Robotica (2002) volume 20, pp. 139–147.

© 2002 Cambridge University Press

DOI: 10.1017/S0263574701003708 Printed in the United Kingdom

Optimal location of a robot path when considering velocity

performance
N. A. Aspragathos and S. Foussias
Mechanical and Aeronautics Engineering Department, University of Patras, Patras (Greece)
E-mail: [email protected]
(Received in Final Form: August 4, 2001)

SUMMARY In these cases, the shape of the path is known since it has
A method for searching the best location of a path in the been specified in the design of the manufactured part.
robot workspace considering the velocity performance of A lot of research efforts have been devoted to minimise
the robot is presented. After a thorough investigation of the the cycle time of robot tasks and this problem involved a
robot performance indices, a new measure for the velocity variation of formulations. In some cases,1–4 the minimisation
efficiency of a robot moving its end-effector along a path is of the cycle time is presented as a version of the well-known
introduced. This measure is an approximation of the Travelling Salesman Problem (TSP), where the task of the
minimum of the Manipulator Velocity Ratio (MVR) along robot is to visit a number of known fixed points. The end-
the path. The minimum is calculated from a finite set of effector of the robot can visit these points in any order, since
MVRs determined at specified points of the path using an there are tasks where the order is not critical such as the spot
algorithm for an approximate motion of the end effector welding. The TSP problem has been adapted to the
along the path. characteristics of the robots mainly by replacing the
The introduced measure is used as the objective function distance between the points with the time required to move
in an optimisation problem, where an optimal location of between two successive points in the configuration space of
the path is searched. The objective function is procedural the robot. Another characteristic taken into account is that
and non continuous, so a Genetic Algorithm is used to the time required to move from one point to the next is not
search the space for the optimal location of the path. The equal to the time required to move back.
proposed method is tested using a simulated PUMA-like In the design of a robotic workcell, the optimal position
robot, which has to move its end-effector along a straight- of the base of the robot has been studied by a number of
line segment. At the end of the paper, the results obtained in researchers.5,6 Dissanayake and Gal5 proposed a technique
these tests are presented and discussed. to optimise the location of the robot, the sequence of the
tasks and the configuration of the robot at each task in order
to minimise the cycle time. In another approach,6 the
1. INTRODUCTION optimal position of the robot is found by minimising the
The design of the layout of a robot workcell is a very distance between the space effectively reached by the robot
important engineering task, which is performed mostly by and the target space if no overlapping occurs between these
trial-and-error and intuition. Using CAD and/or Robot spaces. In the case of overlapping, the volume of the target
Simulation Systems, the designer can obtain considerable space not reached by the robot is minimised.
support in order to put the jigs and fixtures and locate the In the above two categories of methods for minimising
robot tasks in the workspace of the robot. These systems the cycle time, the capabilities of the robot such as the
offer powerful 3D graphics for a realistic representation of velocity efficiency were not taken into account. In some
the objects, collision detection and off-line programming, methods7–11 for trajectory planning or kinematic control of
but do not provide special tools for the determination of the redundant robots, indices or measures characterising the
optimal or, at least, sub-optimal location for a robot task. efficiency of the robot are taken into account. These
Therefore, the workcell designer moves the fixtures in trial measures are related to the Jacobian matrix, so they are
locations into the workcell and uses off-line programming configuration dependant. In some of these papers, an
searches for the best location, taking into account design integral cost criterion is optimised instead of an instanta-
criteria such as low cycle time, reduced wear and energy neous criterion as it has been used in the well-known
consumption, or avoiding obstacles and/or joint velocity methods based on extended or pseudo-inverse Jacobian. In
limits. It is well known, that the capabilities of the robot these papers, the location of the path in the Cartesian space
depend on the robot configuration; for example, the velocity is pre-specified, while the trajectory in the joint space is
efficiency or the mechanical advantage of the robot varies determined using the proposed methods.
considerably in the robot workspace. Therefore, in applica- The optimisation of the robot task location based on the
tions of trajectory planning and generation such as painting, manipulator efficiency attracted little attention of the robotic
welding, cutting (especially in laser cutting), it is very research community.12–14 An algorithm has been proposed
important to search for the location of the trajectory, where by Nelson and Donath12 for optimising the location of an
the velocity efficiency of the robot would be the maximum. assembly task in a manipulator’s workspace. In their

