Soft Comput

DOI 10.1007/s00500-012-0964-8

METHODOLOGIES AND APPLICATION

Multi-objective optimal path planning using elitist non-dominated

sorting genetic algorithms
Faez Ahmed Kalyanmoy Deb

Springer-Verlag Berlin Heidelberg 2012

Abstract A multi-objective vehicle path planning

method has been proposed to optimize path length, path
safety, and path smoothness using the elitist non-dominated
sorting genetic algorithma well-known soft computing
approach. Four different path representation schemes that
begin their coding from the start point and move one grid at
a time towards the destination point are proposed. Minimization of traveled distance and maximization of path
safety are considered as objectives of this study while path
smoothness is considered as a secondary objective. This
study makes an extensive analysis of a number of issues
related to the optimization of path planning task-handling
of constraints associated with the problem, identifying an
efficient path representation scheme, handling single versus
multiple objectives, and evaluating the proposed algorithm
on large-sized grids and having a dense set of obstacles.
The study also compares the performance of the proposed
algorithm with an existing GA-based approach. The evaluation of the proposed procedure against extreme conditions having a dense (as high as 91 %) placement of
obstacles indicates its robustness and efficiency in solving
complex path planning problems. The paper demonstrates

Communicated by F. Herrera.
F. Ahmed  K. Deb (&)
Department of Mechanical Engineering,
Indian Institute of Technology Kanpur, Kanpur, India
e-mail: deb@iitk.ac.in
F. Ahmed
e-mail: faez.iitk@gmail.com
K. Deb
Department of Information and Service Economy,
Aalto University School of Economics,
00100 Helsinki, Finland

the flexibility of evolutionary computing approaches in

dealing with large-scale and multi-objective optimization
problems.
Keywords Multi-objective path planning 
Potential field  Path length  Path safety 
Path smoothness  NSGA-II  Genetic algorithms

1 Introduction
In recent years we have seen much advancement in the
field of industrial robotics, but mobile intelligent machines
have mainly been confined to research labs. For an
autonomous vehicle to be used in real-world applications it
must be able to navigate autonomously and take intelligent
decisions. Vehicle path planning comprises of not only
generating collision-free paths from a given location to its
destination point but also finding an optimized path that
minimizes or maximizes certain critical objectives.
LaValle (2006) and Hwang and Ahuja (1992) provide a
broad coverage of the field of motion planning algorithms
focusing on vehicle motion planning.
There have been many conventional methods for twodimensional path planning using classical optimization
methods (Bisse et al. 1995; Xue et al. 2010), artificial
potential field method (Ge and Cui 2002; Al-Sultan and
Aliyu 2010), visibility graph (Lozano-Perez and Wesley
1979; Murrieta-Cid et al. 2005), Voronoi road-map
(Choset 1996), etc. Many soft computing methods such as
artificial neural networks (Glasius et al. 1995; Yang and
Meng 2000), fuzzy logic method (Oriolo et al. 1997;
Pratihar et al. 1999), genetic algorithms (Gerke 1999;
Burchardt and Saloman 1831) and ant colony optimization
(Sauter et al. 2002) have recently come to forefront for

123

F. Ahmed, K. Deb

path planning. In a path planning task, a majority of

research has been dedicated to finding a feasible shortest
path by formulating a single-objective optimization problem. However, finding the optimized path is never a singleobjective problem, as many other attributes such as path
safety (or path vulnerability) (Ahmed et al. 2011) and path
smoothness are also desirable for a navigation vehicle. The
path vulnerability objective is associated with how close
the vehicle pass through an obstacle. To compute this
objective, we have taken help of the artificial potential field
concept. Each obstacle holds a large potential in its center
and the potential is assumed to reduce away from the
obstacle. Thus, for a given path and a given setting of
obstacles, the overall potential computed for the entire path
accumulated due to the effect of all obstacles is denoted as
the path vulnerability. Path smoothness is directly related
to the amount of turn a vehicle has to take along the path
from the start to the end of its motion.
Besides the single versus multiple conflicting objectives
associated with the path planning problem, there are a
number of other issues that must be resolved during an
optimization process. One important matter is the scheme
used to represent a path within the optimization procedure,
as the performance of an optimization algorithm depends
largely on the chosen representation scheme. In this paper,
we propose four different techniques for path representation based on a binary, mixed or integer coding of a path
description. Second, the issue of a two-objective versus a
three-objective optimization is another matter that will
concern the practitioner and will test the usefulness of the
chosen algorithm. Third, the use of obstacle avoidance as a
hard constraint in the optimization process or handling of it
softly through a penalty term is another matter that will
determine the performance of the chosen algorithm.
Finally, for a path planning problem, the scaling of the
chosen optimization algorithm with the number of obstacles and the grid size must be well understood before the
method can be recommended for a practical use. In this
paper, we address all the above issues related to the path
planning problem one at a time by considering obstacle
scenarios that were adopted in earlier path planning studies.
A bi-objective algorithm minimizing path length and path
vulnerability is proposed based on a popular multi-objective optimization algorithmthe elitist non-dominated
sorting genetic algorithm (NSGA-II) (Deb et al. 2002), but
the algorithm is modified to use the third objective (path
smoothness) to introduce diversity in the population and
also to help in a subsequent decision making. To provide a
stiff challenge to the proposed algorithm, the complexities
of obstacle placement are gradually increased to have as
high as 91 % of the cells occupied by obstacles. For this
purpose, the search domain is also increased from 8 9 8 to
a staggering 128 9 128. These extremities are rarely used

123

in the past computational path planning studies and the

successful working of the proposed algorithm in them
amply demonstrates their practical importance. Although
other evolutionary multi-objective optimization (EMO)
algorithms, such as SPEA2 (Zitzler et al. 2001) or MOEA/D
(Zhang and Li 2007) or others, can very well be used to
obtain similar results, the purpose of this paper is not
establish superiority of this or that algorithm, but to demonstrate that a multi-objective treatment of the path planning
problem with an appropriate representation scheme can
produce better results than an usual single-objective optimization algorithm, particularly in a scenario having a dense
set of obstacles.
In the remainder of the paper, four new path representation techniques are discussed in Sect. 2. Three objectives
used in this study are described next in Sect. 3. Section 4
discusses two different constraint handling methodologies
and compares the three representation schemes suggested
in this paper. Results for bi-objective problems are shown
comparing them to a benchmark problem in Sect. 5. A
comparative study of single and three-objective optimization runs is also done. Thereafter, Sect. 6 makes random
placements of obstacles on differently sized grids and
investigates the efficiency of the proposed algorithm in
finding a feasible path in the presence of an increasing
number of obstacles. Problems having as large as
128 9 128 grids and 91 % coverage of grids with obstacles, thereby resembling extreme practical scenarios, are
considered next. Conclusions of the study are made in
Sect. 7.

