VERIFICATION OFA LINEAR DYNAMIC MODEL FOR FLEXIBLE ROBOTIC MANIPULATORS

*
By Gordon G. Hastings, and WayneJ. Book

of Technology
Department of Mechanical Engineering, Georgia Institute
Atlanta, Georgia

Abstract
This paper describesa linear state-space model for a flexible single link manipulator arm. The resultant
model is compared to an experimental four foot long direct drive manipulator. The method employed to
generate the model utilizes a separable formulation of assumed modes to. represent the tranverse
displacement due to bending. Lagrangian dynamics are applied to determine the kinetic and potential
energies for the system. The resultant dynamic equations are then organized into a state space model
suitable for use in linear cantrol system design procedures. The performance of the model is considered
for different model orders and assumed modes. Several important aspects of candidate mode selection, and
results for different model order are discussed. The final section of the paper provides a brief summary
and describes ongoing and future work.
Introduction of the process when the equations of motion are
The material in this paper describes a linear model formed in terms of time varying variables only.
which forms the basis for investigating the control Next
the
kineticand
potential
energiesare
of flexible manipulators[l]. The initial sections derived. The distributed character of the flexible
discuss the modelling process, and verify manipulator is takenintoaccountviaintegral
parameters and algorithm implementation. The latterexpressions over the mass of the entire system in
sectioncomparessimulation of the model to forming the energy expressions. The integral for
experimental measurements. calculating the kinetic energy, KE,has the
following form;
Model Generation
This sub-sectiondescribestheformationof
linear statespace model for the flexible
manipulator. The process for forming the model will
a KE = 1/2 R-idm
f
where k , the absolute velocity vector, and mass
be outlined in this section, a detailed description range over the entire system. The potential energy,
is contained in appendix A. PE, of the system is stored in the flexible modes
and can be attributed to "modal stifnesses", K.,
The first step of the process is to describe the which are evaluated by integrals over the length
position of every point along the flexible as shown in equation A.12. Lagrange's equations of
manipulator. A linearcombinationofvibratory motion can be formed from the energies;
modes to describe flexible deflections, and a rigid
body motion of the center of mass is selected. A
manipulator witha rigid body rotation and flexible
"pinned-mass" mode is depicted in figure 1.
where the qi are the coordinates, and Qi are the

REFERIEYCE Y
I PAYLCAO. '
.s-
generalized
work
terms
associated
with
each
coordinate. Turning the computational crank on the
/

various differentials and integralsas carried out

in appendix A results in a coupled set of second
4

I
order dynamic equations with familiar form;
\
\

'JOINT
REFERENCE X .
R O T F i T l O N OF CENTER OF MASS

z = [e,$l(t),+2(t) ,....,$,(t)I (5)

Flexible Manipulator M is a mass matrix, K represents stiffness, and Q

Figure 1 the
input.
The
dynamic
equations
are
easily
nized intoa state-space model as;
Theflexibledeflectionsaredescribed by an
infinite series of separable modes. Separability in
this instance refers to describing the flexible
deflections as a series modes which are products of
two functions each a function of a single variable,
one a function of a spatial varible, and the other
a functionof time. This is noted as:

w(x,t)=l+i(x)@i(t) , -for i=l ,2.. .n (1 1

This separabilityisimportantinlaterphases
*
This material is based in part on work supported by the National Science Foundation under grant
MEA-8303539 and by NASA grant NAG-1-623.

