Académique Documents
Professionnel Documents
Culture Documents
AbstractOptimal design is an inevitable step for parallel manipulators. The formulated optimal design problems are generally
constrained, nonlinear, multimodal, and even without closed-form
analytical expressions. Numerical optimization algorithms are
thus applied to solve the problems. However, the optimization algorithms are usually chosen ad arbitrium. This paper aims to provide
a guideline to choose algorithms for optimal design problems.
Typical algorithms, the sequential quadratic programming (SQP)
with multiple initial points, the controlled random search (CRS),
the genetic algorithm (GA), the differential evolution (DE), and
the particle swarm optimization (PSO), are investigated in detail
for their convergence performances by using two canonical design
examples, the Delta robot and the GoughStewart platform. It is
shown that SQP with multiple initial points can be efficient for
simple design problems, while DE and PSO perform effectively
and steadily for all design problems. CRS can be used to generate
good initial points since it exhibits excellent convergence evolution
in the starting period.
Note to PractitionersNumerical optimization algorithms are
generally inevitable in solving optimal design problems of parallel
manipulators. Various algorithms have been applied in literature
and in engineering. This paper provides a thorough comparison on
convergence performance of typical optimization algorithms, SQP
with multiple initial points, CRS, GA, DE, and PSO. Two parallel
manipulators, the Delta robot and the GoughStewart platform,
are used as design examples by maximizing the effective regular
workspace. Computation shows that DE and PSO are good choices
for complicated optimal design problems, while SQP with multiple
initial points is superior for simple problems. CRS performs excellently in the starting period. It can be used to generate good initial
points.
Index TermsControlled random search (CRS), differential
evolution (DE), genetic algorithm (GA), optimal design, optimization algorithms, parallel manipulators, particle swarm
optimization (PSO), sequential quadratic programming (SQP).
Manuscript received December 17, 2012; accepted April 14, 2013. This paper
was recommended for publication by Associate Editor T. D. Murphey and Editor K. Lynch upon evaluation of the reviewers comments. This work was supported in part by the National Natural Science Foundation of China under Grant
51075085 and Grant U1134004 and in part by the Introduction of Innovative
R&D Team Program of Guangdong Province under Grant 2009010051.
Y. Lou, Y. Zhang, and R. Huang are with the School of Mechantronics Engineering and Automation, Harbin Institute of Technology Shenzhen Graduate
School, and the Shenzhen Key Lab for Advanced Motion Control and Modern
Automation Equipments, Shenzhen 518055, China (e-mail: louyj@hitsz.edu.
cn).
X. Chen is with the School of Mechatronics Engineering, Guang Dong University of Technology, Guangzhou 510006, China (e-mail: chenx@gdut.edu.
cn).
Z. Li is with the Department of Electronic and Computer Engineering, Hong
Kong University of Science and Technology, Hong Kong, China, and is also
with the DG-HUST Manufacturing Engineering Institute, Dongguan 523808,
China (e-mail: eezxli@ust.hk).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TASE.2013.2259817
I. INTRODUCTION
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
2
Hereafter,
denotes the set of design parameters of interest, where is the number of design parameters. The velocity
relation can in general be written as
(3)
where
ping joint rate
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
LOU et al.: OPTIMIZATION ALGORITHMS FOR KINEMATICALLY OPTIMAL DESIGN OF PARALLEL MANIPULATORS
D. Problem Formulation
Combining constraints (4)(10), the optimal design problem
for maximization of effective regular workspace is formulated
as following.
1) Problem 1: Optimal Mechanism Design: Find a set of
optimal design parameters such that
(11)
Fig. 1. The distance between two links.
(12)
(13)
(14)
segments is larger than the sum of corresponding radii. Fig. 1 depicts the distance between two spatial cylinder-modeled links.
The following inequalities ensure that no link interference will
occur at a point :
(15)
(16)
(17)
(8)
,
, and links and are not neighfor all
boring, which means there is no joint connecting link and link
. Here,
is the function computing distance between two line segments
and . Note
.