2 Path representation through a grid

To optimize any given path of a vehicle through an optimization technique, the foremost requirement is an efficient path representation that can be incorporated in the
programming environment and importantly that is suitable
to be efficiently used with the chosen optimization algorithm. In this study, we use a genetic algorithm (GA) as an
optimization algorithm. Since a GA allows a flexible way
to represent a solution, we suggest four different representation schemes for the purpose.
To facilitate a discrete representation and also for the
ease of navigation purposes, a grid-based approach is
adopted here. In a given problem of 2-D off-line path
planning, we map the environment on a grid of size
n 9 n, where the grid size n is decided depending on the
accuracy required in the path planning problem. The size of
the vehicle is taken to be smaller than unit cell size in grid
so that the dynamics of vehicle rotation is not to be considered. For simplicity, we always take the start location as
the bottom-left corner, denoting grid location (1, 1). In the

Multi-objective optimal path planning using elitist NSGA-II

Cartesian coordinate system, the destination is then fixed as

(n, n) (or the top-right corner of the grid). This assumption
is a reasonable one since most other path planning scenario
can be mapped on to such a grid by simple linear transformations, such as translation, rotation, and scaling. Now
if we consider the characteristics of the entire path between
a start location and final location then different paths taken
by a vehicle may have different lengths, directions, and
turns at different points. A possible solution is a linked-list
representation (Xiao and Michalewicz 2000) of the entire
path having a variable length gene. Such a representation
of a path by genes of variable length is cumbersome considering various GA operations, like crossover as well as
the absence of knowledge about the bounds of path length.
A generalized representation would considerably increase
the search space and hence slowing down the search process. Considering these difficulties with variable-length
gene representations, we propose a fixed-length path representation here. Motivated by another study (Sugihara and
Smith 1999), we consider only paths which are monotonous along one axis. That is, starting from location (1, 1),
the path moves one grid at a time towards right till it
reaches the destination point (n, n). Since there are n columns in the search domain and the vehicle is already at the
first column, we would require (n - 1) genes to represent a
path. We propose three different techniques of representing
a path involving binary, integer or mixed-valued genes. We
describe these representation schemes in the following
subsections.
2.1 Binary-coded genes
Our proposed path representation technique was inspired
by a recent study (Sugihara and Smith 1999). We first
describe the procedure and then our coding method to
alleviate a difficulty associated with the earlier procedure.
Let us consider the grid as comprised of rows and columns.
To represent a path, following symbols were defined:

A denotes an upper diagonal step from one column to

^
the next parallel to (i^ j).
B denotes a lower diagonal step from one column to
^
the next parallel to (i^ j).
C denotes a horizontal step from one column to the
^
next parallel to (i).
Dy denotes an upward vertical movement by y units
^
within a column parallel to (j).
Ey denotes downward vertical movement by y units
^
within a column parallel to (j).

Graphical notion of the above symbols is given in

Fig. 1, where i^ corresponds to positive x-axis and j^ to
positive y-axis. A gene comprises of (n - 1) movements

Fig. 1 Representation scheme denoting (left) steps corresponding to

alphabetic codewords, (center) decoded binary bits, and (right)
representation schemes of transition followed by vertical movement
step in terms of alphabetic codewords

where each movement is one transition step (A, B or C)

followed by a vertical motion (Dy or Ey) within a column,
where y denotes magnitude of vertical motion. In Sugihara
and Smith (1999), transition from one column to other was
either through A, B or C, but any vertical movement Dy or
Ey within a column was ignored if the last transition was
either A or B. That is, if a diagonal move is taken, the path
cannot go upwards or downwards; it can only go forward,
whereas to move up more than one grid up or down, the
only option is to go horizontally towards right and then
make a vertical move up or down. This is restrictive and as
we shall discuss later it will create paths that are dominated
by diagonal paths. We modify the above coding by
allowing vertical steps after type A or B movements as
well.
In our proposed binary-coding, each gene comprises of
two parts. First two bits represent the direction of transition
to the next column in four possible ways. The vehicle can
move diagonally up by one grid and then a distance y up
along the column. We call this movement A-Dy and it is
represented by a 11 substring followed by a second substring decoding to y. Since n is the maximum grid size,
dlog2 ne bits are needed to represent y. Thus, a movement
by one grid towards right requires (2 dlog2 ne) bits. This
and other three scenarios are depicted in Fig. 1.
Similarly, B-Ey would represent a move diagonally down
by one grid and then a further movement down by y grids.
This process is represented by a (00 followed by dlog2 ne)
bits decoding to y. Substrings (10 or (01 represents a

123

F. Ahmed, K. Deb

horizontal movement by one grid, but each is integrated with

a further upward or a downward movement of y grids. The
upward movement is denoted as C-Dy and downward
movement is by C-Ey. Thus, to represent all (n - 1) grid
movements towards right from start to destination, a total of
n  12 dlog2 ne) bits would be required.
A careful thought will reveal that by the above representation scheme, we have eliminated a diagonally up
movement followed by a downward movement or A-Ey, as
the same can be achieved with a smaller path length and
path safety using a C-Dy or a C-Ey movement. Since in our
multi-objective formulation A-Ey would always be dominated by either C-Dy or C-Ey, we simply do not allow
A-Ey or B-Dy movements. Thus, this representation
scheme reduces the feasible search space and in turn helps
the optimization algorithm to work better.
Next, we discuss an end-fixing mechanism that goes
with the above representation scheme. It is mentioned
above that every path originates from the start point
(1, 1), but following a particular coding of (n - 1) genes
(as mentioned above) every path may not end up reaching
the destination point (n, n). To make sure that every path
reaches the destination point, we modify the (n - 1)-th
gene as follows. The path is followed by the first substring
of the (n - 1)-th gene, but the value of y from the second
substring is ignored. A (11) will take the path one grid
diagonally up, a (00) will take the path one grid diagonally
down, and (10) or (01) will take the path horizontally by
one grid. Thereafter, instead of using y to determine whether to go up or down, the path is simply joined to the
destination point. Since the second substring coded to
represent y for the (n - 1)-th gene is redundant, we use it
to represent any upward movement of the vehicle at the
start grid. We take a couple of examples to illustrate this
representation scheme.
In Fig. 2, two sample paths are shown on a 8 9 8 grid.
Thus, n = 8 in this example. The binary representation of
Path 1 is given as follows:
(000, 11-010, 11-000,
00-000, 10-000, 11)

10-000,

00-001,

Note that the seventh gene is split into two, of which the
first substring indicating a diagonally upward movement
from seventh to eighth grid is represented by the final (11),
and its associated substring (000) is shown at the beginning to indicate that there is no vertical movement of the
path at the start grid. Other movementsfor example, from
first to second grid there is an upward diagonal movement
(decoded by (11)) followed by two grids (decoded value of
(010) of upward movement (A-D2), and othersare
straightforward. Notice how the end-fixing on the rightmost column takes the path to the destination point. The
vertical movement of five grids is not coded in the string,

123

Fig. 2 Example paths with corresponding binary representation

along with a few non-represented steps in path

but is fixed algorithmically, making sure that the overall

path goes from the start to the end grid. On the same
manner, Path 2 can be represented by the following binary
string:
(101, 11-000, 01-000, 11-000, 00-001,
00-000, 11-001, 11)
Here, notice that at the start the vehicle moves by five
grids (decoded value of (101) shown at the front of the
string) at the start and then follows the other grid-wise
movements. At the final grid, the vehicle makes an upward
diagonal movement and reaches the destination point on its
own. Thus, in this path, no end-fixing is needed for the path
to reach the destination point.
In addition to the end-fixing involving the start and the
n-th columns, if at any intermediate column the vehicle
goes out of the search area (to the bottom of first row or to
the top of n-th row), it is brought back to the search area
using an edge-fixing procedure. In this event, the y component of the string is ignored and the vehicle is placed
randomly at a row with a linear probability distribution
having its maximum at the respective edge of the search
area to zero at the current row of the vehicle. In both Paths
1 and 2, this scenario did not occur and no edge-fixing was
required.
The advantage with the above binary coding is that a
binary-coded genetic algorithm can now be used directly to
evolve the population members. The usual single or multipoint crossover operators and the bit-wise mutation operator can be applied directly. Here, we have used a twopoint crossover and a bit-wise mutation operator for this
purpose; however, the mutation probability is linearly
reduced to 1/5-th of its initial value at the final generation.

