Expert Systems With Applications 227 (2023) 120254

Contents lists available at ScienceDirect

Expert Systems With Applications

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

Review

Path planning techniques for mobile robots: Review and prospect

Lixing Liu , Xu Wang , Xin Yang *, Hongjie Liu , Jianping Li , Pengfei Wang
College of Mechanical and Electrical Engineering, Hebei Agricultural University, Baoding 071000, China

A R T I C L E I N F O A B S T R A C T

Keywords: Mobile robot path planning refers to the design of the safely collision-free path with shortest distance and least
Path planning time-consuming from the starting point to the end point by a mobile robot autonomously. In this paper, a sys­
Mobile robots tematic review of mobile robot path planning techniques is presented. Firstly, path planning is classified into
Key technologies
global path planning and local path planning according to the mastery of environmental information. In the
Algorithms
global path planning, environment modeling methods and path evaluation method are introduced. The methods
of environment modeling include grid method, topology method, geometric feature method and mixed repre­
sentation method. In the local path planning, we introduce the sensors commonly used in the detection envi­
ronment, including laser radar and visual sensor. Next, according to the characteristics of algorithms, mobile
robot path planning algorithms are divided into three categories: classical algorithms, bionic algorithms and
artificial intelligence algorithms. Among the classical algorithms, we introduce the cell decomposition method,
sampling based method, graph search algorithm, artificial potential field method and dynamic window method.
In the algorithm based on bionics, we introduce genetic algorithm, ant colony algorithm, gray wolf algorithm,
etc. in detail. In artificial intelligence algorithm, we introduce neural network algorithm and fuzzy logic. Finally,
we compare the key technologies of mobile robot path planning in the form of graphs and charts based on the
classification statistics of the collected literature to provide references for future research.

1. Introduction drive themselves. As long as the target location is entered, they can
provide an optimal route and arrive safely and accurately, subverting
Mobile robots are intelligent devices that can perform specific tasks the way people used to travel; In production, the warehouse transport
independently in complex environments and do not rely on human be­ robot can independently sort goods and transport them to the designated
ings. Today, mobile robots are widely used in many special situations location, which greatly saves manpower; In emergency rescue and
such as deserts, mines, battlefields, and disaster relief (Patle et al., disaster relief, mobile robots can quickly, accurately and safely find
2019). Especially since the outbreak of the COVID − 19 in 2020, we have targets in the changing dangerous environment; In scientific research,
become more aware of the convenience brought by mobile robots. mobile robots can take the place of humans to enter the harsh envi­
Mobile robots can deliver material supplies without assistance or direct ronment to help humans detect unknown planets and complete the
contact of human beings, greatly reducing the risk of epidemic trans­ acquisition task; In military science and technology, mobile robots can
mission while increasing work efficiency. independently carry out reconnaissance activities on the battlefield to
Mobile robot systems are broadly divided into three modules, obtain enemy information. These application fields all use path planning
including information perception, path planning and motion control. technology, so it is of great significance to study path planning tech­
Path planning is the bridge between information perception and motion nology in the development of the 21st century.
control, and it is a significantly fundamental part of a mobile robot Path planning can generally be divided into global path planning and
system. Advanced path planning techniques for mobile robots can local path planning according to the level of information about the
reduce capital investment and robot wear and tear. The continuous environment (Mohanty et al.,2021). Global path planning means that
development of path planning technology has injected fresh vitality into the robot is aware of the environment and can reach the target by
all walks of life. For example, in life, the sweeping robot can take the following a predefined path, based on this feature, global path planning
place of humans to do housework. And driverless cars do not need to is also called offline path planning or static path planning. Local path

* Corresponding author.
E-mail address: [email protected] (X. Yang).

https://doi.org/10.1016/j.eswa.2023.120254
Received 16 August 2022; Received in revised form 20 April 2023; Accepted 22 April 2023
Available online 3 May 2023
0957-4174/© 2023 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-
nc-nd/4.0/).
L. Liu et al. Expert Systems With Applications 227 (2023) 120254

Fig. 1. Key contents of global and local path planning.

environmental acquisition devices. See Fig. 1 for details.

Since the 1950s, researchers have proposed many algorithms to solve
the optimal path of mobile robots. In this chapter, we analyze the path
planning of mobile robot from the perspective of algorithm. According
to the characteristics of algorithms, researchers generally divide them
into three categories: classical algorithms, bionic algorithms and artifi­
cial intelligence algorithms. The detailed classification is shown in
Fig. 2.
The rest of this paper is organized as follows. Section 2 summarizes
Fig. 2. Algorithm classification block diagram.
the characteristics and development trends of global path planning and
local path planning. Section 3 classifies and analyzes the advantages,
planning means that the robot is partially or completely unknown to the
disadvantages and development trends of mainstream algorithms from
environment and carries out real-time monitoring by components and
the perspective of algorithms. Section 4 conducts quantitative statistics
reacts according to the definition, so local path planning is also called
and analysis on the collected literature, and forecasts the development
online path planning or dynamic path planning. Local path planning is
trend of path planning technology.
characterized with good flexibility, while the disadvantage being that
the planned path may be locally optimal but globally without being
2. Global path planning and local path planning
guaranteed, or even the target may be unreachable. Global path plan­
ning, on the other hand, requires the mobile robot to understand and
2.1. Global path planning
build a global map model, using a search and seek algorithm on the map
model to obtain the optimal or suboptimal path to guide the robot to
Global path planning is to design a safely collision-free path for the
move safely toward the target point in the actual environment. In
robot from the starting point to the end point based on the available
practical path planning strategies, it is often necessary to integrate
environmental information. The accuracy of the planned path depends
systems that combine global path planning, which aims to find the
on the accuracy of the environment information. The process of global
global optimal path, and local path planning aims to achieve real-time
path planning can be regarded as finding the optimal solution. There­
obstacle avoidance planning.
fore, in the global path planning in this chapter, we will introduce the
There is no fundamental difference between global path planning
method of environment modeling and the evaluation method of path
and local path planning. Many methods that are applicable to global
quality (Zhang et al., 2018).
path planning can be improved for local path planning, while methods
that are applicable to local path planning can also be improved for
2.1.1. Environment modeling
global path planning. The robot can better plan its path from the start to
The essence of environment modeling is the organization and utili­
the end if they work together. In global path planning, we focus on the
zation of the perceived environment. A proper environment model can
modeling method of the environment and the method for evaluating the
effectively help mobile robots to search the path. Therefore, exploring
merits of the path. In local path planning, we put emphasis on several
environment models that help robots operate efficiently is a key aspect

2
L. Liu et al. Expert Systems With Applications 227 (2023) 120254

Fig. 3. Grid environment.

environment model will be very clear and the planned paths will be safe.
However, it will occupy a lot of storage space of the system and also
generate more interference signals, which will lead to a long path
planning time. If the grid size is larger, the path planning takes less time
and solid real-time performance, but if the environment information is
blurred, which is not conducive to path planning. So the determination
of grid size is very important for the later path-searching.
The default shape in the Grid Method is square, and the search
process mostly uses quadtrees or octrees to represent the workspace. In
recent years, researchers have proposed honeycomb grid in order to
reduce the time and length of path planning and improve the walking
safety of mobile robots. Shao et al. (2021) established a square grid-
based map and a honeycomb grid-based map for the problem of un­
even step length of traditional grid. After analyzing and comparing the
results, it was found that the honeycomb-shaped grid has shorter path
length and higher safety when performing path planning, and it has
smaller obstacle avoidance angle, smoother path, and less energy con­
sumption by the robot when encountering obstacles. In addition, for the
grid method, a 2D occupied grid map creation system based on
boundary detection is proposed by Oršulić et al. (2019) for supporting
robots in path planning and exploration tasks. A multi-level human-like
motion planning method is proposed based on a 2D grid map based on
LRF by considering multiple constraints of the robot and the environ­
ment to guide the robot to navigate autonomously in an unstructured
indoor environment (Zhang et al., 2018a).
GM is characterized by its simplicity and ease of implementation as
Fig. 4. Voronoi diagram.
well as easy extension to 3D environments. A square grid was estab­
lished for 3D path planning of UAV (unmanned aerial vehicle) (Wu et al.,
of improving robot intelligence and driving robot deployment in a va­
2021c). The consistency and normality of the grid simplified the adja­
riety of environments. Zou (2005) focused on four common approaches
cency relations in the grid space. In the Chen’s literature (Chen et al.,
to build environment models, namely, GM, TM, GCM, and MR.
2021), when modeling the environment for a mobile camera, white grids
(1) GM (grid method).
are used to represent the areas where the camera can pass freely and
Howden (1968) proposed GM, which is essentially a partitioning of a
black grids represent the obstacle areas, and other colors represent the
mobile robot’s workspace into a series of network cells with binary in­
starting point or end point of the robot. After giving each grid a passage
formation and of equal size. When there is no obstacle in a grid, we call it
factor, the path planning problem becomes a problem of finding the
a free grid and the mobile robot can walk freely. When there is an
optimal path between two grid nodes on the grid network. However, it
obstacle in a grid, even if the obstacle does not fill the whole area of the
has certain requirements on the size of the working area. If the area is
grid, researchers generally use the expansion method to expand the
too large, the number of grids will increase sharply and there is the
obstacle to fill the whole grid and call the grid an obstacle grid. The free
problem of combinatorial explosion when performing path search. For
grid is usually marked by 0 and the obstacle grid by 1 (as shown in
example, in Fig. 3, we use MATLAB to generate a 10*10 2D environ­
Fig. 3). We generally determine the size of the grid according to the
ment, which takes 0.096 s, and a 10*10*10 3D environment, which
actual size of the robot. If the size of the grid is small, then the

3
L. Liu et al. Expert Systems With Applications 227 (2023) 120254

takes 0.557 s. especially in complex unstructured environments. Liang et al. (2018)