http://journals.cambridge.org Downloaded: 07 Apr 2011 IP address: 117.211.90.154

140 Optimal location of a robot path
algorithm, a gradient function of manipulability has been location of the path can be considered the location, where
developed based on an explicit determination of the the configurations of the robot along the path require the
manipulability function in joint space. The formulated lower joint velocities over the other locations of the path.
problem of optimisation has been solved by a modified This qualitative definition of the best location of a robot
steepest descent procedure, so the explicit calculation of path has to be translated to a quantitative definition of an
det(J(q)) and  | det(J(q)) | /q for a non redundant manip- optimisation problem.
ulator is required. Therefore, this method cannot be used For the definition of the optimisation problem it is
when the determinant of the Jacobian is non-analytic or necessary to define a velocity performance or velocity
non-continuous. Fardanesh and Rastegar13 proposed an efficiency measure along a path for robot manipulators. The
approach for optimal placement of a prescribed task in the proposed velocity efficiency measure has been defined after
workspace of a manipulator workspace. The optimisation a thorough investigation of the robot performance indices
criterion is the cycle time needed to visit a set of points with that appeared in the literature.
specified relative positions, while the location of the set of
points in Cartesian coordinates is not pre-specified but is
2.1. Review of the dexterity measures
determined by the proposed approach. Aspragathos14 pro-
It is well known that the end-effector velocity is related to
posed a method based on Genetic Algorithms to locate a
the joint velocity through the Jacobian matrix J, and the
continuous path in the location with the higher manipul-
proposed measures of the robot velocity efficiency are
ability. An optimisation problem is formulated where an
connected to the properties of this matrix. Yoshikawa15
approximation of the mean value of the manipulability is
introduced the following performance measure called
estimated using a bisection algorithm for a path described
manipulability
by a straight-line segment or by a 3D curve is calculated.
The optimisation problem has been formulated in a w = det J (1)
procedural way and the objective function is non-linear and for the case of non-redundant manipulators, while for the
non-continuous, so its solution is based on a simple Genetic redundant manipulators the manipulability index is given
Algorithm. by
The present paper is an extension of the procedure in
reference [14], since a new optimisation criterion is w = det(JJT) (2)
introduced to characterise the location of path where the Alternatively, the product of the singular values of the
velocity performance of the robot is the best. The proposed Jacobean matrix gives the manipulability index
criterion is based upon the orientation of the manipulability
ellipsoid relatively to the path. It is well known that if the w = 12 . . . n (3)
major axis of the manipulability ellipsoid is parallel to the where i are the singular values of the Jacobian matrix, and
path then the velocity performance of the manipulator is the n is the number of the degrees-of-freedom of the robot. The
best. Therefore, an objective function is formulated to manipulability index can be considered as a measure of the
represent the relative orientation of the manipulability distance between the manipulator configuration and the
ellipsoid to the path. This objective function is not analytical singular ones. It represents the average motion capabilities
but it is procedural and since the search space is quite of the robot at its configuration, where the manipulability
complicated a Genetic Algorithm is used to search for the index is determined. Using this as a criterion for the robot
optimum of the formulated objective function. velocity efficiency it cannot be concluded which direction of
the movement is the best, because it does not include any
2. DEFINITION OF THE OPTIMAL LOCATION OF information concerning the shape of the manipulability
A PATH ellipsoid.
Let us assume that the robot of an industrial workcell has When the manipulability index is equal to zero, the
been selected, and the shape of a path to be followed by the corresponding configuration is singular. However, this index
tip of the robot tool has been defined. This path could cannot be used safely as a practical measure of the Jacobian
represent the welding seam, the pattern to be followed in ill conditioning. Instead the matrix condition number has
glue applications or in cut processes. In these cases, it is been recommended for its accuracy of estimation. Based on
very important to reduce the motion time by exploiting the this remark, another dexterity measure is proposed,16 which
robot transmission capabilities at a maximum. Therefore, a is reciprocal to the condition number of the Jacobian
location of the path where the velocity performance of the matrix
robot is the best, has to be searched, before any minimum
time motion algorithm is applied. max
k(J) = (4)
From an engineering point of view, it is desired to locate min
a task into the workspace where the required joint velocities
are the lowest for a given velocity of the end-effector. In this where max, min are the maximum and minimum singular
case, the energy consumption and the wear of the robot values of the Jacobian matrix respectively. Since the
actuators are reduced. Alternatively, it is required to maximum and minimum singular values of the Jacobian
determine the location of a path, where the maximum end- matrix correspond to the major and minimum axis of the
effector velocity would be obtained under the constraints of manipulability ellipsoid, this dexterity measure gives an
the upper limits of the joint actuators. Therefore as the best indication of the shape of the ellipsoid. When the condition

http://journals.cambridge.org Downloaded: 07 Apr 2011 IP address: 117.211.90.154

Optimal location of a robot path 141
number is equal to unity, the robot is at “isotropic” points, If uv is a unit vector along ẋv , then Eq. (6) can be replaced
where the robot end-effector can move with the same by ¯the following relation for¯ the MRV:
performance to any direction. In other words, the ellipsoid is
transformed into a circle and it presents a directional 1
rv = (7)
uniformity. From the definition of the condition number, it u (Jv JvT )–1uv
T
v
can be concluded that the greater the condition number the ¯ ¯
closer the arm to a singular configuration. The condition It was proved that for a given configuration, the end points
number measures the roundness of the ellipsoid but does not of the vectors rv uv form the surface of an ellipsoid with
give any indication of its volume. principal axes 1¯u1, . . . , 1ui , . . . , nun, where ui is an
A third index is equal to the minimum singular value of eigenvector of the ¯Jv JvT , and that
¯ ¯ ¯
the Jacobian matrix17 and can be used to set an upper bound n ≤ rv ≤ 1 (8)
on the required joint velocities. The minimum singular
value dominates the variation of the manipulability index The above reviewed measures or indices depend on the
and the reciprocal of the condition number, since its value configuration of the robot therefore they characterise locally
changes more sharply close to singularities than the rest of the performance of the robot and can be used as an
the singular values. instantaneous or local criterion. Integral measures are
The Cartesian velocity of a body includes components of proposed7,8 to overcome the local and instantaneous nature
different nature, namely three elements for the translation of these indices. Martin et al.7 chose to minimise the
velocity and three elements for the angular velocity. following path integral:
Therefore the elements of the Jacobian matrix are of