Multi-objective optimal path planning using elitist NSGA-II

Moreover, since the search space is substantially reduced

by not allowing intuitively dominated paths to appear
through the representation scheme, the search is also likely
to be efficient. The fix-ups mentioned above will also make
the paths realizable. The drive towards feasible and eventually optimized paths will then be achieved by the way of
handling the objective function and constraints.
2.2 Mixed-coded genes
The above binary-coded representation makes a significant
improvement from the existing study (Sugihara and Smith
1999) in terms of reducing the dominated paths. However,
some thoughts will reveal that the coding can be made
more efficient by using a mixed binary-integer representation of a path.
Consider, for example, a C-D2 path going from (1, 1)
to (2, 1) to (2, 3). This path is dominated by another path
(A-D1) which makes an upward diagonal movement and
then move up vertically by one grid. This path is shorter
than C-D2 and since the vehicle is considered to be small
compared with the grid dimension, it is also as safe as
C-D2. Thus, on both account, A-D1 is no worse than C-D2
and hence C-D2 can be considered redundant in our search
process. To eliminate such dominated paths to appear
through our coding scheme, here we use one bit to represent upward or downward movement and a direct integer
value to represent the extent of vertical movement, if any.
We discuss this mixed-coding scheme in the following
paragraph.
The mixed-coding representation for a gene has two
components as before. But here, the first component has
just one bit. However, the meaning of the bit depends on
the associated second componentan integer in the range
[0, n] (both inclusive) representing vertical movement y. If
y = 0, a 1 or a 0 in the first substring would mean an
identical movement (C), that is, a movement of one grid
horizontally to the right. However, if y [ 0, then a 1 in the
first component would mean a move diagonally upwards
and then make (y - 1) vertical grids up. Thus, 1-y means
A-D(y - 1) for y C 1. Similarly, a 0 with a non-zero
y would mean a downward diagonal movement followed by
(y - 1) vertical grids down. Thus, 0-y means B-E(y - 1).
This mixed-coding does not allow the path to have a horizontal move followed by any vertical upward or downward
movement, as they were allowed in the binary-coded representation. Any up or downward movement must be
achieved with a non-zero y at any gene. This restriction does
not prevent any feasible non-dominated sub-path to be
represented. This substantially reduce the number of dominated solutions in the search space.
Path 1 shown in Fig. 2 would now be represented by the
following coding:

(0

1-3 1-1 0-0 0-2 0-1 1-0 1)

Like before, the first element (a zero) indicates no vertical movement at the start grid. Next mixed-gene indicates
that an upward diagonal movement and then a vertical
movement of (3-1) or two grids up. The path can then be
followed by the above-mentioned coding principle. At the
(n - 1)-th column, a 1 indicates a diagonal movement
upward. A 0 would have indicated a horizontal movement
to the n-th column. Since the seventh gene is a 1, the
vehicle moves one grid diagonally upward. From this
position, the end-fixing rule is applied and the vehicle is
moved up another five grids up to reach the destination
point. Path 2 can be represented as follows:
(5 1-1 1-0 1-1 0-2 0-1 1-2 1)
The edge-fixing and top-row-fixing rules are applicable
here as well. Although the above gene representation
makes the explanation of the path easier, in a GA, we
concatenate all bits together and form a (n - 1)-bit string
and all integer values are put together to form a (n - 1)size integer vector. The binary string is operated by the
standard single-point crossover and bit-wise mutation
operators and the integer-valued vector is operated by the
discrete versions of the simulated binary crossover (SBX)
(Deb and Agrawal 1995; Deb and Goyal 1998) and the
polynomial mutation operator (Deb 2001).
2.3 Integer-coded genes
It is clear from the above mixed-coded representation that
there is a still a small redundancy in the coding. A horizontal movement by one grid can be represented by using
either 0-0 or 1-0. As mentioned above, as long as y = 0,
the value of the preceding bit was not important. We remove
this redundancy in the mixed-coding scheme by completely
eliminating the binary coding and representing y of each
gene by an integer in the range [-(n - 1), (n - 1)]. Only in
the first gene, y values in the range [0, (n - 1)] is
allowed. Other end-fixing, edge-fixing and top-row-fixing
rules are applicable here as well. Thus, to represent Path 1
in Fig. 2, the following (n - 1)-variable vector would be
used:
(0, 3, 1, 0, -2, -1, 0)
Here, the first element represents the movement at the
start column and the final element represents the movement
at the (n - 2)-th column to (n - 1)-th column. The
movement from (n - 1)-th to n-th column has just two
options before the end-fixing operation. The vehicle can go
either upward and a possible vertical movement thereafter
to reach the destination point or a horizontal movement to
directly reach the destination point. This is because a
horizontal move followed by a vertical move always gets
dominated in our set-up by a direct diagonal movement. At
the end of string coding, we investigate the string for either

123

F. Ahmed, K. Deb

of the two scenarios and arrange the path accordingly. For

Path 1, we need to move diagonally one grid up, as the
current location of the vehicle is below the destination
point and then moves five grids vertically upward to reach
the destination point. Thus, the transition from (n - 1)-th
column to the n-th column is never coded in the proposed
integer-coding scheme and is fixed based on the above
argument. Path 2 can be represented by the following
integer-coding:
(5, 1, 0, 1, -2, -1, 2)
This path also requires a diagonal movement from
seventh column to the eighth column, as the location of the
vehicle at the seventh column is below to that of the destination point.
To demonstrate the difference between the three coding
schemes discussed above, let us consider a couple of
alternate paths shown in Fig. 2. Let us denote
ABCDENFGHIJK as Path 3 and ABCDEFGHOIJK as
Path 4. These paths can be represented by our binary-coded
representation scheme, but cannot be represented by our
mixed or integer-coded scheme. In Path 3, the part ENFG
takes the vehicle from the initial cell E to cell G in the next
column. This can always be replaced by EFG to reach the
same location with a shorter path length and path vulnerability (to be defined in a later section). Similarly, in Path
4, the part HI dominates the part HOI. Therefore, Paths 3
and 4 are dominated by another path and can never be on
the true Pareto-optimal front.
2.4 Absolutely positioned integer-coded genes
As a fourth representation scheme, the integer-valued
genes described in the previous subsection can be replaced
with the absolute location of the grids along each column.
Since a path always starts from the grid location (1,1), a set
of (n - 1) integer values in the range [1, n] will determine
the path exactly. The final transition from (n - 1)-th column to the n-th column can be made using the procedure
described in the previous subsection. A similar representation scheme was suggested elsewhere (Elshamli et al.
2004).
The integer-valued coding in the previous subsection is
sensitive to the initial part of the path. A small change in
the initial part may make a significant change in the final
part of the path, as every gene represents a relative change
in location from the previous gene. The absolute-positioned
integer-coding scheme described here does not have such a
difficulty. However, when we perform simulation runs on
some simpler scenarios, we noticed the inadequacy of this
absolutely positioned integer-coded representation. Due to
the insensitivity of each gene over the other, the relative
movement between one column to the next is not important
in this coding and this leads to a zig-zag path and the path

123

often results in infeasible paths. Since a genetic algorithm

is believed to work by finding lower-order building blocks
to form higher-order building blocks, the relative representation may provide a GA to focus on the initial part of
the path first and then may enable to form the remaining
path grid by grid till the complete path is formed. Such an
implicit ordering in forming the optimized path may turn
out to be an important property for the successful working
of a GA. Discouraged by these proof-of-principle experiments, we have not continued with the absolutely-positioned integer-coded gene representation scheme in this
study.