(2) TM (topological method). proposed a path planning method for UAV based on GCM. The method
TM is an environmental modeling method proposed by Professor focuses on the shape of obstacles and finally finds a collision-free path.
Zhang of Tsinghua University (China) in 1984 (Chien et al. 1984). Its The obstacles are organized in a list of tree. The collision-free path is
basic idea is to use nodes to represent a particular location and edges to generated by querying the blocking obstacles and connecting the sub-
represent the connection of these locations, which can be expressed as G goals of them. Jafarzadeh and Fleming (2018) proposed a new accu­
= (V, E) to characterize the free space, where V denotes the set of rate algorithm to solve the path planning problem, which involves
vertices and E denotes the set of edges connecting the vertices. A Vor­ finding the shortest collision free path from the starting point to the
onoi diagram is usually used to represent the skeleton of the environ­ target point in a 2D environment containing convex and non convex
ment feasible region. The basic idea of this method is to generate obstacles. A linear network connecting the vertex and the starting point
Voronoi boundaries from the environment model, which has equal of the obstacle is constructed, which is smaller than the network
shortest distance to each obstacle, and form the vertices of the graph at generated by the visibility map. Then find the shortest path from the
the junction of the edges. beginning to the target point in the network.
Voronoi diagrams were proposed by the Russian mathematician G. (4) MR (mixed representation).
Voronoi in 1908 and were originally applied to the study of the prox­ In 1998, Thrun (1998) proposed a method for extracting topological
imity problem of points in the plane (Preparata et al., 1990). The features from global metric maps, using distance information from sonar
proximity problem means that if there are certain points in a given sensors to extract features during robot navigation. A new node is
plane, and divide the plane according to these points, then any location generated when the distance information of the sonar happens to be
in the region is closer to the center point than to the other center points significantly different from the features of the known nodes. If the
(as shown in Fig. 4). In the planning of mobile robots, Voronoi diagrams feature information is similar to the features of the known node, a
with obstacles as solid objects are used to describe the networked probabilistic localization is performed using a partially observable
structure of feasible regions. Markov decision-making method and a topological map of the envi­
TM can effectively reduce the dimensionality of the metric space, ronment is generated at the same time. However, generating a topology
Voronoi diagram can convert the two-dimensional metric perceptual map in regions where the obstacle distance features are not significant
information into a network model of the accessible area, making plan­ enough is still not easy and also sensitive to dynamic environment with
ning a one-dimensional optimal node sequence problem, which is an sensor noise. In 1999, Yeap & Jefferies (1999) proposed a method to
effective topological representation that cost less storage space and extract a global topology map from a local metric map. The local maps
computation of planning time, etc. Voronni diagram represents the state are represented by grids, using a topology-like edge set to connect
of the neighborhood partitioning centered on obstacles in the environ­ multiple local maps that already exist. Since the local metric maps are
ment. Therefore, incremental topological environment modeling can be maintained in a small range, the errors can be neglected. Two levels of
achieved by modeling sub-regions. Gomez et al. (2020) proposed a local and global planning are used in path planning, i.e., regional
global topological representation for the large scale indoor scenes, planning based on occupying grids and global planning based on topo­
where each room is modeled as a submap of the topological map for logical connectivity relations, which is the most primitive Mixed
robot navigation and trajectory planning. Yang et al. (2021) presented a Representation.
new method for estimating the geometry of three-dimensional space MR can combine the efficiency of topology path planning with higher
using convex polyhedra. Then, the geometric information is used to modeling accuracy in metric space to achieve multi-level planning. A
divide the space into different regions. And these regions are added to topology-metric map construction method for efficient item search is
the topological map as nodes to guide the search process. Shin and Kim proposed in the study of home-service oriented robots to enable mobile
(2021) presented an algorithm to improve the optimality of planning robots to quickly locate target items in complex home environments
MPP (multi-agent path planning) solutions through parallel pebble di­ (Zhang, 2021), providing an important guarantee for robots to effi­
agrams. Embedding the pebble diagram into an arbitrarily complex ciently perform domestic operations. Unlike the traditional topology-
environment, it is shown that feasibility constraints can be transformed metric map, in this method, the topological nodes on the map are
into differentiable geometric constraints and that these constraints are created by referring to the environmental items, which reduces the
met the requirements of numerical optimization of the constraints. redundant nodes unrelated to the item search with lower searching cost
Stenzel et al. (2021) came up with an algorithm to compute a roadmap of the mobile robot, and improves the path planning efficiency. Gregorio
with intermediate points. The algorithm is able to make full use of as & Stefano (2017) proposed SkiMap, a multi-level environment model
much available space as possible and is able to guarantee the non- framework for robot navigation. SkiMap contains a 3D voxel grid map, a
overlapping between edges and vertices. 2.5D height map, and a 2D occupancy grid map, and supports contin­
TM path planning is efficient (paths may also be sub-optimal), but it uous updates to improve the accuracy of the map during walking.
requires little storage space, so it is suitable for applications in large-
scale environments. However, topological maps are more difficult to 2.1.2. Evaluation method for path quality
create and maintain, and are prone to confusion when two similar lo­ The main task in the path search problem for mobile robots is to find
cations exist in the environment. the correct search strategy. The search strategies, i.e., path planning
(3) GCM (geometric characteristic method). algorithms will be analyzed in detail in Section 3. In order to evaluate
GCM refers to a mobile robot that extracts relatively abstract geo­ the capability of each search strategy and measure their performance in
metric features from the environment perceived information, such as an objective way, in this section seven different evaluation method will
lines, arcs, and circles, to describe the environment (Guivant et al., be presented (Tsardoulias et al., 2016).
2000). This method builds a relatively compact model and stores a small (1) Average path planning time - tm. The average path planning time
amount of information, which makes it easy to perform the estimation of spent by an algorithm, this criterion is crucial because whatever algo­
position. However, the difficulty of extracting stable features from the rithm must plan the path for the mobile robot efficiently and give results
raw data that do not change with environmental changes and robot in the shortest possible time.
movement leads to the method being mainly applicable to structured (2) Relative standard deviation of the average path planning time.
operational environments. In addition, finding an exact match between The relative standard deviation of the average path planning time spent
observed features and continuously updated environmental information by an algorithm, RT, the calculation formula is,
is a technical difficulty. The shortcoming of the geometric representa­
tion is the difficulty of environmental geometric feature extraction,

4
L. Liu et al. Expert Systems With Applications 227 (2023) 120254

Table 1
Summary of LRS basic parameters.
Name Distance Channel Distanceerror Response Measurement Horiz Horiz Vertical Vertical Weight
(m) (cm) time frequency FOV resolution FOV resolution (g)
(ms) (Hz) (◦ ) (◦ ) (◦ ) (◦ )

Velodyne PUCK VLP-16 100 16 ±3 – 5 ~ 20 360 0.1 ~ 0.4 − 15 ~ +15 2.0 830
Velodyne ULTRA PUCK 200 32 ±3 – 5 ~ 20 360 0.1 ~ 0.4 − 25 ~ +15 0.33 925
VLP-32C
SickRS1104C-111010 64 4 ±6 – 5 ~ 15 275 0.25 − 7.5 ~ 0.25 1200
+7.5
SickRS1104C-111011 64 4 ±6 40 5 ~ 15 275 0.25 − 7.5 ~ 0.25 1200
+7.5
Ouster OS-1 16 120 16 ±1.5 10 10 ~ 20 360 ±0.01 − 15.8 ~ ±0.01 330
+15.8
Ouster OS-1 64 120 64 ±3 – 10 ~ 20 360 0.09 − 15.8 ~ 0.52 330
+15.8
LeiShen Ls Lidar C16 150 16 ±3 – 5 ~ 20 360 0.1 − 15 ~ +15 2 1500
3D RB-Lsl-01
LeiShen Ls Lidar C16 50 16 ±7 – 5 ~ 20 360 0.1 − 15 ~ +15 2 1500
RB-Lsl-10
Robosense RS LIDAR 16 150 16 ±2 – 5 ~ 20 360 0.1/0.2/0.4 − 15 ~ +15 2 840
Robosense RS LIDAR 32 200 32 ±3 – 5 ~ 20 360 0.1/0.2/0.4 − 15 ~ +25 0.33 1000
Robosense RS Ruby 200 128 ±3 – 10 ~ 20 360 0.2/0.4 − 15 ~ +25 0.1 3750
RPLIDAR A2 18 1 – – 5 ~ 15 360 – 0 – 190
RPLIDAR A3 25 1 ±5 – 10 ~ 20 360 0.225 0 0.225 190
Sick TIM551 8 1 ±6 67 15 270 1 0 1 250
Hokuyo PBS-03JN 6 1 – 180 – 178 1.8 0 1.8 500
Hokuyo URG-04LX- 5.6 1 ±3 100 10 240 0.36 0 0.36 160
UG01
Sick LMC132-11101 18 1 ±3 20 50 270 0.25 ~ 0.5 0 0.25 ~ 0.5 1100
VdS
Sick LMS111 20 1 ±3 20 25 ~ 50 270 0.25 ~ 0.5 0 0.25 ~ 0.5 1100
Sick LD-MRS400001 300 4 ±30 – 12.5 ~ 50 85 0.125 ~ 0.5 0 0.125 ~ 0.5 1000
SickRS1104C-111010 64 4 ±6 – 5 ~ 15 275 0.25 − 7.5 ~ 0.25 1200
+7.5
SickRS1104C-111011 64 4 ±6 40 5 ~ 15 275 0.25 − 7.5 ~ 0.25 1200
+7.5
Sick LD-MRS400001 30 4 ±30 – 12.5 ~ 50 85 0.125 ~ 0.5 0 0.125 ~ 0.5 1000
Sick LMC132-11101 18 1 ±3 20 50 270 0.25 ~ 0.5 0 0.25 ~ 0.5 1000
VdS
Sick TIM551 8 1 ±6 67 15 270 – 0 – 250

Here “-” represents unknown and FOV means field of view. It should be noted that the channel refers to the number of laser transceiver modules in the LRS. Single-line
LRS can only obtain two-dimensional plane information, while multi-line LRS can obtain three-dimensional information. The more channels, the higher the
comprehensive performance of the LRS, but more laser transceiver modules will also lead to the large volume and high cost of the LRS.

√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
∑
N•K deviation of path length after the calculation of an algorithm, similar to
(ti − tm )2 the relative standard deviation of execution time. The calculation for­
1
N•K
•
mula is,
i=1
RT = (1)
tm
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
In the formula, ti refers to the time spent by an algorithm in planning ∑
N•K
1
• (Di − Dm )2
the path in the ith time; N refers to the target points in the environment;
N•K
i=1
RD = (3)
and K refers to the times of the algorithm is executed. Relative standard Dm
deviation is a metric in statistics that reflects the degree of accuracy and
In the formula, Di refers to the path length obtained by running an
repetition. From the perspective of algorithm planning time, the relative
algorithm for the ith time. This criterion assesses the accuracy of the
standard deviation represents the reliability of an algorithm, that is, the
paths generated by an algorithm, i.e., whether the algorithm generates
degree of deviation of the time of each run of the algorithm from the
the same paths in different map environments.
average time.
(3) Average path length. The average length of a path generated after
Path curvature - RC. path curvature reflects the smoothness of the
an algorithm run - Dm. The average path length is directly related to the
path generated by the algorithm, i.e. the amount of rotational motion
path tracking time of a mobile robot, so this evaluation criterion is
that the robot must perform when tracking the path. Here, according
essential. We assume that each path has many target points and mark the
to the literature (Tsardoulias et al., 2016), we specify that the RC is
ith target point as Pathi. in addition, the path length of the robot to reach
bounded in the range [0,1]. The turning motion of the robot is
the target point PathSize=|Path|. So the formula for Dm is,
measured as it crosses a specific path. To simplify the computation,
∑
PathSize the default vehicle turns only at the target point and the purpose of
Dm = Dist(Path, Pathi+1 ) (2) the turn is to align with the next target point. What’s more, it is
assumed that the robot always rotates the smallest angle when
i=1
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
turning, so the angle range of the turn is [0◦ ,180◦ ]. Obviously, the
The formula Dist(p, q) = (xp − xq )2 + (yp − yq )2 refers to the
best case is that the robot is aligned with the final target and no
Euclidean distance between the point p and the point q on the coordinate
rotational motion is required, so the angle sum is 0◦ . On the other
axis.
hand, the worst case is to perform a 180◦ turn on each sub-target,
(4) Relative standard deviation of path length - RD. The standard

5
L. Liu et al. Expert Systems With Applications 227 (2023) 120254

Table 2
Performance parameters of common VS.
Name TS Pixel PS QE DN SNR AST DR(dB)
(μm) (%) (e) (dB) (γ)

IMX428 1.1′′ 3208*2200 4.5 51.79 2.77 40.44 5.38 70.59

IMX287 1/2.9′′ 720*542 6.9 61.82 3.80 43.51 7.55 74.43
IMX392 1/2.3′′ 1920*1200 3.45 60 2.38 40.32 4.81 71.45
IMX265 1/1.8′′ 2048*1536 3.45 59.66 2.31 40.30 4.93 71.74
AR0134 1/3′′ 1280*960 3.75 57 5.12 37.49 9.73 59.97

Here TS means target size; PS means pixel size; QE means quantum efficiency; DN means dark noise; SNR means signal-to-noise ratio; AST means absolute sensitivity
threshold; DR means dynamic range.

