Application of Direct Methods in Optimal Path

Planning of Redundant Cooperative Robots

Ali Hosseini and Mehdi Keskmiri
Department of Mechanical Engineering and
Center of Excellence for Offshore Engineering
Isfahan University of Technology
Isfahan, Iran
E-mail: Mkeshmiri@istt.org
H. R. Marzban
Department of Mathematics Science
Isfahan University of Technology
Isfahan, Iran
E-mail: hmarzban@cc.iut.ac.ir


AbstractThis paper introduces a numerical technique for
solving the optimal path planning problem of redundant
cooperative manipulators. The technique is based on
converting Abstract Optimization Problems to Parametric
Optimization Problems using Direct Methods. In this
approach, the state and control variables are expanded in
finite numbers of Legendre Polynomials and the kinematics
of cooperative system as well as performance index of
optimization problem are approximated. Application of this
method results in the transformation of nonlinear differential
and integral expressions into nonlinear algebraic expressions.
The optimal states are then determined by solving a set of
nonlinear equations in Legendre coefficients. The approach
is applied to a cooperative manipulator system consisting of
two 3-DOF serial manipulators jointly carrying an object.
The results are compared with those obtained from solving a
Two Point Boundary Value Problem, as an abstract
optimization approach.
Keywords- Direct Methods; Path Planning; Redundant
Cooperative Manipulators; Parametric Optimization; Legendre
Polynomials.
I. INTRODUCTION
The capabilities of single robot arms are insufficient for
some types of complex tasks. Two or more cooperative
robot arms are, therefore, needed to extend the range of
potential applications and to handle complicated and
dexterous tasks skillfully. When two (or more) robots are
cooperating, (a) closed kinematic chain(s) will be
generated. Regarding the kind of object grasping, the
closed chain is usually redundant. Redundant manipulators
can achieve additional tasks by utilizing their degrees of
redundancy. However, as a result of kinematic redundancy,
path planning for redundant manipulators is complicated.
Path planning as redundancy resolution deals with selecting
an individual configuration in each instant of time among
all possible ones. The common idea in redundancy
resolution is that redundancy should be solved in such a
way that the mechanism optimizes a performance index of
the system while carrying out its given duty.
Let
n
q be the joint space coordinates of a
redundant cooperative robot system. Also,
m
X represents the variables of end effectors, which
jointly carry the object and is called end effector space in
this paper. This vector should not be confused with the
vector of object task space. In single robot systems, these
two are usually the same. It is worth noting that the number
n represents the summation of joint variables of all robots
involved, and is different from the number of DOF of the
system.
Kinematic function
m n
: F describes the
relationship of X and q variables as:
) ( ) ( t X q F = (1)
Therefore, the path planning problem can be stated as: Find
a path ) (t q in joint space in a manner that it moves all end
effectors along the specified path ) (t X in the end effector
space ]) , [ (
0 f
t t t .
Equation (1) does not have an easy solution since it
involves triangular nonlinear algebraic equations.
Additionally, due to redundancy, it is a set of under-
determined equations; i.e., the number of unknown
variables exceeds the number of equations ) ( n m < . This
equation can be stated at velocity level as
) ( ) ( t X q q J

= (2)
in which q F J = is the Jacobian matrix of cooperative
manipulator system.
Numerous studies on redundant manipulators have
involved instantaneous redundancy resolution at velocity
level using pseudo inverse of Jacobian matrix [1-3]. A lot
of effort has been put into finding the null space solution
including the approaches of least square joint velocities.
All the methods developed can only rely on local
information, which results in local optimal solutions. An
alternative method involves the optimization of an integral
performance index with the forward kinematics as
constraints. The integral is over the length of the path.
Therefore, the history of cost function is taken into
account, which yields a global optimal. This global method
is based on calculus of variations and leads to a set of
ordinary differential equations (ODEs) with split boundary
conditions (BCs). Accordingly, to obtain the optimal
solution, one should solve a boundary value problem [4-6].
0-7803-8463-6/04/$20.00 2004 IEEE
Proceedings of 2004 IEEE/RSJ International Conference on
Intelligent Robots and Systems
September 28 - October 2, 2004, Sendai, Japan
3619
The solution of this Two Point Boundary Value Problem
(TPBVP) is usually difficult, time consuming, and, in some
cases, not practical to obtain due to iterative methods used
in solving the problem.
Another approach is to find the optimal path by solving
parametric optimization problem extracted from abstract
optimization problem. In these approaches, rather than
solving the differential equations, the optimal path is
approximated by polynomial of finite dimensions and
numerical schemes are used to solve for unknown
coefficients [7-9].
Utilizing orthogonal functions and polynomial series
have received considerable attention in dealing with
various problems of dynamic systems. Typical examples
are Fourier, Taylor, and Herimite Series as well as
Chebyshev, Legendre, and Laguerre polynomials. The
main advantage of using orthogonal functions and
polynomial series is that the differential equations are
reduced to a set of algebraic equations and there is no need
to solve Boundary Value Problems. Therefore, the optimal
(or indeed, the near optimal) solution is directly extracted
by solving algebraic equations. Consequently, these
approximation algorithms are known as Direct Methods
[10].
In this paper, using Direct Methods, the optimal path
planning problem of cooperative robots in joint space is
transformed into a parametric optimization problem. To
this aim, the Legendre polynomials are employed and, via
an appropriate algorithm, the kinematics of the system and
the supposed performance index of optimization are
restated in undetermined coefficients of Legendre
polynomials. These undetermined coefficients are
determined utilizing parametric optimization techniques
and finally, the near optimal paths are extracted.
II. DIRECT METHODS
The key idea of this technique is based on the following
integral property of vector ) (x :
) ( ) (
0
x d
x
P =