3 Objectives of the path planning problem

In this paper, we have considered three specific objectives
of the path planning problem. Of these, minimizing path
length and enhancing safety are the major objectives of the
study while path smoothness is used as a secondary
objective. We describe the objectives in the following
subsections.
3.1 Minimum path length
Every path planning problem must provide some importance to a shorter path. In this study, since a path is represented on a rectangular grid, we simply compute the path
length by summing the distance between consecutive grid
points along the path. The vehicle is always considered to
be traversing through the center of a grid cell and thus, a
distance between (i, j)-th grid to (k, l) grid (where
i; j; k; l 2 1; n) is computed as the Euclidean distance
between coordinate points (i - 0.5, j - 0.5) to (k - 0.5,
l - 0.5).
3.2 Minimum path vulnerability
The flip side of following a shortest path is the danger of
colliding with an obstacle. To estimate how precariously
the vehicle moves close to obstacles, we follow the commonly used potential field approach. The potential value in
a cell represents directly the difficulty in traversing through
it. Since each cell is allotted a potential value according to
obstacle distribution, we sum the potential of each cell
through which the path traverses to compute the total
vulnerability level.
The potential field approach is a widely studied method
(Ahuja and Chuang 1997; Connolly et al. 1990; Khatib
1986). It gained popularity because it displayed a great
reduction in computational demand when compared with
the configuration space approach. In this method, each
obstacle in the system is surrounded by a repulsive

Multi-objective optimal path planning using elitist NSGA-II

potential field while the goal location is surrounded by an

attractive potential field. A vehicle applies the resulting
force experienced at its location as a control input to the
driving system. The approach, however, suffers from a
major problem. This has to do with the formation of local
potential minima that can trap the vehicle before it reaches
its destination point. In this study, we have used practical
approach of assigning a bell-shaped (Gaussian) potential
field at every cell containing an obstacle. The obstacle
containing cells are given extremely high potentials. A
variance of 50 % of the cell size is used here; however,
higher or lower values of variance can be used to have an
effect of more or less conservative paths, respectively.

minimized or a constraint to be satisfied, the optimization

process will drive the search towards paths that are
obstacle-free.
However, in an extended version of our optimization
algorithms, in addition, we count the number of obstacles a
path crosses and call this number the Number of interfering cells. The multi-objective optimization problem is
modified as follows:

3.3 Path smoothness

Minimize Path length or path vulnerability;

Minimize Path length;

Minimize Path vulnerability
Subject to Potential  P0 ;

and the single-objective problem is modified as follows:

Subject to Number of interfering cells  0:

The smoothness of a path is evaluated by summing the

angle of each turn the vehicle has to make in traversing the
path. The smoothness of a trajectory is a desired attribute in
vehicle path planning (Kanayama and Hartman 1997).
Smooth paths decrease unnecessary curvature discontinuity
and possible stops and thus lower the possibility of slippage. It leads to lesser power consumption as well as lesser
travel time. Though smoothness is desired in a path, it need
not be used as an objective function in most path planning
optimization tasks, as the true optimized smooth path that
ignores path vulnerability and other maneuvering issues
may not be what a path planner would be looking for.
In this study, we consider path smoothness as a secondary goal of the path planning task. In a bi-objective
optimization framework of minimizing path length and
path vulnerability, path smoothness can be used as a tiebreaker or as a decision-making aid. We shall discuss the
interaction of path smoothness with the other two objectives in the next section.

4 Optimization methodologies
In this section, we discuss various optimization methodologies developed for solving the path planning problems
from various perspectives.

If, for any path, the number of interfering cells is greater

than zero, we declare the path to be infeasible. Furthermore, in our proposed approach, we use the parameter-less
single (Deb 2000) or multi-objective (Deb et al. 2002)
constraint handling technique, as the case may be, to
handle infeasible and feasible solutions. In the tournament
selection of comparing two solutions, if an infeasible
solution is compared with a feasible solution, the latter
always wins (Deb 2000, 2001). On the other hand, if both
solutions are feasible, the solution with a smaller objective
value or having a better non-dominated front or a larger
crowding distance value wins. Finally, if both solutions are
infeasible, the one with a smaller count on interfering cells
wins. For brevity, we do not present any simulation results
here but state that this extended approach works well for
cases with sparse obstacles, but fails to find a feasible path
in cases with a highly dense set of obstacles in a single
objective study. It has been observed that a better technique
is to penalize the objective(s) with the count of interfering
cells. Initial populations usually have infeasible solutions,
but the drive the search towards minimizing penalized
functions eventually brings the number of interfering cells
to zero. Based on these initial proof-of-principle studies,
we use the penalty function approach for the rest of the
paper here.

4.1 Obstacle avoidance procedures

5 Multi-objective path planning
It is intuitive to keep constraint from interfering with an
obstacle. In our algorithms, we have used a potential
function for this purpose. Each cell having an obstacle has
a high potential at its center and it reduces away from the
center. Thus, any path that crosses a cell having an obstacle
will accumulate a large potential value. Thereafter, using
the overall potential (called as path vulnerability above) in
the optimization problem either as an objective to be

In this paper, we modify the elitist non-dominated sorting

genetic algorithm or NSGA-II algorithm (Deb et al. 2002)
for this purpose; however, other evolutionary multiobjective optimization (EMO) algorithms can also be used
to obtain similar results. For details of NSGA-II, readers
are encouraged to refer to the original NSGA-II study (Deb
et al. 2002). Here we describe the procedure briefly.

123

F. Ahmed, K. Deb

Initially, a population is initialized randomly. Thereafter, the population members are sorted based on their
non-domination level in the population. In the so-called
non-dominated sorting, the first front members are nondominated members of the entire population. The second
front members are those that are non-dominated population
members for which first front members are excluded, and
so on. Each individual in a front is assigned a crowding
distance based on the distance of the neighboring solutions
from it. A solution has a large crowding distance if its
nearest neighbor lies far away from it in the objective
space. In a tournament of comparing two solutions, a
solution is considered better if it lies on a better front or has
a larger crowding distance value (thereby indicating an
isolated solution). After the mating pool is created, they are
used to create a new population using crossover and
mutation operator exactly the same way as they are done in
the case of single-objective GA. In the NSGA-II approach,
both parent and newly created populations are merged and
the combined population is sorted for their non-domination
level again. Population members lying on the better fronts
are chosen one at a time till the new population cannot
accommodate any new front. To maintain the population
size, all members of the last front, which could not be
accommodated as a whole, are compared for their crowding
distance valuesa measure of emptiness in a solutions
vicinity in the objective space (Deb et al. 2002)and the
ones having the largest crowding distance values are selected. This process is continued till the termination condition
is met.
Since an EMO algorithm, including NSGA-II, is stochastic in nature and is likely to produce different sets of
trade-off solutions in each run, we propose a more reliable
multi-objective optimization algorithm here. We perform
10 runs of the proposed NSGA-II starting with different
initial random populations for a particular choice of representation scheme until the specified termination criterion
is satisfied. Thereafter, we combine all 10 sets of nondominated fronts and present the overall non-dominated
front.

path smoothness wins. If two solutions being compared

have an identical non-domination rank and identical path
smoothness value, then the crowding distance value is
checked and the one with a higher crowding distance wins.
We also introduce the path smoothness criterion in another
way to emphasize preferred paths. In the post-processing
step, if two or more paths lie on the same point in the
objective space (length-vulnerability trade-off front), path
smoothness criterion is used to choose the single solution
having the largest smoothness value. Thus, in both cases,
the third objective of path smoothness is used as a decisionmaking aid.