which is unrealistic, but from this it is possible to calculate the upper applications of laser technology, and laser ranging is characterized by
limit of RC, which is calculated as follows. long detection distance and high measurement accuracy. Laser ranging
∑PathSize methods are mainly divided into two categories, one is the pulse ranging
θi method; the other is the continuous wave ranging method (An & Zeng,
RC = i=1
∈ [0, 1] (4)
(PathSize − 1) • 180◦ 2001). LRS is similar to laser rangefinder in that it uses environmental
(6) Estimated path tracking time - ETT. ETT refers to the time spent geometric information as feedback as well as environmental portrayal,
by an algorithm-generated path robot in performing path tracking. with small feature changes under illumination changes as well as
Obviously, the time calculated by this criterion includes the time spent viewpoint changes, dense feedback information, and stable perception
by Dm and the rotational motion that the vehicle must perform. The of the exterior (Chen, 2020). Therefore, it has been applied in many
following assumptions are made to simplify the calculation a) the studies. The classical work on path planning for mobile robots based on
vehicle performs a purely linear or rotational motion, and not a com­ LRS, including the book Probability Robotic by Thrun (2002) and the
bination of both; b) the linear velocity is constant, denoted LS m/sec, AMCL open source localization algorithm package, have a considerable
and the angular velocity is constant, denoted RS rads/sec; c) the tran­ application base.
sition from linear to rotational motion is instantaneous and vice versa, d) LRS is a complex opto electromechanical system in essence. Its ad­
no friction or slippage occurs during each movement; e) the occupied vantages and disadvantages cannot be determined by a single index. It
grid cell size (in meters) is denoted as OGCD, and the accumulated needs to be comprehensively determined by combining actual applica­
∑
rotational movement is denoted as AS = PathSize− 1
θi . The resulting ETT tion scenarios and corresponding functions. In order to facilitate the
i=0
reader’s choice of lidar, we collected some information on the website
is calculated as follows.
and listed some basic parameters of LRS on the market in Table 1.
ETT =
Dm • OGCD AS
+ (5) The geometric structure information of laser feedback is for different
LS RS sensing modes of the same environment. For the problem of difficult
(7) Success rate - SR. SR is the most important measure of a path convergence of laser localization under similar structural features, its
planning algorithm and refers to the success rate of an algorithm plan­ difficulty in environmental modeling work in the face of outdoor open
ning a path that can reach the target point safely and collision-free, i.e., environments, and environments with highly repetitive geometric in­
its ability to compute the path regardless of a) environmental conditions formation, often due to the high repetition of environmental sensing
and b) the pose of the robot and the target. Path planning algorithms can information and the single content of feedback information.
fail due to two main reasons a) due to violation of execution time con­
straints and b) due to environmental constraints, i.e., narrow passages or 2.2.2. VS (visual sensor)
lack of traversable areas. As mentioned above, the number of executions VS is a typical passive sensor, in which the sensing information is
of each method in an environment is N•K. Therefore, by noting S as the obtained from the reflection of light from the surface of the object.
number of successfully created paths, the success rate of each algorithm Compared with LRS, VS has a wider detection range and can acquire rich
in a given environment is calculated as follows. image information. Depending on the number of VS and their functions,
they can be classified into monocular vision detection, stereo vision
SR =
S
• 100% (6) detection, and omnidirectional vision detection (Liu, 2018). In recent
N•K years, obstacle detection methods based on omnidirectional vision have
been widely studied due to the advantages such as wide perception
2.2. Local path planning range of omnidirectional vision system. Shi et al. (2017) built an
omnidirectional vision system on a humanoid robot to achieve real-time
The difference between local path planning and global path planning environment perception and obstacle detection in 360◦ . A feature point
is that the environmental information faced by the mobile robot is extraction system based on deep convolutional neural networks was
partially or completely unknown. Therefore, one of the focuses of local proposed by Daniel et al. in 2017, using a dual-network architecture that
path planning is the environment detection device, and the current open first extracts salient feature points and then computes point locations,
source mapping algorithms rely on two types of sensors, that is, LiDAR which is a real-time system but more streamlined and can easily run on a
sensors and vision sensors. In addition, sensor fusion as a starting point single CPU (Detone et al., 2017). In 2022, Chang et al. (2022) proposed a
to build a highly stable system with multi-sensor combinations is also an Yolov4 micro-network-based visual mapping algorithm. Meanwhile, a
inevitable trend of improving mobile robots (Nian, 2021). dynamic feature point elimination strategy based on traditional ORB­
SLAM is proposed. In order to obtain semantic information, target
2.2.1. LRS (laser radar sensor) detection is performed while extracting feature points of the image. In
LRS is frequently used in mobile robots for map building, before that addition, the polar line geometry algorithm and LK optical flow method
laser echo-based laser rangefinders were the main instrument for indoor are used to detect dynamic targets. Dynamic feature points are removed
engineering distance measurement. Its reliability is higher with more in the tracking thread and only static feature points are used to estimate
mature development, and its robustness to light environment is also the camera position. After evaluation on the TUM dataset, this method is
higher, but it is more expensive, and it is LRS that is configured in the featured by well real-time performance.
sweeping robots for home cleaning. Laser ranging is one of the earliest Since the VS carried by the mobile robot is not directly provided to

6
L. Liu et al. Expert Systems With Applications 227 (2023) 120254

the human eye for observation, but used for environmental modeling, well constrained to calculate and acquire more accurate positional data.
multiple parameters need to be considered when evaluating its image The earliest visual-inertial odometry (VIO) scheme was proposed by
quality. To facilitate readers’ choice of VS, Table 2 lists the performance Mourikis in 2007 as a Kalman filter-based real-time vision-assisted in­
parameters of common VS. ertial measurement unit navigation algorithm (Mourikis & Roumeliotis,
However, vision-based environment modeling still faces many 2007). After this, in 2014, Dr. Li from the University of California pro­
problems, mainly: (1) when using vision sensors for environment posed improved algorithms for fusion of vision sensors with IMUs in his
detection, due to uneven lighting and other factors, there may be object dissertation, and the paper used some small, inexpensive systems with
shadow interference in the environment, which may easily cause more limited sensing resources to do motion tracking and develop new
obstacle miss detection, wrong detection and other situations, making commercial opportunities (Li, 2014). In addition to vision and inertial
the subsequent construction of the navigation map lack of reliability, guidance fusion solutions, there are also a number of excellent solutions
how to adaptively and efficiently remove environmental shadow inter­ in the field of laser and vision fusion. V-LOAM is the system that excels in
ference based on environmental brightness information, so as to all aspects and was proposed by Zhang’s group in 2015. He is a self-
enhance the reliability of environmental perception information, has motion estimation method combining monocular camera’s and 3D
important research significance. (2) Binocular vision system can recover LiDAR, capable of accurately estimating the motion of 6 degrees of
parallax based on 2D image information and use triangulation principle freedom as well as the metric representation of the environment space,
to get 3D information in the scene, but the three-dimensional matching among others (Zhang & Singh, 2015). The entire system is divided into
link is of low accuracy and serious time consuming, so it is important two sequentially interleaved processes, one of which uses a visual
research value to improve the accuracy and real-time of binocular vision odometer at 60 Hz for motion attitude estimation; the other uses LiDAR
measurement. (3) Mobile robot navigation tasks require target locali­ at a low frequency of 1 Hz to assist in improving motion estimation and
zation, static target position acquisition is relatively simple and can be eliminating the cumulative error and point cloud distortion introduced
set in the map in advance, dynamic targets have motion characteristics, by the visual odometer, a method that does not require closed-loop
there will be background interference, partial occlusion, target rotation detection and corroborates the accurate estimation of the odometer
and rapid motion blur and other challenging scenarios, how to design during map building. The LIMO algorithm proposed by Johannes and
dynamic target tracking algorithms with real-time, robustness and ac­ Alexander et al. in 2018 is also typical for laser vision fusion, mainly
curacy has a greater challenge. (4) VS-based path planning, because the applied in autonomous driving, but does not use camera sensors to ac­
target is moving in real time, a more concise and real-time dynamic path quire images in real time, but uses images, LIDAR to acquire depth in­
planning algorithm is needed, and further research is needed on how to formation for RGB images, and the optical flow method to estimate
fuse multiple vision sensors information and combine image constraints camera motion, and the semantic method for the annotation of some
and spatial location constraints to better plan the path. landmark points, which is also able to determine the weights of target
points and remove outliers at the same time. The method has been open-
2.2.3. MIF (multi-sensor information fusion) sourced and has received high rankings on the KITTI dataset for large
MIF refers to the fusion of data from multiple sensors by some al­ scale applications in mobile robots (Graeter et al., 2019).
gorithm for mobile robots in complex, dynamic and uncertain envi­ For mobile machines applied to simple scenarios, 2D flat environ­
ronments, so as to make up for the deficiencies of a single sensor in some ment models are very efficient, which can accurately scan the distance of
aspects and reduce the redundancy of a single sensing information, in obstacles for the robot and accomplish real-time navigation tasks. For
order to make full use of the information from multiple sensors, so that mobile robots in complex scenarios, 2D maps are no longer sufficient.
the system can more reliably and accurately reflect the characteristics of From the viewpoint of security and practicality, it is necessary to
the external environment to which the robot is exposed in a shorter construct 3D maps. For the previously introduced LiDAR sensors and
period of time at a smaller cost (Chen, 2020). vision sensors, combined with the related research on multi-sensor
According to the above analysis of the pros and cons of LiDAR sen­ fusion mobile robot environment modeling by scholars at home and
sors and computer vision sensors, LiDAR has high perceptual stability in abroad, the complementary sensing capabilities of multiple sensors and
small indoor environments, but in large environments, the cumulative the application of information fusion methods can not only make the
error and the lack of closed-loop detection make the robot build maps mobile robot have the ability to perceive the 3D environment, but also
with low accuracy. On the other hand, scene recognition localization in the ability of mutual noise suppression between sensors to improve the
complex environments is not possible due to the small amount of in­ robot sensing accuracy, making the acquired information more accurate
formation collected by sensors. Computer vision modeling techniques and effective (Su et al., 2017c).
with low data acquisition difficulty and large amount of information are Mohanmadhossein investigated the close coupling of vision and
widely favored by researchers. However, the large amount of two- LIDAR applied to 3D environment perception for mobile robots (Daraei,
dimensional image information acquired by vision sensors can be used 2018), mainly including LIDAR and image temporal inconsistency dur­
for the implementation of closed-loop detection. However, vision solu­ ing LIDAR and vision fusion, and studied the modeling of occlusion due
tions lack the perception of distance, while laser sensing happens to have to parallax during the fusion of both information. LIC-Fusion proposed
a unique advantage in this area (Liang, 2016). Due to the complexity of an efficient tightly coupled multimodal laser-inertial-guidance-visual
the working environment and the uncertainty of the robot’s own state, it odometry fusion method within the framework of MSCKF (Zuo et al.,
is difficult to perceive and model the external environment with just one 2019; Ma et al., 2019), which can achieve robust self-motion estimation
sensor. To solve the problems of low information and poor robustness of under different environments and strenuous motions. Chinese university
single sensor, the fusion technology of multi-sensor information has research institutes and others have also carried out various researches
received more and more attention and research from scholars, making it and achieved some results. Meanwhile, a number of traditionally
a feasible method to improve the environment perception of robot. renowned schools and research institutes in robotics, mechanics, and
There are several types of solutions for sensor fusion, the more automation in China, such as Tsinghua University, Harbin Institute of
common one is the fusion of vision and inertial navigation, which re­ Technology, and Shenyang Institute of Automation, have made a series
quires the help of an inertial measurement unit (IMU), which measures of achievements in multi-sensing fusion algorithms and corresponding
the angular and linear velocities of moving objects and is a powerful hardware and software development (Ding et al., 2016; Zhao et al.,
sensor tool to calculate the positional attitude through the inter-frame 2017). There is a promising trend and huge market of this research both
velocity transformation of the robot (Zhang, 2013). However, the at home and abroad.
poses acquired by the inertial measurement unit alone can have signif­ However, the development of sensor fusion schemes is relatively
icant drift, and the image information acquired with the camera can be short and there are still many research results to be explored, and there

7
L. Liu et al. Expert Systems With Applications 227 (2023) 120254

Fig. 5. Schematic diagram of the RCD.

be made applicable to uncertain environments by different optimization

methods, but some researchers are prone to generate problems such as
high computation cost and low computation efficiency.

