.0 e-3

12.5 e-3
74.8 e-3
7.3 e-3
324.0 e-3
17.0 e-3
C 86.0 e-3 75.6 e-3 324.0 e-3 Table VIII. Comparison of average ex-
perimental values and manufacturer data
4.. 1.80 e-3 5.32 e-3 3.83 e-3 (Nm/A).
'„. 1.80 e-3 5.20 e-3 3.83 e-3
2.50 e-3
1.30 e-3 3.37 e-3 Source Kwrilt
300 e-6 487 e-6 216 e-6
Armstrong 0.203 0.069
300 e-6 482 e-6 216 e-6
400 e-6 572 e-6 348 e-6 CSIRO 0.230 0.069
Paul81 0.238 0.076
150 e-6 123 e-6 437 e-6 MU 0.235 0.095
150 e-6 123 e-6 437 e-6 RCCL 0.254
40 e-6 58 e-6 13 e-6 Tarn 0.259
Unimation 0.260 0.090
Moments of inertia about the center of gravity—load Kawasaki 0.258 0.097
referenced (kg m 2 ).
PUMA 560 Dynamics 257

Table IX. Motor inertia—motor referenced (kg m2).

Parameter Armstrong Vnimation Tarn Kawasaki CSIRO

291e-6 200e-6 198e-6 200e-6

409e-6 200e-6 203e-6 200e-6
299e-6 200e-6 202e-6 200e-6
35e-6 18e-6 18.3e-6 20e-6
35e-6 18e-6 18.3e-6 20e-6
35e-6 18e-6 18.3e-6 20e-6 26e-6

