Robotics and Autonomous Systems 123 (2020) 103320

Contents lists available at ScienceDirect

Robotics and Autonomous Systems

journal homepage: www.elsevier.com/locate/robot

A Boundary Node Method for path planning of mobile robots

∗
R.A. Saeed a , , Diego Reforgiato Recupero a , Paolo Remagnino b
a
Department of Mathematics and Computer Science, University of Cagliari, Cagliari, Italy
b
The Robot Vision Team, Kingston University, London, UK

article info a b s t r a c t

Article history: In this paper we propose a new method for solving the path planning problem in a static environment
Received 7 September 2018 to find an optimal collision-free path between starting and goal points. First, the grid model of the
Received in revised form 9 September 2019 robot’s working environment is constructed, and then the potential value of the grid cells is calculated
Accepted 9 October 2019
based on the new proposed potential function. This function is used to guide the robot to move toward
Available online 15 October 2019
the desired goal, it has the lowest value at the goal position and the value is increased as the robot
Keywords: moves further away. Second, we developed an efficient method, called the Boundary Node Method,
Robot path planning to find the initial feasible path. In this method, the robot is simulated by a nine-node quadrilateral
Path optimization element, where the centroid node represents the robot’s position. The robot moves in the working
Simulation model environment toward the goal with eight-boundary nodes based on the potential value of the boundary
Autonomous mobile robot
nodes. The initial feasible path is generated from a sequence of waypoints that the robot has to traverse
Potential function
as it moves toward the goal point without colliding with any obstacles. However, the proposed method
Boundary Node Method
Path Enhancement Method can generate the path safely and efficiently, but the path is not optimal in terms of the total path
length. Therefore, in order to construct an optimal or near-optimal collision-free path, an additional
method, called the Path Enhancement Method, is developed. Finally, the cubic spline interpolation is
adopted to generate a continuous smooth path that connects the starting point to the goal point.
The proposed method has been tested in several working environments with different degrees of
complexities. The results demonstrated that the proposed method is able to generate near-optimal
collision-free path efficiently. Moreover, we compared the performance of the proposed methods with
the other path planning methods in terms of path length and computational time. The results revealed
that the proposed method can solve the robot path planning problem more efficiently. Finally, in order
to verify the performance of the developed method for generating a collision-free path, experimental
studies were carried out on the real robot.
© 2019 Elsevier B.V. All rights reserved.

1. Introduction addition, path planning approaches are useful for repeatable tasks
in static environments where optimality is essential (e.g. indus-
The aim of path planning is to find a collision-free path for trial applications) [7]. These factors make the path planning an
a mobile robot to move from a starting point to a goal point interesting and challenging subject for researchers [6].
in a given working environment based on certain optimization The path planning problem started around the sixties, but the
criteria, such as, the walking distance, the walking time, the interest in the path planning area for mobile robot grew after the
energy consumption, and so on [1–3]. It is expected that the work of authors in [8] after which many methodologies have been
robot reaches the final destination point safely through the short- proposed [2,9]. The existing methods are mainly categorized into
est walking path within the minimum computational time. Path classical and heuristic path planning [2,6]. The classical methods
planning has been widely applied in many robotic applications to include cell decomposition, potential field method, subgoal net-
perform various tasks that humans could not accomplish in sev- work and road map [10]. They involve finding a set of defined
eral domain such as nuclear facilities [4], for space exploration [5], steps to search for a path starting from an initial position to a
for rescue mission, landmines and enemies in war field [6]. In goal position. In classical methods only deterministic actions are
considered [10,11]. However, it has been found that the classical
methods have some disadvantages such as the high computa-
∗ Correspondence to: Department of Mathematics and Computer Science,
tional cost, trapping into local minima, and high time complexity
University of Cagliari, Via Ospedale 72, 09124 Cagliari, Italy.
E-mail addresses: [email protected] (R.A. Saeed),
in high dimensions [1,6,9]. As classical search methods fail to
[email protected] (D.R. Recupero), [email protected] find exact solutions, many heuristic methods have been proposed,
(P. Remagnino). i.e. Genetic Algorithm (GA) [7,12], Particle Swarm Optimization

https://doi.org/10.1016/j.robot.2019.103320
0921-8890/© 2019 Elsevier B.V. All rights reserved.
2 R.A. Saeed, D.R. Recupero and P. Remagnino / Robotics and Autonomous Systems 123 (2020) 103320

4. In computational complexity theory, path planning is clas-

Abbreviations sified as a non-deterministic polynomial time problem NP-
BNM Boundary Node Method complete, and the required computational time grows ex-
PEM Path Enhancement Method ponentially as the complexity of the path-planning prob-
lem increases [9].
IFP Initial Feasible Path
5. There are many methods that use random operation to pro-
C Workspace
duce a set of solutions for each independent run. Then, in
Cobs Space occupied by obstacles
order to find the optimal, all of these different solutions are
Cfree Free Space, Cfree = C − Cobs selected, combined and replaced. This process requires a lot
Cs Start Point, Cs (xs , ys ) ∈ Cfree of computational time, therefore, reducing the variation of
Cg Goal Point, Cg (xg , yg ) ∈ Cfree the final solution is important. [2].
CBGC Boundary Grid Cells, CBGC ⊂ Cobs
Cr Robot Position, Cr (xr , yr ) ∈ Cfree Based on the limitations and research gaps, as previously
explained above, we investigate a novel off-line path planning
p1 (t) Current Location, p1 (t) = [x1 (t) ; y1 (t)]
method for a mobile robot in a two-dimensional (2D) static envi-
p2 (t) Updated Location, p2 (t) = [x2 (t) ; y2 (t)]
ronment. In the developed method, that we called the Boundary
sy Variation Potential Value between p(4) Node Method (BNM), the robot is simulated by a nine-node
and p(6) quadrilateral element, where the centroid node represents the
sx Variation Potential Value between p(2) robot’s location and it moves with eight-boundary nodes in the
and p(8) working environment. The robot is exploring the environment
pbest Best Node Position with the help of the node’s potential value at each location, where
d Distance between the centre and the the potential value is calculated based on the proposed potential
edge of the obstacle, d = 0.5 unit function. In this method, we have considered only 8-generated
e Motion Directions, e(u), (u = 1...8) grid points that are overlapping with the eight-boundary node,
w Set of waypoints rather than considering all the generated points which lead to less
computational time. Moreover, the proposed method is capable
of generating an efficient path for a mobile robot safely and
quickly and it can also overcome the local minima problem. We
(PSO) [6], Artificial Neural Networks (ANNs), Ant Colony Opti- also developed an additional method, that we called the Path
mization (ACO) [9,13], and Fuzzy Logic (FL) [9]. Surveys works Enhancement Method (PEM), to construct an optimal path by
in [1,13] showed that the heuristic path planning methods are reducing the number of waypoints (w ) and path length.
computationally more efficient in terms of path distance, obstacle The term BNM has already been used in one of the meshless
avoidance, and elapsed time [9]. Heuristic methods attempt to boundary integral equation methods that combine the Moving
find a good solution to the path planning problem in a short Least squares (MLS) interpolation with the Boundary Integral
amount of time, but these methods are not guaranteed to pro- Equations (BIEs) to solve boundary value problems in potential
vide an optimal solution [10,11]. The combinatorial path plan- theory and engineering.
ning methods in continuous space can solve many path planning To evaluate the contribution of the proposed approach, a com-
problem and construct optimal solution efficiently [10,14–17]. parison study has been conducted between the proposed method
Many of the existing methods for robot path planning are and the other path planning methods, namely, PSO, GA, A-Star,
able to find a path for the robot, but in most of the cases, the and Artificial Potential Field (APF ). The comparison results are
quality of the generated path is not accurate enough or their
presented and discussed in Section 5.3. The PSO algorithm is
efficiency is not sufficient [18]. Researchers have always been
widely used in path planning problems [19,23,29] as it is fast and
seeking for a better solution to improve the performance of the
simple [29], easy to implement [30], and a powerful means [14]
existing path planning methods. A list of goals that researchers
to solve mobile robot navigation problems [23]. Moreover, A-Star
of several earlier works have pursued is the following: improve
algorithm is an effective and direct method to search paths [22],
the accuracy [5,19,20], improve the efficiency [18,21], increase
which was used for many path planning applications [31], and
safety [4,5,22], increase the capability [23], reduce the processing
it is mainly employed on an environmental grid [20]. In addition,
time [24,25], overcome the non-reachable goal problem [26], pass
through narrow passages [27], overcoming the local-trap prob- GA is known as a robust optimization method among the existing
lems, and improve the quality of planned paths [18]. However, approaches for robot motion planning problem [18]. By taking
several important gaps and limitations still need to be addressed, advantage of its strong optimization ability, the GA has been
as outlined in the following: widely used in previous study to generate an optimal path [32].
The potential field method is a fast [13,29,31], simple [26,31],
1. In several works, the computational time is still too high easy to implement [31] method, and it has good results for
because the process of a large number of unnecessary path planning [13]. The disadvantages of artificial potential field
points. Moreover, the search for an optimal path might not method are related to the local minima problem that it can
succeed [2]. incur [26,29,31].
2. In many previous studies, the considered environments Therefore, the main contributions of this paper can be sum-
are relatively too simple and unusual for testing the effi- marized in the following points:
ciency of the proposed method [28,29]. Obviously, the path
planning problem in a complex environment can be very 1. The proposed method, BNM, is capable of finding the initial
difficult and it is still a challenging issue. feasible path (IFP) for a mobile robot without colliding with
3. As the range of a robot’s application is expanding over time, any obstacles even if the complexity of the environment is
the complexity of the path planning problem and working increased.
environment are increasing as well. For this reason, it be- 2. The proposed method uses an optimization technique
comes much more difficult to find an optimal path within based on the lowest potential value to accelerate the robot
a reasonable amount of time [2]. to find the path safely and quickly in reasonable time.
 R.A. Saeed, D.R. Recupero and P. Remagnino / Robotics and Autonomous Systems 123 (2020) 103320 3