3.1.1. CD (cell decomposition approach)

CD first utilizes the grid method introduced in Section 2 to divide the
configuration space in which the mobile robot is located into a free grid
and an obstacle grid, and uses a connectivity map to traverse from one
cell to the next, forming a collision-free path from the starting point to
the target point (Šeda, 2007). The cell decomposition approach is clas­
sified into regular cell decomposition algorithm, approximate cell
decomposition algorithm, and exact cell decomposition algorithm based
on the strategy of dividing grids.
(1) RCD (regular cell decomposition).
RCD commonly pre-sets the size of the cell and uses a rectangular
shape, thus dividing the environment of the mobile robot into small cells
Fig. 6. Schematic diagram of the ECD.
of the same shape and equal size. For the safety of the mobile robot, we
also use inflation to enlarge the obstacles, i.e., whenever there is an
is no more mature scheme in this field for complex indoor environment obstacle in a small cell, no matter how large it occupies, this cell is
mapping. The fusion algorithm does not provide a closed-loop detection classified as an obstacle cell (as shown in Fig. 5). By counting the
module (Zhang & Singh, 2015), so the accuracy of the mapping needs to number of collected literature, the regular cellular decomposition al­
be improved, and the algorithm requires high hardware equipment, gorithm is the most common compared to the other two cellular
which is not fully applicable to small and lightweight indoor service- decomposition algorithms, which can plan paths for mobile robots well
oriented robots; some dense visual point cloud mapping can greatly in environments without localized minimal obstacles, which is very
restore Some dense visual point cloud mapping can greatly recreate applicable to drones and unmanned vehicles. Bao et al. (2022) added A*
indoor scenes, but the computation is huge and usually needs to be algorithm with artificial potential field method for unmanned path
accelerated by cuda and GPU, which consumes more computational planning under the shaft of a trackless robot based on RCD. Zhou et al.
resources in large scale scenes and is difficult to run well in embedded (2018) proposed a variable cell strategy based on RCD applied to the
systems. path planning of UAVs. Challita et al. (2018) incorporated a deep
learning algorithm in a honeycomb-shaped RCD for UAV path planning
3. Path planning algorithm during obstacle avoidance.
Although RCD is very applicable, it still has some drawbacks. On the
Since the 1950s, many algorithms have been proposed by researchers one hand, when a barrier is much smaller than the cell, it will still make
to seek the optimal path of a mobile robot. In this chapter, we analyze the whole cell become a barrier cell, this may lead to a path scheme that
mobile robot path planning from the perspective of algorithms. Ac­ is not optimal for system planning. On the other hand, if the cell is set too
cording to the characteristics of the algorithms, researchers generally small, it will lead to too much computation and become less efficient
classify them into three categories: classical algorithms, bionic algo­ (Debnath et al., 2021).
rithms and artificial intelligence algorithms. The classical algorithms (2) ECD (exact cell decomposition).
including cell decomposition method, sampling based method, graph The shape and size of cells in ECD are indeterminate; they are
search algorithm, artificial potential field, and dynamic window determined by the obstacle and the environment. The free space is first
approach. Among the bionic algorithms include genetic algorithm, ant decomposed into multiple non-overlapping small cells, which are usu­
colony optimization algorithm, particle swarm algorithm, firefly algo­ ally rectangular, trapezoidal or triangular. Then give each small cell a
rithm, bacterial foraging algorithm, cuckoo search algorithm and arti­ number. These small cells are precisely equal to the original space
ficial bee colony algorithm. Artificial intelligence algorithms mainly composed of free cells (as shown in Fig. 6). For the case where the
include fuzzy control algorithms and neural network algorithms. boundary of the obstacle in the environmental space is curved, Jung
et al. (2019) proposed an exact cell decomposition algorithm based on
3.1. The classical algorithm opposite angle (OAECD) based on ECD. OAECD algorithm is also focused
on the static environment and has poor robustness for dynamic envi­
Initially the classical path planning algorithms are very popular in ronment. In the Gonzalez’s literature the performance of the conven­
the field of autonomous navigation for mobile robots because they all tional algorithm was improved by changing the cell decomposition, the
have a common advantage, easily observed computation results. In graphical weights of small cells and the numbering order (Gonzalez
contrast, classical algorithms are more suitable for application in global et al., 2017). Park et al. (2015) described a method to represent the
path planning with a deterministic environment, but of course they can generation of local path using the ECD method.

8
L. Liu et al. Expert Systems With Applications 227 (2023) 120254

Fig. 7. Schematic diagram of the ACD.

division is performed for all slightly larger cells, and the procedure is
Table 3 then repeated among the smaller cells formed within the final bound­
Comparison of the different CD algorithms.
aries of the division until the limits of the solution are reached (as shown
Method Optimal Path Real time Memory Safety in Fig. 7).
RCD N N N Y Guruprasad & Ranjitha (2021) applied ACD to the field of full
ECD N N Y N coverage of mobile robot paths. Samaniego et al. (2019) proposed
ACD N N Y Y Recursive Rewarding Modified Approximate Cell Decomposition
Here Y stands for “Yes” and N stands for “No”. (RRMACD) to construct small cubic cells in the 3D environment in which
the UAV is located, and considers various constraints such as UAV
maneuverability, geometry, static and dynamic environmental obsta­
Table 4 cles, without invoking other algorithms to reduce the calculation and
The pseudo-code of the RRT’s algorithm. complexity.
Input: М,Qinit,Qgoal The sum space of the small cells formed by ACD does not match the
Result: A path Γ from Qinit to Qgoal free space of the original robot, and the amount of available space
1.
around the generated path is controlled by setting the size of the smallest
Γ.init();
2. for i = 1 to n do
3. Qrand ← Sample(M); cell. The free space of the approximate cell decomposition algorithm is
4. Qnear ← Near(Qrand,Γ); locally adapted to the shape of the obstacle in the vicinity of the
5. Qnew ← Steer(Qrand,Qnear,StepSize); obstacle, and the degree of adaptation is related to the predetermined
6. Ei ← Edge(Qnew,Qnear); minimum cell size. Thus, the minimum cell size setting is crucial.
7. if CollisionFree(M,Ei) then
8. Γ.addNode(Qnew);
Therefore, most methods operate in a hierarchical manner by generating
9. Γ.addEdge(Ei); an initial coarse decomposition and then locally refining that decom­
10. end if position until a free path is found or the decomposition becomes too
11. if Qnew = Qgoal then meticulous to continue (Latombe, 1991).
12. Success();
To sum up, CD is intuitive and can ensure the minimum local colli­
13. end if
14. end for sion probability of mobile robot. However, if the cells formed are rela­
tively rough, CD cannot guarantee that the path is optimal. On the
Here M stands the environment map, Qinit stands the starting point of the mobile
contrary, for RCD and ACD, if the cell division is too careful, it will
robot, and Qgoal stands the target point.
consume more computing time. For RCD, when an obstacle is far smaller
than the cell size, according to the traditional division scheme, it will
still cause the entire cell to become an obstacle cell, which may lead to
that the path planned is not optimal, and even cannot guarantee to find
an existing path. If we divide this cell into free cells, the length of the
path may be shortened, but the completeness of the planned path cannot
be guaranteed. In order to visually compare CD algorithms, we have
done so in Table 3 (Debnath et al., 2021).

3.1.2. SBM (sampling based method)

(1) The RRT (rapidly-exploring random tree) algorithm and the
optimization algorithm based on it.
The RRT is an active sampling-based algorithm proposed by LaValle
in 1998. The method is suitable for solving path planning problems with
different motion constraints and has the ability to handle multi-degree-
of-freedom problems. The principle of the RRT algorithm is to quickly
Fig. 8. Schematic diagram of the RRT algorithm. search the bit-shaped space nodes to generate a path connecting the
starting and target nodes (as shown in Table 4 and Fig. 8) (Abdallaoui
et al., 2022).
(3) ACD (approximate cell decomposition).
Usually, researchers can set an upper limit on the running time or an
All small cells of the ACD are of a predetermined shape, usually
upper limit on the number of searches in order to make the algorithm
rectangular. The whole environment is divided into several larger
manageable. If the target point cannot be reached within the limited
rectangles, each of which is continuous with the other, typically by the
times, the algorithm returns will fail. In the Palmieri’s literature, a
“quadtree” method and the “octree” method, and if the large rectangle
combination of (discrete) arbitrary angle search and (continuous) RRT
contains obstacles or boundaries, it is divided into four small cells. This

9
L. Liu et al. Expert Systems With Applications 227 (2023) 120254

Table 5 the path cost, and the search path strategies are both based on random
The pseudo-code of the RRT - Connect’s algorithm. sampling of the search.
Input: М,Qinit,Qgoal The asymptotically optimal RRT* algorithm, which improves the
Result: A path Γ from Qinit to Qgoal parent node selection on the original RRT algorithm, uses a cost function
1. Γa.init(); Γb.init(); to select the node with the smallest cost in the domain of the expanded
2. for k = 1 to K do
nodes as the parent, and at the same time, reconnects the nodes on the
3. Qrand ← Sample(M);
4. Qrand’←Sample(M); existing tree after each iteration, thus ensuring the computational
5. Qnear ← Near(Qrand,Γa); complexity and the asymptotically optimal solution (Karaman & Fraz­
6. Qnear’←Near(Qrand,Γb); zoli, 2011). Its principle is the same process as RTT before finding the
7. Qnew ← Steer(Qrand,Qnear,StepSize); point Qnew, and there are two stages after finding Qnew. (as shown in
8. Qnew’←Steer(Qrand’,Qnear’,StepSize);
Table 6 and Fig. 10) After these two steps, the mobile robot reduces
9. if CollisionFree(M,Ei) then
10. Γ.addNode(Qnew); unnecessary search and storage, which largely improves the search
11. Γ.addNode(Qnew’);
12. Γ.addEdge(Ei);
13. end if Table 6
14. if Qnew’=Qnew then The pseudo-code of the RRT*’s algorithm.
15. Success();
16. end if Input: М,Qinit,Qgoal
17. end for Result: A path Γ from Qinit to Qgoal
1. Γ.init();
Here М,Qinit,Qgoal have the same meaning as Table 4. 2. for i = 1 to n do
3. Qrand ← Sample(M);
4. Qnear ← Near(Qrand,Γ);
motion planning is applied to a small time controllable Nonholonomic
5. Qnew ← Steer(Qrand,Qnear,StepSize);
Wheeled Mobile Robot (Palmieri et al., 2016). 6. if CollisionFree(qnew) then
The RRT algorithm can only find the path, but it cannot guarantee 7. Qnear ← NearC(Γ,qnew);
that the path is optimal. And it is a purely random search algorithm 8. qmin ← ChooseParent(Qnear,qnear,qnew);
insensitive to the environment, and when the environment contains a 9. Γ.addNodEdge(Qnew,qnew);
10.
large number of obstacles or narrow passages, the convergence of the
Γ.rewire();
11. end if
algorithm is slow and the efficiency drops dramatically. Next, we pre­ 12. end for
sent several classical solutions for optimizing the RRT algorithm. The
Here М,Qinit,Qgoal have the same meaning as Table 4.
first method is the RRT - Connect algorithm, based on the blindness of
the RRT search space, the shortcomings of the node expansion link lack
of memory, in order to improve the search speed in space. The RRT al­
gorithm is coupled with a two-tree bi-directional exploration guidance
strategy, and a greedy strategy is added to the growth method to speed
up the searching and reduce the invalid search of the blank area, and
save time (as shown in Table 5 and Fig. 9).
A flexible multi-directional fast exploration tree generation method
was used in the Qian’s literature (Qian et al., 2020). Based on the
principle that trees grow centripetally, new multi-directional trees will
be constructed as needed to achieve a specific coverage of space. Then,
based on the previous path exploration vertices, the complete paths are
formed and optimized by the tree fusion method, and finally the local
optimal paths are generated. Experimental results show that the method
can effectively improve the search efficiency at a low computational
cost, ensuring that each sampled point is not useless information and
minimizing the search times of the whole map. the RRT-Connect algo­
rithm adds heuristic strategies, and greedy ideas, but the common
drawback of both the RRT algorithm and the RRT-Connect algorithm is
that neither of their paths is optimal. No function is added to evaluate Fig. 10. Schematic diagram of the RRT* algorithm.

Fig. 9. Schematic diagram of the RRT-Connect algorithm.