5.1 Third objective as a decision-maker

Higher population sizes and the number of generations

may be essential to handle a higher grid size problem and a
more densely packed obstacles than that used in this study,
but we use the above setting for all problems irrespective of
density of obstacles and grid size. The bi-objective path
planning scheme elaborated above is generic and before
proceeding to show results on test cases, performances of
three representation schemes are compared.
The purpose of this paper was not to compare one EMO
algorithm with another method, but to demonstrate the
importance of an EMO algorithm in obtaining trade-off
path planning solutions. Hence, in this paper, we refrain

In the path planning problem, we have used both path

length and path vulnerability as two objectives. The number of interfering cells is used as a constraint and is handled
using the penalty function approach. To introduce the path
smoothness as a criterion to prefer smooth paths, we
modify the above NSGA-IIs selection scheme somewhat.
For two paths being compared in a tournament or in
choosing the final front members, solutions of the same
rank are checked for their smoothness values, instead of
their crowding distance values. The solution with a better

123

5.2 Parameter settings

A little thought will reveal the fact that the complexity of
the path planning problem would depend on the relative
location of the obstacles in a grid system. For an identical
grid system, a fixed number of obstacles may be placed in
such a manner that finding a valid path from start to finish
may be difficult or in some cases impossible. Thus, the
parameter settings for any path planning task are dependent
on placement of obstacles and also on the grid size.
Increasing the grid size requires more computation due to
increase in the number of variables. It is also observed that
higher obstacle density usually requires more function
evaluations to converge, but location of the obstacles is an
important matter.
Based on initial studies on many such path planning
problems prior to this study, we choose the following
NSGA-II parameter settings for all multi-objective test
cases of this paper (n is the size of the grid):
Population size (N) = 500,
Generations (t) = 500,
Generations (t) = 500,
Crossover rate (pc) = 0.9,
Crossover distribution index (gc) (Deb and Agrawal
1995) = 10,
Mutation rate (pm) = 1/n,
Mutation distribution index (gm) (Deb 2001) = 20.

Multi-objective optimal path planning using elitist NSGA-II

from comparing the chosen NSGA-II algorithm with any

other EMO algorithm; however, such an application
remains as an important future study.
5.3 Comparison of three representation schemes
In this section, we use the three representation schemes
discussed earlier, namely, the binary-coded, mixed-coded,
and integer-coded representation schemes. To compare,
here we choose a 32 9 32 obstacle arrangement shown in
Fig. 4. 173 obstacles were randomly kept in different cells.
As mentioned above, we choose a population of size 500
and run the bi-objective NSGA-II for 500 generations. Ten
runs with different initial populations are taken for each
scheme to obtain the trade-off solutions. Thereafter, all the
solutions of a particular scheme are combined to study
convergence behavior. Figure 3 shows all the trade-off
points obtained in the final trade-off fronts obtained in ten
different runs for three different schemes. It can be seen that
the obtained trade-off frontiers for integer coding are clearly
better than that obtained with the mixed coding or binarycoded representation schemes. Binary-coded scheme is
found to be consistently performing poorly. The trade-off
solutions for integer-coded scheme dominate the other two.
Similar results are also observed on a few other test problems, which are not shown for brevity.
The superior performance of integer-coded representation scheme is explained here. As discussed in Sect. 2, the
search space in integer-coded scheme to represent the same
problem is substantially smaller than the other two
schemes. Thus, it is relatively easier to find valid and
meaningful paths using the integer-coded scheme. This
leads to improvement in the performance of the optimization algorithm. Three paths are marked on the
Fig. 3 Trade-off points
obtained using three
representation schemes

non-dominated trade-off front, corresponding to the minimum path distance solution, minimum path vulnerability
solution, and an intermediate compromised solution in
Fig. 4. It shows how the intermediate solution negotiates
the distance and safety and becomes a compromised
alternative solution.
5.4 Bi-objective path planning
Having found that the integer-coded representation scheme
produces a consistently better performance compared with
other two potential coding schemes, we use the integercoded scheme from now on. In this section, we take two
sample problems with grid sizes 16 9 16 and 32 9 32,
respectively, to demonstrate the trade-off paths obtained by
the proposed modified NSGA-II approach. First, we consider a benchmark problem shown in Fig. 5b which helps
to compare our method directly with another existing
study. The study also used a genetic algorithm approach.
This scenario was the most complex scenario considered in
the path planning study, while, as we shall see later, the
current scenario is much simpler than some of the later
scenarios which our proposed algorithm could handle.
The obstacle profile is as shown in Fig. 5b along with
three paths obtained by the integer-coding representation
scheme. The modified NSGA-II is used with the path
smoothness criterion used as a decision-making aid as
discussed earlier with 500 population members and run for
500 generations. Ten runs with different initial populations
are taken and the final solution set of each run is combined.
Thereafter, we find the non-dominating points from this set
to locate the trade-off front. The trade-off front is shown in
Fig. 5a and the paths marked on it correspond to minimum
path length, minimum path vulnerability, and an

450
Binary Coded
Mixed Coded
Integer Coded
Path 1
Path 2
Path 3

400

Path Difficulty

350

300

250

200

150

100
45

50

55

60

65

70

75

80

85

90

Path Length

123

F. Ahmed, K. Deb
32
30
28
26

Path 2

24
22

Path 3

20
18
16
14
12
10
8

Path 1

6
4
2
0

10 12 14 16 18 20 22 24 26 28 30 32

Fig. 4 Obstacles on a 32 9 32 grid and few solution paths obtained

using integer coded representation scheme

intermediate compromised solution, respectively. It is

interesting to note that how the minimum path length
solution nudges close to some of the obstacles to reduce the
overall path length and how the minimum path vulnerability solution is keeping the path away from the obstacles
to ensure adequate safety. The intermediate solution is a
compromise between these two extreme solutions. It is clear
that the compromised solution lies on a knee region of this
trade-off front. When there exists such a knee point on a
trade-off frontier, it is often a preferred solution (Deb and

Gupta 2001), as there is not enough motivation to move

away from the knee point due to unfavorable trade-offs.
The solution reported in the existing GA study (Castillo
and Trujillo 2005) is also marked on Fig. 5a for a comparison purpose. The study did not show multiple paths
clearly with interconnecting grids; however, we extract
the minimum path length configuration from the reported
figure and compute path vulnerability using our computational method. Figure 5a shows that the earlier reported
solution (indicated by Path X) gets dominated by our
minimum path length solution and two other solutions,
thereby indicating efficacy of our proposed procedure.
We now discuss the results obtained with this multiobjective optimization algorithm on a sparse obstacle scenario presented in Fig. 6b. Here also, we use the integercoding representation and a population of size 500 and run
the modified multi-objective NSGA-II algorithm for 500
generations to obtain trade-off solutions. Ten runs are taken
and final non-dominated set of points is found. The trade-off
solutions are shown in Fig. 6a. The ranges of path vulnerability and path length values obtained by this study show
that a diverse set of solutions are found by the proposed
multi-objective optimization algorithm. Figure 6b shows
three paths picked from the obtained trade-off frontier. The
solid line represents the shortest path, the dotted line shows
the least vulnerable path, and the dashed line shows an
intermediate path. It seems that the chosen intermediate path
makes a good compromise between the two extreme solutions (not too long and not too close to obstacles) and can be
a preferred solution for an application. Interestingly, the
shortest and least difficult paths are similar to those obtained
by the single-objective optimization study in Fig. 7a.