I024
0 1986 IEEE
CH2282-2/86/0000/P02iC~01.00
The input, u, into the systemis torque applied by motoroutputtoact as acurrentsource. The
the motorat the joint, the generalized work terms, physical configuration of the flexible arm, -torque
Q , are then related to the rotation of the joint motor, and sensors is represented in figure 2.
wfth each variable. Examinationof the form of the Table 1 identifies the important parameters of the
model reveals the expected result that the coupling beam.
between the modes, and the rigid body motion occurs
from inertial terms of the mass matrix. System Parameters
Table 1
Equation (6) depicts a 2(n+l) order linear model
where n is the number of included modes. Non-linear Flexible Beam-
terms arise from the evaluationof equation 2 for Material -
Aluminum 6061-T6
the kinetic energy, and the specific assumptions Form - Rectangular 314 x 3/16in
employedtoresultinonlylineartermsis Length - 48 in
discussed in Apppendix A. Moment of Inertia - 4.1 2E-4 in4
Product
E1 - 4120 1bf-in2
Mode Selection and Frequency Determinant
The approach being followed in this developement is Parameter and Program Verification
called assumed modes methods[2], and the remaining This section describes exDeriments conducted to
task in generating a trial model is the selection verify
system
parameters
and
program
of the flexible modes to be used in forming the implementation of the model generation process.
constant mass and stiffness matricies. Initallythe
frequencies
determinedvia
the
Bernoulli-Euler beam equations with clamped-mass,
The pathchosen in thisworkistoselect and pinned-mass boundary conditions are compared to
admissible functions as
candidates which
are measured eigenvalues of the beam. ,This examines
solutionsto
closelyrelatedprobems.
These beam length, modulus, and density parameters, as
solutions are eigen-functions for selected well as, the suitability of the chosen boundary
flclamped-massfT, andlIpinned-massTfboundaryvalue conditions.
problems. Clamped describes a boundary condition
where the joint is fixed against rotation, pinned Figure 3 shows the measured frequency response of
describesajointfreetorotate,and-mass strain at the base of the beam compared to random
describes the condition of the payload at the other t-orques applied by the motor. The peaks correspond
beam boundary. The admissible functions will then to "clamped-mass modes", while the valleys can be
satisfy the differential Equation, the essential or identified by the ffpinned-maasR modes.
geometricboundaryconditions,andthenatural
boundary conditionsof the free vibration problem.

Appendix B describesthedevelopementofthe
differential equation for a Bernoulli-Euler beam
and solution of selected boundary value problems.
The problem is formulated in terms of a frequency
determinant for determination of the eigen-
functions and the associated frequencies.
I

Experimental Setup 1- ?1>.85Hi 70.56Hz

This section describes the experimental system usedO
in examining the model. The system consists of a
flexible arm with payload, DC torque motor with
servo-amp, signal conditioning withA/D conversion
for data acqusition, 16 bit computer system for
implementation of controlalgorithms,and
conversion for torque signal output.
/JOINT ANGLE
SENSOR
1 TORQUE MOTOR
/ BEAM
<\ S T R A I N GAGES
\ PAYLOAD
MOUNTINGBASE
L-
F r e q u e n c y Hz 100
Frequency responsef o r clamped beam
figure 3
D/A The vibratory modes were additionalg calculated by
the frequency determinant described in AppendixB.
Table 2 compares the measured modal frequencies to
those computed using the Bernoulli-Euler Beam.The
application of the Bernoullid-Euler formulation to
for the "clamped massTf case agrees very well with
themeasuredfrequencies,however,the"pinned-
/
mass" conditions were not as accurate.
Comparison of Modal Frequencies(Hz)
Table 2

Pinned-Mass
Clamped-Mass
Flexible Beam Apparatus Mode
Measured
Calculated Measured
Calculated
Figure 2
The processor is equipped for hardware computation 12.096 2.08 10.01 9.732
of floating point operations with a characteristic 2 13.389
13.92 31.608
33.45
time for32 bit multiplicationsof 19 microseconds. 3 41.324
41.38 62.683
70.56
A torque motor is driven by a high internal DCgain 4 a1 .la148.76881.225
servo-amp configured with a sense resistor on the
216.0485 136.352

1025
 The the poorer agreement for the pinned case is presented in the figure, while not used in the
attributedtothefrictionfoundinthejoint controller,itprovidesanindicationofthe
hardware, this is a difficult condition to model relative modal amplitudes.
andmayhavesignificanteffect for thesmall
amplitude motions used during the tests.