3. An additional method, PEM, is developed to find an optimal integrating fuzzy theory and genetic algorithm. To solve path
or near optimal path from the IFP by reducing the number planning problem, researchers in [34] suggested another method
of waypoints and the overall path length. named SACOdm Based on Simple Ant Colony Optimization Meta-
4. The proposed method does not work through random op- Heuristic (SACO–MH). One of the main contributions of SACOdm
erations and there is no uncertainty in generating points, is the inclusion of memory capabilities to the artificial ants to
which leads to finding the final solution for the problem prevent stagnation. Another contribution of this method is the
without variation in solution. use of the fuzzy cost function to evaluate the best path. An
5. The proposed method generates a safe path for a mobile additional methodology has been proposed in [16] by integrating
robot to navigate in a complicated environment within a the Artificial Bee Colony (ABC ) algorithm with the evolutionary
relatively short computational time. programming algorithm. In this method ABC algorithm has been
6. The concept involved in the proposed method is simple and studied and applied to generate a feasible path, then the fea-
can be applied in a grid environment efficiently. sible path is enhanced by using an evolutionary programming
7. The computational time required to solve path planning algorithm.
problem by using BNM does not increase significantly with Additionally, many researchers built on top of existing meth-
the increase of the environment’s complexity. ods to improve their performance and to overcome their limita-
8. The comparison between the developed approach and tions. In fact, authors in [22] proposed an improved version of
other path planning methods reveals that the BNM can A-Star algorithm to overcome inherent drawbacks of the original
solve the path planning problem effectively and efficiently A-Star. One of the main improvements of the proposed method is
in terms of path computational times and path length. that the local path is planned before the next search in the current
node’s neighbourhood. A-Star algorithm calculates heuristic func-
The remainder of this paper is structured as follows: Section 2 tion’s value at each node on the work area and checks adjacent
includes background work within the domain of mobile robot nodes in order to find the optimal solution with zero proba-
path planning. Section 3 introduces the path planning problem bility of collision. However, its time complexity is too high. To
in a static and completely known environment. In Section 4, the overcome this problem, [24,25] introduced a number of improve-
details of the novel method and potential function are described ments to the A-Star algorithm to reduce the computational time
with several illustrative examples. In Section 5, the application and to increase the overall performance. Another improvement
results of the developed method is presented and discussed for is the minimization of the resulting path length by reducing the
several working environments with different complexity. It also number of local paths. With the aim of reducing the chances of
presents and discusses the comparison results between BNM collisions between robot and obstacles, researchers in [20] pre-
and the other path planning method. The experimental study is sented a new approach of path planning technique, where they
presented in Section 6. Final conclusions and prospective future assumed that the virtual obstacle’s size increased approximately
research are provided in Section 7. (2n + 1) times the size of the cell in the workspace. The capa-
bilities of GA for solving the path planning problem for mobile
2. Related work robots in static and dynamic environments have been investi-
gated by several researchers, who have extended the method.
The path planning problem has attracted many researchers’ For example, the authors in [12] proposed a new fitness function
attention due to the uncertainties, complexities and real-time for GAs whereas authors in [21] proposed Knowledge-based GA
nature of the problem [28], and it has been a very active area of and authors in [18] proposed an adaptive GA. Furthermore, [35]
research over the last few decades. In the literature, the problem presents the new variant of GA using the binary codes through
of path planning for mobile robots has been widely discussed and matrix. Calculation of artificial potential values is another solution
various solutions and approaches have been proposed to solve it. to obtain a collision-free path. This method was first used [36]
For example, the authors in [2] proposed a new methodology to for a collision-free robot motion planning problem. This method
solve the path planning problem in two steps. First, they generate is based on attraction and repulsive values which are considered
the IFP based on the surrounding point-set (SPS), which refers two fields produced by the target point and obstacles, and the
to a set of points that surround the obstacles. Then, they applied robot is considered as a moving object in these fields. The robot
the path improvement algorithm to get the optimal path by using moves toward the target based on the negative slope of the po-
the outcome of the first step. As stated in [2], this method has a tential function. The problem with this approach is that the robot
low-level of randomness that reduces the variation of solutions, can get stuck in local minima of the potential field [18]. Conse-
and also this method is able to generate points in narrow or small quently, various techniques have been proposed to avoid the min-
spaces in the map. Another method is the Bacterial Potential Field ima, i.e. authors in [37] tried to solve the problem by using har-
(BPF ), developed by [15] to compute an optimal path for a mobile monic potential functions around obstacles. In order to solve the
robot in a real-world scenario with static and dynamic obstacles. problem of non-reachable goals with obstacles nearby (GNRON)
As reported in [15], the path planning with the BPF allows a robot in potential field method, a new repulsive potential function is
to navigate in an autonomous way without being trapped in local proposed by [26]. In order to overcome the local minima and
minima. heavy computational time for robotic path planning, probabilistic
Furthermore, there are a number of researchers who com- sampling-based algorithms such as the rapidly-exploring ran-
bined algorithms to improve path planning performance. For dom tree (RRT ), and the probabilistic roadmap (PRM) algorithms
example, the authors in [14] presented a hybrid meta-heuristic are introduced due to their remarkable practical performance
GA–PSO algorithm for mobile robot navigation to find an op- and strong theoretical properties [38,39]. Such algorithms work
timal path between starting and ending point in a grid envi- by computing multiple distributed random points in the free
ronment. The proposed algorithm avoids time complexity and workspace and connect them to construct a tree or graph, after
premature convergence in conventional GA and PSO algorithms. that a search method is used to find a path [22]. In RRT , the most
The hybrid GA–PSO is used to generate the IFP, then a cubic B- important factor that affects the overall efficiency of path plan-
spline technique is applied to construct a near-optimal collision- ning is how to select a tree to extend or connect. In the literature,
free path. To reduce the complexity of robot path planning, au- the Rapidly-exploring Random Trees (RRT ) algorithm has been
thors in [33] proposed a hierarchical path planning method by widely used. [40] proposed a novel learning-based multi-RRTs
4 R.A. Saeed, D.R. Recupero and P. Remagnino / Robotics and Autonomous Systems 123 (2020) 103320

(LM–RRT ) approach for robot path planning in complex environ- moves toward the target. For such a reason, the proposed method
ments with narrow passages. As stated in [40], this approach is known as off-line path planning, and generates the entire path
can guarantee the efficiency of global path planning and enhance to the goal before the motion begins.
the local space exploration ability of each tree. The investigation In the proposed method, the robot is simulated by a nine-
of the shortest path with the minimum time required for the node quadrilateral element p(q), (q = 1...9). The centroid node
global path planning is carried out in [23] by using Modified p(5) is considered as the robot’s location and the other nodes
PSO. Authors in [41] has made a comparison between PSO and Q- (p(1 → 4)&p(6 → 9) represent the eight-boundary nodes which
learning, a Reinforcement Learning-type algorithm. For the single are distributed uniformly around the robot’s location, as shown
robot case, they showed that the final performance obtained with in Fig. 1a. The robot moves forward and changes its direction
Q-learning approach is very similar to the one obtained with PSO. based on the potential values and features of boundary nodes.
Some optimal path planning algorithms are presented in [27] for The potential value E(q), (q = 1...9) for the robot and boundary
navigating mobile robot among obstacles and weighted regions. nodes are equivalent to the potential value of the corresponding
These algorithms can search an optimal path and also intelligently generated points in the workspace. All the visited waypoints w ,
rotate the robot configuration to pass through narrow passages. starting from Cs and ending at Cg , represent the obtained initial
Researchers provided great effort for real application to solve feasible path IFP. Fig. 1b shows the motion directions, and Fig. 1c
path planning problem for mobile robots, and they proposed shows the exploration location in the workspace.
many new methods. For example, the authors in [7] presented
preliminary results of the application of two-Kinect cameras sys- 4. Proposed method
tem on a two wheeled indoor mobile robot for off-line optimal
path planning and execution. To solve path planning problem for This section describes the proposed method used in the pre-
rovers, authors in [5] presented a new algorithmic improvement. sented study to find the optimal or near optimal collision-free
In this study, they proposed OUM–BD over the Ordered Upwind path. The proposed method consists of four main steps:
Method (OUM) to include a bi-directional search. They stated
that the proposed method OUM–BD is faster than the existing 1. Construct a 2D grid model of the robot’s working envi-
OUM. Authors in [6] proposed a multi-objective path planning ronment, and then calculate the potential value of the
algorithm based on improved PSO for robot navigation where grid cells based on the new proposed potential function.
the robot often involves various danger sources, such as a fire This function has the lowest potential value at Cg and the
in a rescue mission, landmines and enemies in the war field. potential value is increasing as the robot moves further
For emergency evacuation simulation, authors in [19] proposed away.
a new path planning approach. In this approach, the Extended 2. Develop an efficient method, BNM, to generate the IFP for
Social Force Model (ESFM) is combined with the improved ABC a mobile robot.
algorithm to improve the efficiency of crowd evacuation. Another 3. Develop an additional method, PEM, to construct an opti-
approach called Grid-Based Random Tree Star (GB–RRT ) has been mal or near optimal path from IFP, (as the obtained IFP is
developed by [4] to provide minimum dose path for occupa- not optimal path in terms of the total path length).
tional workers in nuclear facilities in complex environments. The 4. Generate a continuous smooth path that connects the start-
probabilistic roadmap (PRM) method has been applied by [42] to ing point to the goal point by using cubic spline method.
optimize the walking path, and to reduce the radiation exposure
of the staff in a radioactive environment of nuclear facilities. Fig. 2 shows an overview of the four steps above mentioned.
A research study revealed that in the radioactive environment In the next subsections we will detail each of the four steps.
of nuclear facilities, the proposed method has a good effect on
path-planning, and it can make a route in a very short time. 4.1. Modelling of the workspace

3. Problem formulation In the proposed method, all the grid cells of the given
workspace meet the following equation:
In this section, we state the path planning problem, that is n
∑ m
∑
moving the robot from a starting position and tracking it through C = C (x, y) (1)
all the intermediate waypoints until it reaches the goal position in x=1 y=1
a two-dimensional environment with static obstacles. The robot
does not have to collide any obstacles and must optimize the path where n and m represent, respectively, the width and height
from starting position to the goal position. of the workspace, and C (x, y) represents the grid cells in the
Let us consider a 2D workspace C = R2 for a mobile robot, workspace. After constructing the workspace model, the potential
the region of space occupied by obstacles is denoted by Cobs , value of each grid cell is calculated based on the new proposed
and the obstacle-free region is represented by Cfree = C − Cobs . potential function, as explained in the following paragraphs.
The continuous workspace is divided into square grid cells. The
grid cells have integer coordinates in the form C (x, y) ∈ C , with 4.1.1. Potential function PF
1 ≤ x ≤ n, and 1 ≤ y ≤ m. A given cell can either correspond to a This section presents a new proposed potential function to
navigable area C (x, y) ∈ Cfree or to a space occupied by obstacles calculate the potential value of grid cells in the workspace C .
C (x, y) ∈ Cobs . Each grid cell C (x, y) in Cfree has a potential value The procedure of calculating the potential value E(k), with (1 ≤
E(x, y) ∈ E, which is calculated according to the potential func- k ≤ N) and N is the number of grid cells (N = n × m) based
tion. The boundary grid cells of the workspace are also considered on the proposed potential function is illustrated in Algorithm 1.
as obstacles. Grid cells are represented by CBGC ⊂ Cobs . The robot Two examples of the new proposed potential function are shown
position in the workspace is denoted by Cr (xr , yr ) ∈ Cfree , and the in Fig. 3a and b. In these figures, the cell’s colour represents the
starting point Cs (xs , ys ) ∈ Cfree and the goal point Cg (xg , yg ) ∈ Cfree . potential value, i.e. the blue cell corresponds to cells with the
We assume that all the information related to the workspace is lowest potential value whereas the yellow cell corresponds to
known in advance, as well as the obstacles which are assumed cells with the highest potential value. As shown in the figures, the
to be fixed, meaning that they do not change while the robot shape of the potential function is conic and the global minimum
 R.A. Saeed, D.R. Recupero and P. Remagnino / Robotics and Autonomous Systems 123 (2020) 103320 5

Fig. 1. A nine-node quadrilateral element (a) along with its motion directions (b) and exploration location in the workspace (c).

of the total potential is located at the goal position. Because the Table 1
lowest potential value of the goal point, it attracts the robot. Characteristics of three different workspace scenarios of example.