Fig. 5 Obstacles on a 16 9 16
grid and obtained results

(a)

123

(b)

Multi-objective optimal path planning using elitist NSGA-II

Fig. 6 Results from multiobjective optimization study are
shown

(a)

(b)

Fig. 7 Single-objective path

planning results for path length
and path vulnerability: (a) ten
single-objective runs are taken
and path corresponding to best
run is shown, (b) singleobjective runs fail to find any
obstacle-free path in the dense
obstacle scenario. Path 1
solution corresponds to path
length objective while Path 2 is
for vulnerability

(a)

5.5 Single-objective path planning

In the previous subsection, we have shown results from the
bi-objective study first; here we show results from a singleobjective study in which integer-coded representation
scheme is used and either the path length or the path vulnerability is minimized one at a time. The number of
interfering cells is used as a constraint. Parameter settings
are identical to that used in NSGA-II run, except that
gc = 2 is used (Deb and Agrawal 1995).

(b)

In Fig. 7 we have shown two single-objective optimization resultsone with path length as an objective and the
other with path vulnerability as an objectivefor twoobstacle scenarios. In the sparse obstacle scenario on a
16 9 16 grid (Fig. 7a), both individual optimized paths are
shown. It is interesting to observe that the minimum path
length solution (solid line) traverses almost in a straight
line from start to finish. This is expected particularly if the
straight line path is feasible. On the other hand, the minimum path vulnerability solution (dotted line) takes a

123

F. Ahmed, K. Deb

somewhat longer path, traversing through a region with no

or a few obstacles. This is again expected, as a minimum
vulnerable path is the one that avoids the obstacles as much
as possible. The trade-off between the two objectives
becomes clear from this example. It is worth mentioning
here that the same problem was solved using a multiobjective optimization algorithm in Sect. 4 involving both
path length and path vulnerability as objective functions in
a bi-objective set-up and two very similar extreme
solutions were obtained by the proposed multi-objective
optimization algorithm as well. The outcome of both
single- and multi-objective optimizations supports each
others results and provides an idea that the obtained
solutions may be close to being truly optimal.
We shall now discuss the results obtained using the
single-objective optimization on the larger grid size problem having a more dense obstacle scenario and also having
64 9 64 grid (shown in Fig. 7b). There are 791 obstacles
in the grid and the problem is more difficult than the previous one. The single-objective GA (with path length as the
sole objective) fails to find a valid path after 500 generations run with 500 population members. The best solution
obtained in 10 runs collides with one of the obstacles.
Similarly, the single-objective optimization to minimize
path vulnerability objective alone also fails to find any
feasible path in 10 runs with identical parameter settings.
The best solution for this objective collides with three
obstacles. Since no valid path is obtained by both singleobjective optimizations, we show the least constraint-violated solutions from each run in Fig. 7. The same problem
is solved later (in Sect. 6) with a multi-objective formulation. As shown in Fig. 11a, feasible paths are found in
80 % runs and each solution does not collide with any of
the 791 obstacles.
Based on these two simulation studies and our experience in solving a few other scenarios, we conclude that the
proposed single-objective GA can be utilized to find the
extremities of trade-off front for simple cases with a low
density of obstacles. A single-objective GA is unable to
perform well on large grid size and dense obstacle cases.
However, the multi-objective GA proposed here with both
path length and path vulnerability is able to find a feasible
path in different grid sizes and on both sparse- and dense
obstacle scenarios.
The reason for superior performance of multi-objective
GA lies in the maintenance of diversity in the population.
Diversity preservation is an important matter for genetic
operations to work properly. When multiple conflicting
objectives are used, a multi-objective GA does not allow
solutions to converge towards any individual optimized
solution. Thus, a natural diversity is maintained in such
an application, besides the ability of the algorithm to
find multiple trade-off solutions. A single-objective GA

123

attempts to find a single optimized solution and it becomes

easy for the algorithm to lose its diversity, particularly in
complex and constrained problems. Robot path planning
problem is highly constrained and in the presence of a
dense set of obstacles it becomes difficult to find a feasible
solution. Diversity preservation is a key to the creation of
feasible solutions and the proposed multi-objective GA is
able to perform well in the difficult version of the problem.
5.6 Three-objective path planning
So far, we have used a two-objective NSGA-II in which in
addition to path length and path vulnerability objectives,
we have introduced the path smoothness objective as a
decision-maker. Thereafter, we have shown that a singleobjective GA is not able to maintain diversity in the population on its own and fails to find a feasible solution in
complex constrained problems.
In this subsection, we study the effect of a truly threeobjective NSGA-II with path smoothness as the third
objective. For this purpose, we used the original NSGA-II
procedure with crowding distance as the sole diversity
preserving metric (Deb et al. 2002). We consider the same
sparse obstacle scenario as considered in Fig. 6 and use the
integer-coding scheme to find trade-off paths. Again, a
population of size 500 is used for 500 generations. The path
smoothness is denoted by the total turn (summation of
angles turned in the entire path) which is minimized to
improve smoothness. Ten runs are taken and non-dominated three dimensional trade-off front is found from the
combined solution set. Figure 8a shows the obtained threedimensional trade-off front and Fig. 8b shows its projection on length-vulnerability axes for a comparison with the
two-objective front obtained earlier in Fig. 6a.
The trade-off between the three objectives are clear
from the figure. A solution marked H in both figures
indicates that although this solution is dominated with
respect to path length and path vulnerability objectives, it is
the best solution for the path smoothness objective.
Path smoothness is an objective, which on its own may
not lead to an interesting path. To avoid twists and turns, an
optimized path smoothness solution may be the one that
prefers concatenation of a few straight-line segments (to
avoid turns) to reach the destination. Figure 9 shows the
path corresponding to path H marked in Fig. 8.
Although in this simple scenario path H seems to be a
reasonable solution, in more complex obstacle scenario the
smoothest path may not even pass through the intermediate
portion of the grid, but instead may prefer to lie along grid
axes, depending on obstacle scenario. Such a path may
become a trivial solution and of not much interest to an
user. But in the presence of other two objectives, path
smoothness may reveal solutions that are smooth as well as

Multi-objective optimal path planning using elitist NSGA-II

(b)

(a)

Fig. 8 Three-objective optimization results with smoothness as the third objective

the trade-off between one, two, and three-objective optimization approaches that we have observed here.
1.

2.

3.