The next step checked the model generation

algorithm. Normalization
of
the
modal
masses
allowssthechecking of thecomputations
by
examining the diagonal componentsof the stiffness
matrix. The stiffnesses should be the square of the
modal frequencies input to the process.

The algorithm was checked for both the clamped-mass

modes and the pinned-mass modes. Table 3 presents
a comparison of the modal frequencies input to the
algorithm to the square root of the stiffnesses.
The results are very good, however it was necessary
to use higher precision computations for the higher
~ ~ ~ " I ~ ' " I ' ' " ~ " ~ ~ ~ ' " ' " ~
modes. 0 0.2 0.4 0.6 0.8 1 12 t4
Time (seconds)
Comparison of Frequencies Determined Measured Step Response
by Stiffness Computations Figure 4

Clamped-mass
Input Stiffness.
Table 3
,2
Input
Pinned-Mass
Stiffness
,,2 The dynamic model was discretized, and simulated
for the step angle change. Small amountsof damping
2.096Hz 2.096 9.732Hz 9.732 based on frequencyresponsemeasurementswere
13.989 13.989 31.608 31 .608 introduced into the model for the flexible modes.
40.552 40.524 62.683 62.683 Additionally, a simple model for hysteretic joint
81 .225 81.225 148.768 148.768 friction was included in the digital simulation.
216.048
136.344
136.352 214.621 Inclusion of modal damping and hysteresis in the
simulations improved the agreement of the models
Dynamic Response Comparison especially in the time interval after the large
The previous section provides confidence that the initial transients had occured.
beam parameters have been properly identified and
modeled by the Bernoulli-Euler beam. The Figure 5 shows the resultsfor a model implemented
computational
procedure
has
additionally
been withfiveclamped-massmodes,whilefigure 6
checked. The major questions concerning the model presents a model using two clamped-mass modes.The
can now be investigated: last case simulated used five pinned-mass modesas
inputs to the modeling process. This is presented
Choosing the Modal Candidates in figure7.
- Required Model Order
Is a Linear Model
of the Coupling Adequate

The following paragraphs describe simulations and

experimentsconductedtogaininsightintothe
answer to these questions.

The simplest and best understood controller for

flexible arms is a collocated controller, thatis,
a control system where the measurement and
actuation is located at the same point.A
collocatedcontrollerwasimplemented for the
experimental system which applied a position gain
to joint angle measurements, and a rate gain to
angular velocity measurements.

The position gain was selected to provide the rigid 0 0.2 0.4 0.6 0.8 1 t2 1.4
body mode with a characteristic time of one second, Time (seconds)
and a rate gain providing a damping ratio of 0.7 Simulated Response- Five Clamped-Mass Modes
was selected. Higher gains could be selected which Figure 5
stress the impact of flexibility on the control
strategy.
-. . however, the chosen gains provide a good The simulations based upon clamped-mass modes agree
starting point wellwithin the operating parameters the best with.measured responses. Suprisingly the
of the system. model implemented with only two clamped-mass modes
agrees almost as well if not better than the higher
Figure 4 displays the measured response of the ordermodel. This presentationmaybesomewhat
experimental system to a step change in desired misleading as better determination of damping for
angle.Thestrain atthebeambase is also the higher modes could provide better results. It

1026
is apparent that a dominant portion of the response APPENDIX A
is adequately characterized by
as few as two modes. FORMULATION OF DYNAMIC EQUATIONS

This section describes the generation of a dynamic

model via application of Lagrange's equations to
the flexible system[AI,A2]. The first step in this
process isto selecta suitable setof coordinates.
Theapproachutilizedselectsonerigidbody
coordinate associated with the joint rotation, and
flexible transverse displacements Prom a set of
axes attached to the joint. This is depicted in
figure 1. Then a position vector R to every point
of the systemcan be can be constructed;

where i,j are unit vectors in the

x, y directions.
The absolute velocity of the position vector;
Simulated Response- Two Clamped-Mass Modes = s+ixR (A.2a)
Figure 6 at

i = xi
&(x,t)j+ + i(xj-w(x,t)i)
(A.2b)
at
The kinetic energy of the system KE, can then be
computed by integrating this expression over the
entire mass of the flexible system Ms;
(A.3a)
KE = I/2/ff-$dm