In Algorithm 1, the computed E(k) represents the potential Workspace No. Cs (x, y) Cg (x, y) Workspace [cells] Obstacles [cells]

value of each grid cell C (h, k), with (h = 1...2), and (k = 1...N) 1 (5,5) (38,45) 2226 770
2 (5,5) (38,45) 2226 345
in the workspace C . The minimum potential value is formulated
3 (5,5) (65,105) 7303 904
at the goal point Cg (xg , yg ). The distance between the start point
Cs (xs , ys ) and the goal point Cg (xg , yg ) is represented by D, where
the slope of a straight line D is denoted by m. The distance
between the goal point Cg (xg , yg ) and surrounding point C (h, k) shown in Fig. 4 are divided into square grid cells, where each cell
in the workspace is represent by dp (1, k). is considered as either an obstacle Cobs or a non-obstacle Cfree . The
potential value of the grid cells in the Cfree is calculated based on
Algorithm 1 The calculation of potential value of grid cells in the the proposed potential function, as illustrated in Algorithm 1. The
workspace. grid cells of the workspace use a different colour to differentiate
1: Inputs: between Cfree and Cobs , where the black cells represent Cobs , and
Cg and C (h, k), (h = 1...2), and (k = 1...N) the coloured cells represent the potential value in the Cfree . The
2: Initialize: safety zone is represented by a number of grey square grid cells
E(k) ← 0, (k = 1...N) of the same size (1 × 1) square unit around the obstacles.
3: D = sqrt((xs − xg )2 + (ys − yg )2 ) For the first (see Fig. 4a) and second (see Fig. 4b) designed
4: m = ((ys − yg )/(xs − xg )) scenario, the obstacles represent about 34.6% and 15.5% of the
5: c = (ys − m ∗ xs ) workspace, respectively. In the third scenario (see Fig. 4c), a
6: ll = sqrt(m2 + b2 ), (b = −1) more complex environment with a higher number of obstacles
7: for k = 1 to N do
of different size is considered, and here the obstacles represent
8: dp (1, k) = sqrt((C (1, k) − xg )2 + (C (2, k) − yg )2 )
about 12.4% of the workspace. After constructing the workspaces
9: L(1, k) = m × C (1, k) + b × C (2, k) + c
with obstacles and calculating the potential value of the grid cells,
10: dl (1, k) = |L(1, k)|/ll
11: E(k) = sqrt(dl (1, k)2 − dp (1, k)2 ) the robot’s path needs to be determined.
12: end for

4.2. Proposed method BNM

4.1.2. Obstacles representation
After constructing the workspace model for a mobile robot, a
number of static obstacles are distributed at different locations The BNM method consists of three steps:
in the workspace. To reduce the complexity of the proposed
1. Simulate the robot,
method, we assume that the obstacles form a set of square cells
2. Exploration process, and
(1 × 1 unit). The centre of the obstacle’s cells are denoted by
3. Obstacle avoidance
a matrix Cobs (h, l), with (h = 1, 2), and (l = 1...O), where O
represents the number of obstacles. The distance d between the
centre and the edge of the obstacle is constant, d = 0.5unit. As
4.2.1. Simulate the robot
the robot might move very close to the obstacle, they should keep
a certain margin for safety. In this study, to avoid the possibility of In the simulated model, the nodes are denoted by p(q), (q =
overlapping the paths traced by the robot with obstacle boundary, 1...9), and their locations are formulated by Eq. (2). At itera-
we have created a safety zone around the obstacles. tion t, the current location of nodes denoted by p1 (t). The x,
An example of three different workspace scenarios with dif- y coordinates of the nodes’ location represent by two vectors
ferent obstacle layouts are shown in Fig. 4. The characteristics of x1 (t) = (x11 , x12 , . . . , x19 ) and y1 (t) = (y11 , y12 , . . . , y19 ), respec-
these three scenarios are illustrated in Table 1. The workspaces tively. Therefore, the current location of nodes p1 (t) is formed by
6 R.A. Saeed, D.R. Recupero and P. Remagnino / Robotics and Autonomous Systems 123 (2020) 103320

Fig. 2. Flow diagram of the proposed method.

vertically concatenating x1 (t) and y1 (t), p1 (t) = [x1 (t); y1 (t)]. 4.2.2. Exploration process
⎧ In each iteration t, the current location of the robot and
⎪
⎪ x, y q=5 boundary nodes move in one particular direction. The new up-
v x y), (x, y + vy ), (xvx , y), (x, yvy )
, dated location of nodes p2 (t) are calculate according to the fol-
⎪
(x +
⎪
⎪
⎨
lowing equations:
p(q) = q = 2, 4, 6, and 8 (2)
⎪(x + vx , y + vy ), (xvx , y + vy ), (xvx , yvy ), x2 (t) = x1 (t) + ∆x (3)
⎪
⎪
⎪
⎩ (x + v , yv )
⎪
x y q = 1, 3, 7, and 9 y2 (t) = y1 (t) + ∆y (4)
p2 (t) = [x2 (t); y2 (t)] (5)
where x and y represent the coordinate of the robot location
pr . Moreover, vx and vy represent the horizontal and vertical where x2 (t) and y2 (t) represent the coordinate of the new up-
distances between pr and boundary nodes, vx = vy = 1unit. dated nodes’ location.
The values of ∆x and ∆y are computed by using Algorithm 2.
The boundary nodes p1 can only move in eight-possible directions
This algorithm is used to find the new updated location p2 (t) of
e(u), (u = 1...8) (see Fig. 1b), which we will explain in the next the current location of nodes p1 (t). In this algorithm, the value
subsection. gx and gy represent the distance between the current location of
 R.A. Saeed, D.R. Recupero and P. Remagnino / Robotics and Autonomous Systems 123 (2020) 103320 7

Fig. 3. The potential value of grid cells in the workspace in 3D view with contour plot. The size of the workspace is 50 × 50, and the Cg is located at (a) (40, 45)
and (b) (25, 25). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

Fig. 4. 2D model of the robot’s workspaces. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this
article.)

the robot pr (xr , yr ) and the goal point Cg in x and y directions, Algorithm 2 Compute the values of ∆x and ∆y
respectively. The variables sx and sy represent the variation of
1: Inputs:
the potential value between p(2)&p(8) and between p(4)&p(6),
Cg , pr (t)
respectively, and the value of sx and sy are calculated by using
2: E(q), (q = 1...9) ← E
Eqs. (6) and (7).
3: sx , sy ← Eqs. (6) and (7)
sx (t) = E(p1 (1, 8), p1 (2, 8)) − E(p1 (1, 2), p1 (2, 2)) (6) 4: gx = xr (t) − xg
5: gy = yr (t) − yg
sy (t) = E(p1 (1, 6), p1 (2, 6)) − E(p1 (1, 4), p1 (2, 4)) (7)
6: if sx < 0 then
∆x and ∆y have the same sign as the variation of the potential 7: compute gx = −1 ∗ gx
value (both positive or both negative). The coefficients α and β 8: end if
are constant, and these two coefficients will influence the conver- 9: if sy < 0 then
gence behaviour. The distance between pr (t) and Cg is decreasing 10: compute gy = −1 ∗ gy
step by step until the robot reaches the global minimum at the 11: end if
goal position. 12: if gx = 0 then
The proposed method uses an optimization technique based 13: compute ∆x = 0 and ∆y = β ∗ gy
on the lowest potential value to accelerate the robot to find the 14: else if gy = 0 then
path and yield to fast convergence. Among all boundary nodes, 15: compute ∆x = α ∗ gx and ∆y = 0
the node with the lowest potential value is chosen as the best 16: else
position and denoted by pbest . At each iteration t, the robot update 17: compute ∆x = α ∗ gx and ∆y = β ∗ gy
its position to the best position pbest . The boundary nodes, their 18: end if
position and potential value, guide the robot to move toward the
goal location and help the robot to avoid obstacles, which we will
discuss in the next section.
8 R.A. Saeed, D.R. Recupero and P. Remagnino / Robotics and Autonomous Systems 123 (2020) 103320

