2011 8th International Multi-Conference on Systems, Signals & Devices

OPTIMAL CONTROL OF AN HOLONOMIC SYSTEM:

FEEDFORWARD DESIGN AND LQR CONTROL FOR A “PICK AND
PLACE” SYSTEM

Asma Salah 1 and Abderrazak Chatti 2

Research in Automatic, Mechatronics and Intelligent Systems (RAMIS)
INSAT Centre Urbain Nord, BP N°676, 1080 Tunis, Cedex, Tunisie.
1
E-mail : [email protected]
2
E-mail : [email protected]
2
E-mail : [email protected]

ABSTRACT the system and one showed that these laws make it
possible to cancel the static error of the system. In second
The goal of this paper is to optimize the performances of place, we synthesized two types of feedforwards by
holonomic systems in particular “pick and place” calling upon the methods of inversion of models in
systems: time execution, precision and energy. The particular the concept of flatness of systems. Finally, we
command generation is divided into two main functions: had recourse to the techniques of optimal control in
The elaboration of a movement law for the system states particular LQR control to work out an indirect optimal
to control minimizing the time execution via the bang- order.
bang control, which is used as a reference for controls
(indirect control using the LQR approach); the definition 2. ARCHITECTURE OF THE OPTIMAL
of a feedforward transforming the movement law into a CONTROL
reference signal for the input of the physical system
(direct control). Control of a CNC axis or a Cartesian robot arm is
decomposed into three phases:
Index Terms— Holonomic systems, pick and i A time-dependent path planning stage which
place, optimal control, command generation, motion law, should, of course, take into account for
flatness, feedforward, contour tracking, LQR control. saturations, but, which will be shown later,
should be designed to minimize time execution
1. INTRODUCTION and also cancel undesirable vibrations.
i A feedforward  () transforming the
The increasing demand in terms of productivity leads the movement law into a reference signal for the
holonomic systems in particularly “pick and place” and input of the physical system (direct control).
high-dynamics systems to the maximum of their i A servo controller () which should help to
technological and structural capacities, invalidating the reject disturbances and to track the reference
classical assumption of a rigid body dynamics. trajectory (indirect control).
Consequently, the system control cannot neglect the
influence of the vibratory phenomena on the accuracy of
the contour tracking, but it has to modulate the command
efforts in order to realize the adequacy between the
intrinsic system characteristics and the speed and
accuracy criteria [1]. This paper aims at improving the
dynamic behavior of these machines by a direct action on
the command generation. The command generation is
divided into two main functions: the elaboration of a
movement law for the system states to control, which is
used as reference for controls (indirect control); the Fig. 1. Control strategy
definition of a feedforward transforming the movement
law into a reference signal for the input of the physical Where:
system (direct control). If correctly managed, these two x (t) : is the reference trajectory
ref

functions make it possible to dissociate the regulation and ε(t) : is the error tracking
the contour tracking problems. Indeed, one could u (t) : is the reference control
ref

generate through the order bang-bang, a law of movement δu(t) : is the signal of regulation
in optimal time by taking of account the energy limits of

978-1-4577-0411-6/11/$26.00 ©2011 IEEE

u(t) : is the control’s signal that is why one fixes the limit of the snap with this
x(t) : is the response of the system expression [7]:
 =   (2)
2π
T =
3. MODELING OF A SINGLE LINK FLEXIBLE ω#
JOINT
Where  is the amplitude of snap profile, $ is the
Single-mode models, such as two-mass spring models are amplitude of acceleration profile, %& is the resonant
widely employed to approximate the dynamic behavior of frequency,  is the duration of the jerk pulse (fig 3).
mechanical systems, [2, 3], and their relation to the
physical parameters depends mostly on the modeling
method. In Fig.2, basically, m1 accounts for the motor
modal mass, m2 accounts for the mass of the load. The
two masses are coupled with a spring stiffness Kt and an
internal damping μt. The effort F ordered by the variator
is transmitted by the engine to the moving part of the
axis. One also holds account frictions of the viscous type,
noted f1 and f2 on the level of guidance between the
mobile masses and the reference frame fixes associated
with the frame of the machine.