1;.
0
, ,
0.2 0.4 0.6
-
, ,, ,,,,,.,,, ,,, ,, ,,, ,,,,,
0.8
S T R A I N A I BASE

1 14
I , ,

12
, .,,.
Nextintroducetheassumedmodeseriesrep-
Time (seconds) resentation for the transverse deflections
w(x,t);
-
Simulated Response Five Pinned-Mass Modes
Figure 7

Summary and Future Work where 5 = x/Lanormalizedlengthvariable.

A modelling Drocess to generate a linear model for Substitution for the transverse deflections results
use
in
controllingFlexible
manipulators
was in;
presented,
and compared to experimental
measurementsforaposition,andratefeedback
controller. The model agreed favorably with the
measured response for a selection of clamped-mass
assumed modes. The dominant parts of the transient
response were chararcterized by inclusion of as few (A.5)
as two assumed flexible modes.
This integral can be separated into three integrals
The material collected here is part of a more over the primary components of the beam, joint
comprehensive effort focusing on the control of mass,
beam
mass,and
payload.
Evaluation of
flexible manipulators using feedback based on the equation A.5 over the joint mass results in;
modal variables, as well as joint position and
velocity
data. The
selection of
appropriate
assumed modes must consider the feedback law, as
the applied torque dominates the boundary condition
at the baseof the beam. Clamped-mass modes yielded Evaluation of equation (A.5) over the mass of the
good results for the simple collocated controller,beam results in;
however this may not prove true f o r more sophis-
ticated controllers.

Current work examines the impact of optimal control

lawsonthemodelaccuracy,andmethods to
integrate the feedback laws into the modal
candidate selection. Notice that in evaluating;

1027
 rotary inertia of the cross section was neglected M =
and
the
squared
flexible
deflections
assumed negligible compared to the axial dimension
squared. Finally the integral is evaluated over the
mass oPthe payload as;

Next it is convenient to introduce an ortho-normal

condition on the spatial mode functions.

K =

1/25 Cd~i(l)dJli(t)Cd$.(l)dJl.(t) = I , for i=j

de dt cJ
dtJ = 0 , for i=j
(A.8)
Thepotentialenergy, PE, f o r thesystemis This systemiseasilyorganizedinto a linear
evaluated by the following integral expression; state-space modelas shownin equation6.

Appendix B
Bernoulli-Euler Beam Equations

Applying the orthogonality condition on the mode Thissectiondescribes thedevelopementof a

functions, and substituting the normalized length frequencydeterminantfromBernoulli-Eulerbeam
variable yields; theory which was used to derive candidate mode
frequencies and the associated shapes. The
PE = 1/29 l[&i(S)]2$2it)dc (A.lO) homogeneousdifferentialequation is presented
L3' dS2 first, followed by a discussion of the boundary
conditions utilized. Lastly the frequency
Notation in the following sections can be greatly determinant is derived.
simplified if the following definitions are made
for a "modal stiffness", Ki, for equation (A.10), Differential Equation
and"momentofmodalmass", Wi, forthelast The transverse displacement of the beam, w(x,t),
integrand in equation (A.5b)as; shown in figure A-1 is a function of both the
spatial variable along the beam, andtime.
(A.11 ) Following the analysis attributed to Bernoulli and
Euler gives rise to following fourth order partial
differential equation.
Wi = pAb2L$i!t)lS$i(S)dS
(A.12)
Then the kinetic energy for the .system can be
expressed as;
where : t=x/L
KE =1/2i2[J +J +M LZ]+QldJl.(t)[Wi+LMp$(l ) +
O P P dtl The next step applies the separability of equation
(A.13)
)]+1/2~[~(t)]*
LJp$(l (1) to obtain the following result;
dt