4.2.3. Obstacle avoidance minimum Algorithm 3 is used, for instance the workspace with a
In the workspace that contains no obstacles, the robot will U-shaped obstacle as shown in Fig. 8. As shown in Fig. 8a, the
reach the goal point along a straight line from any starting point. robot starts to move at position (3, 15) toward the goal point
As obstacles exist, the robot interfere with obstacles when the at (23, 3). Similarly, in Fig. 8b, the robot moves from (23, 15) to
distance between the robot and the obstacles is less than the (3, 3). The robot uses two different modes while moving in the
distance d. Therefore, the robot and boundary nodes require to simulated environment, namely the ‘‘normal mode’’ and the ‘‘lo-
avoid obstacles and to change their moving direction by selecting cal minimum recovery mode’’. In the normal mode, for iteration
a new position in the Cfree . (t)t =1,...,6 , as the robot moves from the point p1 (t) toward the
To explain the obstacle avoidance, consider an example shown point p2 (t), the line between p1 (t) and p2 (t) does not intersect
in Fig. 5. The boundary nodes p(1 → 4) and p(6 → 9) are with the obstacles (see step (1) Fig. 8). In order to check the
generated around the robot position p(5) by using Eq. (2). As feasibility of the path represented by each line segment between
shown in Fig. 5a, the red object represents the robot, and the blue corresponding points in p1 (t) and p2 (t), we create a new row ma-
objects represent the boundary nodes. At iteration t, the robot trix (chk(q),(q=1...9) ). The value of each element of the row matrix
and boundary nodes are changing their positions from the current is equal to ‘‘0’’ or ‘‘1’’. At iteration t, for the 1st element (q =
position (see Fig. 5a) to the new updated position (see Fig. 5b) 1), if the line between the first node of the simulated model
by using Eqs. (3), (4), and (5). As a result, the nodes p(7), p(8), in p1 (t) and p2 (t) intersects obstacles, then the value chk(q) =
and p(9) interfere with the obstacles (see Fig. 5b). Therefore, the 1, otherwise chk(q) = 0. The same procedure is then repeated
robot needs to investigate the workspace to find next position for the 2nd element (q = 2), 3rd element (q = 3) until the last
without colliding obstacles. In this case, the robot will move in y- element(q = 9). In the normal mode, the values of chk(q),(q=1...9)
direction either to upward or to downward direction. The motion are equal to ‘‘0’’, and the robot travels with the help of Algorithm
direction depends on the value of sy (see Fig. 5c). The robot moves 4. In the recovery mode, the robot switches to Algorithm 3, as
backward when sy (t) is negative, and it moves forward when sy (t) shown in steps (2 → 9) in Fig. 8. The proposed method gives
is positive. the highest priority to the obstacle avoidance processes and the
Furthermore, in order to demonstrate how the robot avoids lowest priority to the potential value. In t = 7, as the robot
the obstacles and changes its motion direction with the help moves forward from p1 (t) to p2 (t), the line segment connecting
of boundary nodes, consider an example shown in Fig. 6. As corresponding nodes (3, 6, and 9) intersects the obstacles (see
illustrated in Fig. 6a and b, in the iterations t = 1 − 4, the Fig. 8). In this case the values of chk(3) , chk(6) , and chk(9) are equal
robot starts to move from Cs and it moves forward toward the to ‘‘1’’, then the robot moves to the right (see Fig. 8a) or to the
goal point from p1 (t) to p2 (t) using Eqs. (3), (4), and (5). At each left (see Fig. 8b) with the help of the Algorithm 3. Once the robot
iteration, all obstacles in the working environment are examined comes out from a local minimum, it can move smoothly again by
for possible collisions with the direct path from p1 (t) to p2 (t). using Algorithm 4 (see step (10)). In the step(10) the BNM method
As the robot moves toward the goal point Cg in the iteration gives the highest priority to the potential value until the robot
t = 5 − 6, nodes p(1), p(2), and p(3) interfere with obstacles reaches the goal point. As it can be seen in Fig. 8, the robot is not
(see Fig. 6c). This implies that the robot can only move in the blocked by the U-shaped obstacles, it always finds the path (if it
y-direction, either upward (when sy is positive) or downward exists) to reach the final destination point.
(when sy is negative). The next position of the robot must be in The proposed potential function is similar to the attractive
upward direction, because the value of sy is positive (E(6) > E(4)). potential field in the sense that both guide the robot to move
The same procedure is repeated for the iteration t = 7 − 10 toward the desired goal location, but differ in calculating the
by shifting the robot upward until the robot passes the block of potential value E (see Algorithm 1), where the potential value
obstacles, as shown in Fig. 6d and e. For the iterations t = 11 − 16, E(1,k) is calculated by using Eq. (8).
the BNM method directs the robot to move forward (see Fig. 6f
and g) until the robot reaches its final destination point at the Cg E(1,k) = f (Cg , C(h,k) ) (8)
(see Fig. 6h).
As illustrated in the figure, in the step (1) the robot starts to
Suppose that the long horizontal set of obstacles block the
move toward the goal until it collides with obstacles. When the
robot path as demonstrate in Fig. 7a. As the robot moves toward
simulated robot detects a collision, the position of the interfered
the goal point, nodes p(1), p(4), and p(7) interfere with obstacles.
points in the boundary nodes is computed by using Eq. (9).
Therefore, the robot needs to change its motion direction along
the x-direction to avoid the obstacles. The motion direction de- p(t ,h) = [x2 (t); y2 (t)], p(t ,h) ∈ R2 : p(t ,h) = (pr ∩ Cobs ) (9)
pends on the value of sx . The robot moves to the right when sx (t)
The new updated location of nodes p2 (t) is calculated accord-
is positive, and it moves to the left when sx (t) is negative. In this
ing to Eqs. (10) and (11), to avoid obstacles, as follow:
case, the nodes at p(8) and p(2) have the same level of potential
value E(8) = E(2). This implies that the variation of the potential x2 (t) = x1 (t) + f (E(1,k) , p(t ,h) , p1 , p2 , Cobs ) (10)
value between p(2) and p(8) is equal to zero sx = 0. To solve this
y2 (t) = y1 (t) + f (E(1,k) , p(t ,h) , p1 , p2 , Cobs ) (11)
problem, the robot moves along both direction (see Fig. 7b). As
shown in the figure, nodes p(7), p(8), and p(9) move one step to For the step (2) to step (9), the proposed method gives the high-
the left and nodes p(1), p(2), and p(3) move one step to the right est priority to the obstacle avoidance processes and the lowest
at the same time. Two temporary sets, which can be described priority to the potential value. Afterward, in the step(10) the BNM
as a ‘‘waiting list’’, of visited grid cells on the left and right-side method gives the highest priority to the potential value until the
are stored. As the simulated robot reaches the end of obstacles in robot reaches the goal point. As it can be seen in Fig. 8 the robot
left-side earlier (see Fig. 7c), then the BNM method chooses the is not blocked in the U-shape obstacles, it always finds the path
stored set of the left-side and disregards the stored points of the (if it exists) to reach the final destination point.
right-side. In this study, the BNM method is used to find IFP for a mobile
In order to solve the local minima problem, we introduced an robot to move from Cs to Cg in the workspace without colliding
algorithm (see Algorithm 3). This algorithm executes a sequence with any obstacles. The IFP is generated from a set of waypoints
of steps that pulls the robot out of a local minimum. In order to w that the robot visits before reaching the final destination point.
illustrate the steps required by the robot to come out of a local For better clarity, the waypoints are connected into a continuous
 R.A. Saeed, D.R. Recupero and P. Remagnino / Robotics and Autonomous Systems 123 (2020) 103320 9

Fig. 5. Obstacle avoidance in a static environment using BNM. (For interpretation of the references to colour in this figure legend, the reader is referred to the web
version of this article.)

Fig. 6. Demonstration of robot exploration in a two dimensional environment using BNM.

Fig. 7. Workspace contains long horizontal set of obstacles that block the path of the robot.

path. The line segment that connects two waypoints in sequence mobile robot by using the proposed method is summarized in
is represented by Pl,l+1 , and the length of all line segments that Algorithm 4.
connect all waypoints sequentially to each other is representing According to Algorithm 4, the robot starts to move at the point
the length of IFP. A complete path IFP is formed by concatenation Cs (xs , ys ) toward the goal point Cg (xg , yg ). The current nodes’ loca-
of all inter-line segments Pl,l+1 , 1 ≤ l ≤ w − 1 as follows: tion p1 (t) of all nodes p(q), (q = 1...9) at iteration t is formulated
IFP = [P1,2 , P2,3 . . . , Pw−1,w ]. The main steps to find IFP for a by Eq. (2), where the x and y coordinate of the robot location
10 R.A. Saeed, D.R. Recupero and P. Remagnino / Robotics and Autonomous Systems 123 (2020) 103320

Fig. 8. Simulation results of local minima problem solution using the proposed algorithm for a simple environment with U-shape obstacle.

Algorithm 3 local minima problem location of nodes p2 (t) = [x2 (t); y2 (t)] is calculated by Eqs. (3)–
(5). The variation of the potential value sx and sy is calculated by
1: Inputs: Cobs , sx , sy , p1 (t), p2 (t)
using the Eqs. (6) and (7). Afterwards the line segments between
2: Check line segments between p1(q),(q=1...9) (t) and
p1 (t) and p2 (t) check for feasibility. So, if collision is not found,
p2(q),(q=1...9) (t) for feasibility
then a new set of E(q), (q = 1...9) and pbest need to be calculated,
3: If line between p1(q) (t) and p2(q) (t) interfered Cobs then chk(q) =
as previously explained. Subsequent, the current location p1 (t)
1 otherwise chk(q) = 0
updates to the new location p2 (t), and the robot pr (t) updates its
4: Construct matrix chk(q),(q=1...9)
position to the best position pbest . The proposed method stores the
5: while sum(chk) > 0 do
robot’s location pr (t) in a waypoints w list. On the other hand, if
6: if chk(1) , chk(2) , and chk(3) = 1 then
the line segments between p1 (t) and p2 (t) collides with obstacles,
7: p2x (t)=p2x (t)-c1 , c1 is constant
another updated location p2 (t) needs to be found, as previously
8: if sy > 0 then p2y (t)=p2y (t)+c2 otherwise p2y (t)=p2y (t)-
explained in Section 4.2.3. This procedure will continue until the
c2
mobile robot reaches the final destination point at Cg (xg , yg ) or
9: repeat steps 2, 3, and 4
the maximum number of iterations is reached.
10: update p1y (t) ← p2y (t)
11: store p5 in a way-points w list
12: end if Algorithm 4 BNM
13: if chk(1) , chk(4) , and chk(7) = 1 then 1: Inputs:
14: p2y (t)=p2y (t)-c2 , c2 is constant Cs , Cg , Cobs , and C (x, y), (x = 1...n, y = 1...m),
15: if sx > 0 then p2x (t)=p2x (t)+c1 otherwise p2x (t)=p2x (t)-c1 maximum iteration number M
16: repeat steps 2, 3, and 4 2: Initialize:
17: update p1x (t) ← p2x (t) p1 ← Eq. (2), x = xs , y = ys
18: store p5 in a way-points w list 3: E(k), (k = 1...N) ← Algorithm 1
19: end if 4: E(q), (q = 1...9) ← E
20: if chk(7) , chk(8) , and chk(9) = 1 then 5: pbest ← minimum E(q)
21: p2x (t)=p2x (t)+c1 6: sx , sy ← Eqs. (6) and (7)
22: if sy > 0 then p2y (t)=p2y (t)+c2 otherwise p2y (t)=p2y (t)- 7: while (xr ̸ = xg or yr ̸ = yg within a M) do
c2 8: p2 (t) ← Eqs. (3), (4) and (5)
23: repeat steps 2, 3, and 4 9: sx , sy ← Eqs. (6) and (7)
24: update p1y (t) ← p2y (t) 10: Check the line segment between p1 (t) and p2 (t) for
25: store p5 in a way-points w list feasibility
26: end if 11: if p(t) interfered with Cobs then
27: if chk(3) , chk(6) , and chk(9) = 1 then 12: p2 (t) ← Obsticle Av oidance
28: p2y (t)=p2y (t)+c2 13: end if
29: if sx > 0 then p2x (t)=p2x (t)+c1 otherwise p2x (t)=p2x (t)-c1 14: E(q), (q = 1...9) ← E
30: repeat steps 2, 3, and 4 15: pbest ← minimum E(q)
31: update p1x (t) ← p2x (t) 16: p1 (t) ← p2 (t)
32: store p5 in a way-points w list 17: pr (t) ← pbest
33: end if 18: pr (t) in a way-points w list
34: end while 19: end while
35: return p1 (t), p2 (t), w 20: IFP ← way-points w list
21: Popt ← Algorithm 5
22: U ← Eq. (12)
23: End
pr at the first iteration coincide with the xs and ys of the start
point Cs (xs , ys ). The node with the lowest potential value among
Time complexity is the computational complexity that esti-
all boundary nodes is chosen as the best position and it is denoted mates the run-time of an algorithm. In the developed method, the
by pbest , where the potential value E(q), (q = 1...9) of nodes is computational time to find a set of waypoints w of the IFP can be
computed by using Algorithm 1. For iteration t, the new updated calculated by summing the time needs for each line from 8 to
 R.A. Saeed, D.R. Recupero and P. Remagnino / Robotics and Autonomous Systems 123 (2020) 103320 11