Fig. 3: Bang-bang profile with snap limited

5. FEEDFORWARD DESIGN

The finite-time transition between stationary set points of

Fig. 2. Two-mass spring model nonlinear SISO systems is considered as a scenario for
the presentation of a new design approach for inversion-
According to the figure 2, the equations of the model based feedforward control. Design techniques which are
are described by: based on a stable system inversion result in input
trajectories with pre- and/or post-actuation intervals. The
  main idea of this approach uses the concept of system’s
⎧
 =  −  − 
⎪  flatness to linearize the model of the system and one can
 =   consequently generate a feedforward which transforms
⎪
⎪ / =  −  the movement law into input for the physical system.
⎪ 1 _ One will generate two types of feedforward: one
= = / obtains the first when one considers the system as one
 
⎨ masses (rigid model) and the second when one consider
 =  /
⎪ (1) that our system is constituted by two mass (flexible
 = _ +  model) as it is presented in the previous paragraph.
⎪
⎪
  =  − 
   5.1. Rigid feedforward
⎪ 
⎩  = 
 
The feedforward is called rigid when the model is
constituted with one mass and it is represented in the
4. PATH PLANNING following figure:

The primordial criterion to optimize for the “pick and

place” systems is the time execution. The solution
involves “bang-bang” control-i.e, it requires that maximal
forward thrust should be applied up to the midpoint, then
an instantaneous switching to maximal reverse thrust.
Fig. 4: Rigid model of flexible axis
Bang-bang control is characteristic of least time
execution of trajectories by physical systems of bounded () , (3)
'̈ () = − '̇ () .
motive force.


In this paper, one aims not only at minimizing the
time execution but also at inducing the system vibrations, Where :
, is frictions of the viscous type
ms is the total mass
 x is the position of the load ∀k: 0 ≤ k ≤ n − 1, ℒ Q ℒRSh(x) = 0
F is the appliqued effort K around xJ
The feedforward corresponding to this model is given by (9)
ℒ Q ℒRSh(xJ ) ≠ 0,
the following expression:
WX
345 =
 '̈ 67 + , '̇ 67 (4) With ℒ ℎ(') = ∑&Z[ W7 Z (') is the derivative of h
Y

Where '̈ 67 8 '̇ 67 correspond respectively to the along f. Concretely, r is the number of derivations to
maximal acceleration and velocity. carry out on y before finding u.
If the relative degree of the output is equal to n, it is
5.2. Flexible feedforward possible to put the system in the following canonical
form:
? = \ ,
5.2.1. Flatness approach ⎧ \̇ = \,
The notion of flatness can be mathematically defined in a ..
differential algebra setting [4] or in the context of . (10)
⎨
& Z^
differential geometry of infinite jets and prolongations \̇
⎩ & = ℒ ℎ(') + ℒ ℒ
]  ℎ(')3,
[5]. In this section, a less formal, more intuitive definition ? (&) − ℒ&ℎ(')
of flatness is given. A finite-dimensional dynamic ⟹3= .
ℒ] ℒ&^ ℎ(')
system.
'̇ () = 9'(), 3();, '() ∈ ℝ& , 3() ∈ ℝ . (5)
Our aim here is to investigate some of the
Is called differentially flat, or simply flat, if there consequences of the above approach, which turned out to
exists a fictitious output ?() ∈ ℝ which satisfies the be so fruitful in the nonlinear context, for linear systems.
following conditions: It will be illustrated by a concrete case-study, i.e.,
The output ?() can be expressed as a function of the controlling a flexible axe and a cartesian robot.
system state '() and input u(t) and finitely many time
derivatives of the input. 5.2.1.2. Feedforward design of flexible axe
?() = @ A'(), 3(), 3̇ (), … 3 (B) ()C . (6) The flexible axe presented in the figure (2) is governed
Reversely, the system state and input can be by the following equations:
expressed as functions of the output y(t) and finitely
 '̈  () = b9' () − ' (); −  ('̇  ()
many of its time derivatives. − '̇  ()) −  '̇  () + () (11)
'() = D A?(), ?̇ (), . . , ? (E) ()C
 '̈  () = b9' () − ' (); −  ('̇  ()
(7) − '̇  ()) −  '̇  ()
3() = D A?(), ?̇ (), . . , ? (EF) ()C .
The fictitious output ?() is then called a flat output. This model can be put in the form given by (8).
It completely determines the dynamic behavior of system '̇ 
(5). If the optimal trajectory of the flat output (fast motion '
Where ' = c  e
which transits from point to point and cancel vibrations) '̇ 
is given, the trajectories of the entire system state '() '
and the system input 3() are determined by (6) and (7), ( F ) ^j  j
respectively. They can be computed without solving a − i 
i i i i
differential equation. Therefore, it is obvious that flatness ⎛ 1 0 0 0 ⎞
(') = ⎜  ⎟x
is property which facilitates the problem of trajectory j
−
(l F ) j
planning and design of feedforward control considerably. l l l l