(3)
where, the elements ) ( ..., ), ( ), (
1 0
x x x
s
of vector
) (x are orthogonal on a certain interval and P is a square
s s constant matrix, which is known as operational
matrix for integration. In this study, operational matrix for
Shifted Legendre polynomials is used to solve the
variational problems by means of direct methods.
A. Approximation Process:
The well-known Legendre polynomials
{ } ,... 2 , 1 , ) ( = k x L
k
, which are orthogonal in the interval
[-1,1] with the weight function of unity, satisfy the
recursion formula:
x x L x L
k x L
k
k
x L x
k
k
x L
k k k
= =
|
.
|

\
|
+
|
.
|

\
|
+
+
=
+
) ( , 1 ) (
1 , ) (
1
) (
1
1 2
) (
1 0
1 1
(4)
In order to use these polynomials on the time interval [0,1],
we define the so-called Shifted Legendre Polynomials by
introducing the change of variable 1 2 = t x which yields:
1 2 ) ( , 1 ) (
1 , ) (
1
) ( ) 1 2 (
1
1 2
) (
1 0
1 1
= =
|
.
|

\
|
+
|
.
|

\
|
+
+
=
+
t t L t L
k t L
k
k
t L t
k
k
t L
k k k
(5)
The property of orthogonality is given by:

=
+

l k
k
l k
d L L
l k
;
1 2
1
; 0
) ( ) (
1
0
(6)
A function ) (t f can be expressed in terms of shifted
Legendre polynomials as follows:
) ( ) (
0
t L t f
k
k
k

=
= (7)
where,
dt t L t f k
k k

+ =
1
0
) ( ) ( ) 1 2 ( (8)
To determine an unknown function ) (t f , one should
determine infinite unknown coefficients of
k
, which is
actually impossible. So, it is approximated by truncating
(7) up to its first p terms.
) ( ) ( ) (
0
t t L t f
T
k
p
k
k
L =

=
(9)
where, the shifted Legendre coefficient vector and the
shifted Legendre vector ) (t L are
T
p
] ,..., , [
1 0
= ,
T
p
t L t L t L t )] ( ),..., ( ), ( [ ) (
1 0
= L (10)
Using orthogonal functions for approximating ) (t f has
the advantage that by increasing the order of approximation
function, the unknown coefficients
k
do not change and
just new coefficients should be calculated and added.
B. Shifted Legendre Operational Matrix:
The shifted Legendre polynomials ) (t L
k
satisfy the
following recursion formula:
3620
| | 1 , ) ( ) (
) 1 2 ( 2
1
) (
1 1

+
=
+
k t L t L
k
t L
k k k

(11)
0 , ) 1 ( ) 0 ( , ) (
2
1
) (
1 0
= = k L t L t L
k
k

(12)
Integrating both sides of (11) from 0 to t and using (12), we
have:
| | 1 , ) ( ) (
) 1 2 ( 2
1
) (
1 1
0

+
=
+

k t L t L
k
d L
k k
t
k
(13)
Also from (12), we get:

| | | |
| | ) ( ) (
2
1
1 ) (
2
1
) 0 ( ) (
2
1
) (
1 0
0 0 0
0
0
t L t L
t L L t L d L
t
+ =
= =