18 in Algorithm 4. The time complexity of the developed method As shown in Fig. 10a, these line segments did not collide with
BNM can be analysed as following: when the size of simulated obstacles. Therefore, u2 of the second line segment is placed in Cg .
model is q(q = 9), the number of iterations is M, the problem The total length of the shortest path U is calculated by summing
size is N (N = n × m), and the number of iterations needed by the length of all the line segments U(i) in the path between Cs
the robot to pass the block of obstacles is M1 . and Cg , as follows:
1. In step 2, the time complexity of computing p1 (q), (q = I
∑
1...9) is T1 = O(q). U = (sqrt(u1x (i) − u2x (i))2 + (u1y (i) − u2y (i))2 ) (12)
2. In step 3, the time complexity of computing E(k), (k = i=1
1...N) is T2 = O(N). where I represents the number of the line segment, which is
3. In steps 4–6, the time complexity of calculating E(q), pbest , equal to 2 in this example. u1x (i), u2x (i), u1y (i), u2y (i) represent
sx , and sy , is T3 = O(q).
the coordinates of the line segment U(i). The general procedure
4. In steps 4–6, 8–9, the time complexity of calculating p2 (t) of PEM is illustrated in Algorithm 5.
is T4 = O(M ∗ q).
5. In step 10, the time complexity of collision checking the
line segment between p1 (t) and p2 (t) for feasibility, can be Algorithm 5 PEM method
done by T5 = O(M ∗ N ∗ q). 1: Inputs:
6. In steps 11–13, in case the line between p1 (t) and p2 (t) Cobs , and w (j), (j = 1...J)
collides with obstacles, the computing time spent in these 2: j←1
is considerable longer, so the time complexity of these 3: while j ≤ J do
steps is T6 = O(M ∗ N ∗ M1 ∗ q). 4: u2 = w (j)
7. In steps 14–18, the time complexity of determining the 5: check the line Ui for feasibility between u1 and u2
new set of E(q), (q = 1...9) and pbest , together with up- 6: if U(i) collide with Cobs then
dating p1 and pr , and store pr in a way-points w list is 7: store u1 , u2 = w (j − 1)
T7 = O(q). 8: u1 ← w (j − 1)
9: end if
The total time complexity of the developed method is: T = T1 +
10: j←j+1
T2 + T3 = T4 + T5 + T6 + T7 T = O(q) + O(N) + O(q) + O(M ∗ q) +
11: end while
O(M ∗ q ∗ N) + O(M ∗ N ∗ M1 ∗ q) + O(q) = O(N ∗ M ∗ M1 )
12: insert Cs and Cg to the beginning and to the end of the new
The obtained IFP for a mobile robot is a safe path, however,
waypoints list.
it is not a shortest path between Cs and Cg . In order to reduce
the overall path length, a new method called path enhancement
method PEM is developed, as we will explain in the following 4.4. Path smoothing using interpolation technique
subsection.
The path we obtained so far may contain sharp turns. This
4.3. Path enhancement method PEM goes against many real-world applications where smooth paths
are preferred [43]. Moreover, the robot may not be able to make
This section introduces the PEM method to generate the short- a sharp turn due to its momentum [10,14]. Finally, the cubic
est path (see Fig. 9b) from IFP (see Fig. 9a). The PEM method is spline interpolation is adopted to generate a continuous smooth
used to reduce the number of waypoints of the IFP between Cs path that connects the starting point to the goal point. The spline
and Cg obtaining an optimal or close-to-optimal path. As shown method is one of the most efficient curve interpolating methods
in Fig. 9, the waypoints of the IFP are represented by red circle which has many applications in robotics, signal processing, and
objects, and the obtained shortest path is represented by a thick
computer graphics [14,18].
red line. In order to explain the basic idea of PEM, consider an
From Fig. 10c, consider the generated shortest path by using
example shown in Fig. 10.
PEM. The path consists of two line segments U1 and U2 between
In this example, the robot starts to move from the starting
Cs and Cg in the form of X and Y vectors, where X = [x1 x2 x3 ]
point and passes through all the intermediate waypoints until it
and Y = [y1 y2 y3 ]. We use the cubic spline interpolation to
reaches the goal point. As illustrated in Fig. 10b, the IFP consists
calculate the spline for three waypoints (w = 3). Therefore,
of 14 waypoints w , and they are connected by line-segments. In
a new vector t of about 200 points is generated between the
the figure, a line segment U has two end points, u1 and u2 . For
starting point at (x1 , y1 ) and the goal point at (x3 , y3 ). Vectors of
the first line segments U1 , the starting position of u1 coincides
interpolated values xsp and ysp are calculated based on equations
at the Cs . In order to determine the starting position of u2 , the
xsp = Spline(tn , x, t) and ysp = Spline(tn , y, t), where tn =
PEM method connects u1 with w (j), (j = 1...J), J = 14 iteratively.
First, u1 is connected with the first waypoint w (1), then the line
[1 2 3]. As illustrated in Fig. 10c, the constructed path passes
smoothly through the way-points thus eliminating the sharp turn.
between these two points is checked for feasibility. If a collision
is not found, then u1 is connected to w (2). Afterwards the line
between u1 and w (2) is checked for feasibility. If the line does not 5. Simulations
collide with any obstacles, then u1 is connected to w (3), and this
procedure continues in the same way until j = 12. When j = 12, In this study, the proposed methods are implemented in
u1 is connected to w (12); in this case the line between these two MATLAB and run on a laptop with Intel(R) core(TM) i5 − 2450M
points collides with obstacles, as shown in Fig. 10a. Therefore, u2 CPU 2.5 GHz 6 GB RAM. The performances of the developed
of the first line segment is placed in w (11). For the second line method have been tested on many different workspace scenarios
segment U2 , the left-hand end u1 is coincides at u2 of the first with different obstacle layouts. In all the tested scenarios, the
line segment. In order to find u2 of the second line segment, the workspace size and the number of obstacles scattered in the
PEM connect u1 with w (j), (j = 12...14), iteratively. Therefore, u1 workspace have been varied. Additionally, the starting Cs and
is connected with w (12), w (13), and w (14) one after another, and the goal Cg points have been positioned in different locations in
the lines between u1 and these points are check for feasibility. the free space Cfree . An example of three workspace scenarios are
12 R.A. Saeed, D.R. Recupero and P. Remagnino / Robotics and Autonomous Systems 123 (2020) 103320

Fig. 9. Example of path planning for a mobile robot. (a) The obtained solution of IFP by using BNM, where the sequence of the red circle objects is represents the
IFP. (b) The shortest path found by using PEM, where the solid red line represents the shortest path.

Fig. 10. Construction of the shortest-path from 14 waypoints in the 2D workspace, where the waypoints are marked with the red circle objects. (a)PEM is used to
find the shortest path between start and goal point. (b)IFP is generated by using BNM. (c) The shortest line-segment path (U) found by using PEM and the smooth
path constructed by using spline method.

shown in Fig. 4. The proposed method is examined to find an op- shown in Fig. 4. The achieved result of the IFP is represented by a
timal or near optimal path from Cs to Cg . The simulation results of set of waypoints w (j), (j = 1 → J). Each new position of waypoint
BNM and PEM are presented in Sections 5.1 and 5.2, respectively. w(j + 1) is allocated after the current waypoint position w(j) on
Additionally, the performance of the proposed method is com- the IFP, where J represents the time in which the robot is reaching
pared with the other path planning methods in Section 5.3. Then, the goal point.
the proposed method is applied to the multi-robot path-planning The simulation results for all the tested scenarios are pre-
problem, and the results presented in Section 5.4. sented in Fig. 11, and the summary of the obtained results is
provided in Table 2. From the figure, it is observed that the
5.1. Simulation results of BNM obtained IFP allows the robot to move from Cs to Cg without
colliding with any obstacle in the workspace. The waypoints of
This section presents the results of the BNM method for gen- the path are represented by red circle objects and for better
erating the IFP between Cs and Cg for all the workspace scenarios clarity, these waypoints are connected into a continuous path.
 R.A. Saeed, D.R. Recupero and P. Remagnino / Robotics and Autonomous Systems 123 (2020) 103320 13