10
L. Liu et al. Expert Systems With Applications 227 (2023) 120254

Table 7 the one hand, making it more efficient. The RRT* algorithm based on
The pseudo-code of the PRM’s algorithm. heuristic information-guided local sampling was also proposed (Ryu &
Input: n,k Park, 2019), where the principle is to use heuristic information to guide
Result: A road map G= the search to find the optimal path quickly after RRT* finds the feasible
(V,E) path for the first time.
1. V← Ø
The paths obtained by the above methods are often convoluted, so
2. E← Ø
3. while |V|＜n do based on the RRT* algorithm, the RRT*-smart algorithm was proposed
4. repeat in Islam’s literature (Islam et al., 2012). RRT*-Smart is identical to RRT*
5. q ← a random configuration in Q in the first stage of operation, but after finding a feasible path from the
6. untile q is collision-free starting point to the end point it starts to consider optimizing the path
7. V ← V∪{q}
and turning the curve into a straight one. The process actually starts
8. end while
9. for all q ∈ V do from the starting point and keeps looking for whether it can connect
10. Nq ← the k closest neighbors of q chosen from V directly to the preceding nodes without any obstacles. If you connect one
according to dist layer directly forward there is one more straight line and one less curve.
11. for all q’∈Nq do
In addition to the above classical optimization schemes, the opti­
12. if (q,q’) ∕
∈ E and Δ(q,q’) ∕= NIL then
13. E ← E∪{(q,q’)}
mization about RRT series of algorithms has been very popular in recent
14. end if years. In the Xu’s literature, a simplified map-based regional sampling
15. end for RRT* algorithm (simplified map-based regional sampling RRT*, SMRS-
16. end for RRT*) was proposed for the shortcomings of RRT* algorithm such as low
Here n stands the number of nodes to put in the roadmap, k stands the number of search efficiency (Xu et al., 2020), slow convergence speed and exces­
closest neighbors to examine for each configuration. sive memory consumption. In the Wang’s literature, a path planning
method based on an improved RRT* algorithm was proposed for solving
the path planning problem of mobile robots based on articulated
Table 8 structures and drifting environmental conditions with vectorized maps
Comparison of the different SBM algorithms. and dynamic constraints to achieve robot motion (Wang et al., 2022a).
Method Optimal Path Real time Memory Safety In the Li’s literature, an improved motion planning algorithm (fast-
RRT N N N Y RRT*) based on a hybrid sampling strategy and a backtracking-based
RRT-Connect N N N Y adaptive GA is proposed (Li et al., 2022a). Researchers have been
RRT* Y N N Y working on the optimization of the RRT algorithm, focusing on three
PRM N N N Y
main aspects: (a) the random sampling approach. (b) The strategy of
branch expansion. (c) The combination with mobile robot kinematics.
These methods have the advantages of fast search of free space and real-
Table 9 time collision detection. However, this class of methods produces un­
The pseudo-code of the Dijkstra’s algorithm. stable paths with heavy dependence on the nearest node, and addi­
Input: G,s,t tionally requires complete environmental information for node collision
Output: a path from s to t with the detection (Wang, 2020).
minimum weight
1. for each vertex v do
2. {status[v] = 0;wt[v] = -1;dad[v] = -1;} PRM (probabilistic roadmap method)
3. status[s] = 2;wt[s]=+∞；
4. for each edge [s,w] do The PRM proposed by Overmars et al. in the early 1990s (Chen et al.,
5. {status[w] = 1;wt[w] = weight(s,w);dad[w] 2021a). This method is based on the sampling algorithm, which solves
= s;}
the difficulty of constructing an effective path graph in high-
6. while there are fringes do
7. v = the fringe with the min wt-value dimensional space. This algorithm represents the connectivity of path
8. status[v] = 2; graph by sampling in configuration space, collision detection of sam­
9. for each edge[v,w] do pling points, and testing whether adjacent sampling points can be con­
10. case 1. status[w]==0;
nected (pseudo-code as shown in Table 7).
11. {status[w] = 1;wt[w] = wt[v] + weight(v,w);
dad[w] = v;} One of the great advantages of this method is that its complexity
12. case 2. (status[w]==1) and (wt[w]＞(wt[v] mainly depends on the difficulty of finding the path, and has little
+ weight(v,w)); relationship with the size of the whole planning scene and the dimension
13. {wt[w] = wt[v] + weight(v,w);dad[w] = v;} of the configuration space. However, when the planned path needs to
Here G stands a graph, s stands the starting point of mobile robot, t stands the pass through dense obstacles or narrow passages, the efficiency of PRM
target point. becomes low. Researchers have done a lot of research to improve these
deficiencies. Zou et al. (2019) introduced an improved strategy of
efficiency. adding nodes on the basis of the basic PRM algorithm to optimize the
The RRT* algorithm has been shown to be an asymptotically optimal sampling points trapped in obstacles, but the increase in the number of
path planning method. In other words, when the number of samples nodes makes the computing cost higher, which affects the real-time
tends to infinity, the solution obtained by RRT* is the optimal solution performance of the algorithm; Ravankar et al. (2020) used hierarchi­
that converges with the probabilistic solution. Recent research de­ cal hybrid PRM algorithm and APF method to carry out global path
velopments have proposed many improved RRT* algorithms that use planning, and used a node distribution decomposition method to divide
different methods to improve the performance of RRT*. Qi et al. (2020) the road map into high potential and low potential areas, which
extended the RRT* algorithm to an unknown dynamic environment and improved the efficiency of the algorithm in searching the path; Inspired
verified the effectiveness of the new algorithm from both simulation and by the adjacent correlation matrix, Esposito and Wright (2019) proposed
real experiments. Brunner et al. (2013) used biased sampling to improve a processing algorithm to optimize the probability roadmap, simplifying
RRT* by controlling the direction of the RRT* algorithm iteration by the calculation required for processing the number of convex elements
biasing the operation to distribute the samples in our desired region on and nodes in free space.
To sum up, SBM has stronger search ability when the spatial

11
L. Liu et al. Expert Systems With Applications 227 (2023) 120254

Table 10 three main approaches to improve the traditional Dijkstra algorithm.

The pseudo-code of the A*’s algorithm. One is to analyze and improve the space complexity of the algorithm to
1. Mark P[star] as openlist. improve storage efficiency and save space. For example, Fink et al.
2. while openlist ∕ = empty do (2019) applied Dijkstra’s algorithm to a planetary rover in a 3D envi­
3. Select the node P[i] from the openlist whose value of evaluation function F(P ronment, Balado et al. (2019) applied Dijkstra’s algorithm in combina­
[i]) is smallest.
tion with point cloud maps for path planning in urban 3D environments,
Mark P[i] as closelist.
4. if P[i] = P[end] then and Yu et al. (2018) introduced Dijkstra’s algorithm in the field of path
5. return “path is found’. planning for driverless cars. The second one is to analyze and improve
6. else the time complexity of the algorithm. The traditional Dijkstra’s algo­
7. Select the successor node Pi[j] around the node P[i], and calculate F(Pi[j]). rithm is inefficient with a long running time, in order to improve the
8. if Pi[j] belons to obstacle or closelist node then
operational efficiency and reduce the time complexity, Dijkstra’s algo­
9. continue;
10. end if rithm is improved and applied to dynamic scenarios (Guo et al., 2019;
11. Mark Pi[j] as openlist. Zhang et al., 2018b). The third one is to apply the algorithm in different
12. if Pi[j] belons to openlist and F(Pi[j])＜F(Pm[j]) when P[m] was marked as fields to develop the application space of the algorithm and enrich the
closelist then
application scope of the algorithm, which has been studied by many
13. Set parent node of P[j] as P[i],F(P[i]) = F(Pi[j]).
end if
scholars. Souza et al. (2019) applied Dijkstra’s algorithm to trans­
14. end if portation cost calculation in the mining industry, Santos et al. (2019)
15. end while introduced Dijkstra’s algorithm in the field of power transmission, and
16. return “the path cannot be found”. Tang et al. (2018) introduced Dijkstra’s algorithm in broadband
17. Mark P[star] as openlist.
technology.
18. while openlist ∕ = empty do
(2) A* algorithm and the optimization algorithm based on it.
Here P[star] stands the starting point of mobile robot, P[end] stands the target The A* algorithm is an improved version of Dijkstra’s algorithm,
point. which uses a heuristic function H(n)to filter the extended nodes, greatly
reducing the number of total extended nodes and improving the effi­
ciency of trajectory planning. The A* algorithm is a heuristic search
Table 11
Comparison of the different cell decomposition algorithms. algorithm for solving the shortest source path in a static environment
(Wang et al., 2022c). The principle of the algorithm is shown in
Method A* Dijkstra
Table 10.
Advantage Global optimum; Introducing Global optimum In the Grid map environment, the A* algorithm can effectively
heuristic information for
represent arbitrarily shaped obstacles in two-dimensional space. How­
higher efficiency
Disadvantage Lack of adaptability Low efficiency; Lack of ever, in the three-dimensional environment, the valuation function be­
adaptability comes complicated with the increase of dimensionality, so the A*
Personalized Optimize heuristic function; Hierarchical search; algorithm is less applied in the high-dimensional space. To overcome
improvement Change environment Improve data structure; Add this drawback, many studies have proposed optimization schemes based
scheme modeling method heuristic information
on A* algorithms, such as Dynamic A* (D*), Field D*, Theta*, Anytime
Common Two way search, expand the search angle; Introducing path
improvement optimization rules repairing A* (ARA*), and so on (Wang, 2020). The D* algorithm was
scheme proposed by Stentz (1997), which is based on the principle of first
generating an initial path with the help of the A* algorithm, and then
relying on the on-board sensors to detect the external environment
dimension is high or the size is large. However, SBM is sampled in the
during the operation of the mobile robot to correct the initial path in real
whole space, so the amount of calculation is large, and there is still much
time. The Field D* algorithm was proposed by D. Ferguson and A. Stentz
room for improvement in planning efficiency. For PRM, if the sampling
in 2006 (Ferguson & Stentz, 2006), which is based on the D* algorithm
parameters are not set properly, even if there is a reasonable path in the
and introduces a linear interpolation function as a way to generate
space, the complete path may not be found. In order to visually compare
smooth paths. Daniel et al. (2010) proposed the Theta* algorithm in
SBM algorithms, we have done so in Table 8.
2007, which also aims to smooth the paths generated by the A* algo­
rithm as a way to save the energy consumption of the mobile robot. The
3.1.3. GSA (graph search algorithm)
ARA* algorithm, on the other hand, automatically adjusts the search
(1) Dijkstra algorithm.
results based on the available planning time (Ziegler et al., 2008),
Dijkstra’s algorithm was proposed by Dutch computer scientist
reusing the previous search results in the process of search path opti­
Edsger Wybe Dijkstra in 1959, and it has been successfully applied to 2D
mization, and is therefore more efficient than other algorithms. In
path planning for mobile robots, computer science, geographic infor­
addition to these classical optimization strategies, researchers have done
mation science, and transportation (Wei et al., 2019; Alyasin et al.,
a lot of work in recent years to improve the A* algorithm. In the liter­
2019; Yu & Ge, 2017). The iterative process of path planning of tradi­
ature (He et al., 2022), an obstacle avoidance algorithm for multi-vessel
tional Dijkstra algorithm is shown in the Table 9 (Sun, 2017a).
encounters is proposed, which generates a quaternion ship domain
Huang (2018b) introduced the concept of equivalent path and
based on automatic identification system data and calculates the ship
determined the conversion formula between equivalent path and actual
navigation risk cost by combining the quaternion ship domain and the
path, thus improving the traditional Dijkstra’s algorithm. Dijkstra’s al­
potential field to ensure the safety of the ship. Martins et al. (2022)
gorithm is a graph search algorithm for single-source optimal path. It
designed an improved multi-objective A* (IMOA*) algorithm for path
performs node expansion by first ranking the total path cost F(n) of each
planning of mobile robots in large workspaces. Wang et al. (2022b)
node, and then selecting the node with the lowest cost to expand until
introduced road condition indicators in the evaluation function of the A*
reaching the target location. The ranking property of the algorithm
algorithm to reduce the road cost of the mobile robot while walking.
determines that the algorithm must be extended to all nodes whose total
In order to visually compare CD algorithms, we have done so in
cost is less than that node before extending to a certain node, which
Table 11.
results in the low planning efficiency of Dijkstra’s algorithm, so the
method is often used as global path planning for mobile robots or local
3.1.4. Artificial potential field algorithm (APF)
path planning for scenes with few nodes Sun (2021). So far, there are
Platonov in 1974 presented the APF approach. The method is to