(14)
Hence,
) ( ) (
0
t d
t
L H L

(15)
where,
(
(
(
(
(
(
(
(
(
(
(

=
0
1 2
1
0 0 0 0 0
1 2
1
0
1 2
1
0 0 0 0
0 0 0
5
1
0
5
1
0
0 0 0 0
3
1
0
3
1
0 0 0 0 0 1 1
2
1
p
p p
"
"
# # # # # # #
"
"
"
H
(16)
is a constant matrix and is called the Operational Matrix
for integrating shifted Legendre polynomials. This matrix
plays an important role in direct method for solving
variational problems and restates the indefinite integral of
Shifted Legendre polynomials in Shifted Legendre
polynomials.
C. Shifted Legendre Direct Method:
Consider the problem of finding the minimum of the
function

=
1
0
0
) , , ( P dt t x x f (17)
The necessary condition for ) (t x to minimize P is that it
should satisfy Euler-Lagrange equations:
0
0 0
= |
.
|

\
|

x
f
dt
d
x
f

(18)
with appropriate boundary conditions. This equation can be
integrated easily only for simple cases. Therefore,
numerical methods and direct methods have been used to
solve variational problems. Now a direct method using
Shifted Legendre Polynomials is developed as follows.
First, we assume ) (t x can be expressed as a truncated
shifted Legendre series as:
) ( ) ( ) (
0
t t L c t x
T
k
p
k
k
L c =

=
(19)
where
k
c 's are unknown as shifted Legendre coefficients
of ) (t x . So, ) (t x can be represented as

) ( ] 0 ,..., 0 , 0 ), 0 ( [ ) (
) 0 ( ) ( ) 0 ( ) ( ) (
0 0
t x t
x dt t x dt t x t x
T
t
T
t
L HL c
L c
+
+ + =


(20)
Considering the first row in (15), t can also be expressed
in terms of ) (t L as:
) ( ] 0 ,..., 0 ,
2
1
,
2
1
[ t t
T
L = (21)
Substituting (19), (20), and (21) in (17), we earn P as a
function of c . Hence, minimizing P requires:
p k ,..., 1 , 0 , 0
c
P
k
= =

(22)
In the other words, instead of solving the differential
equation (18) which results from abstract optimization
techniques it will be enough if one solves the algebraic
equations (22) which result from parametric optimization
algorithms. When the
k
c 's are obtained, the function ) (t x
is determined accordingly.
It the case where
0
f is also a function of control law u ,
one can approximate it by defining an unknown vector
b corresponding to ) (t L
) ( ) ( t t u
T
L b = (23)
Besides, in cases where an optimization problem is
associated with constraints 0 ) , ( = t x g , the constraints can
be restated in vector c through utilizing the approximation
of ) (t x and then appended to index ) , ( P b c using
Lagrange's multipliers:
3621
) ( ) , ( P ) , , ( P
a
c g b c b c
T
+ = (24)
and the following sets of algebraic equations must be
solved:
0

0
b
0
c
=

a a a
P
,
P
,
P
(25)
Employing the Equation Approximation technique,
different kinds of constraints can be considered such as
differential, integral, or inequality constraints. In all cases,
the solution procedure is the same. Only the number of
unknown coefficients will be different. The important point
to be noticed is that when using direct methods, the
resulting optimum is in fact a near optimal solution, rather
than an exact solution. The higher the order of
approximation functions, the closer the near optimum to its
exact solution.
III. APPLICATION TO OPTIMAL PATH PLANNING
PROBLEM
In this section, we consider the optimization problem of
path planning in joint space with time integral of norm two
of velocity vector as the performance index of the system
and the kinematic equations of the cooperating system as
the constraints of the problem. This problem can be stated
as:

0 X q f q g
q q
= =
=

) ( ) ( ) , ( Subject to
P Minimize
0
t t
dt
f
t
t
T

(26)
where,
n
q and
m
X .
Assume that the path of each joint can be expressed
with a polynomial of order p. ) (
~
t q
i

is defined as the
approximation of the function ) (t q
i
with the first 1 + p
terms of shifted Legendre polynomials in the interval of
[0, 1].
) ( ) (
~
t t q
T
i i
L c =

(27)
where,
i
c is the vector of unknown coefficients of the i
th

joint corresponding to the vector of the shifted Legendre
polynomials ) (t L .