Table 2 On the other hand, the cumulative length of the smooth paths
The total computational time and path length of the IFP and final path by using shown in Fig. 13 are longer than the cumulative length of the line-
BNM and PEM.
segment path presented in Fig. 12, and the length of the paths are
Workspace Total computational time [s] Total path length [unit]
increased by 7%, 4.4%, and 4.5% for all three scenario, respectively.
No. IFP Final path IFP Final path
The aim of the cubic spline method is to generate a smooth
1 0.955601 1.043110 112.9662 83.4301 path for an initial feasible path that connects the starting point to
2 0.896196 1.004691 110.1285 81.0895
3 1.100025 1.138494 201.3783 146.3850
the goal point. However, in some cases, the constructed smooth
path can bring the robot close to the safety zone around the
obstacles or the robot collides with the safety zone, which is
undesirable in practice. To avoid the possibility of overlapping
As it can be seen from the results, the BNM method is able the paths traced by the robot with safety zone, additional way-
to overcome the local minima problem. From Table 2, we can points can be inserted between the original waypoints until no
clearly see that the developed method provides the collision- safety zone or obstacles were found along the resulting path, as
free path for the robot in short time, in particular for the high explained [44]. Alternatively, Piecewise Cubic Hermite Interpo-
complex environment shown in Fig. 11c. As presented in the third lating Polynomial (PCHIP) can be used to construct a continuous
scenario, the total computing time to find a IFP is less than 1.1 s. smooth path, as illustrated in [7,45]. The PCHIP is like cubic spline
The results show that the BNM method has been well applied
interpolation, but PCHIP interpolation ensures a shape-preserving
to generate the IFP for a mobile robot, and also this method has
interpolation and avoiding the overshoots and oscillations that
achieved good results in terms of safety and short computational
could arise from spline interpolation. The generated path from
time. However, the generated path is not optimal in terms of the
the X and Y vectors of the waypoints w is a zigzag line; we
total path length. In order to reduce the overall path length, a new
generate a new vector xi of about 1000 points from start point
method called PEM is developed as explained in Section 4.3, and
to goal point. yi = PCHIP(X , Y , xi ) returns a vector of interpolated
the results are presented in the following subsection.
values yi containing elements corresponding to the elements of
5.2. Simulation results of PEM xi . Resulting xi versus yi give the smoothed path. The obtained
results of smoothed paths by using PCHIP is shown in Fig. 14. The
This section presents the obtained results of the PEM method length of the paths are increased by 4.4%, 2.4%, and 2.6% for all
to find optimal or near-optimal path for the three workspace three scenario, respectively. We can see the difference between
scenarios. The best-obtained results are presented in Fig. 12, and the interpolation results produced by PCHIP and cubic spline in
the results of computational time and path length for all the Figs. 13 and 14.
scenarios are provided in Table 2. As shown in Fig. 12, the PEM In the proposed approach, the grid-based method is used to
method can find the collision-free path that covers the least create a workspace environment. In this method, the workspace
number of waypoints, where the solid red lines represent the best environment is divided into a number of small square grid cells
solution found so far. Additionally, Table 2 revealed that the total of the same size (1 × 1 unit). Each grid cell can either correspond
path length for all the three designed workspaces is significantly to a navigable area or to a space occupied by obstacles. Different
reduced, and the percentage of enhancement of the path length obstacle shapes can be generated, such as circular or non-convex
for all the three scenarios are 26.2%, 26.4% and 27.3%, respectively. obstacles, by approximating the shape of the obstacles and divid-
Obviously, the geometrical complexity of the workspace is ing it into square grid cells. The completeness of the obstacles’
the main factor affecting the computational time. However, the shape depends on the resolution of the grid environment. Fig. 15a
results show that the computational time required to obtain and e show two examples of different workspace scenarios. In
the IFP and the final path by using BNM&PEM is not increased these scenarios, the workspace consists of (50 × 50) grid cells,
significantly with the growing complexity of the workspace. For and the number of the obstacles in the workspace is 312 and
example, the size of the workspace is increased 3.2, times and the 316 grid cells, respectively. The starting Cs and goal Cg points
number of obstacles are increased 2.6 times from the second (see are positioned in the free space Cfree at (5, 5) and (45, 45),
Fig. 12b) to the third scenario (see Fig. 12c), accordingly the total respectively. The proposed method is examined to find an optimal
computational time to find the IFP and the final path is increased or near optimal path from Cs to Cg . The simulation result of
only by about 1.5 and 1.4 times, respectively (see Table 2). On the the BNM method for generating the IFP between Cs and Cg is
other hand, in the second (see Fig. 12b) and first (see Fig. 12a)
presented in Fig. 15b and f . From the figures, it is observed that
scenario, the required computational time to find the IFP and
the obtained IFP can successfully drive the robot towards the
the final path is increased by 6.6% and 3.8% respectively, as the
goal while avoiding obstacles in the highly complex environment.
number of obstacles increased by 123.2% for the same workspace.
The robot location is represented by red circles object at each
This is because, for each iteration t during the search process, all
iteration. The obtained results of the PEM method to find optimal
obstacles in the workspace are examined for possible collisions
or near-optimalpath are presented in Fig. 15c and g. As shown
with the direct path from the current p1 (t) to updated p2 (t) nodes’
location. in the figures, the PEM method can find the short path, where
In the simulations results presented in Fig. 12, we observe that the solid red lines between Cs and Cg represent the best solution
the proposed method generates a path consisting of straight lines found so far. Additionally, the generated path from the PEM is
between waypoints with sharp turns. In real applications when smoothed by using the cubic spline method and the result are
the robot follows a path in the workspace, it may not be able to presented in Fig. 15d and h.
make a sharp turn and also it is not the safest path for the robot. The proposed method can easily be extended to include alti-
In order to improve the path with respect to the robot dynamics, tude as a third coordinate to solve the path planning problems
the proposed method applied MATLAB cubic spline to construct a in three-dimensional (3D) workspace. The method was imple-
continuous smooth path that connects the starting point to the mented with several 3D scenarios and the results were found
end point for all three designed workspaces, and the results are to be satisfactory. An example of the workspace scenario is pre-
presented in Fig. 13. sented in Fig. 16. The workspace is discretized into uniform cubic
The results demonstrate that the spline method can be used grid cells (1 × 1 × 1 unit), and the generated path is a sequence
to generate a continuous smooth path to eliminate sharp turns. of cubic cells in a 3D grid model.
14 R.A. Saeed, D.R. Recupero and P. Remagnino / Robotics and Autonomous Systems 123 (2020) 103320

Fig. 11. The simulation results to generate IFP for all three workspace scenarios using BNM.

Fig. 12. The simulation results to generate an optimal or near-optimal path for all three workspace scenarios using PEM.

Fig. 13. The simulation results to generate a smooth path by using spline method for all three workspace scenarios.
 R.A. Saeed, D.R. Recupero and P. Remagnino / Robotics and Autonomous Systems 123 (2020) 103320 15

Fig. 14. The simulation results to generate a smooth path by using PCHIP for all three workspace scenarios.

Fig. 15. Examples of grid cells with obstacles (a and e) simulation result of BNM (b and f ), PEM (c and g) and cubic spline method (d and h).

Fig. 16. The simulation result of the BNM (a) and PEM (b) to solve the path planning problems in three-dimensional (3D) workspace.

5.3. Comparison results shown in Fig. 17. The size of the workspace is set to 43 × 68,
where the space occupied by obstacles Cobs consists of 1078
This section presents the performance evaluation of the grid cells and the obstacle-free space Cfree consists of 1846 grid
BNM&PEM method in comparison with PSO, A-Star, and APF . cells. After constructing the workspace with obstacles, all four
Therefore, a simple example of a 2D workspace is designed as methods namely BNM&PEM, PSO, A-Star, and APF are used to find
16 R.A. Saeed, D.R. Recupero and P. Remagnino / Robotics and Autonomous Systems 123 (2020) 103320

Table 3 Table 5
A summary of the obtained results of the computational time and path length Mean and standard deviation of the computational time and path length for
by using BNM, PSO, A-Star, and APF . 1000 independent runs to find feasible path using proposed method with PSO,
Method Total computational time [s] Total path length [unit] GA, and A-Star.

BNM 0.82 53.49 Methods Computational Time, CT [s] Path Length, PL [unit]
PSO 1.51 57.70 MeanCT StdCT MeanPL StdPL
A-Star 2.57 57.11 PEM 0.0142 0.0072 30.7710 15.9340
APF 0.66 61.00 A-Star 0.0489 0.0640 30.0907 14.6124
PSO 0.0217 0.0052 34.3788 15.2495
GA 0.1188 0.2122 33.0144 11.1463
Table 4
The total computational time required to find shortest path using
BNM and improved GA.
Method Total computational time [s]
near-optimal path grows exponentially as the complexity of the
Improved GA Ref. [46] 1.03 path-planning problem increases (see Section 1, point 4) even
Improved GA Ref. [18] 4.07
Improved GA Ref. [32] 1.68
in some circumstance the path planning methods cannot find a
Improved GA Ref. [47] 0.85 feasible path, whereas the proposed method BNM solve these
BNM 0.964 problems.
In order to validate the proposed method BNM&PEM, and
compare its performance with the A-Star, PSO, and GA, a 2D
environment of the static robot’s workspace is created. The size
the shortest path between Cs at (8, 10) and Cg at (32, 56). The
of the workspace is set to 60 × 60, and the space occupied by
obtained results of BNM&PEM, PSO, A-Star, and APF are shown
obstacles Cobs consists of 136 grid cells. Thereafter, all methods
in Fig. 17a, b, c, and d, respectively. A summary of the obtained
are implemented simultaneously to find a feasible path for 1000
results of the computational time and path length is provided in
independent runs. At each time in this test, the starting point
Table 3. By comparing the results presented in Table 3, it can be
Cs , the goal point Cg , and obstacles are placed randomly in the
seen that the proposed method is able to find the shortest path
working environment, each random placement of the obstacles
within less than one second, and it requires less than 55% and 32%
led to different workspace layout. Two measures of evaluation are
of the computational time to find shortest path by using PSO and
used for comparison among path planning methods: the length
A-Star, respectively. In terms of the total path length, the shortest
path achieved by BNM&PEM is about 7.2% and 6.3% shorter than of the obtained feasible path as well as the execution time of the
the path length generated by PSO and A-Star, respectively. In this method. The mean and the standard deviation (Std) of the com-
workspace, the computational time required to find the shortest putational time and the path length are calculated and presented
path by using APF is lower by 20% compare to BNM&PEM. In in Table 5. The results shown in the table reveal that the proposed
contrast, the shortest path achieved by BNM&PEM is 12% shorter method achieved the best solution within a reasonable computa-
than the path length generated by APF . tional time. Moreover, the mean value of the computational time
The PEM method can also be used to optimize the paths to find a feasible path is decreased significantly compared with
obtained by using PSO, A-Star, and APF as shown in Fig. 17b, c, other path planning methods. In comparison with PSO and GA,
and d, respectively. The best-obtained results are presented in the proposed method showed noticeable improvement in terms
Fig. 18b, c, and d, respectively. As shown in the figures, the of the path length. Additionally, the mean value of the path length
PEM method can find the collision-free path that covers the least obtained by the proposed method is smaller than that obtained
number of waypoints, where the solid red lines represent the with PSO and GA by %11.73 and %7.3, respectively. However,
best solution found so far. The results revealed that the length the mean value of the path length generated by BNM&PEM is
of the paths obtained from PSO, A-Star, and APF reduced by slightly larger than that obtained with A-Star by %2.21. The PSO
7.6%, 5.1% and 9.9%, respectively. Furthermore, the cubic spline method had the least variance of the computational time, and GA
interpolation is used to generate a continuous smooth path that better than the other method in terms of variance of the path
connects the starting point to the end point bu using BNM, PSO, length. The comparative study shows that heuristic algorithms
A-Star, and APF , and the results are presented in Fig. 19a, b, c, did not yield optimal results, and the results agree with [48]. The
and d, respectively. graphical representation of the simulation results of all methods
In order to make an extra comparison and to demonstrate is illustrated in Fig. 21.
the ability of BNM for solving robot path planning problem in
the workspaces that have previously been used in [18,32,46,47], 5.4. Path planning of a multi-robot system
a 2D workspace is created as shown in Fig. 20. The size of
the workspace is set to 67 × 67, where the space occupied In this section, the implementation of the proposed method
by obstacles Cobs consists of 1520 grid cells and the obstacle- for collision avoidance in multi-robot systems is presented. We
free space Cfree consists of 2969 grid cells. After constructing conducted different simulations with different multi-robot sys-
the workspace with obstacles, the proposed method is used to tem parameters, i.e. the number of robots, the initial and goal
generate a IFP (Fig. 20a), shortest path (see Fig. 20b), and smooth positions for each robot, and the positions of static obstacles.
path (see Fig. 20c) from the Cs at (64, 4) to the Cg at (4, 64). The Fig. 22 shows an example of the simulation results for a multi-
obtained computational results of the BNM and an improved GA robot system, in which there are 4 robots, 4 different start Cs
is provided in Table 4. By comparing the obtained results of BNM and goal Cg positions corresponding to each robot, and 304 static
with an improved GA in the previous studies (see Table 4), it is obstacles. The problem formulation is to determine the path of
observed that the computational time of the proposed method is each robot in the simulated environment by avoiding the collision
remarkably reduced. with static obstacles and other moving robots in the system.
The comparison results demonstrate the effectiveness and ef- Each robot moves from a starting position Cs , through all the
ficiency of the proposed method for solving robot path planning intermediate waypoints w until it reaches the goal position Cg .
problem. For the comparison test, a simple workspace scenario Each robot uses the BNM to find IFP to move from Cs to Cg in the
has been selected because the required time to find optimal or workspace without colliding with any obstacle (see Fig. 22a and
 R.A. Saeed, D.R. Recupero and P. Remagnino / Robotics and Autonomous Systems 123 (2020) 103320 17

Fig. 17. The simulation results of the BNM&PEM (a), PSO (b), A-Star (c), and APF (d) to solve the path planning problem in two-dimensional (2D) workspace.