12
L. Liu et al. Expert Systems With Applications 227 (2023) 120254

Table 12 opposite in direction, and the robot falls into a local stable state and
The pseudo-code of the APF’s algorithm. cannot continue to move forward, which is called a local optimal solu­
1. x(0) = xstart tion (as shown in Fig. 12). In addition, due to the mismatch between the
2. y(0) = ystart step size of the mobile robot and the control distance of the obstacle and
3. θ = θstart the discretization of the solution process, APF is prone to the problem of
4. i=0
path oscillation. In order to overcome the shortcomings of the artificial
5. while ∇U(x(i)) ∕ = 0 and ∇U(y(i)) ∕
= 0 do
6. θ = arctant(∇U(x(i))/∇U(y(i))) potential field algorithm, Du et al. (2021) proposed an improved APF
7. x(i + 1) = x(i) + γ*cos(θ) path planning algorithm for unknown environments and dynamic ob­
8. y(i + 1) = y(i) + γ*cos(θ) stacles, and simulation experiments showed that the improved algo­
9. i++ rithm could resolve the path planning failure with a success rate of 95%.
10. end while
The existing literature on the application of APF to mobile robot path
Here xstart stands the X-axis coordinate of mobile robot starting point, planning is mainly focused on solving the APF local minima problem by
ystart stands the Y-axis coordinate, θstart stands the included angle be­ optimizing and improving the gravitational and repulsive potential
tween the line between the starting point and the coordinate point and functions or adding other additional conditions. Yun and Tan (1997)
the positive direction of the X axis. proposed a simple and practical algorithm for escaping local minima in
the potential field based motion planning. The algorithm switches be­
tween wo control modes: the potential field based control mode and the
wall-following control mode. The new algorithm switches to a wall­
following control mode when the robot falls into a local minimum.
Szczepanski et al. (2022) proposed top quark-based mechanism which
can prevent the robot from stagnation in a local minimum, and allow to
bypass the local minimum with a smooth movement.
A MTAPF (multiple sub-target artificial potential field) deterministic
algorithm with improved APF algorithm is proposed by Sang et al.
(2021), where the global optimal path is first generated by the improved
heuristic A* algorithm, and the optimal path is divided into multiple
sub-target points to form a sequence of sub-target points. MTAPF can
greatly reduce the probability of a mobile robot falling into local minima
and help the robot to get out of local minima by switching target points.
An optimization framework based on the APF model is proposed by Pang
et al. (2021) for changing crossing waypoints, route segments, and
spatial topological maps of prohibited, restricted, and hazardous areas.
Based on this framework, an adaptive approach is proposed to dynam­
ically control the optimization process with the objective of minimizing
Fig. 11. Inaccessibility of targets. the total route cost. Using the APF algorithm, a neural network training
scheme is proposed by Wang et al. (2021), the robot’s workspace is
divided into two parts, Global safety area and local hazard area. In the
global safety zone, the robot receives only the gravitational force from
the target and moves directly toward it. In the danger region, in addition
to attractive forces, the robot receives repulsive forces from obstacles.
The repulsive force and the angle between the obstacle and the target are
then used as inputs to the fuzzy control system to output the deflection
angle of the robot, and the final direction of movement of the robot is
determined by summing this deflection angle and the direction of
attraction. An improved APF was proposed by Luo et al. (2019b). The
repulsive force of obstacles outside a certain angle range in the direction
of robot movement is excluded to reduce the computational effort; the
distance factor between robot and target point is introduced in the
Fig. 12. Optimum local solution.
repulsive force function to solve the target unreachability problem; the
tangent method is used to solve the problem of local minima formed by a
construct virtual gravitational force and virtual repulsive force in the
single obstacle, and the search method is used to solve the problem of
environment where the robot is located. The virtual gravitational force
local minima formed by multiple obstacles acting simultaneously. The
is generated by the target point and the virtual repulsive force is
adaptive step size adjustment algorithm is proposed considering the
generated by the obstacle, and the robot moves toward the target point
complexity of path planning. A MTAPF (multi-sub-target artificial po­
under the joint action of gravitational force and repulsive force (Zhang
tential field) deterministic algorithm based on improved APF is pro­
and Ming, 2021d; Liu et al., 2022). The core of APF is gradient descent
posed by Shin and Kim (2021). MTAPF belongs to local path planning
method, which is a common method to solve optimization problems.
algorithm, which refers to a global optimal path generated by an
From the initial state, move in the opposite direction to the gradient. The
improved heuristic A* algorithm, by which the optimal path is divided
pseudo code of APF is shown in the Table 12.
into multiple sub-objective points to form a sequence of sub-objective
Through analysis, it can be known that when the mobile robot is very
points. Zheyi & Bing (2021) added an angle function to the conven­
close to the target point, the gravitational force is small. If there is an
tional algorithm to match with the Force Field function to improve the
obstacle nearby at this time, the repulsive force will be large, and the
effectiveness of the conventional algorithm. Feng et al. (2021) intro­
combined force will cause the robot to deviate from the target point, and
duced a safe distance in APF so as to generate a safe path in a simulated
such a situation is called target unreachable (as shown in Fig. 11). On the
traffic scenario. Sepehri & Moghaddam (2021) combined the advan­
other hand, in the process of robot movement, there may be a situation
tages of the RRT algorithm and APF, first using RRT to generate a
where the combined force and the repulsive force are equal in size and
feasible path with attractive and repulsive potential fields implemented

13
L. Liu et al. Expert Systems With Applications 227 (2023) 120254

Table 13 these simulated trajectories through scoring rules, transforming the

The pseudo-code of the DWA’s algorithm. obstacle avoidance problem of the robot into the optimal speed execu­
1. BEGIN DWA(robotPose,robotGoal,robotModel) tion problem, which has obvious advantages in feasibility and motion
2. desiredV = calculateV(robotPose,robotGoal) continuity. DWA pseudo code is shown in Table 13.
3. laserscan = readScanner() In DWA, the evaluation function consists of three weighting items:
4. allowable_v = generateWindow(robotV, robotModel)
the guide angle of the target point, the distance from the obstacle and
5. allowable_w = generateWindow(robotW, robotModel)
6. for each v in allowable_v the moving speed of the robot. Compared with other obstacle avoidance
7. for each w in allowable_w algorithms, DWA pays more attention to the limitation of the maximum
8. dist = find_dist(v,w,laserscan,robotModel) speed and acceleration of the robot. Therefore, from this point of view,
9. breakDist = calculateBreakingDistance(v) the control command of DWA seems to be more in line with the move­
10. if (dist > breakDist) //can stop in time
ment of the mobile robot in the actual situation. However, the distri­
11. heading = hDiff(robotPose,goalPose, v,w)
12. clearance = (dist-breakDist)/(dmax - breakDist) bution of obstacles in the robot environment is usually complex, and the
13. cost = costFunction(heading,clearance, abs(desired_v - v)) weight parameters in the objective function of DWA need to be adjusted
14. if (cost > optimal) to improve the flexibility of the algorithm in different environments. The
15. best_v = v
traditional DWA and its optimization algorithm usually use constant
16. best_w = w
17. optimal = cost
weight parameters when facing different obstacle distributions. In the
18. set robot trajectory to best_v, best_w past few years, many optimized DWAs have also been proposed. Zhang
19. END et al. (2015) proposed the combination of DWA and fuzzy logic to
Here robotPose stands the current position of the robot, robotGoal stands the analyze the information about targets and obstacles, and use fuzzy logic
target point, robotModel stands the robot model. to adjust the weight combination. The proposed method can make the
robot move more smoothly and safely. On this basis, Huang (2018a)
proposed a parameter fuzzy control DWA based on radar information to
Table 14 improve the adaptability of mobile robots to complex environments in
The pseudo-code of the GA’s algorithm for path planning. local path planning. Li et al. (2022b) proposed an improved A* and DWA
INPUT: map,start point,target point, algorithm, using laser radar to build maps, and realized path planning
population,maximum number of iteration and automatic navigation of orchard robots. Wu et al. (2022) proposed a
OUTPUT: optimal path hybrid dynamic path planning algorithm based on improved A* and
1. SEE ← See(map,target point) //Traverse
improved DWA. The improved DWA evaluation function can ensure that
the set of visual points
2. X ← Initial(map,starting point,target point, the local path of the mobile robot is closer to the global optimal path.
population) //Initial population Zhenyang & Wei (2022) combined A* algorithm with adaptive DWA,
3. fx ← fitness(X) //Calculate initial fitness and used optimization A* to guide the global path planning of adaptive
4. gen ← 1
DWA. This algorithm is more consistent with the motion characteristics
5. while gen＜maximum number of iteration
do of mobile robots.
6. fit ← 1./(fx + 1) //Calculate the average
fitness. 3.2. Bionic algorithm
7. Xsel ← select(X,fit) //Select operation.
8. Xsel ← crossover(Xsel,fit) //Crossover
operation(Xsel).
Bionic algorithm is a random algorithm inspired by the phenomenon
9. Xsel ← mutation(Xsel,fit,map) of biological swarm intelligence in nature. Based on the principle and
//Mutation operation. path search mechanism of bionic algorithm, it can be divided into
10. X ← evolving(Xsel,SEE,map) //Evolving functional bionic algorithm, informational bionic algorithm and
operation.
component bionic algorithm. Functional bionic algorithm is an algo­
11. fx ← fitness(X) //Calculate the new fitness
12. gen ← gen + 1 rithm proposed to simulate the role of a specific physiological function
13. end while of an organism in foraging, social living, escape or other biological be­
14. return optimal path haviors. For example, genetic algorithm, cuckoo search algorithm,
Here map stands the environment map, start point stands the starting point of the sparrow search algorithm. Information bionic algorithm is an algorithm
mobile robot, target point stands the target point, population stands initial pop­ proposed to simulate the basic principles of information transmission
ulation and maximum number of iteration stands iterations. and intelligent behavior of organisms in organs, individuals and groups.
For example, Particle swarm optimization, bacterial foraging optimi­
at all points along the path and used in a gradient optimization algo­ zation, artificial bee colony algorithm, grey wolf optimizer. component
rithm to determine joint trajectories to reach the desired position. In the bionic algorithm simulates the behavior and function of organisms from
literature by Liu et al. (2021d), the two potential fields fused adaptive the perspective of chemistry, and realizes the activity behaviors such as
path planning system consisting of two parts, a potential fields fusion mate seeking and aggregation by simulating the secretion and release of
controller and an adaptive weight assignment unit is proposed. trace chemicals of insects, such as ant colony optimization algorithm
In summary, the Artificial Potential Field Method (APF) has its and firefly algorithm (FA) (Li et al., 2021).
unique advantages in local path planning for mobile robots, which does
not need to generate a complete trajectory in advance, but only de­ 3.2.1. GA (genetic algorithm)
termines the next stage of motion based on the current state and the GA was formally proposed by Holland and his students in 1975，
surrounding environment, and this algorithm plans smooth trajectories which and applied to optimization, machine learning and other issues.
with less computational effort and can update real-time data to adapt to Its core is to use genetic operators to simulate selection, crossover and
the environment (Zhang, 2021b). variation in natural evolution and produce new populations that are
more adaptable to the environment (pseudo-code as shown in Table 14)
3.1.5. DWA (dynamic window approach) (Xu et al., 2021; Lin et al., 2021; Mahmud et al., 2019).
DWA is an algorithm proposed by Fox et al. (2002) that combines Lamini et al. (2018) proposed an improved crossover operator,
robot curvature and speed. This algorithm will simulate the trajectory of which is used to solve the path planning by GA in static environment.
the robot in a period of time according to the speed, and then it will score The proposed crossover operator avoids premature convergence and
provides a better and more feasible path than its parent generation.