 
different units and thus these indices could be meaningless.
Cosellin18 and Angeles19 proposed alternative definitions of tf 1 T –1
the Jacobian matrix to overcome this problem. Another q̇ W q̇ + g(q) dt (9)
t0 2¯ ¯ ¯
approach to resolve this problem is to normalise the
velocities of the joints and of the end-effector, which is
followed in this paper. subject to the constraints of the kinematic equations of the
The mutual orientation of the generalised kinematic robot:
ellipsoid (manipulability ellipsoid), the generalised dynamic x = f (q) (10)
ellipsoid and the generalised stiffness ellipsoid is given by ¯ ¯
the Hybrid Manipulator Measures (HMM) introduced by W is a diagonal matrix whose elements reflect the relative
Ermolov and Podurajev.20 The HMM are frame-invariant cost in terms of energy or time, of utilising each joint axis,
and can be used to optimise the trajectory placement within and g(q) is a function of the robot configuration such as the
¯
manipulability measure or the distance of some obstacle.
the workspace in the case of constrained motions.
Dubey and Luh21 used the Manipulator Velocity Ratio Uchiyama et al.8 proposed the following integral measure
(MVR) in control algorithms of redundant robots as a for optimal trajectory planning:
performance criterion to be maximised along a desired
trajectory. The Manipulator Velocity Ratio is defined as the
ratio of the end-effector velocity norm to the joint velocity
vector norm
 P1

P0
ds
det J
(11)

where P0 and P1 are the end points of the desired trajectory


in the Cartesian space, and ds is the differential distance
ẋTv ẋv along the trajectory.
rv = ¯ ¯ (4)
q̇Tv q̇v From the review of the proposed dexterity measures so
¯ ¯ far the following conclusions can be derived:
where ẋ and q̇ are the weighted vectors of the end-effector (i) The manipulability index, the reciprocal of the condi-
¯ velocities,
and joint ¯ respectively. By replacing the well tion number, and the minimum singular value, or any
known equation combination of them can represent the motion ability
of the robot. Using these criteria the ability of the robot
q̇v = Jv+ ẋv (5) to move in a given direction cannot be derived.
¯ ¯ (ii) The MVR is an index of the robot ability to move in a
where Jv+ is the inverse or the pseudo-inverse of the given direction at a specific configuration.
weighted Jacobian Jv for non-redundant or redundant (iii) The Hybrid Manipulator Measures are suitable for
manipulators, respectively, the following relation for the combined tasks, where velocity and applied force are
MVR it can be obtained specified simultaneously.
(iv) These integral measures proved superior compared to


the instantaneous ones, but they are used to determine
ẋTv ẋv a trajectory in the joint space for given boundaries in
rv = ¯ ¯ (6) the Cartesian space. However, even if the manipul-
(Jv+ ẋv )T(Jv+ ẋv )
¯ ¯ ability of the robot is taken into account in these

http://journals.cambridge.org Downloaded: 07 Apr 2011 IP address: 117.211.90.154

142 Optimal location of a robot path
integral measures, the directional variation of the robot From the presentation of the MVR, it can be concluded
velocity efficiency is not considered. that it includes information concerning the ability of the
(v) The integral measures are time consuming and can fail robot to move the end-effector along a specific direction,
when the integrated function presents large variations. and it gives the velocity efficiency of the robot in this
A very high value of the integrated index may give a direction. Therefore, it is the most suitable criterion for the
poor performance along the rest of the path. free motion of a robot; however, it is a local and
instantaneous criterion. In the following, various forms of
global criteria for the movement of a robot along a path with
2.2. The proposed optimisation criterion its maximum velocity efficiency are suggested and com-
From the investigation of the proposed measures, it can be pared. The first is the minimum of the MVR and the second
concluded that the velocity ratio (MVR) depends not only is the mean value of the MVR along a path.
on the robot configuration but also on the end-effector The problem of estimating an optimal location of a
velocity vector. In addition, it could be proved that for a desired robot path of a given shape is defined as follows:
given configuration the MVR presents a measure of the Determine the location of the path represented by a curve in
deviation between the end-effector velocity and the major 3D space so that the minimum of the Manipulator Velocity
axis of the ellipsoid. This can be understood better by the Ratio (MVR) along the path is maximum. The minimax
following example, where a 3D ellipsoid is used to illustrate form of optimisation gives the best worst-case performance
the relation between the principal axes of the ellipsoid and throughout the workspace. Therefore, the location of the
the end-effector velocity. In Figure 1, a 3D ellipsoid is path is determined where the robot presents high perform-
shown, where the vectors a, b and c correspond to the major, ance at every point of the path.
secondary and minor axes¯ of¯ the 3D
¯ ellipsoid, respectively.
The method is presented using a straight-line segment as
If | q̇ | = 1 then the desired shape of the path for simplicity and then it can
¯ be extended to any 3D curve. The local Cartesian co-