The set of points satisfying (6)(8) constitute the workspace
reachable by the resultant parallel manipulator. Therefore, any
point
should satisfy (6)(8).
C. The Manipulability Constraints
In order to guarantee the regular workspace to be effective,
i.e., the manipulator is able to conduct tasks effectively within
the regular workspace, constraints on the manipulability index
are introduced to characterize quality of the regular workspace.
A frequently used measure for manipulability is the inverse condition number of the kinematic Jacobian matrix, which is defined as
where
denotes the inverse condition number function of
matrices, and
and
its minimal and maximal
singular value functions, respectively. Thus,
. Here,
we treat separately orientation and position manipulability by
rewriting the differential kinematics (3) as
where and are linear velocity and angular velocity, respectively. Thus,
and
, respectively, give measures for
position and orientation manipulability. To guarantee position
and orientation manipulability, they are applied in design by imposing the following constraints:
(9)
(10)
and
are, respectively, lower bounds for position
where
and orientation manipulability, which are constants assigned according to practical design requirements.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
4
is applied to adaptively modify the standard deviation of the normal probability distribution for every
random variable in each iteration. It is actually the standard
deviation for the vector of random variables
. Therefore,
control comes by adjustment of the standard deviation of the
distribution, which explains the name of the method. Compared
with standard optimization techniques, the random variable
can be regarded as a search direction, while the standard
deviation serves as a kind of step-length, which is adjusted
automatically during the search in two situations.
(a) Each time a successful trial has been made. In this case,
standard deviations are set according to
,
, where
is a positive quantity describing the
distance between the variables current value
and the
nearest bound of the variable.
is a compression
factor to reduce search interval and maintain searches in
the neighborhood of the best previous point.
(b) After a specified number, typically 100, of consecutive
failure. Failure means that no improvement is made with
respect to the objective function. When this occurs, for
instance, as the optimum is approached, the standard deviations are reduced by
where
is a positive number.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
LOU et al.: OPTIMIZATION ALGORITHMS FOR KINEMATICALLY OPTIMAL DESIGN OF PARALLEL MANIPULATORS
5) Termination. Similar to GA, common terminating conditions include a sufficiently good fitness or a maximum
number of iterations (generations).
In the real computation, the code by Buehren was used [37].
E. The Particle Swarm Optimization (pso)
PSO is originally proposed in [38] and was first intended for
simulating social behavior. PSO does not use the gradient of the
problem being optimized either. It was modified and improved
by researchers, e.g., [39]. Let the position and velocity of th
particle be, respectively, and , and
and be, respectively, the best fitness value the th particle has achieved so far
and corresponding location. Assume that the best fitness value
achieved by any particle of the population is called
and
the corresponding location is denoted as . The fundamental
process for implementing of PSO is described as follows [40],
[41].
1) Initialization. A population array of particles with positions
and velocities are randomly generated on dimensions in
the search space.
2) Evaluation. Each particle is evaluated by the fitness function (the objective function) in variables.
3) Comparison and selection. The particles fitness value is
compared with particles
. If current value is better
than
, then set
value equal to the current value,
and the
location equal to the current location in
-dimensional space. Then, the particles fitness value is
compared with the populations overall previous best. If
current value is better than
, then reset
to the
current particles array index and value.
4) Adjustment. The velocity and position of the particle is
changed by the following equation:
(20)
5) Loop to step 2) until a criterion is met, usually a sufficiently
good fitness or a maximum number of iterations.
In the (20),
, where
is specified by
users.
and
are two acceleration constants weighting of
the random acceleration terms and the functions
and
are two random number generators uniformly distributed in
. They are randomly generated at each iteration
and for each particle.
In the real computation, the code by Sam was used [42].
F. Termination Criterion and Discussions
In the early years, gradient-based algorithms were the main
approach to solve optimal design problems of parallel manipulators. Since the algorithms could only find local minima, global
optimization algorithms were gradually introduced to deal with
the multimodal problems.