the Coulomb and viscous friction parameters are fi and B, Table XI. Degree of variability in reported values of the
respectively, and 6 is the joint angle. Those parameters PUMA 560 parameters.
are given in Table X. The data give some impression of
the variability encountered, though it should be stressed Normalized Ratio of
Standard Extremal
that the experimental approaches used may vary widely. Parameter Table Deviation Values
The break-away friction level (the static friction) has
been observed to be roughly 120% of the Coulomb Kinematics 2 2.7% 1.20
friction level in the PUMA 560.23 Link Mass 3 29% 2.93
Center of Gravity 4 124% 10.70
Gravity Coefficients 5 27% 6.20
6. DISCUSSION Moments of Inertia 6 66% 4.52
The degree of variability found among reported values Motor Torque 7 11% 1.53
for each of the tabulated parameters is seen in Table XI. Armature Inertia 9 29% 2.06
For each parameter, the normalized standard deviation is Friction Parameters 10 40% 4.42
given by the square root of the variance between
published values divided by the mean of those values.
The value in Table XI is then the mean of the several different measurements that would give con-
normalized standard deviations within a table. fidence in the correctness of the data. Regarding Tables
Whereas the normalized standard deviation is an L 2 II through X, it is clear that consensus values are not
measure, the ratio of extremal values is an L» measure. available, even for such basic parameters as link mass or
It is the ratio of the largest value given for a parameter coefficients of the gravity loading model. In some cases,
to the smallest, non-neglible value given. For each table, such as the electrical parameters, inter-robot variability
the largest such ratio is presented in Table XI. has been identified and may account for the variations
At the outset of this project, it was hoped that by among reports. In other cases, however, the observed
comparing the data from several reports, consensus and variability stems from challenges to accuracy underlying
thereby more reliable data should be obtained. And the parameter measurement and estimation methods
indeed, for each parameter and from each report there applied to robots.
are data which are inconsistent with the others, and may As robots become more compliant the demands on the
be regarded as outliers. Thus, by combining data from accuracy of the dynamic model will become more
the 11 available reports, a more complete and reliable stringent, and obtaining accurate dynamic model
model is achieved. parameters will be a priority. When the time comes that
Rejecting outliers, however, is not equivalent to the literature can provide multiple reports of dynamic
establishing consensus. And it is consensus among parameter measurements taken with a new robot, these
too should be translated into a single coordinate frame
for comparison. Doing so will make possible assessment
Table X. Measured friction values for three PUMA 560s.
of consensus that exists and give confidence in the
Parameter Armstrong16 CSIRO MU Motor correctness of the data.
B7 3.45 6.27 3.85
B: 4.94 6.40 3.20
-29.8 -6.74 -6.14 Acknowledgments
/r
n
-8.26
8.43 27.0 7.24 6.14 The authors appreciate the efforts of Vaughan Roberts
B; 8.53 8.89 22.1 and Malcolm Good, who performed some of the
7.67 11.7 24.7 parameter identification work referred to for the CSIRO
-11.34 -8.30 -13.0 -10.6 PUMA robot; and Ron Perez, who provided many
n 12.77 14.7 15.9 10.6
helpful comments on a draft of this manuscript.
B; 3.02 5.31 5.59
b; 3.27 2.91 4.33
References
n -5.57 -5.87 -4.56 -5.26
n 5.93 7.37 4.19 5.26 1. P. Corke and M. Good, "Dynamic effects in high-
performance visual servoing" Proc. IEEE Int. Conf.
Coulomb friction, ft (Nm), viscous friction, B, (Nms/rad). Robotics and Automation (1992) pp. 1838-1843.
Superscripts indicate direction of rotation. Load referenced. 2. B. Armstrong, O. Khatib and J. Burdick, "The explicit
258
dynamic model and inertial parameters of the Puma 560
arm" Proc. IEEE Int. Conf. Robotics and Automation 1
(1986) pp. 510-518.
3. K.S. Fu, R.C. Gonzalez and C.S.G. Lee, Robotics, Control,
Sensing, Vision and Intelligence (McGraw-Hill, New York,
1987).
4. C.S.G. Lee, "Robot arm kinematics, dynamics and
control" IEEE Computer 15, 62-80 (1982).
5. R. Paul, M. Rong and H. Zhang, "Dynamics of Puma
manipulator" American Control Conference (1983) pp.
491-496.
6. R.P. Paul and H. Zhang, "Computationally efficient
kinematics for manipulators with spherical wrists" Int. J.
Robot. Res. 5, No. 2, 32-44 (1986).
7. T.J. Tarn, A.K. Bejczy, S. Han and X. Yun, "Inertia
parameters of Puma 560 robot arm" Tech. Rep.
SSM-RL-85-01 (Washington University, St. Louis, MO.,
1985).
8. R. Vistnes, "Breaking away from VAL" Tech. Rep.
(Unimation Inc. 1981).
9. J. Lloyd, "Implementation of a robot control development
environment" Master's Thesis (McGill University, 1985).
10. M. Liu, "Puma 560 robot arm analogue servo system
parameter identification" Tech. Rep. ASR-91-1 (Dept.
Mechanical and Manufacturing Engineering, University of
Melbourne, 1991).
11. A. Izaguirre, M. Hashimoto, R. Paul and V. Hayward, "A
new computational structure for real time dynamics" Tech.
Rep. MS-CIS-87-107 (University of Pennsylvania, 1987).
12. C.H. An, "Trajectory and force control of a direct drive
arm" Tech. Rep. 912 (MIT Artificial Intelligence
Laboratory, 1986).
13. P. Khosla, "Estimation of robot dynamics parameters:
Theory and application" IEEE Trans. Robot. Autom. 3,
35-41 (1988).
14. P. Khosla and T. Kanade, "An algorithm to estimate the
manipulator dynamics parameters" IEEE Trans. Robot.
Autom. 2,127-135 (1987).
PUMA 560 Dynamics
15. G. Seeger and W. Leonhard, "Estimation of rigid body
models for a six-axis manipulator with geared electric
drives" Proc. IEEE Int. Conf. Robotics and Automation
(1989) pp. 1690-1695.
16. H. Mayeda, M. Maruyama, K. Yoshida, N. Ikeda and O.
Kuwaki, "Experimental examination of the identification
methods for an industrial robot manipulator" In: (R.
Chatila & G. Hirzinger) Experimental Robotics II
(Springer-Verlag, London, 1993) pp. 546-560.
17. C.H. An, C.G. Atkeson, J.D. Griffiths, and J.M.
Hollerbach, Model based control of a robot manipulator
(MIT Press, Cambridge, Mass, 1989).
18. M. Leahy, D. Bossert and P. Whalen, "Robot model-based
control: An experimental case study" Proc. IEEE Int.
Conf. Robotics and Automation (1990) pp. 1982-1987.
19. R.S. Hartenberg and J. Denavit, "A kinematic notation for
lower pair mechanisms based on matrices" J. Applied
Mechanics 77, 215-221 (1955).
20. J.J. Craig, Introduction to Robotics (Addison Wesley,
Reading, Mass., 1986).
21. R.P. Paul, Robot Manipulators: Mathematics,
Programming, and Control Massachusetts (MIT Press,
Cambridge, Mass., 1981).
22. P. Corke and B. Armstrong-He'louvry, "A search for
consensus among model parameters reported for the
PUMA 560 robot" Proc. IEEE Int. Conf. Robotics and
Automation (1994) pp. 1608-1613.
23. B. Armstrong-Helouvry, Control of Machines with Friction
(Kluwer, Amsterdam, 1991).
24. B. Armstrong-Helouvry, "Stick slip and control in
low-speed motion" IEEE Trans. Autom. Control 38, No.
10, 1483-1496 (1993).
25. V. Hayward and R.P. Paul, "Robot manipulator control
under UNIX—RCCL: a Robot Control C Library" Int. J.
Robot. Res. 5, No. 4, 94-111 (1986).
26. B. Armstrong, "Dynamics for Robot Control: Friction
Modelling and ensuring Excitation During Parameter
Identification PhD Thesis (Stanford University, 1988).