Fig. 18. Simulation results for generating an optimal or near-optimal path for BNM (a), PSO (b), A-Star (c), and APF (d) using PEM method.

Fig. 19. Simulation results for generating an optimal or near-optimal path for BNM (a), PSO (b), A-Star (c), and APF (d) using cubic spline method.

d). IFP is generated from a set of waypoints w that the robot visits 6. Experimental results
before reaching the final destination point. Then for each robot,
In this section, a real robot is employed to test the perfor-
the PEM method is used to reduce the number of waypoints of the mance of the developed method BNM&PEM and illustrate how
IFP between Cs and Cg to obtain an optimal or close-to-optimal the robot can navigate along a collision-free path. An e-puck
robot, shown in Fig. 23a, is used for the experimental test, and
path (see Fig. 22b and e). Finally, the cubic spline interpolation the experimental set-up shown in Fig. 23b. First, the developed
is applied to construct a continuous smooth path that connects method is used to generate a collision-free path to direct the
robot to move among the static obstacles from the starting point
the starting position to the goal position (see Fig. 22c and f ). Cs toward the goal point Cg as shown in Fig. 24a. As illustrated in
The simulation results show that all robots reached to their the figure, the waypoints w are represented by red circle objects,
and the obtained shortest path is represented by a thick red
final destination positions successfully without any collision with
dashed-line. The obtained shortest path by BNM&PEM consists of
either static obstacles or other robots. the waypoints, w (j), (j = 1...J), J = 5, whose x, y coordinates
18 R.A. Saeed, D.R. Recupero and P. Remagnino / Robotics and Autonomous Systems 123 (2020) 103320

Fig. 20. The simulation result of the BNM (a), PEM (b), and cubic spline (c) method for solving robot path planning problem in the workspace that is previously
have been used in [18,32,46,47].

Fig. 21. Performance of final evaluations for 1000 independent runs to find feasible path using proposed method with PSO, GA, and A-Star, the obtained results for
the path length data presented in (a) and the computational time data presented in (b).

Fig. 22. The simulation results for the multi-robot path planning problem.

are known with respect to the simulated environment. Based on is used to control e-puck in MATLAB. Let w1x and w1y be the cen-
the generated data of the obtained path, the e-puck robot motion troid coordinates of the first waypoint w1 of the generated path,
and w2x and w2y those of the centroid of the second waypoint
data is determined. Next, the e-puck robot is connected to the
w2 . Then, the orientation of the robot is calculated in MATLAB by
computer via Bluetooth and the generated motion data are trans- using atan2 (w2y -w1y , w2x -w1x ). Subsequently, in order to move
mitted to the robot via a toolbox ePic(v 2.1.2), where ePic(v 2.1.2) the e-puck robot toward the second waypoint w2 , the angle of the
 R.A. Saeed, D.R. Recupero and P. Remagnino / Robotics and Autonomous Systems 123 (2020) 103320 19

Fig. 23. The robot used in the experiment (a) and the experimental set-up (b).

Fig. 24. The simulation and experimental results. (a) the simulation results to generate IFP by using BNM&PEM. (b) to (f ) locations of the e-puck robot at different
waypoints in the robot’s working environment.
20 R.A. Saeed, D.R. Recupero and P. Remagnino / Robotics and Autonomous Systems 123 (2020) 103320

w2 with respect to the robot is calculated. Thereafter, the e-puck Acknowledgements

robot starts to move from w1 to w2 and so on until it reaches
the goal point. Fig. 24(b → f ) show the robot’s positions at The authors gratefully acknowledge Sardinia Regional Govern-
different locations in the robot’s working environment during the ment, Italy for the financial support (Convenzione triennale tra
experimental test. The test results demonstrate that the proposed la Fondazione di Sardegna e gli Atenei Sardi Regione Sardegna
method is able to generate the shortest path to direct the e-puck – L.R. 7/2007 annualità 2016 – DGR 28/21 del 17.05.2016, CUP:
robot to final destination point. F72F16003030002). Moreover, this work has been supported by
Sardinia Regional Government, Italy (P.O.R. Sardegna F.S.E. Oper-
7. Conclusions and future works ational Programme of the Autonomous Region of Sardinia, Euro-
pean Social Fund 2014-2020 - Axis IV Human Resources, Objec-
tive l.3, Line of Activity l.3.1.)
In this paper a novel off-line path planning method called
Boundary Node Method is developed for solving the path plan-
ning problem of a mobile robot in a two-dimensional working Appendix A. Supplementary data
environment. The developed method is used to find collision-
free path for a mobile robot through a sequence of way-points Supplementary material related to this article can be found
that the robot has to traverse from the starting point to the goal online at https://doi.org/10.1016/j.robot.2019.103320.
point without colliding with any obstacles. The concept involved
in the developed method is simple and can be applied in a References
grid environment efficiently. Additionally, this method does not
[1] N. Leena, K. Saju, A survey on path planning techniques for autonomous
work through random operations and there is no uncertainty in mobile robots, IOSR J. Mech. Civil Eng. 8 (2014) 76–79.
generating points, which leads to finding the final solution for [2] J. Han, Y. Seo, Mobile robot path planning with surrounding point set and
the problem without variation in solution. Moreover, this method path improvement, Appl. Soft Comput. 57 (2017) 35–47.
uses an optimization technique based on the lowest potential [3] P. Victerpaul, D. Saravanan, S. Janakiraman, J. Pradeep, Path planning of
autonomous mobile robots: A survey and comparison, J. Adv. Res. Dyn.
value to accelerate the robot to find the path safely and quickly
Control Syst. 9 (2017) 1535–1565.
in reasonable time. The simulation results show that the Bound- [4] N. Chao, Y.-k. Liu, H. Xia, A. Ayodeji, L. Bai, Grid-based RRT for minimum
ary Node Method can successfully find an initial feasible path, dose walking path-planning in complex radioactive environments, Ann.
and generates a safe path for a mobile robot to navigate in a Nucl. Energy 115 (2018) 73–82.
complicated environment within a relatively short time. And also [5] A. Shum, K. Morris, A. Khajepour, Direction-dependent optimal path
planning for autonomous vehicles, Robot. Auton. Syst. 70 (2015) 202–214.
the computational time required to find shortest path does not [6] Y. Zhang, D.-w. Gong, J.-h. Zhang, Robot path planning in uncer-
increase significantly with the increase of the environment’s com- tain environment using multi-objective particle swarm optimization,
plexity. Furthermore, the results have verified that the boundary Neurocomputing 103 (2013) 172–185.
node method solves the local minima problem effectively. An [7] A. Bakdi, A. Hentout, H. Boutami, A. Maoudj, O. Hachour, B. Bouzouia,
Optimal path planning and execution for mobile robots using genetic
additional method that we called Path Enhancement Method, algorithm and adaptive fuzzy-logic control, Robot. Auton. Syst. 89 (2017)
has been applied on top to build an optimal or close-to-optimal 95–109.
collision-free path by reducing the number of waypoints and the [8] T. Lozano-Pérez, M.A. Wesley, An algorithm for planning collision-free
overall path length. In order to validate the performance of the paths among polyhedral obstacles, Commun. ACM 22 (10) (1979) 560–570.
[9] M. Brand, M. Masuda, N. Wehner, X.-H. Yu, Ant colony optimization
developed method in comparison with existing path planning
algorithm for robot path planning, in: 2010 International Conference on
methods, several different scenarios with different complexity Computer Design and Applications, Vol. 3, IEEE, 2010, pp. V3–436.
have been tested. The comparison reveals that the Boundary [10] S.M. LaValle, Planning Algorithms, Cambridge University Press, 2006.
Node Method can solve the path planning problem effectively [11] J. Rintanen, Introduction to Automated Planning, in: Lecture Notes of the
and efficiently in terms of the computational times and the path AI Planning Course, Albert-Ludwigs-University Freiburg, 2006.
[12] C. Lamini, S. Benhlima, A. Elbekri, Genetic algorithm based approach for
length. Finally, the cubic spline method has been used to generate autonomous mobile robot path planning, Procedia Comput. Sci. 127 (2018)
a continuous smooth path that connects the starting point to the 180–189.
end point. [13] T.T. Mac, C. Copot, D.T. Tran, R. De Keyser, Heuristic approaches in robot
The developed method is used to solve the multi-robot path path planning: A survey, Robot. Auton. Syst. 86 (2016) 13–28.
[14] H.-C. Huang, C.-C. Tsai, Global path planning for autonomous robot nav-
planning problem, and the simulation results showed that the
igation using hybrid metaheuristic GA-PSO algorithm, in: SICE Annual
developed method effective and useful for collision avoidance Conference (SICE), 2011 Proceedings of, IEEE, 2011, pp. 1338–1343.
in multi-robot systems. Additionally, the performance of the de- [15] O. Montiel, U. Orozco-Rosas, R. Sepúlveda, Path planning for mobile robots
veloped method for generating a collision-free path is tested using Bacterial Potential Field for avoiding static and dynamic obstacles,
on a real robot. The experimental test shows that the proposed Expert Syst. Appl. 42 (12) (2015) 5177–5191.
[16] M.A. Contreras-Cruz, V. Ayala-Ramirez, U.H. Hernandez-Belmonte, Mo-
method is able to generate shortest path, and direct the real robot bile robot path planning using artificial bee colony and evolutionary
to the final destination point. programming, Appl. Soft Comput. 30 (2015) 319–328.
In the future work, we will address a number of research [17] W. Song, H.-X. Li, Y.-N. Zhang, Path planning of mobile robot based on
issues related to autonomous navigation of mobile robots in genetic bee colony algorithm, DEStech Trans. Comput. Sci. Eng. 16 (5)
(2016) 615–620.
unknown environments, where the deployed robot does not have
[18] A.H. Karami, M. Hasanzadeh, An adaptive genetic algorithm for robot
full knowledge about its environment. Another possible direction motion planning in 2D complex environments, Comput. Electr. Eng. 43
will explore an extension to the proposed method in order to deal (2015) 317–329.
with a dynamic scene. [19] H. Liu, B. Xu, D. Lu, G. Zhang, A path planning approach for crowd
evacuation in buildings based on improved artificial bee colony algorithm,
Appl. Soft Comput. (2018).
Declaration of competing interest [20] J.K. Goyal, K. Nagla, A new approach of path planning for mobile robots, in:
Advances in Computing, Communications and Informatics (ICACCI), 2014
International Conference on, IEEE, 2014, pp. 863–867.
The authors declare that they have no known competing finan-
[21] Y. Hu, S.X. Yang, A knowledge based genetic algorithm for path planning of
cial interests or personal relationships that could have appeared a mobile robot, in: Robotics and Automation, 2004. Proceedings. ICRA’04.
to influence the work reported in this paper. 2004 IEEE International Conference on, Vol. 5, IEEE, 2004, pp. 4350–4355.
 R.A. Saeed, D.R. Recupero and P. Remagnino / Robotics and Autonomous Systems 123 (2020) 103320 21