SQP is a deterministic and local optimization technique,
while DE, PSO, GA, and CRS are all probabilistic and global
optimization algorithms. It is necessary to use a suitable criterion to evaluate them. In order to better estimate an algorithms
ability to locate the true global optimum, a trial can be classified
as a success when the best objective function value reaches
a predetermined limit known as the value-to-reach, or VTR
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
6
where
Algorithmic Settings
The algorithms, DE, GA, PSO, CRS and SQP, were applied
to solve the optimization problem. The probability-based algorithms, DE, GA, PSO, and CRS, were tested five times. For all
algorithms, VTR was given 0.3701, which is a close estimate of
the global optimum 0.3702. The maximum allowable time
was set 3600 s, i.e., 1 h. For CRS, DE, GA, and PSO, the initial
feasible points were randomly generated in the feasible region.
The population sizes used in DE, GA, and PSO were identically
set 20, 10 times the number of optimization variables. Since algorithmic parameters play important roles in the convergence
performance, they were set according to the algorithm properties and the problem features as follows.
DE.
and
. Default values were used for
other algorithmic parameters.
GA. In the realization, the elite count was set 2, the
crossover fraction was set 0.8, the migration fraction was
set 0.2, and the interval was taken 20.
PSO. The acceleration constants
and
,
and
.
CRS. The compression factors
and
.
SQP. Since SQP is intrinsically a local optimization technique, a sufficiently large number of initial points should be
provided to locate the global optimum. For a constrained
nonlinear optimization problem, it is usually difficult, even
impossible, to determine which initial point will lead to the
global optimum by simple observation. Here, a uniform
discretization of the feasible region was used to generate
the set of initial points. Each discretized point was applied
as an initial point in the algorithm. The generation of initial points starts from a coarse discretization of the feasible
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
LOU et al.: OPTIMIZATION ALGORITHMS FOR KINEMATICALLY OPTIMAL DESIGN OF PARALLEL MANIPULATORS
TABLE I
COMPUTATION RESULTS OF OPTIMAL DESIGN OF THE DELTA ROBOT
Fig. 4. Computation evolution of DE, CRS, GA, and PSO: the Delta robot case.
TABLE II
COMPUTATION RESULTS WITH TERMINATION TIME 400 S
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
8
where
and
are the base ball joint axes at the current configuration and at the home position, respectively. The vectors
and
are the moving platform ball joint axes, respectively, at
the current configuration and at the home position. The functions
and
compute pivot angles of base ball joint and moving platform ball joint on the th
leg, respectively. We take
in the simulation.
Constraints due to leg interference are given as
where
represents the line segment of the th leg. Here, it is
assumed that the radius of the minimal cylinder enveloping a
leg is 0.02.
Combining the objective and constraints together, the optimal
design problem of a GoughStewart platform is formulated as
follows.
1) Problem 3: Optimal Design of the GoughStewart Platform: Find a set of optimal design parameters such that
(21)
(22)
(23)
(24)
(25)
(26)
(27)
and
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
LOU et al.: OPTIMIZATION ALGORITHMS FOR KINEMATICALLY OPTIMAL DESIGN OF PARALLEL MANIPULATORS
TABLE III
COMPUTATION RESULTS OF OPTIMAL DESIGN OF THE GOUGH-STEWART PLATFORM
TABLE IV
COMPUTATION RESULTS WITH TERMINATION TIME 60000 S
Fig. 6. Computation evolution of DE, CRS, GA, and PSO: the GoughStewart
platform case.
and DE became slow since they reached a small neighborhood of the true optimum. GA still ranked the last although
it got good improvement. Finally, the convergence rank is
.
It is also worthy noting that CRS converges very fast in the
starting period while it improves rather slow in the later periods.