Since the flat output constitutes an algebraic ⎝ 0 0 0 1⎠

parameterization of the system’s dynamics, flatness is 
also a particularly useful property for the solution of the sv

solution of dynamic optimization problems. and H(') = ⎛ 0 ⎞

0
⎝0⎠
5.2.1.1. Flatness of a non linear system
It is possible to call upon to the techniques of In fact, the flat output x corresponds to x  and the
linearization introduced into [11] to determine a flat feedforward is given by the following expression:

output. 345 = (jF 4) [z

 ' + { (
 +

Definition
A non linear system, single input, single output 
 )' +  (b(
 +
 ) +  ( + ) + (11)
defined by: 12'|8+1+2'|8]
'̇ = (') + H(')3 (8)
G
? = ℎ(') When x ~R corresponds to the trajectory reference, so
to the motion law of the mass’s position m2.
Have a relative degree in x J if and only if: From the expression of the feedforward given by the
previous equation, one can conclude that if we wish to
transit from point to point as soon as possible, the optimal
motion law in time and precision corresponds to the
saturation the fourth derivative of x and consequently to
the snap law.

5.3. Experimental results

5.3.1. Prototypes test

The experimental simulations are carried out in a first
time on a 3-axes robot (Fig. 8): it has been equipped with
a real-time "dSPACE 1103" control card. The available
measurements on the motor part come from the actuator
axis encoders. When the horizontal axis is only moving, it
will be assumed that the axis stiffness remains almost
constant (with an actual variation of 30 %) and the
system can be considered as linear.

Fig. 8: Overview of the first test setup prototype (stroke

[mm]:X-1000 Y-400 Z-800, max, federate:120 m.min-1, max,
acceleration: 4m.s-2, max)

We have used also, as a support of validation, a Fig. 10: Comparison of the different bang-bang
monoaxis system: it has been equipped with a linear profiles designed coupled to the adequate
motor LSP120C – P=12KW and a real-time “dSPACE feedforward (Cartesian robot)
1005” control card.
The simulation was undertaken for a displacement on
the X axis, with 'J = 0 to ' = 1000

Fig. 9: Overview of the second test setup prototype (stroke

[mm]:X-600, max, feederate:100 m.min-1, max, acceleration: 20
m.s-2, max, m1=260kg, m2=20kg)

5.3.2. Results
One can compare the differents feedforwards: the flexible
feedforward coupled to the snap law and the rigid
feedforward coupled to respectively the acceleration law
and the jerk law. The first feedforward is quite optimal in
term of precision respect to the second feedforward. In
fact, the flexible feedforward presents an error tracking
which is equal to 10-15m while the rigid feedforward
presents a maximal error greater than the previous.

Fig. 11: Comparison of the error tracking of the

different bang-bang profiles designed coupled to the
adequate feedforward (Cartesian robot)
 6. INDIRECT CONTROL: LQR APPROACH

The objective of this paragraph is to generate a control

which minimizes the error tracking (the static and
dynamic error) and the energy at the same time. To
guarantee these objectives, one used the principles of
optimal control in particularly the LQ approach.
The fig. 12 represents the structure of LQR control.
This approach consists in finding the coefficients of the
vector K of the buckled system which will condition the
response of the system consequence of the minimization
of a criterion fixed in advance.
xref represents the reference and kref a compensator of
setting in scale.

Fig. 14: the structure of the LQR control

In fact, this approach is based on the formalism of
state’s representation.
The state-space model of the process is the following :
0 1 0 0
Fig. 12: : Comparison of the different bang-bang profiles ⎡ b  +  b  ⎤
designed coupled to the adequate feedforward (system ⎢− − ⎥
monoaxis)