14
L. Liu et al. Expert Systems With Applications 227 (2023) 120254

Table 15 transition rules, which leads to the lack of flexibility and the failure of
The pseudo-code of the ACO’s algorithm for path planning. solving the premature convergence of the algorithm.
1. procedure ACO For the mobile robot path planning problem, many scholars pro­
2. for each edge posed new improvement methods based on existing improvement stra­
3. set τ0. tegies. Jiao et al. (2018) proposed a polymorphic ACO for path planning,
4. end for
adaptive state transfer strategy and adaptive pheromone update strategy
5. while not stop
6. for each ant k was applied to ensure the relative importance of pheromone intensity
7. randomly choose an initial city. and heuristic information in the iteration of algorithm. Miao et al.
8. for i = 1 to n (2021) proposed an IAACO (improvement adaptive ant colony algo­
9. choose next city j with the probability rithm). Firstly, in order to speed up the real-time and safety of robot path
10. given
planning in IAACO, angle guiding factors and obstacle exclusion factors
11. end for
12. end for were introduced in the transition probability of ACO; secondly, adaptive
13. compute the length Ck of the tour constructed by the kth ant. adjustment factors and adaptive pheromone volatility factors were
14. for each edge introduced in the pheromone update rule of ACO to balance the
15. update the pheromone value
convergence and global search capability of ACO; finally, the multi-
16. end for
17. end while
objective performance index was introduced in ACO, which trans­
18. end procedure formed the path planning problem into a multi-objective optimization
problem, realizing the global integrated optimization of robot path
Here τ0 stands the initial pheromone value, n stands the number of ants.
planning. Ji & Liu (2022) used ACO to find a fastest path for the robot,
traversing all the target points. Yang et al. (2022) improved the tradi­
Therefore, the convergence rate of this algorithm is faster. Nazarahari tional ACO in terms of both pheromone updates and heuristic functions,
et al. (2019) proposed an improved GA, which can improve the original providing a strategy for solving the deadlock problem. Combined with
path in continuous space and find an optimal path between the original other methods, ACO is featured by strong robustness, fast optimization
position and the destination position, as well as extended to the path search, and strong convergence. Although these fusion strategies can
planning of multiple mobile robots. Shivgan & Dong (2020) proposed a learn from others to effectively improve the ACO’s optimization-seeking
path planning model of UAV based on the acceleration, deceleration, ability, the algorithm become more complex and require more optimi­
hovering and turning of UAV, considering the total energy consumption zation time. Chen et al. (2021b) proposed an improved ACO-APF algo­
of mission. It can reduce the energy consumption of UAV by reducing the rithm based on grid maps, which is used in local and global path
number of turns of UAV. Rahmaniar & Rakhmania (2022) proposed an planning of unmanned boats in dynamic environments. On the basis of
optimal path planning method based on GA. Firstly, find the feasible ACO, a single-step search strategy was placed by a multi-step search
path points by performing local search. Then, optimize these points to strategy, the pheromone update mechanism was redesigned by Xue et al.
find shortest path. When calculating the optimal path, the position of the (2021), and path smoothing was disposed to improve the performance of
points on the path will be smoothly move to avoid obstacles in the the algorithm. Yan (2021) improved the initial pheromone, transition
environment, which makes this method have advantages in dynamic probability and pheromone update strategy. An adaptive ACO path
environment. Chen et al., (2021c) developed a new fitness function of planning method was proposed by Li et al. (2021). The algorithm opti­
GA, which can consider the constraints, such as obstacles of freedom, mized the initial pheromone matrix based on the environment map,
length of the path, smoothness of the path, and visibility in camera which reduced the blindness of the initial ant colony in path finding and
roaming process. In addition, a new evolution operator is introduced in improved the convergence speed. Then an adaptive heuristic function
GA, greatly reducing the iteration times and improving the efficiency of was used to adjust adaptively according to its different proportions in
GA. In order to improve the ability of receiving GPS (global positioning the process of the algorithm, preventing the algorithm from falling into
system) signals for wall-inspecting robot, Tao et al. (2022) introduced local optimum. The pheromone was updated according to the corner
signal influencing factors into the fitness function of GA. Germi et al. points of the planned route to reduce the sharp angles and unnecessary
(2018) fused GA with APF and realized the adaptability of the algorithm turns of the route and further optimize the route. Zhu et al. (2021)
by changing the population generation method of GA in each iteration. proposed an algorithm that can adaptively adjust the pheromone co­
Liu et al. (2021b) designed an AGA (adaptive GA) to solve the task efficients, which can allocate pheromone with more suitable proportion.
allocation, and a fitness function considering the current iteration times Zhang et al. (2021c) used the non-uniform distribution to construct the
and the maximum iteration times is proposed in order to improve the initial pheromone, reducing the blindness of the search in the initial
convergence performance of AGA. stage, as well as a diffusion model was used to enhance the exploration
and collaboration among ants. Wang et al. (2021c) proposed the heu­
3.2.2. ACO (ant colony optimization) ristic information by the directional information of the starting, ending
In 1992, Dorigo et al. (1991) proposed ACO in his doctoral disser­ points and the turning angle. The improvement of the heuristic infor­
tations, which was inspired by the behavior of ants discovering paths mation increased the search direction and reduce the turning angle of
when they search for food (Dai et al., 2019; Chu et al., 2021)(pseudo- the robot. Meanwhile, the MMAS was used to limit the pheromone
code as shown in Table 15). concentration, preventing from falling into the local optimal path. Chen
ACO is featured by positive feedback, parallel computation, and & Zhou (2021) proposed an improved ACO to search and select the path
good robustness performance, but suffers from such disadvantages as autonomously according to the number of pheromones, which will be
slow convergence, easy to fall into local optimum, and premature improved by adjusting the heuristic factor function. An efficient gain-
convergence (Zhang et al., 2020b). In order to address the problems of based dynamic green ant colony optimization was proposed by San­
basic ACO, many scholars have proposed many effective improvement geetha et al. (2021a). When path planning is performed in dynamic
measures in terms of algorithmic framework and structure, such as ACS scenarios, the energy consumption will be significant due to its char­
(ant colony system) (Dorigo & Gambardella, 1997), MMAS (max–min acteristics. The proposed algorithm reduced the total energy consumed
ant system) (Stützle & Hoos, 2000), ASelite (ant system with elitist in the path planning process through an efficient pheromone enhance­
strategy), ASrank (ant system with elitist and ranking) (Bullnheimer & ment mechanism based on gain function.
Hartl, 1997), and other classical improvement algorithms. In spite of
effective improvement of optimization capability, these improved al­ 3.2.3. PSO (particle swarm optimization)
gorithms adopt a fixed pattern to update the pheromone and probability PSO was firstly proposed by Eberhart & Kennedy (Jain et al., 2022),

15
L. Liu et al. Expert Systems With Applications 227 (2023) 120254

Table 16 minimum and improve the convergence speed of the algorithm. Then,
The pseudo-code of the PSO’s algorithm for path planning. use fitness variance to measure the diversity of the particles. The in­
1. for each particle i crease in diversity helps to overcome the prematureness. The diversity of
2. Initialize velocity Vi and position Xi for particle i the particles is increased by the defined extended Gaussian distribution.
3. Evaluate particle i and set pBesti = Xi Finally, apply the smoothing principle to path planning. Phung & Ha
4. end for
(2021) proposed a new spherical vector-based particle swarm
5. gBest = min{pBesti}
6. while not stop optimization.
7. for i = 1 to N (2) Combining PSO and other optimization algorithms can form a
8. Update the velocity and position of particle i hybrid PSO algorithm. Qiuyun et al. (2021) proposed an improved
9. Evaluate particle i particle swarm optimization algorithm, which can better solve the
10. if fit(Xi)＜fit(pBesti)
problem of path planning. We propose a new path planning method, and
11. pBesti = Xi;
12. if fit(pBesti)＜fit(gBest) design a crossover operation in GA based on the encoding method of this
13. gBest = pBesti; algorithm to update the locations of particles. In addition, a variational
14. end for mechanism is used in order to prevent the algorithm from local opti­
15. end while
mum. Wu et al. (2021b) proposed an effective compensation method
Here Vi and Xi stand the velocity and position of particle i, N stands the based on reinforcement learning and particle swarm optimization,
number of ants. which provided a strategy for real-time rescue and assignment of multi-
AUV systems in a 3D underwater environment.

Table 17 3.2.4. BFO (bacterial foraging optimization)

The pseudo-code of the BFO’s algorithm for path planning.
In 2002, Passino proposed BFO based on the simulation of foraging
1. Initialize the parameters and reproduction behavior of bacterial population (Yan et al., 2018;
2. Initialize a random population {xi} for i ∈ [1,N]
Passino, 2002). Beginning from initializing a set of random solutions,
3. For l = 1 to Ne
4. For k = 1 to Nr
regarding the locations of bacteria as potential solutions to the problem,
5. For j = 1 to Nc the algorithm change the locations of bacteria by simulating the chemo-
6. For each individual xi,i ∈ [1,N] taxis, reproduction, and elimination-dispersal behaviors of the bacterial
7. Generate a random n-dimensional unit vector Δ population to optimize them(pseudo-code as shown in Table 17).
8. For m = 1 to Ns(cost reduction steps)
Due to its simplicity and efficiency, BFO has been widely used in the
9. x i←xi + cΔ
̂
10. path planning of mobile robots. Muni et al. (2021) used the foraging
If f (̂
x i) < f (xi) then
′ ′

11. xi←̂ xi quality of the Escherichia coli to obtain the shortest path between two
12. else locations in the least amount of time. The Gaussian cost function
13. m←Ns(exit the cost reduction loop) assigned to the bacterial attractor and repellant profiles plays a major
14. End if
role in obtaining the best path between two locations. Long et al. (2020)
15. Next m
16. Next individual
proposed a grid partition-based BFO algorithm. And the chemotactic
17. Next j operation that using A* optimize this algorithm allow this algorithm to
18. For each individual xi,i ∈ [1,N] be applied in unmanned surface mobile robots. Abdi et al. (2020)
19. Fi ← average value of f (xi) during Nc steps of chemotaxis loop
′
generated virtual obstacles at local minima to make the robot shrink and
20. Next individual
regenerate a safe path. This method can be easily extended to coordi­
21. Eliminate the worst N/2 individuals based on {Fi}
22. Clone the best N/2 individuals based on {Fi}
nated movements between multiple robots. The information associated
23. Next k with the virtual obstacle is shared with the whole population, which
24. For each individual xi,i ∈ [1,N] allowed them to escape the same local minima, thus saving time and
25. Random number r ← U[0,1] effort. Long et al. (2021) proposed a new multi-subpopulation bacterial
26. If r＜pe then
foraging optimisation algorithm for path planning of sleep robots to
27. xi ← random point in the search space
28. End if improve the search performance of mobile robots in complex environ­
29. Next individual ments. From the above-mentioned optimized algorithm improvements
30. Next l and applications of bacterial community, there have been significant
Here N stands the number of iterations, Ne stands the number of dispels, Nr advances in the research of bacterial optimization since the past decades
stands the number of copies, Nc stands the number of chemotaxis. or so. Parameter improvement, multi-algorithm hybridization, and
operator improvement will remain important directions for algorithm
and originated from the study of foraging behavior of birds(pseudo-code improvement research. However, there are still some deficiencies in the
as shown in Table 16). existing research, which deserve to be further studied. Wang et al.
Particle swarm algorithm is a representative algorithm that combine (2022d) proposed an IBFOA (improved bacterial foraging optimization
particle swarm algorithm and other algorithms or factors. The final algorithm). In addition, in order to overcome the problems in optimi­
optimization goals of hybrid particle swarm algorithm are: (a) to zation algorithms and traditional interpolation algorithms, such as large
improve the convergence and diversity of the population, and prevent amount of calculation, unsmooth trajectory, etc., adaptive factors and
premature convergence; (b) to obtain the optimal solution quickly and elite retention strategies were used to implement IBFOA. Wang et al.
reduce the time cost; (c) to improve the local search performance and (2021d) designed an adaptive decreasing fractal dimension chemical
accuracy of PSO. At present, there are usually two improvement di­ step instead of fixed step to achieve adaptive step length adjustment,
rections to achieve the above objectives (Yang & Cai, 2021). proposed an optimal swimming search method to solve the problems of
(1) Adjust the parameters of the PSO to balance the global detection traditional BFOA, such as invalid search and repeated search. In addi­
and local exploitation capabilities. For example, Chai et al. (2021) tion, an adaptive migration probability was developed to replace the
applied a locally improved multi-objective particle swarm optimization fixed migration probability, which can prevent elite individuals from
method to deal with the problem about constrained ballistic path being lost in BOF.
planning. Zhang et al. (2020a) improved the inertia weight, acceleration
factor, and localization to prevent the algorithm from falling into local 3.2.5. ABC (artificial bee colony algorithm)
ABC is an algorithm inspired by bee colony behavior, proposed by