Comparing the computation performances of the algorithms
in the two examples, it indicates that (1) for a simple design
problem (a simple mechanism and/or simple kinematics), SQP
with multiple initial points can be efficient. (2) For a complicated design problem, however, SQP becomes very inefficient.
DE and PSO are the best choices to solve such optimization
problems. CRS exhibits good convergence evolution performance in the starting period. It can be used to generate good
initial points for other algorithms.
VI. CONCLUSION
In this paper, the convergence performances of five typical
algorithms, CRS, GA, PSO, DE, and SQP with multiple initial
points, are evaluated by optimal design of two parallel manipulators, the Delta robot and the GoughStewart platform.
Conclusions are obtained based on the problem complexity.
(1) For a simple design problem (a simple mechanism and/or
simple kinematics), SQP with multiple initial points performs
best. (2) While for complicated design problems, SQP performs
the worst and DE and PSO are the best choices. Taking into
account the efficient convergence of CRS in the starting period,
we may combine it with other algorithm, e.g., DE or PSO. CRS
is used to generate good initial points while the other algorithm
is applied to continue later search.
REFERENCES
[1] C. A. Klein and B. E. Blaho, Dexterity measures for the design and
control of kinematically redundant manipulators, Int. J. Robot. Res.,
vol. 6, no. 2, pp. 7283, 1987.
[2] T. Yoshikawa, Manipulability of robotic mechanism, Int. J. Robot.
Res., vol. 4, no. 2, pp. 39, 1985.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
10
[3] C. Gosselin and J. Angeles, Singularity analysis of closed loop kinematic chains, IEEE Trans. Robot. Autom., vol. 6, pp. 281290, Jun.
1990.
[4] F. C. Park, Singularity analysis of closed kinematic chains, Trans.
ASME J. Mech. Design, vol. 121, no. 1, pp. 3238, 1999.
[5] G. Liu, Y. Lou, and Z. Li, Singularities of parallel manipulators: A
geometric treatment, IEEE Trans. Robot. Autom., vol. 19, no. 4, pp.
579594, Aug. 2003.
[6] D. Chablat and P. Wenger, Architecture optimization of a 3-dof
translational parallel mechanism for machining applications, the
orthoglide, IEEE Trans. Robot. Autom., vol. 19, no. 3, pp. 403410,
Jun. 2003.
[7] Y. Lou, D. Zhang, and Z. Li, Optimal design of a parallel machine
based on multiple criteria, in Proc. IEEE Int. Conf. Robot. Autom.,
2005, pp. 32193224.
[8] C. Gosselin, Stiffness mapping for parallel manipulators, IEEE
Trans. Robot. Autom., vol. 6, no. 3, pp. 377382, Jun. 1990.
[9] J.-P. Merlet, Jacobian, manipulability,condition number, and accuracy
of parallel robots, Trans. ASME J. Mech. Design, vol. 128, no. 1, pp.
199206, 2006.
[10] J. Meng, D. Zhang, and Z. Li, Accuracy analysis of parallel manipulators with joint clearance, Trans. ASME J. Mech. Design, vol. 131,
no. 1, pp. 011013-1011013-9, 2009.
[11] K. Zanganeh and J. Angeles, Kinematic isotropy and the optimum
design of parallel manipulators, Int. J. Robot. Res., vol. 16, no. 2, pp.
185197, 1997.
[12] J. A. Carretero, R. P. Podhorodeski, M. A. Nahon, and C. M. Gosselin, Kinematic analysis and optimization of a new three degree-offreedom spatial parallel manipulator, Trans. ASME J. Mech. Design,
vol. 122, no. 1, pp. 1724, 2000.
[13] S.-G. Kim and J. Ryu, New dimensionally homogeneous Jacobian
matrix formulation by three end-effector points for optimal design of
parallel manipulators, IEEE Trans. Robot. Autom., vol. 19, no. 4, pp.
731736, Aug. 2003.