$=⎢ ⎥
⎢ 0 0 0 1 ⎥
⎢ b  b  +  ⎥
− −
⎣



 ⎦
0
⎡⎤ (12)
1 0 0 0
‡ = ⎢ i ⎥ ;  = ˆ ‰
⎢0⎥ 0 0 1 0
⎣0⎦
Where the state vector is:
‹ = ['56Z6] '̇ 56Z6] 'Œ6Ž '̇ Œ6Ž ] for the
cartesian robot.
‹ = [' '̇  'Œ6Ž '̇ Œ6Ž ] for the system
mono-axis.
In order to minimize the error tracking and the energy,
the criteria can be written by the following expression:
1 “
 =  (‘  ’ ‘ + 3  •3) (13)
2 ”
where:
‘ = A77̇–——˜.™š›^7 œ—™ C : Vector of position’s error and
–——˜.™š› ^7̇ œ—™
velocity’s error.
And since a stationary system is treated, the criterion
can thus be written following form:
1 ž
 =  [(8 − ‹) ’(8 − ‹) (14)
2 ”
+ 3  •3]]
Fig. 13: : Comparison of the tracking error of different bang-
bang profiles designed coupled to the adequate feedforward
And the optimal control is given by the following
(system monoaxis) expression:
3 ∗() = • ^ ‡ [ '() + b] (15)
 K is the solution of the Ricatti equation:
$ + $  + ‡• ^ ‡  −   ’ = 0 (16) 7. CONCLUSIONS
And k is the solution of the following algebraic
equation: A design of optimal control to the holonomic systems in
($ + ‡• ^ ‡ )b +   ’' = 0 (17) particularly the “pick and place” systems is a hard
mission. In fact, a classical control causes the dependence
of the regulation and the contour tracking problems. The
($ + ‡• ^ ‡ )b +   ’' = 0 (18) solution adopted in this paper is to dissociate these two
aspects and the control results of the sum of a
¾ Experimental results: feedforward that assure the contour tracking and indirect
The figures (15) and (16) represent the load’s control which reject the perturbances.
position and the error tracking respectively of the system In order to minimize the time execution, one used the
monoaxes and the cartesian robot consequence of LQR bang-bang profiles which indeed minimized the
control. One notes that the load’s position follows the vibrations of the system and with a feedforward adequate
trajectory reference (trajectory with snap limited in order to this profile, one minimized the error tracking. One also
to minimize the vibrations and the time execution). used the LQR approach to design an optimal control
which minimizes in the same time the error tracking and
energy.
8. REFERENCES

[1] Bearee, R., Barre P.J., Bloch S., Influence of high feed rate
machine tool control parameters on the contouring accuracy, J.
Intell. Robotic Systems, Vol. 40, 2004, pp 321-342.

[2] Clarke, D. W., Mohtadi, C., & Tuffs, P. S. “Generalized

predictive control. Part I: the basic algorithm. Part II:
Extensions and interpretations.” Automatica, 23, 137– 160 ,
1987.

[3] Colas, F., Dieulot, J.-Y., & Barre P.-J. “A Quantitative

Evaluation of the Non-Minimum Phase Phenomenon for a
Robot Arm”, IMACS-IEEE-SMC Conference on Computational
Engineering in Systems Applications, Beijing, 2006.

[4] Fliess, M., Lévine, J., Martin, P. et Rouchon, P. (1995).

Flatness and defect of non-linear systems: introductory theory
and examples. International Journal of Control, 61, pp. 1327-
Fig. 15 : The motor’s position and the error tracking of the 1361.
monoaxis system (with LQR control)
[5] Fliess, M., J. Lévine, P. Martin and P. Rouchon (1999). A
Lie-B cklund approach to equivalence and flatness of
nonlinear systems. IEEE Trans. Automatic Control 44, 922-937.

[6] Isidori, A. (1989). Nonlinear Control Systems. New York,

Springer.

[7] Lau M. A., Pao, L. Y. , Input shaping and time-optimal

control of flexible structures, Automatica, Vol. 39, 2003, pp
893-900

[8] Richalet, J., Estival, J.L., & Fiani, P. “Industrial applications

of predictive functional control to metallurgical industries”,
Proc. 4th IEEE Conf. Control Appl., 934 – 942, 1995.

[9] Rossiter, J. A. “Predictive functional control: more than one

way to prestabilise”. 15th IFAC World Congress. Barcelona,
Spain, 2002.

[10] Susanu, M., & Dumur, D. “Using Predictive Techniques

within CNC Machine Tools Feed Drives”, 44th IEEE Conf.
Decision and Control, 2005.

Fig. 14 : The motor’s position and the error tracking of the [11] Vivas, A., & Poignet, P. ‘Predictive functional control of a
cartesian robot (with LQR control) parallel robot’, Control Engineering Practice, 13, 863 – 874,
2005.