To form the dynamic equation we form the Lagrangian

of the energy expressions;
Searchif&for periodic time functions of the form
(A.14) $(t)- e leads to the following formulation;

-pLrA$(S)w2]@(t) = 0 (B.3a)
Where the qi are the coordinates, and Q. represents
the work done by the input torque at the joint by
each coordinate. The resultant equations can then This implies that the term in brackets must be
be organized in matrix form; equal .to zero for all time t. This is expressed as:

1028
 The second boundary condition balancing the moment
at the joint forms a relation between three of the
of the constants:
where the new parameter,
2DJ0 =-pAbL3B3(A+C) (B.12)
B"= pA L*w2
-b
E1 The shear force balance at the payload relates all
the constants of the solution:
has been substituted.
A(M BsinB-cosB)+B(sinB+M BcosB) +
$(e):
This is readily solved for
C(M
P n - m = 0 (B.13)
@ ( ~ ) = A s i n ( ~ ~ ) + B c o s ( ~ ~ ) + C s i n h ( ~ ~ ) + D c o s (B.5)
h(B~~

Boundary Conditions The moment balance at the payload forms a similar

The solution for the spatial mode function $(c) relation;
requires four independent boundary conditions be
provided. The first and most obvious results from
noting that there cannot be transverse displacement
at the pinnedjoint, this takes the form;

$ ( E ) = 0 , for 5 = 0 (B.6)
A i
b
Y-
AbL
'
C(sinh8-J B3coghB)+D(-J B3ginhB+coshB) = 0 (B.14)

Theexpressions (B.ll-B.14)involveonlythe
The second condition is provided from a moment constants from the solution for the mode function
balance at joint, this is expressed as; and the parameter B. This can be configured in
matrix form as:

wherethefollowingsubstitutionwasmadeto
eliminated the dependence on the frequency;
F =sinp+sinhB+ J ,B3(cosg-coshB) (8.16)
11 P
F12=sinB-sinhB+2coshB+M * B( cosB-cosh5)+-B
P JO*B3 (B.17)
Jo*6
Using this boundary condition results in pinned P*B(sinB-sinhB)
F21=cosB+cosh€i-M (8.18)
mode shapes for small joint inertias, additionally
clamped mode shapes can be determined by inputing F22=-c~~B-co~hB+2sinhB+J
verylargejointinertias. This providesmore P*B'(sinhB-sinB)-
J,*B
programming versatility than supplying one (B.16)
formulation of the frequency determinant for each
type of boundary condition.
Where the starred subscripts indicate modification
The third boundary condition is derived by by the appropriate area and length terms. rootsThe
resolving the shear force at the end of the beam of the frequency determinant det[F(Bi)]' = 0 yield
against the inertial forces of the payload mass. the characteristic values for the mode functions
This takes the following form; $i(S), and associated frequncies wi.

Table of Symbols
-
J- Joint Inertia J -Payload Inertia
MU -Payload Mass -'E Modhus of Elasticity
The last boundary condition arises from a moment :A -Beam Area I - Beam Area Inertia
of the
balance against the angular inertial forces L - Beam Length p - Density per Length
payload.
Bibliography
1. C. Hastings, W. Book,"Experiments in the Optimal
(B.lO)
Control of a Flexible Manipulatorf1, Proceedings
ACC, Summer 1985.
Frequency Determinant L.2. Meirovitch,I!Elements Vibration
of
Applicationoftheboundaryconditionstothe Analysis",Mc-Craw Hill, 1975.
solutionfor $(E) willresultinafrequency Al.J. Neto,"AutomaticControl of aVibrating
determinant for the eigen-values 0. Application of TransverseBeam1I,MS Thesis,MIT,January 1972.
the
first
boundary
condition
for
transverse A2. V. Sangveraphunsiri,"The Optimal Control and
displacementatthejointrelatestwo of the Design of a Flexible Manipulator Arm", PhD Thesis,
constants in the solution; Georgia Institiute of Tech., May 1984.

(B.11) B = -D

1029