[22] B. Fu, L. Chen, Y. Zhou, D. Zheng, Z. Wei, J. Dai, H. Pan, An improved A* [48] I. Chaari, A. Koubaa, H. Bennaceur, A. Ammar, M. Alajlan, H. Youssef,
algorithm for the industrial robot path planning with high success rate Design and performance analysis of global path planning techniques for
and short length, Robot. Auton. Syst. (2018). autonomous mobile robots in grid environments, Int. J. Adv. Robot. Syst.
[23] N.A. Shiltagh, L.D. Jalal, Optimal path planning for intelligent mobile robot 14 (2) (2017) 1729881416663663.
navigation using modified particle swarm optimization, Int. J. Eng. Adv.
Technol. 2 (4) (2013) 260–267.
[24] F. Duchoň, A. Babinec, M. Kajan, P. Beňo, M. Florek, T. Fico, L. Jurišica, Path
planning with modified a star algorithm for a mobile robot, Procedia Eng. Raza Saeed graduated from the University of Sala-
96 (2014) 59–69. haddin and he starts his academic carrier at Sulaimani
[25] A.K. Guruji, H. Agarwal, D. Parsediya, Time-efficient A* algorithm for robot University, Sulaimani, Iraq since 1995. He is currently
path planning, Proc. Technol. 23 (2016) 144–149. a Ph.D. student in Computer Science at the University
[26] S.S. Ge, Y.J. Cui, New potential functions for mobile robot path planning, of Cagliari, Italy. His research interests include mobile
IEEE Trans. Robot. Autom. 16 (5) (2000) 615–620. robot, path planning algorithms, Artificial Intelligent,
[27] G.E. Jan, K.Y. Chang, I. Parberry, Optimal path planning for mobile robot and Computational Mechanics.
navigation, IEEE/ASME Trans. Mechatronics 13 (4) (2008) 451–460.
[28] R. Kala, A. Shukla, R. Tiwari, Robotic path planning in static environment
using hierarchical multi-neuron heuristic search and probability based
fitness, Neurocomputing 74 (14–15) (2011) 2314–2335.
[29] M. Davoodi, F. Panahi, A. Mohades, S.N. Hashemi, Clear and smooth path
planning, Appl. Soft Comput. 32 (2015) 568–579.
[30] P. Raja, S. Pugazhenthi, Optimal path planning of mobile robots: A review, Diego Reforgiato Recupero has been an Associate
Int. J. Phys. Sci. 7 (9) (2012) 1314–1320. Professor at the Department of Mathematics and Com-
[31] O. Souissi, R. Benatitallah, D. Duvivier, A. Artiba, N. Belanger, P. Feyzeau, puter Science of the University of Cagliari, Italy since
Path planning: A 2013 survey, in: Industrial Engineering and Systems December 2015, where he is co-director of the Se-
Management (IESM), Proceedings of 2013 International Conference on, mantic Web laboratory (http://swlab.unica.it/). He is a
IEEE, 2013, pp. 1–8. programmer, software developer, automation and ICT
[32] A. Tuncer, M. Yildirim, Dynamic path planning of mobile robots with expert. He holds a double bachelor from the University
improved genetic algorithm, Comput. Electr. Eng. 38 (6) (2012) 1564–1572. of Catania in computer science and a doctoral degree
[33] Z. Bi, Y. Yimin, Y. Wei, Hierarchical planning approach for mobile robot from the Department of Computer Science of University
navigation under the dynamic environment, in: Industrial Informatics, of Naples Federico II. He got the National qualification
2008. INDIN 2008. 6th IEEE International Conference on, IEEE, 2008, pp. for Computer Science Engineering and he is a Computer
372–376. Science Engineer. In 2005 he was awarded a 3 year Post Doc fellowship with
[34] M.P. Garcia, O. Montiel, O. Castillo, R. Sepúlveda, P. Melin, Path planning the University of Maryland where in 2006 he won the Computer World Horizon
for autonomous mobile robot navigation with ant colony optimization and Award in the USA for the best research project on OASYS, an opinion analysis
fuzzy cost function evaluation, Appl. Soft Comput. 9 (3) (2009) 1102–1110. system that was commercialized by SentiMetrix (US company he co-founded).
[35] B. Patle, D. Parhi, A. Jagadeesh, S.K. Kashyap, Matrix-Binary Codes based In 2008, he won a Marie Curie International Grant that allowed him to come
Genetic Algorithm for path planning of mobile robot, Comput. Electr. Eng. back in Italy and was able to fund a 3 year Post Doc fellowship within the
67 (2018) 708–728. Department of Electrical, Electronic, and Computer Science Engineering (DIEEI)
[36] O. Khatib, Real-time obstacle avoidance for manipulators and mobile at the University of Catania where he won the ‘‘Best Researcher Award 2012’’ for
robots, in: Robotics and Automation. Proceedings. 1985 IEEE International a project about the development of a green router nearing commercialization. In
Conference on, Vol. 2, IEEE, 1985, pp. 500–505. the same year he got to the winning podium of the ‘‘Startup Weekend’’ event
[37] J.-O. Kim, P.K. Khosla, Real-time obstacle avoidance using harmonic held in Catania and was a winner of Telecom Italia Working Capital Award
potential functions, IEEE Trans. Robot. Autom. 8 (3) (1992) 338–349. with a grant of 25k euros for the ‘‘Green Home Gateway’’ project. In 2012
[38] A.H. Qureshi, Y. Ayaz, Potential functions based sampling heuristic for he co-founded the Italian company R2M Solution s.r.l. In 2013 he won a Post
optimal path planning, Auton. Robots 40 (6) (2016) 1079–1093. Doctoral Researcher position within the Semantic Technology Laboratory (STLAB)
[39] L. Janson, B. Ichter, M. Pavone, Deterministic sampling-based motion of the Institute of Cognitive Science and Technologies (ISTC) of the National
planning: Optimality, complexity, and performance, Int. J. Robot. Res. 37 Research Council (CNR) where he worked on Semantic Web and Linked Open
(1) (2018) 46–61. Data; he is still an associated researcher at STLAB where he collaborates. In
[40] W. Wang, L. Zuo, X. Xu, A learning-based multi-RRT approach for robot 2013 he published a paper on Science related to the energy efficiency techniques
path planning in narrow passages, J. Intell. Robot. Syst. 90 (1–2) (2018) in Internet. Besides SentiMetrix inc. (US company founded in 2007) and R2M
81–100. Solution s.r.l. (Italian company founded in 2012), he co-founded R2M Solution
[41] E. Di Mario, Z. Talebpour, A. Martinoli, A comparison of PSO and reinforce- ltd. (UK company founded in 2014), La Zenia s.r.l. (Italian company founded in
ment learning for multi-robot obstacle avoidance, in: 2013 IEEE Congress 2014 for management of sports and recreational events), B-UP (Italian company
on Evolutionary Computation, IEEE, 2013, pp. 149–156. founded in 2016 together with colleagues of CNR related to development of
[42] Z. Wang, J. Cai, Probabilistic roadmap method for path-planning in ra- software within the Semantic Web domain) and VISIOSCIENTIAE (spin-off of the
dioactive environment of nuclear facilities, Prog. Nucl. Energy 109 (2018) University of Cagliari founded in 2018 and related to the application of artificial
113–120. intelligence in the financial domain). He serves on the board of directors for
[43] K. Yakovlev, E. Baskin, I. Hramoin, Grid-based angle-constrained path R2M Solution and SentiMetrix. He is a patent co-owner in the field of data
planning, in: Joint German/Austrian Conference on Artificial Intelligence mining and sentiment analysis (20100023311). He has industrial and consulting
(K´’unstliche Intelligenz), Springer, 2015, pp. 208–221. experiences in several national and international industries. He has research
[44] A. Ravankar, A. Ravankar, Y. Kobayashi, Y. Hoshino, C.-C. Peng, Path experience across a wide array of industrial, Italian, FP7 and H2020 research
smoothing techniques in robot navigation: State-of-the-art, current and projects most recently in the fields of ICT, energy saving in telecommunication
future challenges, Sensors 18 (9) (2018) 3170. networks, robotics and semantic web.
[45] J.M. Hyman, Accurate monotonicity preserving cubic interpolation, SIAM J.
Sci. Stat. Comput. 4 (4) (1983) 645–654. Prof. Remagnino leads the Robot Vision Team (ROVIT),
[46] Q. Li, W. Zhang, Y. Yin, Z. Wang, G. Liu, An improved genetic algorithm of in the Department of Computer Science at Kingston
optimum path planning for mobile robots, in: Intelligent Systems Design University. Prof. Remagnino research is mainly con-
and Applications, 2006. ISDA’06. Sixth International Conference on, Vol. 2, cerned with the development of innovative methods
IEEE, 2006, pp. 637–642. for image and video interpretation, making wide use
[47] C. Liu, X. Yan, C. Liu, G. Li, Dynamic path planning for mobile robot based of pattern recognition, machine and deep learning and
on improved genetic algorithm, Chin. J. Electron. 19 (2) (2010) 2010–2014. distributed intelligence techniques. Prof. Remagnino
has published over 120 scientific articles in interna-
tional conferences and high impact journals. Currently,
Prof. Remagnino is the principal investigator of four
projects on video analytics for security with fixed and
mobile cameras.