T
ip i i i
c c c ] ,..., , [
1 0
= c (28)
Defining ) 0 (
i i
q d = , the function ) (
~
t q
i
is expressed as
) ( ] 0 ,..., 0 , 0 , [ ) ( ) 0 ( ) ( ) (
~
0
t d t q dt t t q
i
T
i i
t
T
i i
L HL c L c + + =

(29)
Substituting ) (
~
t q
i

in index P:

= = =
= =
n
i
i
T
i
n
i
i
T T
i
n
i
i
dt t t dt q
1 1
1
0
1
0
1
2
) ( ) ( P Dc c c L L c (30)
where, the matrix

=
1
0
) ( ) ( dt t t
T
L L D is defined as:

+
= ]
1 2
1
,...,
5
1
,
3
1
, 1 [
p
diag D (31)
with regard to orthogonality of (6). Accordingly, the index
P is approximated and expressed in constant parameters of
i
c and
i
d .
In order to approximate the kinematic equations of the
system as the equality constraints of the optimization
problem, we should also express the vector of path ) (t X in
end effector space in terms of shifted Legendre
polynomials up to order p.
) ( ) ( X
~
i
t t
T
i
L a = (32)
with
T
ip i i i
a a a ] ,..., , [
1 0
= a as the vector of constant
coefficients, which is calculated from a relation similar to
(8). The constraint equations are rewritten in terms of
i
c 's,
i
d and ) (t L by substituting
i
q
~
's and
i
X
~
's in equation
) ( ) ( t X q f = . If the left hand side of the kinematic equation
can be stated as a linear combination of components of
vector L , the vector ) (t L can be omitted on both sides of
the constraint equations with regard to the orthogonality of
Legendre polynomials such that these equations are
restated only in terms of the constant parameters
i
c and
i
d .
The presence of nonlinear sine and cosine terms in the
kinematic equations makes it difficult to easily eliminate
) (t L on both sides of the equation. Therefore, a different
technique should be applied to remove t from constraint
equations. One possible solution is to evaluate kinematic
equations in definite nodes. A good suggestion to choose
appropriate nodes is Legendre-Gauss-Lobatto nodes [11],
which have a fixed value at the first and last nodes and are,
therefore, best fit to solving BVP. These nodes are given
by:
0
0
= t , 1 =
p
t and
m
t , 1 1 p m , the zeros of ) (t L
p

Employing these 1 + p nodes, the constraint equations
are represented by 1 + p sets of equations with constant
coefficients of
i
c and
i
d with no term " t ":
0 c c c g = ) ,..., , , ,..., , (
~
2 1 2 1 n n k
d d d (33)
with
m
k
g
~
. Appending these constraints to the
performance index by a set of undetermined vectors of
Lagrange's multipliers, we define
3622

k
p
k
k a
g
~
P
~
P
~
1
1

+
=
+ = (34)
where,
m
k
and 1 1 + p k . The necessary
conditions for the extremum are given by:

1 ,..., 1 ,
P
~
,..., 1 ,
P
~
,
P
~
a
a a
+ = =

= =

p k
n i
d
k
i i
0

0 0
c
(35)
Equations (35) are ) 1 ( ) 2 ( + + + p m p n nonlinear
equations and can be solved for the unknown
i
c ,
i
d ,
and
k
; subsequently, functions ) (
~
t q
i
are determined.
IV. SIMULATION RESULTS
In this section, the performance of the proposed method
will be evaluated through two simulations. Two different
manipulation systems have been tried: a 3-DOF planar
single system and a 5-DOF cooperative system.
A. 3-DOF Single System
First, we consider a 3-DOF planar manipulator shown
in Fig. 1 with arms l
1
, l
2
, and l
3
of 0.5, 0.4, and 0.3 (m),
respectively. The trajectory of the end effector is known in
Cartesian space as:
] 1 , 0 [ , )] cos( 1 [
4
1
) ( , 4 . 0 ) ( = = t t t y t x (36)
and so, the system has one degree of redundancy (DOR).
q1
q2
q3
m1 , l1
m2 , l2
m3 , l3

Figure 1. 3-DOF single robot system
The path ) (t q
i
is approximated using Shifted Legendre
polynomials and, consequently, the equations obtained
from (35) are solved for undetermined coefficients of
i
c 's
and
i
d . The results for , 2 = p 3 = p and 4 = p are reported
in Table I. These results are compared with the exact
solution of the problem by solving the corresponding
TPBVP [6]. Error of near optimal solution in joint space
using DM is compared with the exact optimum obtained
from TPBVP and is shown in Fig. 2. It could be observed
that the results tend toward the exact results as we proceed
to higher order approximations and, hence, the end effector
tracks its desired path more accurately.
TABLE I. COMPARISION OF THE RESULTS OF DIRECT METHOD
WITH TWO POINT BOUNDARY VALUE PROBLEM (3-DOF SINGLE ROBOT )
p No. of Unknowns
Near Optimal P
(DM)
Optimal P
(TPBVP)
2 18 1.2643
3 23 1.2438
4 28 1.2736
1.2928