Vx
a
2
+
Vy
b
2
+
Vz
c
2
=1 (12)
ordinate frame {FS} and the length of the segment define the
position and the orientation of the straight-line segment. As
it is illustrated in Figure 2, the starting frame {FS} coincides
with the staring point of the path, while the x-axis is along
where V = [Vx, Vy, Vz]T is the vector of the translational the path and the z-axis is perpendicular to the straight-line
velocity¯ of the end-effector referenced to the Cartesian segment.
system defined by the axes of the manipulability ellipsoid. The frame {FS} is defined in a parametric form by
Since the elements of the end-effector velocity vector are xs = [x y z   ]T,where x, y, z are the Cartesian co-ordinates
equal to the projection of the velocity vector to the principal ¯ the starting point of the segment defined in the universal
of
axes of the ellipsoid, Eq. (12) can be written as: frame {U} and , ,  are the Euler angles of the frame
1 {FS} referenced to the frame {U}. Alternatively, the location

     
|V|= (13) of the frame {FS} can be defined through the transforma-
¯ ua T
uTb 2
2 uTc 2
¯ 2¯ + ¯ 2¯ + ¯ 2¯ tion
a b c U
TFS(qs), qs = [q1 . . . qn]T and qLS ≤ qS ≤ qUS (14)
where u is the unit vector along the end-effector velocity. ¯ ¯ ¯ ¯ ¯
¯ direction of the end-effector velocity coincides where qS is the vector of the joint co-ordinates of the
When the ¯
manipulator, when the frame of the tool coincides with the
with the major axis of the ellipsoid, the magnitude of the frame {FS}. For simplicity, it is supposed that the tool is
end-effector velocity is maximum, since the scalar products perpendicular to the segment. The Cartesian or the joint co-
of u with the vectors corresponding to the other two axes ordinates defining the local frame {FS} are the variables of
¯
become equal to zero due to the orthogonality of the the objective function representing the minimum value of
ellipsoid axes. For the same reason, the end-effector
velocity becomes minimum when its direction coincides
with the minor axis.

Fig. 1. The axes of the 3D manipulability ellipsoid and the end-

effector velocity. Fig. 2. The intermediate points of the straight-line segment.

http://journals.cambridge.org Downloaded: 07 Apr 2011 IP address: 117.211.90.154

Optimal location of a robot path 143
the MVR along the path. The global measure, which is the the joint co-ordinates of the manipulator are determined and
objective function, is approximated in a procedural way then the corresponding Jacobian matrix
based on Taylor’s22 bounded deviation method for planning
J(qi(UTFi(qs)) (17)
and execution of straight-line manipulator trajectories. It is ¯ ¯
well known, that given two configuration frames of the For a genetic algorithm search, it is not necessary to obtain
hand, the starting frame {FS} and the frame {Ff } corre- an analytic formula for the path MVR index so the inverse
sponding to the final point of the path, the Taylor’s kinematic problem can be solved numerically if it is
algorithm calculates the corresponding joint co-ordinates qS impossible to obtain an analytic solution. Using Eq. (17) the
¯
and qf . Subsequently, the Cartesian co-ordinates of the joint MVR at every intermediate knot point and at the two end
¯
midpoint are calculated and are compared to the Cartesian points of the segment are calculated as functions of qS, and
configuration corresponding to the midpoint of the straight- finally the measure rm is calculated using Eq. (15). So¯ at the
line segment. If the two configurations “deviate” more than end of the analysis the MVR index along the path is
an allowed amount, the final point configuration is replaced determined as a function of the joint coordinates of the
by the Cartesian midpoint configuration and the algorithm is frame {FS}.
applied recursively to this straight-line segment. This The number and the position of the intermediate knot
generation of midpoints terminates, when the deviation is points depends upon the location of the path into the
smaller than the specified amount. The arm motion is workspace so that the objective function is non-continuous.
linearly interpolated between successive midpoints. Taylor In a previous work by Aspragathos,22 the Taylor’s method
proposed the Euclidean metric as a position deviation has been generalised to be applicable to a Cartesian
function. This technique reduces substantially the real-time trajectory represented by any 3D curve. In this case too, the
computation. algorithm calculates enough intermediate points in order to
Let suppose that the Taylor’s algorithm determines Pi generate a bounded position deviation trajectory of the tip of
(i = 1, k  2) intermediate knot points, then a global index the manipulator hand. Therefore, the proposed formulation
for the MVR along the path is given by of the index rm can be applied to a 3D curve using the
generalised Taylor’s method.
rm = min{(rv)i , (rv)s, (rv)f }, for i = 1, k  2 (15) For comparison and final selection of the best objective
function in order to search for the optimal location of a path
where (rv)i , (rv)s, and (rv)f are the MVR indices at an from the point of the velocity efficiency of the robot, an
intermediate point, the starting point, and the final point of index equal to the average of the MVR along the path is also
the segment, respectively. The calculated minimum MVR proposed.