[14] X.-J. Liu, J. Wang, and J. Kim, Determination of the link lengths for
a spatial 3-DOF parallel manipulator, Trans. ASME J. Mech. Design,
vol. 128.
[15] Y. J. Lou, G. F. Liu, and Z. X. Li, Randomized optimal design of
parallel manipulators, IEEE Trans. Autom. Sci. Eng., vol. 4, no. 2, pp.
625649, Apr. 2008.
[16] L. W. Tsai and S. Joshi, Kinematics and optimization of a spatial
3-upu parallel manipulator, ASME J. Mech. Design, vol. 122, no. 4,
pp. 439446, 2000.
[17] M. Stock and K. Miller, Optimal kinematic design of spatial parallel
manipulators: Application of linear delta robot, Trans. ASME J. Mech.
Design, vol. 125, no. 2, pp. 292301, 2003.
[18] S.-G. Kim and J. Ryu, Optimal design of 6 DOF parallel manipulator using three point coordinates, in Proc. IEEE/RSJ Int. Conf. Intell.
Robot. Syst., 2001, pp. 21782183.
[19] K.-S. Hong and J.-G. Kim, Manipulability analysis of a parallel machine tool: Application to optimal link length design, J. Robot. Syst.,
vol. 17, no. 8, pp. 403415, 2000.
[20] Y. X. Su, B. Y. Duan, B. Peng, and R. D. Nan, A real-coded genetic
optimal kinematic design of a Stewart fine tuning platform for a large
radio telescope, J. Robot. Syst., vol. 18, no. 9, pp. 507516, 2001.
[21] D. Zhang, Z. Xu, C. M. Mechefske, and F. Xi, Optimum design of
parallel kinematic toolheads with genetic algorithms, Robotica, vol.
22, no. 1, pp. 7784, 2004.
[22] J. Collard, P. Fisette, and P. Duysinx, Contribution to the optimization of closed-loop multibody systems: Application to parallel manipulators, Multibody Syst. Dynamics, vol. 13, no. 1, pp. 6984, 2005.
[23] N. M. Rao and K. M. Rao, Dimensional synthesis of a spatial 3-RPS
parallel manipulator for a prescribed range of motion of spherical
joints, Mech. Mach. Theory, vol. 44, no. 2, pp. 477486, 2009.
[24] P. S. Shiakolas, D. Koladiya, and J. Kebrle, On the optimum synthesis
of six-bar linkages using differential evolution and the goemetric centroid of precision positions technique, Mech. Mach. Theory, vol. 40,
no. 3, pp. 319335, 2005.
[25] R. R. Bulatovi and S. R. Dordeevi, On the optimum synthesis
of a four-bar linkage using differential evolution and method of
variable controlled deviations, Mech. Mach. Theory, vol. 44, no. 1,
pp. 235246, 2009.
[26] Q. Xu and Y. Li, Error analysis and optimal design of a class of
translational parallel kinematic machine using particle swarm optimization, Robotica, vol. 27, no. 1, pp. 6778, 2009.
[27] Y. Li and Q. Xu, Design and analysis of a totally decoupled flexurebased XY parallel micromanipulator, IEEE Trans. Robot., vol. 25, no.
3, pp. 645657, Jun. 2009.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.
LOU et al.: OPTIMIZATION ALGORITHMS FOR KINEMATICALLY OPTIMAL DESIGN OF PARALLEL MANIPULATORS
Xin Chen received the B.S. degree in manufacturing engineering from Changsha Railway Collage,
Changsha, China, in 1982, the M.Sc. degree in
manufacturing engineering from the Harbin Institute
of Technology, Harbin, China, in 1988, and the
Ph.D. degree in mechanical engineering from the
Huazhong University of Science and Technology,
Wuhan, China, in 1995.
He is a Professor with the School of Electromechanical Engineering, Guangdong University
of Technology, Guangzhou, China. His research
interests include manufacturing industrial informationlizing and collaborative
design, mechanical design theory and method, and microelectronic packaging
technology and equipment.
11