0 0.2 0.4 0.6 0.8 1
-0.1
-0.08
-0.06
-0.04
-0.02
0
0.02
0.04
0.06
0.08
q
1
e
r
r
o
r

(
r
a
d
.
)
Time (Sec.)
p = 2
p = 3
p = 4
0 0.2 0.4 0.6 0.8 1
-0.06
-0.05
-0.04
-0.03
-0.02
-0.01
0
0.01
0.02
q
2
e
r
r
o
r

(
r
a
d
.
)
Time (Sec.)
p = 2
p = 3
p = 4
0 0.2 0.4 0.6 0.8 1
-0.06
-0.04
-0.02
0
0.02
0.04
0.06
0.08
q
3
e
r
r
o
r

(
r
a
d
.
)
Time (Sec.)
p = 2
p = 4
p = 3

Figure 2. Optimal Joint Space Errors of DM Solution with Respect to
TPBVP (3-DOF single robot system).
B. 5-DOF Cooperative System
Optimal joint space variables are generated employing
a cooperative robot system (Fig. 3), which consists of two
3-DOF serial robots carrying a common object with the
following characteristics:
l
1
= l
4
= 0.5 (m) , l
2
= l
5
= 0.4 (m) , l
3
= l
6
=0.3 (m)
The distance between the bases of the two robots = 1 (m)
The object is also modeled as a bar with a length of 0.2
(m). The trajectory of the mass center of the object is
known in Cartesian space as
T
)] ( ), ( ), ( [ t t y t x with respect
to the reference coordinate frame attached to the base of
Robot 1. is defined as the angle between the line
connecting the two end effectors, and the horizontal axis of
the reference coordinate frame. The object is connected to
each end effector by a revolute joint where ) , (
1 1
y x and
) , (
2 2
y x are the coordinates of the end effectors P
1
and P
2

in Fig. 3, respectively. The joint space variables are defined
as
T
t q t q t )] ( , ... , ) ( [ ) (
6 1
= q . Given the trajectory of the
mass center of the object, the task space of each serial
robot, i.e., ) , (
1 1
y x and ) , (
2 2
y x can be specified.
Hence, 6 = n and 4 = m . And so, it has 2 degrees of
redundancy. The mass center of the object is supposed to
move along the following trajectory:
] 1 , 0 [ , 0 ) ( , )] cos( 1 [
4
1
) ( , 5 . 0 ) ( = = = t t t t y t x (37)
3623
The velocity of each joint angle is approximated with a
shifted Legendre series of the order 3 = p . Equations (35)
lead to 46 triangular equations to be solved. Solving these
nonlinear equations for unknown coefficients, the near
optimal paths for each joint are extracted. Table II shows
the near optimal performance index obtained from DM and
its exact value which is extracted through solving the
corresponding TPBVP. Also, in Fig. 4, the history of each
joint angle obtained from DM is compared with its exact
path resulting from TPBVP. It is observed that a good
convergence is obtained. Obviously, the higher the order of
approximation, the closer the convergence of the near
optimal paths to their exact values, despite the requirement
for more complicated sets of nonlinear equations to be
solved. Therefore, the order of polynomials should be
chosen with regard to the accuracy required and,
consequently, a balance should be made between solution
accuracy and its computational cost.
q1
q2
q3
q4
q5
q6 Trajectory of object
Object
Robot 1 Robot 2
m2 , l2
m3 , l3
m1 , l1
m6 , l6
m5 , l5
m4 , l4
P1
P2

Figure 3. 5-DOF cooperative robot system
0 0.2 0.4 0.6 0.8 1
-0.7
-0.6
-0.5
-0.4
-0.3
-0.2
-0.1
0
0.1
q
1
(
r
a
d
.)
Time (Sec.)
Direct Method
TPBVP
0 0.2 0.4 0.6 0.8 1
3
3.1
3.2
3.3
3.4
3.5
3.6
3.7
3.8
3.9
q
4
(
r
a
d
.)
Time (Sec.)
Direct Method
TPBVP