along the path is an approximation of the real minimum of
i=k2
the MVR along the path, and it depends upon the permitted
maximum position deviation used in Taylor’s algorithm. ra = ((rv)s + (rv)f + (rv)i)/k for i = 1, k  2 (18)
i=1
This dependence is complicated enough so it is difficult to
estimate the error. However, using the proposed approxima-
tion, a considerable amount of computational time is saved. In this paper, two additional velocity efficiency indices
It is well known that the Taylor algorithm converges based on the robot manipulability are investigated and
quickly and a deviation can be obtained as small as it is compared. The index wm presents the minimum manipul-
desired. The relation between the position deviation and the ability along the path and it is defined by
respective deviation of the MVR is very complicated and it wm = min(wi, ws, wf ) for i = 1, k  2 (19)
can be assumed that this function is incremental at least for
small position deviation as it concluded experimentally. where wi, ws, and wf are the manipulability indices at an
However, near a singular point for a small position deviation intermediate point, the starting point, and the final point of
a very large deviation of the corresponding MVRs is the straight-line segment, respectively. The forth measure of
expected. Since the proposed optimisation algorithm pushes the velocity efficiency along the path, which is investigated
the location of the path far from singular points then such in this work, is the path manipulability defined by
problems are avoided. Aspragathos.14 This measure represents the approximate
Every intermediate knot position orientation of the tool average of the manipulability along the path and it is given
on the straight-line segment is given as a function of the by

 
parameters defining the frame at the starting position i=k2
orientation by
wa = wi + ws + wf k (20)
U i=1
TFi = UTFS(qS)[Li 0 0]T (16)
¯
From the presented analysis of the proposed indices
where Li is the distance of the ith intermediate point from
concerning the performance of the robot along a path, the
the origin of the coordinate system {FS}
following conclusions can be derived:
The Cartesian co-ordinates xS can replace the joint co-
ordinates qS, but in the following¯ the joint coordinates are (i) The definition of the objective function based on the
¯
used. Solving the inverse kinematic problem at every minimum of a local index along the path is better than
intermediate and the final position/orientation of the tool, an objective function representing the mean value,

http://journals.cambridge.org Downloaded: 07 Apr 2011 IP address: 117.211.90.154

144 Optimal location of a robot path
since in the mean value index the result can be multi-modal and very complex non-linear search space. On
dominated by a very high value of the local index at a the other hand, classical optimisation methods depending
point of the path. upon restrictive requirements of continuity and derivative
(ii) The bisection method used to determine the local existence, and due to their inherent local scope of search,
indices in the intermediate points of the path saves are unsuitable for the solution of this problem. Second, GAs
computational time compared to that of integral do not require an analytic form of the objective function and
indices. of the constraints. In our case the objective function and the
(iii) The MVR index gives a measure of the robot velocity constraints are procedural. Third, GAs allow additional
performance in the direction of the path, while the constraints to be specified easily, so the introduced method
manipulability index gives an average of the robot can be extended to work in an environment cluttered with
velocity performance and the reciprocal of the con- obstacles or other constraints.
ditioning number represents the degree of isotropy. The simple GA used in the presented optimisation
problem is composed of the following three operators:
Based on these conclusions the measure rm is proposed as
reproduction, crossover and mutation. This simple GA has
the most suitable objective function, in order to determine a
been proved effective for a variety of optimisation prob-
location of the path where the robot presents the best
lems. A critical aspect in designing a GA is the basic
velocity performance. To strengthen this qualitative justifi-
mechanism that links the GA to the real problem, which has
cation of the proposed measure, the presented four indices
to be solved. This mechanism is twofold: First, an
rm, ra, wm, and wa are compared experimentally in the section
evaluation of the function (fitness) that returns a measure of
of the results of this paper and it is verified that the first
the expected goodness of the solution is defined, and,
index gives more coherent relation between the index and
second, a way of encoding solutions to the real problem on
the required maximum joint velocity along the path.
artificial chromosomes is proposed.
Therefore, the first index is used in the following as an
GAs are essentially unconstrained search procedures
objective function in searching for the optimal location of a
within the given representation space. Goldberg24 proposed
path in the robot workspace.
a penalty function concept to convert the constrained
By maximising the global MVR index rm subject to the
problem to an unconstrained one. Another formulation is to
constraints of optimisation variables given by Eq. (14), one
design and apply special recombination operators that
can find the location of the path where the MVR of the robot
maintain feasibility. In the GA used to solve the proposed
is the higher. This optimisation problem cannot be solved
optimisation problem, a different approach introduced by
using a classical gradient method (see Cooper and Stein-
Nearchou and Aspragathos25 has been adopted, so when a
berg23) because the objective function has a procedural
candidate solution violates the constraints, the fitness
nature and in the most general case is non-analytic, non-
function takes a value equal to zero and has almost zero
continuous and non-differentiable. The selection of the
possibility to survive and no hope to get copies in the next
Genetic Algorithms to solve this problem is based on the
generations.
advantages of the GAs in search of very complex and
Therefore the fitness function related to the objective
unstructured environments.
function given by Eq. (15) and the joint coordinate
constraints (see Eq. (14)) is given by
3. THE GENETIC ALGORITHM SOLUTION