Fig. 9 Path H corresponding to optimized path smoothness is shown

reasonable with respect to other two objectives. When the

obtained trade-off solutions are projected on the path-vulnerability plane (Fig. 8b), the non-dominated front is found
to be similar to that obtained using the two-objective study.
This gives us confidence about the accuracy of both twoand three-objective NSGA-II optimization approaches.
However, although a few solutions with smooth paths
are now obtained, this happens at the cost of the minimal
path length solution which was found in the two-objective
study. With a focus on all three objectives, the effect of a
finite population size makes an emphasis to the intermediate portion of the trade-off front, thereby losing the
individual optimized solutions in some occasions. This is

Single-objective optimization cannot keep diversity

due to a singular focus and hence cannot find a feasible
solution in complex obstacle scenarios.
Three-objective optimization can find a three-dimensional trade-off frontier, but this may come at the
expense of some individual optimized solutions.
Two-objective optimization with third objective as a
decision-maker seems to be adequate in obtaining the
entire set of trade-off solutions.

Thus, if path smoothness is not an absolute necessity, we

suggest using the modified two-objective NSGA-II
approach, whereas if the all three objectives are important,
the three-objective original NSGA-II provides a good set of
trade-off solutions.
Having established the fact that (i) obstacle avoidance
constraint handling is better through the penalty function
approach, (ii) multi-objective optimization is better than
single-objective path planning approach, and (iii) a twoobjective approach coupled with path smoothness as a
decision-making aid is a better approach than considering all
three objectives in the optimization problem, we are now
ready to test the algorithm on increasingly difficult cases.

6 Path planning under extreme conditions

An extreme test of a path planning algorithm proposed here
is made for attempting to solve problems having a large

123

F. Ahmed, K. Deb

grid size and a dense set of obstacles. When adequate

spaces are available, multi-objective optimization algorithms can be used to find a set of trade-off solutions.
However, when there are more obstacles than the available
space in a grid, finding a single feasible path is often a
challenging task. Thus, in this section, we employ the twoobjective NSGA-II proposed above, but terminate the run
as soon as a feasible path is found.
We create increasingly higher grid size and increasingly
dense obstacle scenarios in which obstacles are placed
randomly on a two-dimensional grid. However, we make
sure that there exists at least one feasible path from start to
the destination point, so that the application of an optimization algorithm is meaningful. To achieve this, we first
generate a random valid path on the entire grid. Then, we
randomly create obstacles with p0 probability of occurrence of obstacle in every cell except those lying on the
above-generated path. Hence, a probability of p0 = 1
would mean that all grid cells have obstacles on them
except the cells lying on the random path initially created.
Such a problem would be a very difficult path planning
problem to solve for any algorithm. To make the problems
of different size, we use grid size that is varied to 8, 16, and
32. The probability p0 is varied from 0.1 to 1, as tabulated
in Table 1.
For all problems, the integer-coded representation
scheme is used. For each problem, a population of size 200
is used and the NSGA-II is run for a maximum of 500
generations. The comparatively low computational cost is
chosen here since the main objective of the study is only to
find a feasible solution and not the Pareto-optimal solutions. For each problem, 100 runs are taken and the percentage of successful runs in finding a feasible path from
start to destination point, mean minimum path length
obtained, and the median of generation numbers in which
the first feasible path is obtained are noted. The first entry
indicates the likelihood of finding a feasible path, the
second entry denotes the ability to further minimize path
length for feasible path while the third entry denotes how
quickly the first feasible path is found on an average.
Figure 10 shows a few of the obtained solutions.
Figure 10d, for example, shows how in a large grid size
and among a large density of obstacles, the proposed
methodology is able to locate a feasible path, a task which
looks apparently daunting.
There are several observations to be made from the
table. First, as expected, as the grid size increases, the
success percentage reduces; however, the proposed NSGAII is able to still find the feasible path in all problems with
8 9 8 and 16 9 16 grids. The number of generations
needed to find a feasible path is also reasonably small.
The second matter to note is that with 16 9 16 grid
problems; the proposed methodology is able to find a

123

Table 1 Path planning under different volumes of dense obstacles

# of
obstacles

Success
percentage
(%)

Median
generations

Mean
min.
path
length

Grid
size

Obstacle
density,
p0

898

0.1

100

9.90

0.2

10

100

10.49

0.3

15

100

10.49

0.4

20

100

10.49

0.5

26

100

11.07

0.6

31

100

11.07

0.7

40

100

11.07

0.8

43

100

11.07

0.9

50

100

11.07

1.0

54

100

11.07

0.1

24

100

21.21

0.2

47

100

22.14

0.3

65

100

23.3

0.4
0.5

80
115

100
100

18
28

22.34
22.37

0.6

150

100

22

22.57

0.7

164

100

25

22.3

0.8

286

100

40

22.13

0.9

214

100

41

22.48

1.0

233

100

46

22.4

0.1

81

100

47.08

0.2

173

100

16

53.4

0.3

253

97

38

53.72

0.4

386

85

73

53.61

0.5

470

36

230

54.38

0.6

556

74

144

53.64

0.7

666

41

295

53.9

0.8

767

52

298

54.31

0.9
1.0

870
967

39
30

378
339

54.13
54.19

16 9 16

32 9 32

feasible path in a problem having a density of obstacles as

large as 233/(16 9 16) or 91 %. This is a remarkable
result. The third aspect to note is that as the density of the
obstacles increase, the problem gets harder to solve and
more generations are needed to locate a feasible path.
Third, for 32 9 32 grid, Table 1 shows that the success
percentage is less than 100 % for obstacle density of 0.3 or
more. Since a population of size 200 is used, it is not
adequate for 32 9 32 grid problems. A larger population
size would have yielded better performance, but instead of
showing results with a higher population size for 32 9 32
grid problems, we use a population size of 500 for higher
grid problems.
To investigate how the proposed bi-objective procedure
will fair on even larger grid size problems, finally we apply
our methodology on a few 64 9 64 and 128 9 128 grid

Multi-objective optimal path planning using elitist NSGA-II

Fig. 10 Feasible paths for cases
shown in Table 1

(a)

(b)

(c)

(d)

(a)

(b)

Fig. 11 Feasible paths for large

grid size problems obtained
using the proposed bi-objective
procedure

123

F. Ahmed, K. Deb

problems with 500 population size and 500 generations.

Figure 11a shows two runs and the obtained feasible paths
in both cases. Clearly, when 4,979 obstacles are scattered
randomly on a two-dimensional grid (as shown in Fig. 11b,
it is impossible to obtain a feasible path manually.
Although many other solutions may be possible in this
case, our proposed algorithm is able to find one feasible
path.