0 0.2 0.4 0.6 0.8 1
1.05
1.1
1.15
1.2
1.25
1.3
1.35
1.4
q
2
(
r
a
d
.)
Time (Sec.)
Direct Method
TPBVP
0 0.2 0.4 0.6 0.8 1
-1.4
-1.35
-1.3
-1.25
-1.2
-1.15
-1.1
-1.05
q
5
(
r
a
d
.)
Time (Sec.)
Direct Method
TPBVP

0 0.2 0.4 0.6 0.8 1
1.6
1.7
1.8
1.9
2
2.1
2.2
q
3
(
r
a
d
.)
Time (Sec.)
Direct Method
TPBVP
0 0.2 0.4 0.6 0.8 1
-2.3
-2.2
-2.1
-2
-1.9
-1.8
-1.7
-1.6
q
6
(
r
a
d
.)
Time (Sec.)
Direct Method
TPBVP

Figure 4. Comparision of Optimal solutions of Joint Angles Using DM
and TPBVP Method (5-DOF Cooperative Manipulator System)
TABLE II. COPMRISION OF RESULTS (5-DOF COOPERATIVE
MANIPULATOR SYSTEM)
Optimal Value of P
p
Direct Method (near Optimal) TPBVP (Exact Optimal)
3 2.5820 2.5852

V. CONCLUTION
In this paper, a Direct Method based on Legendre
polynomials is utilized to obtain near optimal trajectories
for redundant cooperative manipulator systems. In the
proposed scheme, the abstract optimization problem is
converted to a parametric optimization problem. In this
manner, the differential equations of the two point
boundary value problem are transferred into nonlinear
algebraic equations and the need to solve TPBVP is
eliminated. The complexity of the presence of
trigonometric functions in the constraint equations is
relaxed satisfying the equations in Legendre-Gauss-
Lobotto nodes. The near optimal results have been
presented through numerical simulation for a system
containing two 3-DOF cooperating manipulators and are
compared with exact solution of the corresponding
TPBVP. The results show that desired accuracy can be
obtained using only four terms in the approximation
process.

REFERENCES
[1] D. E. Whitney, "Resolved motion rate control of manipulators and
human prostheses," IEEE Trans., on Man-Machine Sys., vol.
MMS-10, no. 2, pp. 4753, June 1969.
[2] A. Liegeois, "Automatic supervisory control of the configuration
and behavior of multibody mechanisms," IEEE Trans. Sys., Man,
Cybern., vol. SMC-7,no. 12, pp. 868871, December 1977.
[3] J. M. Hollrebach and K. C. Suh, "Redundancy resolution of
manipulators through torque optimization," IEEE J. Robotics and
Automation, RA-3, pp. 308316, August 1987.
[4] O. P. Agrawal and Y. Xu, "On the global optimum path planning
for redundant space manipulators," IEEE Trans. on Sys., Man,
Cybern., vol. 24, pp. 13061316, September 1994.
[5] K. Gotlih, "The use and comparison of a special global optimal
control algorithm with a local optimal control solution for
redundant mechanisms," Proc. IEEE Int. Symp. on Industrial
Electronics, vol. 2, pp. 872876, 1999.
[6] A. Hosseini and M. Keshmiri, "Optimal path planning of
cooperative robots under equality and inequality constraints," Proc.
ASME Int. Mech. Eng. Congress & Expo., Washington, D.C.,
November 2003.
[7] M. L. Nagurka and V. Yen, "Optimal design of robotic manipulator
trajectories," Technical Report CMU-RI-TR-87-12, The Rob.
Institute, Carnegie Mellon University, Pittsburgh, PA 15213, April
1987.
[8] H. Qiu, and H. Ozaki, "Expression and optimization of joint
trajectory of manipulator using Uniform B-Spline," IEEE Int. Conf.
on Sys., Man, Cybern., vol. 2, pp. 12671272, October 1995.
[9] V. Yen and M. Nagurka, "Linear quadratic optimal control via
Fourier_based state parameterization," Trans. ASME J. Dynmic
Sys., Meas., Control, vol. 113, pp. 206214, June 1991.
[10] H. Yen, F. Fahroo, and I. M. Ross, "Optimal feedback control laws
by Legendre Pseudospectral approximation," Proc. American
Control Conf., Arlington, VA, June 2001.
[11] C. Canuto, M. Y. Hussaini, A. Quarteroni, and T. A. Zang, Spectral
Methods in Fluid Dynamics, Addison-Springer-Verlag, 1988.
3624