Genetic Algorithms (GA) are powerful domain independent
search mechanisms, which emulate the process of genetic rm(qS), if qLS ≤ qS ≤ qUS
fitness = ¯ ¯ ¯ ¯ (21)
evolution found in nature as a means of progressing towards 0, if qS ≤ qLS or qUS ≤ qS
the optimum. Combining an artificial survival of the fittest ¯ ¯ ¯ ¯
with special genetic operators, GA provides a robust search The variables of the objective function are the joint
mechanism that is suitable for a variety of search problems. coordinates qS determining the local frame {FS}, which
A Genetic Algorithm works simultaneously from a number ¯
defines the position and the orientation of the path. This set
of points scattered in the search space and examines many of variables is coded as a finite length binary string, which
local optima in parallel, thus the probability to determine represents one chromosome of the population.
the global optimum is high, when it works in a multi-modal In this work, the chromosome selected is an m-bit string
space. Genetic Algorithms require only the objective with the following syntax: the (m/n)-bits correspond to the
function, while in many optimisation methods auxiliary first joint of the robot, the next (m/n)-bits correspond to the
information is necessary. For example, the gradient methods second joint, etc. Therefore the ith joint co-ordinate can be
need the derivatives of the objective function. calculated from the values of the corresponding (m/n)-bit
In the following, a solution to the formulated optimisa- using the following mapping:
tion problem based on a simple Genetic Algorithm is
qi = qLi + [Bitvalue /2m/n]qUi (22)
introduced. There are at least three basic reasons that make
GAs attractive in solving such a problem: First, they are where Bitvalue is the decimal number corresponding to the
proven to provide a robust search in complex spaces by string of the ith joint, and qLi, qUi are the limits of the ith joint
finding nearly global optima. In general, the bisection variable.
recursive method used to define the knot points under It must be underlined that a single m-bit string represents
bounded position deviation provides a non-continuous, one of the 2m alternative solutions to the problem. The value

http://journals.cambridge.org Downloaded: 07 Apr 2011 IP address: 117.211.90.154

Optimal location of a robot path 145
m/n, which denotes a number of bits that correspond to a
parameter of the original optimization problem, is not
selected arbitrarily but depends on the adequate “resolution”
while preventing the search space from becoming too large
for easy and effective searching. The GA generates
successive populations of the strings in the form of natural
selection. The purpose of the searching is to find the best-fit
string (chromosome). Concerning the convergence of the
algorithm, the following termination condition is used:
(average fitness)/(best fitness) > , where is an arbitrarily
selected real number close to one and represents the degree
of the uniformity of the chromosome values (alleles).

4. DISCUSSION OF THE RESULTS

In this paper, two categories of simulated experiments are
carried out using a PUMA-like robot. The first category
includes the experiments for the investigation of the
Fig. 4. The maximum joint velocity q̇max
i versus the average of the
suitability of the proposed measures for the determination of MVR along the path ra.
the location of the path where the velocity efficiency of the
robot is the best. In the second category, the efficiency of the
to reciprocal relation can be concluded. The scattering of
proposed method for estimating the location of a straight-
the points is quite limited. In Figure 4, the maximum joint
line path with the higher minimum MVR along the path is
velocity q̇max
i versus the average of the MVR along the path
demonstrated.
ra is shown, for all the randomly selected locations of the
path. For low and medium values of the measure ra the
4.1. Comparison of the proposed measures
scattering of the points is much greater than for high values
For an arbitrary location of a straight-line segment the end-
of ra.
effector of the robot moves with constant velocity along this
Figure 5 presents the relation between the maximum joint
path using the Resolved Motion Rate Control.25 For this
velocity q̇max
i and the minimum of the manipulability along
location of the path the four measures are determined and
the path wm, for all the randomly selected locations of the
the maximum joint velocity appeared during the motion
path. The high scattering of the points appears for low
along the path is calculated. The joint velocities are
values of the measure wm. The higher scattering appears in
normalised to obtain a common base for comparison. This
the last diagram shown in Figure 6, where the relation
experiment is repeated for a large number of randomly
between maximum joint velocity q̇max i and the average
selected locations of the same straight-line segment where
manipulability wa is presented.
the end-effector moves with the same constant velocity. The
From the comparison of the results presented in Figures 3
results of these experiments are presented in the diagrams of
to 6, it can be concluded that the measure of the velocity
Figures 3 to 6.
efficiency along a path rm defined by the minimum MVR is
Figure 3 illustrates the relation between the maximum
superior, since statistically it corresponds to the lowest
joint velocity q̇max
i and the minimum of the MVR along the
maximum joint velocity for the same location of the path
path rm. It is clear that the points of the diagram presenting
and the same constant end-effector velocity.
the pairs {q̇max
i , rm} are very well correlated and a very close

Fig. 3. The relation between the maximum joint velocity q̇max

i and Fig. 5. The relation between the maximum joint velocity q̇max
j and
the minimum of the MVR along the path rm. the minimum of the manipulability along the path wm.

http://journals.cambridge.org Downloaded: 07 Apr 2011 IP address: 117.211.90.154

146 Optimal location of a robot path

Fig. 7. Variation of the MVR, of the maximum joint velocity and

of the manipulability along the path for the 2nd case of the
Table I.