7 Conclusions
In this paper, we have dealt with several issues of a
two-dimensional path planning problem having various
complexities using a soft computing approach.
First, we have suggested three path representation
schemes for optimal path planning tasks to be solved by
using a multi-objective genetic algorithm. The flexibility in
their usage enables an efficient application of an evolutionary multi-objective optimization technique for the path
planning problem. Of the three schemes, the integer-coded
gene representation scheme that directly codes the relative
movements of a vehicle from one column to the next has
been found to yield best results.
Second, it has been observed that instead of considering
three different criteriapath length, path vulnerability, and
path smoothnessas objectives, the use of the first two as
objectives and the third as a decision-making aid is a better
option. This approach makes a good compromise between
a single-objective approach which has been found to suffer
from lack of diversity in a GA population and a threeobjective NSGA-II approach which suffers from its
concentration in the intermediate portion of the threedimensional front.
Third, the above multi-objective consideration has been
found to be better than a single-objective treatment of say
path length or path vulnerability alone.
Fourth, compared with existing studies, the proposed
two-objective approach has been shown to be more accurate in finding feasible and trade-off paths. Particularly, the
proposed approach has been demonstrated to work on
large-sized grids having a dense set of obstacles (as high as
91 % of the space), thereby demonstrating the efficacy of
the proposed multi-objective optimization and novel representation scheme. Such results are rarely demonstrated in
the literature.
While the existing past studies limited their studies to
small-size grids, here we have shown that A GA-based
two-objective approach can find feasible paths in grids as
large as 128 9 128 and having as large as 4,979 obstacles
occupying more than 90 % of the space by obstacles. These
results remain as the hallmark achievement of this paper.

123

This paper can be extended to include non-monotonic

paths along both axes which will make it a more practical
scheme for real-world applications. Three-dimensional
path planning can also be tried with minor modifications in
the proposed integer-coded representation scheme. Nevertheless, extensive simulation results of this study indicate
that the current multi-objective technique is a reliable tool
for a point-to-point, off-line path planning task and must be
pursued further.
Acknowledgments The study is supported by JC Bose National
Fellowship to Prof. K. Deb and Department of Science and Technology, Government of India, under SERC-Engineering Sciences
scheme (No. SR/S3/MERC/091/2009). Funding from Academy of
Finland under grant 133387 for executing a part of this research is
also appreciated.

References
Ahmed F, Deb K (2011) Multi-objective path planning using spline
representation. In: Proceedings of the IEEE international
conference on robotics and biomimetics (IEEE-ROBIO 2011),
Piscatway
Ahuja N, Chuang J (1997) Shape representation using a generalized
potential field model. Pattern Analysis and Machine Intelligence.
IEEE Transactions on 19(2):169176
Al-Sultan KS, Aliyu MDS (2010) A new potential field-based
algorithm for path planning. Journal of Intelligent & Robotic
Systems 17(3):265282
Bisse E, Bentounes M, Boukas EK (1995) Optimal path generation
for a simulated autonomous mobile robot. Autonomous Robots
2(1):1127
Burchardt H, Saloman R (2006) Implementation of path planning
using genetic algorithms on mobile robots. In: Proceedings of the
World Congress on evolutionary computation (WCCI-2006),
pp 18311836
Castillo O, Trujillo L (2005) Multiple objective optimization genetic
algorithms for path planning in autonomous mobile robots.
International Journal of Computers, Systems and Signals 6(1):
4863
Castillo O, Trujillo L, Melin P (2007) Multiple objective genetic
algorithms for path-planning optimization in autonomous mobile
robots. Soft Computing-A Fusion of Foundations, Methodologies and Applications 11(3):269279
Choset H (1996) Sensor based motion planning: the hierarchical
generalized Voronoi graph. PhD thesis, Citeseer
Connolly C, Burns J, Weiss R (1990) Path planning using laplaces
equation. In: Proceedings of IEEE international conference on
robotics and automation. IEEE, pp 21022106
Deb K (2000) An efficient constraint handling method for genetic
algorithms. Comput Methods Appl Mech Eng 186(24):311338
Deb K (2001) Multi-objective optimization using evolutionary
algorithms. Wiley, Chichester
Deb K, Agrawal RB (1995) Simulated binary crossover for continuous search space. Complex Systems 9(2):115148
Deb K, Goyal M (1998) A robust optimization procedure for
mechanical component design based on genetic adaptive search.
Transactions of the ASME: Journal of Mechanical Design
120(2):162164
Deb K, Gupta S (2011) Understanding knee points in bicriteria
problems and their implications as preferred solution principles.
Engineering Optimization 43(11):11751204

Multi-objective optimal path planning using elitist NSGA-II

Deb K, Agrawal S, Pratap A, Meyarivan T (2002) A fast and elitist
multi-objective genetic algorithm: NSGA-II. IEEE Trans Evol
Comput 6(2):182197
Elshamli A, Abdullah HA, Areibi S (2004) Genetic algorithms for
dynamic path planning. In: Proceedings of IEEE Canadian
conference on electrical and computer engineering (CCECE-04),
pp 677680
Ge SS, Cui YJ (2002) Dynamic motion planning for mobile robots
using potential field method. Autonomous Robots 13(3):207222
Gerke M (1999) Genetic path planning for mobile robots. In:
Proceedings of the American control conference, vol 4. IEEE,
pp 24242429
Glasius R, Komoda A, Gielen S (1995) Neural network dynamics for
path planning and obstacle avoidance. Neural Networks 8(1):
125133
Hwang Y, Ahuja N (1992) Gross motion planninga survey. ACM
Computing Surveys (CSUR) 24(3):219291
Kanayama Y, Hartman B (1997) Smooth local-path planning for
autonomous vehicles. Int J Robot Res 16(3):263
Khatib O (1986) Real-time obstacle avoidance for manipulators and
mobile robots. Int J Robot Res 5(1):90
LaValle S (2006) Planning algorithms. Cambridge University Press,
Cambridge
Lozano-Perez T, Wesley M (1979) An algorithm for planning
collision-free paths among polyhedral obstacles. Commun ACM
22(10):560570
Murrieta-Cid R, Tovar B, Hutchinson S (2005) A sampling-based
motion planning approach to maintain visibility of unpredictable
targets. Autonomus Robots 19(3):285300
Oriolo G, Ulivi G, Vendittelli M (1997) Fuzzy maps: a new tool for
mobile robot perception and planning. Journal of Robotic
Systems 14(3):179197

Pratihar D, Deb K, Ghosh A (1999) Fuzzy-genetic algorithms and

time-optimal obstacle-free path generation for mobile robots.
Engineering Optimization 32:117142
Sauter J, Matthews R, van Dyke Parunak H, Brueckner S (2002)
Evolving adaptive pheromone path planning mechanisms. In:
Proceedings of the first international joint conference on
autonomous agents and multiagent systems, part 1. ACM,
pp 434440
Sugihara K, Smith J (1999) Genetic algorithms for adaptive planning
of path and trajectory of a mobile robot in 2D terrains. IEEE
Transactions on Information and Systems 82(1):309317
Xiao M, Michalewicz Z (2000) An evolutionary computation
approach to robot planning and navigation. Soft Computing in
Mechatronics, pp 117128
Xue Q, Sheu PCY, Maciejewski AA, Chien SYP (2010) Planning of
collision-free paths for a reconfigurable dual manipulator
equipped mobile robot. Journal of Intelligent & Robotic Systems
17(3):223242
Yang SX, Meng M (2000) Real-time collision-free path planning of
robot manipulators using neural network approaches. Autonomous Robots 9(1):2739
Zhang Q, Li H (2007) MOEA/D: A multiobjective evolutionary
algorithm based on decomposition. Evolutionary Computation,
IEEE Transactions on 11(6):712731
Zitzler E, Laumanns M, Thiele L (2001) SPEA2: improving the
strength Pareto evolutionary algorithm for multiobjective optimization. In: Giannakoglou KC, Tsahalis DT, Periaux J,
Papailiou KD, Fogarty T (eds) Evolutionary methods for design
optimization and control with applications to industrial problems. International Center for Numerical Methods in Engineering (CIMNE), Athens, Greece, pp 95100

123