and 1.16, due to the low accuracy of the Genetic

Algorithms. However, all the results are very close to the
Fig. 6. The relation between maximum joint velocity q̇max
i and the maximum of the fitness function, so the solutions obtained
average manipulability wa along the path. by the proposed method are considered as sub-optimal ones.
Even for the first three cases, where the obtained maximum
4.2. Search for the best location of the path of rm is equal to 1.28, the obtained locations of the path
The second category of experiments includes the demon- differ considerably. This means that there are more than one
stration of the efficiency of the proposed method in areas in the workspace of the robot, where a sub-optimal
searching for the best location of a straight-line path of location can be found. This conclusion helps the designer of
length L. Suppose that the simulated PUMA-like robot has a workcell, because there are more alternative “very good”
to move the end-effector along the path presented by a solutions for the location of a task.
straight-line segment with the tool axis perpendicular to the For further investigation of the results obtained by the
path. It is required to search for a location of the straight- proposed method, the variation of the MVR (solid line), of
line segment into the workspace of the robot where the the maximum between the absolute of the joints velocity
velocity efficiency of the robot will be the best. The velocity (dotted line), and the manipulability index (dashed line)
efficiency of the robot along the path is defined by the along the path are presented in the diagrams of Figures 7
fitness function given by Eq. (21). and 8 corresponding to the 2nd and 6th case of the Table I
The Genetic Algorithm was run several times and some of respectively. The presented diagrams show that the variation
the results are presented in Table I. In the first column of the of the MVR presents the minimum value at one of the ends
Table I, the maximum of the rm (the minimum of the MVR of the path and the maximum MVR around the middle of
along the path) obtained by searching the workspace of the the path. The maximum between the absolute velocities of
robot is presented. In the second column, the maximum the joints for each configuration of the robot moving its end-
joint velocity q̇max
i is shown, which is obtained by moving effector with constant velocity presents a minimum around
the end-effector along the path with constant velocity using the middle of the straight-line path. This correspondence
the Resolved Motion Rate Control.25 The third and fourth between the maximum of the MVR and the minimum of the
columns shows the Cartesian coordinates of the starting and maximum joint velocity was expected from the definition of
end point of the path in the determined optimal location, the MVR. Similar results are obtained in all the runs of the
respectively. GA, so the proposed method can be considered as quite
From the results shown in the Table I, the following robust even the accuracy of the results is not high.
conclusions can be derived: The maximum fitness obtained
by the runs of the Genetic Algorithm ranges between 1.28

Table I.

rm q̇max
i Starting point Final point

1.28 0.68 (1.69, 0.02, –0.32) (1.57, –0.92, –0.00)

1.28 0.73 (1.18, –1.44, –0.56) (0.27, –1.58, –0.94)
1.28 0.69 (1.45, –1.25, –0.25) (1.65, –0.40, –0.75)
1.275 0.68 (–0.32, 1.79, 0.30) (–1.14, 1.24, 0.47)
1.26 0.69 (0.95, –1.52, –0.13) (1.49, –0.71, –0.34)
1.25 0.77 (0.16, 1.93, 0.53) (–0.58, 1.51, 1.05)
1.24 0.78 (1.5, 0.2, 0.74) (1.63, 0.86, –0.01)
1.22 0.72 (0.68, 1.42, 0.21) (1.48, 0.89, –0.05)
1.20 0.79 (–0.93, 1.4, –0.57) (–1.6, 1.05, 0.09) Fig. 8. Variation of the MVR, of the maximum joint velocity and
1.16 0.60 (–0.93, –1.62, –0.59) (0.04, –1.87, –0.55) of the manipulability along the path for the 6th case of the
Table I.

http://journals.cambridge.org Downloaded: 07 Apr 2011 IP address: 117.211.90.154

Optimal location of a robot path 147
The proposed method can be implemented and incorpo- 6. P. Chedmail and Ph. Wenger, “Design and Positioning of a
rated in software packages for robot simulators and for Robot in an Environment with Obstacles using Optimal
Research”, IEEE Conference on Robotics and Automation
robotic workcell design to help the designer in the layout of (1989), pp. 1069–1074.
the workcell. The designer can locate the path in the 7. D. Martin, J. Baillieul and J. Hollerbach, “Resolution of
determined location using the proposed algorithm and then Kinematic Redundancy Using Optimization Techniques”,
he can apply another minimum time algorithm to reduce the IEEE Transactions on Robotics and Automation 5, No. 4,
cycle time of the task by exploiting the velocity transmis- 529–533 (1989).
8. M. Uchiyama, K. Shimizu, and K. Hakomori, “Performance
sion capabilities of the robot. Evaluation of Manipulators Using the Jacobian and its
Application to Trajectory Planning”, Proceedings of 2nd
International Symposium on Robotics Research, Kyoto
(August 1984), pp. 447–454.
5. CONCLUSIONS 9. Y. Nakamura and H. Hanafusa, “Task Priority Based Redun-
An approach for the determination of an optimal or sub- dancy Control of Robot Manipulators”, Proceeding of 2nd
optimal location of a path into the workspace of a robot is International Symposium Robotics Research, Kyoto (August
presented. The optimisation criterion is based on the robot 1984), pp. 155–162.
10. A. Ghosal and B. Roth, “A New Approach for Kinematic
velocity efficiency or the capability of velocity transmission Resolution of Redundancy”, Robotics Research 7, No. 2,
of the robotic mechanism. A global measure describing the 22–35 (1988).
velocity efficiency of the robot along the path is introduced. 11. D. Cho, B. Cho and M. Chung, “Optimal Conditions for
Among the four suggested global measures, which are Inverse Kinematics of a Robot Manipulator with Redun-
compared experimentally, the minimum Manipulator Veloc- dancy”, Robotica 13, Part 1, 95–101 (1995).
12. B. Nelson and M. Donath, “Optimizing the Location of
ity Ratio (MVR) was proved superior. This result was Assembly Tasks in a Manipulator’s Workspace”, Robotic
expected since the MVR depends on the direction of the System, 7(6), 791–811 (1990).
velocity end-effector, so this criterion is a measure of the 13. B. Fardanesh and J. Rastegar, “Minimum Cycle Time
deviation of the path direction to the direction of the Location of a Task in the Workspace of a Robot Arm”, Proc.
manipulability ellipsoid. of the 27th Conference on Decision Control, Austin, Texas
(1988) pp. 2280–2283.
The proposed measure is calculated in a procedural way, 14. N. A. Aspragathos, “Optimal Location of Path Following
which does not guarantee the continuity of the objective Tasks in the Workspace of a Manipulator using Genetic
function of the formulated optimisation problem. Mainly for Algorithms”, ARK’96, Portroz, Slovenia (June 1996). Pub-
this reason, a Genetic Algorithm is used to search the lished in Recent Advances in Robot Kinematics (Kluwer
complex space for a global optimal or sub-optimal location Academic Publishers, 1996) pp. 179–188.
15. T. Yoshikawa, “Manipulability of Robotic Mechanisms”, Int.
of the path. By successive runs different sub-optimal J. Robotics Research 4, No. 2, 3–9 (1985).
solutions are determined in the workspace of a PUMA-like 16. J.K. Salisbury and J.J. Craig, “Articulated Hands, Force
robot. The maximum values of the proposed criterion are Control and Kinematic Issues”, Int. J. Robotics Research 1,
very close to each other while the corresponding locations No. 1, 4–17 (1982).
are scattered into the workspace. 17. C.A. Klein, “Use of Redundancy in the Design of Robotic
Systems”, Robotics Research: The Second International
After a thorough verification and validation, the proposed Symposium (Eds H. Hanatusa and H. Inoue) (MIT Press,
algorithm can be incorporated into software packages for 1985) pp. 207–214.
robot simulation. Using this algorithm the best location of a 18. C.M. Gosselin, “Dexterity Indices for Planar and Spatial
task can be determined, which is a considerable help in Robotic Manipulators”, IEEE Conference on Robotics and
optimal design of the layout of a robotic workcell. Therefore Automation (1990) pp. 650–655.
19. J. Angeles, “A Scale-Independent and Frame-Invariant Index
the proposed algorithm can save design time and cost, and of Kinematic Conditioning for Serial Manipulators”,
reduce the energy consumption and wear of the robot or the Advances in Robot Kinematics (Eds. S. Stiffer & J. Lenarcic)
cycle time of a task. (Springer-Verlag, Vienna, 1991) pp. 2–9.
20. I. Ermolov and Podurajev, “Intelligent Robot Motion Plan-
ning Based on Hybrid Manipulator Measures Calculus”, IFAC
LSS’98, L̀arge Scale Systems: Theory and Applications, Patras
References (1998) pp. 1132–1137.
1. S. Dubowsky and T. Blubaugh, “Planning Time-Optimal 21. R. Dubey and J.Y. Luh, “Redundant Robot Control Using
Robotic Manipulator Motions and Work Places for Point-to- Task Based Performance Measures”, Robotic Systems 5(5),
Point Tasks”, IEEE Transactions on Robotics and Automation 409–432 (1988).
5, No. 3, 377–381 (June, 1989). 22. N.A. Aspragathos, “Cartesian Trajectory Generation Under
2. L. Abdel-Malek and Z. Li, “The Application of Inverse Bounded Position Deviation”, Mechanisms and Machine
Kinematics in the Optimum Sequencing of Robot Tasks”, Int. Theory 33, No. 6, 697–709 (1988).
J. Production Research 28, No. 1, 75–90 (1990). 23. L. Cooper and D. Steinberg, Introduction to Methods of
3. Y. Edan, T. Flash, U. Peiper, I. Shmulevich and Y. Sarig, Optimisation (W.B. Saunders Company, 1970).
“Near-Minimum Time Task Planning for Fruit-Picking 24. D.E. Goldberg, Genetic Algorithms in Search, Optimisation
Robots”, IEEE Transactions on Robotics and Automation 7, and Machine Learning (Addison-Wesley Publishing, 1989).
No. 1, 48–56 (1991). 25. A.C. Nearchou and N.A. Aspragathos, “Obstacle Avoidance
4. J. Petiot, P. Chedmai and J. Hascoet, “Contribution to the Control of Redundant Robots using Genetic Algorithms”, 3rd
Scheduling of Trajectories in Robotics”, Robotics and IEEE Mediterranean Symposium Directions in Control and
Computer-Integrated Manufacturing 14, 237–251 (1998). Automation, Limassol, Cyprus (1995) pp. 60-67.
5. M. Dissanayake and J. Gal, “Workstation Planning for 26. D.E. Whitney, “The Mathematics of Co-ordinated Control of
Redundant Manipulators”, Int. J. Production Research 32, Prosthetic Arms and Manipulators”, J. Dynamic Systems
No. 5, 1105–1118 (1994). Measurement, and Control 94, 303–309 (1972).

http://journals.cambridge.org Downloaded: 07 Apr 2011 IP address: 117.211.90.154