16
L. Liu et al. Expert Systems With Applications 227 (2023) 120254

Table 18
The pseudo-code of the ABC’s algorithm for path planning.
Input:N,FN,lb,ub,limit
Output:
bestSolutionFound
1. INITIALIZATION PHASE:
2. for m = 1 to FN do
3. Random selection of food source xm
4. end for
5. for iteration = 1 to N do
6. EMPLOYED BEES PHASE:
7. for m = 1 to FN do
8. Select rm randomly
9. Produce new solution vm
10. if fitness(vm)＞fitness(xm) then
11. Set xm to vm
12. Set trialsm to 0
13. else
14. Increment trialsm
15. end if
16. end for
17. ONLOOKER BEES PHASE:
18. for m = 1 to FN do
19. Calculate selection probability pm
20. end for
21. for m = 1 to FN do
22. Select rm
23. Produce new solution vm
24. if fitness(vm)＞fitness(xm) then Fig. 13. Levy flight path.
25. Set xm to vm
26. Set trialsm to 0
nectar sources, the onlooker bees wait in the hive and select nectar
27. else
28. Increment trialsm sources, and the scout randomly search for other nectar sources. The
29. end if location of the nectar source corresponds to the solution of the problem,
30. end for and the amount of nectar is related to the fitness of the solution.(pseudo-
31. SCOUT PHASE:
code as shown in Table 18).
32. if iteration modulus SPP = 0 then
33. Select food source with the highest trials value
Although ABC is not mature yet and its research is still in the
34. if trialsm ≥ limit then exploration stage, it has been widely concerned by experts and scholars
35. Set xm to random solution in range (lb ÷ ub) due to its powerful global search capability. In recent years, a large
36. end if number of researchers have improved the performance of ABC algo­
37. end if
rithm in three aspects: improvement of the position update equation, the
38. SAVE RESULT:
39. Update bestSolutionFound with position of food source with dimensional update strategy, and the overall search strategy (Li, 2020).
the best fitness From the perspective of improving the location update equation, Shah
40. end for et al. (2018) introduced the gbest mechanism applied in the particle
Here N stands the number of iterations, FN stands the number of honey sources, swarm algorithm into the traditional ABC iterative formulation, accel­
lb,ub stand the lower and upper limit of search space and limit stands the same erating the convergence speed. Muntasha et al. (2021) proposed that an
honey source is limited to mining times. anti-collision and path planning system for swarm UAVs can be designed
by ABC. The proposed system uses ABC algorithm to optimize the speed
of UAVs that can reach the destination efficiently in the shortest path
Table 19 without collisions with others. Chengli et al. (2018) added variable
The pseudo-code of the CSA’s algorithm for path planning. neighborhood search factors to search equation to increase the popu­
1. begin lation diversity and proposed a new empirical strategy to further
2. Objective function f(x), x=(x1,…,xd)T improve the global search efficiency. From the perspective of improving
3. Generate initial population of
the dimensional update strategy, Chu et al. (2016) changed ABC’s
4. n host nests xi(i = 1,2,…,n)
5. while (t＜MaxGeneration) or (stop criterion) single-dimensional updating strategy to full dimensional, and generated
6. Get a cuckoo randomly by Levy flights two solutions each time, and chose the better one to improve the search
7. evaluate its quality/fitness Fi efficiency. From the perspective of improving the overall search strat­
8. Choose a nest among n (say, j) randomly egy, Szczepanski & Tarczewski (2021) proposed a hybrid approach for
9. if (Fi＞Fj),
10. replace j by the new solution;
path planning of variable-space mobile robots. It consists of an offline
11. end global path optimization algorithm, in which ABC was applied, and an
12. A fraction (pa) of worse nests are abandoned and new ones are built; online path planning scheme. In the path planning layer, Zhao et al.
13. Keep the best solutions (or nests with quality solutions); (2021) developed MABCM (multiple-sub populations ABC with mem­
14. Rank the solutions and find the current best
ory) algorithm and proposed a new exit evaluation strategy. Hao et al.
15. end while
16. Postprocess results and visualization (2021) proposed the ABC based on balanced search factor for UAV path
17. end planning. It used the search strategy based on balanced search factor to
perform a deep search while maintaining population diversity. Liu et al.
Here f(x) stands the objective function, n stands the number of host nests,
MaxGeneration stands the number of iterations.
(2021) calculated the optimal path planning between adjacent mission
points through improved ABC.
Karaboga’s group for optimizing algebra problems in 2005 (Karaboga &
3.2.6. CSA (cuckoo search algorithm)
Basturk, 2007). The ABC algorithm divides bees into three categories:
Yang & Deb (2009) proposed CSA which is an emerging heuristic
employed bees, onlooker bees, and scout. The employed bees search for

17
L. Liu et al. Expert Systems With Applications 227 (2023) 120254

Table 20 Table 22
The pseudo-code of the FA’s algorithm for path planning. The pseudo-code of the SSA’s algorithm for path planning.
1. Objective function f(x), x=(x1,…，xd)T Input: G,PD,SD,R2,n
2. Generate initial population of firefies xi(i = 1,2,…,n) Initialize a population of n sparrows and
3. Light intensity Ii at xi is determined by f(xi) define its relevant parameters.
4. Input γ 1. Output:Xbest,fg.
5. while (t＜MaxGeneration) 2. while(t＜G)
6. for i = 1:n all n fireflies 3. Rank the fitness values and find the
7. for j = 1:i all n fireflies current worst individual.
8. if (Ij＞Ii), move firefly i towards j in d-dimension; 4. R2 = rand(1)
9. end if 5. for i = 1:PD
10. Attractiveness varies with distance r via exp[-γr] 6. Using equation update the sparrow’s
11. Evaluate new solutions and update light intensity location;
12. end for j 7. end for
13. end for i 8. for i=(PD + 1):n
14. Rank the fireflies and find the current best 9. Using equation update the sparrow’s
15. end while location;
10. end for
11. for i = 1:SD
12. Using equation update the sparrow’s
Table 21 location;
The pseudo-code of the GWO’s algorithm for path planning. 13. end for
14. Get the current new location;
1. Begin
15. If the new location is better than
2. Initialize the grey wolf population Xi(i = 1,2,…,n)
before,update it;
3. Initialize a,A,and C
16. t=t+1
4. For all Xi do
end while
5. Calculate fitness F(Xi)
17. return Xbest,fg.
6. End for
7. Get the first three best wolves as Xα,Xβ,and Xδ Here G stands the maximum iterations, PD stands the number of producers, SD
8. While(t＜Max number of iterations) stands the number of sparrows who perceive the danger, R2 stands the alarm
9. For each search agent value and n stands the number of sparrows.
10. Update the position of the current search agent
11. End for
12. Update a,A,and C avoiding the defect of premature in traditional algorithm. Song et al.
13. For all Xi do (2020) proposed an improved CSA based on compact parallel technol­
14. Calculate fitness F(Xi) ogy for three-dimensional path planning. Fan et al. (2021) established
15. End for
16. Update Xα,Xβ,and Xδ
the collision avoidance model of unmanned craft and obstacles based on
17. t=t+1 the principle of speed obstacle method, and proposed a control method
18. End while for dynamic collision on the basis of improved CSA.
19. Return Xα
20. End
3.2.7. FA (firefly algorithm)
Here n stands the number of population, a decreases linearly from 2 to 0, A and In 2009, professor Yang (2010) of the University of Cambridge
C stand coefficient vector. proposed FA by simulating the natural phenomenon of firefly swarming
at night (pseudo-code as shown in Table 20).
algorithm in 2009. It mainly originated from the brood parasitism Here f(x) stands the objective function, γ stands the light absorption
reproduction of cuckoo and Levy flight search mechanism (pseudo-code coefficient, MaxGeneration stands the number of iterations.
as shown in Table 19 and Fig. 13) (Wu, 2021). Although FA has the same advantages as intelligent algorithms,
Levy flight is a kind of wandering foraging method in the animal currently it has an imperfect theoretical foundation, weak interpret­
realm. The wandering step length satisfies a stable distribution of heavy ability, and suffers from common defects of other swarm intelligence
tails. And during the wandering process, short-distance exploration is stochastic optimization algorithms. For this reason, many improvements
interspersed with occasional long-distance walking, which is a most have been made to FA in the academic world. These improvement
ideal way of foraging. Using Levy search in intelligent algorithms can methods can be summarized into the following categories (Song, 2021):
prevent from easily falling into local optimal solutions (Liu, 2019). the first is elitism that prevents fireflies form moving to worse positions
Therefore, CSA has many advantages, such as avoiding local optimum, with poor fluorescein (Wang et al., 2017a; Wang et al., 2016); the second
few and simple parameters, and fast convergence, etc. Once proposed, it is, random moving mode that uses the random moving part of the al­
has gained the common attention of domestic and foreign scientific gorithm to avoid local optima and enhance the global search of the al­
scholars. And through continuous improvement and optimization, it has gorithm (Dhal et al., 2016); the third is adaptive parameter method that
become an active bionic intelligent optimization algorithm nowadays. change the parameters in the standard FA so that they can adapt to the
Wu (2021) proposed a fusion obstacle avoidance algorithm of cuckoo algorithmic search process (Yu et al., 2016; Wang et al., 2017b; Chou &
and pigeon flock. Firstly, we use the Levy flight mechanism in CSA to Ngo, 2017); and the fourth is mixing other algorithms for improvement.
perform long-short alternate search to obtain the latest point of nest As FA has good global search capability but weak local search capability,
location, and then obtain the suboptimal obstacle avoidance path; sec­ combining with other intelligent algorithms can make up for its short­
ondly, we introduce the map-compass operator and the ground surface comings (Pei et al., 2019). Improvements in the field of mobile robot
operator of the pigeon flock optimization algorithm on the basis of the path planning are as follows: Patle et al. (2018) applied FA in mobile
suboptimal path to perform the secondary obstacle avoidance path robot path planning under uncertain situations. Sadhu et al. (2018)
planning demonstration, update the latest point of nest position again, proposed a strategy of Q-learning FA was for light absorption parame­
and then obtain the optimal obstacle avoidance path. Zhang et al. ters. Each firefly learned through a Q-learning strategy and then applied
(2021a) proposed an adaptive cuckoo search algorithm, which has high what they learned into the execution phase. An optimal FA-based path
efficiency and good stability under strict dynamic constraints. Sharma planning method with adaptive population size was proposed by Li et al.
et al. (2021) changed the random function in the algorithm based on the (2020). Firstly, an evaluation method for collision degree was estab­
traditional cuckoo algorithm, obtaining more alternative solutions and lished at the cost of collision avoidance. Based on the collision degree of

18
L. Liu et al. Expert Systems With Applications 227 (2023) 120254

Table 23
Comparison of bionics-based path planning algorithms.
Method Parallel Global Convergence Whether the search direction Escape ability of local Robustness Scope of application
search optimization time can be adjusted optimal solution
capability capability

GA Good Good Long Yes Weak Good Global path planning

2. TSP
ACO Good Good Long Yes Weak Good Global path planning
2. TSP
PSO Good Good Short Yes Weak Medium Global path planning
Local path planning
BFO Good Good Long Yes Good Medium Global path planning
Multi-robot combinatorial
optimization
ABC Good Good Medium Yes Good Good Global path planning
Multi-robot combinatorial
optimization
Path continuous
optimization
CSA Good Good Medium Yes Weak Medium Global path planning
Path continuous
optimization
FA Weak Good Long Yes Good Medium Local path planning
TSP
GWO Weak Good Long Yes Weak Medium Global path planning
Multi-robot combinatorial
optimization
SSA Good Weak Short Yes Weak Good Global path planning
Multi-robot combinatorial
optimization

prey(Lu, 2017)(pseudo-code as shown in Table 21).

Table 24 From the current research on GWO, GWO has been continuously
Comparison of bionics-based path planning algorithms. improved in theoretical research and has a broad application prospect in
BNN FL the field of mobile robot path planning. Although the GWO algorithm
Identical points It has strong fault tolerance and can process inaccurate has many application advantages, it still has some disadvantages (Zhang
information in an imprecise way. & Wang, 2019; Tang, 2019): (1) reducing population diversity due to
When dealing with and solving problems, the precise randomly generated initialized populations; (2) slow convergence at a
mathematical model of the control object is not required.
later stage caused by the search mechanism of the GWO algorithm; and
Differences It can directly handle It can not deal with
structured knowledge. structured knowledge
(3) tend to fall into local optimum. Zhang et al. (2016) proposed a new
It has poor ability to adapt to directly, and needs to learn meta-heuristic GWO for two-dimensional path planning of the un­
environmental changes and samples. manned combat air vehicle that can find a safe path by connecting
acquire knowledge. It has learning ability and selected nodes in a two-dimensional coordinate system while keeping
easy knowledge
away from threat areas and saving fuel. Albina & Lee (2019) proposed a
acquisition.
Advantages of Knowledge can be acquired through learning. multi-robot exploration stochastic optimization method that simulates
combination of Strong self-organization and adaptability. the collaborative predatory behavior of grey wolves. The motion of the
the two Strong robustness. robots is computed by a combination of deterministic and meta-heuristic
techniques, using a new method of collaborative multi-robot search and
GWO as a mixed stochastic search. Jamshidi et al. (2021) proposed an
the population, two nonlinear functions were proposed to determine the
improved GWO for UAV 3D path planning considering UAV dynamics,
population size. Then, individuals were added or deleted from the firefly
in which a variable weight called “alignment coefficient” was defined to
population. Individuals can add randomly. The feasible and infeasible
handle the scattering of waypoint. The application of parallel GWO re­
solutions were distinguished in the firefly population, and the fireflies in
duces the computation time and makes real-time implementation more
the infeasible solutions were first deleted when performing the elimi­
possible. Zhang et al. (2021e) proposed an adaptive convergence factor
nation operation. Finally, based on the existing methods for handling
adjustment strategy and adaptive weighting factors to update individual
infeasible paths, a coefficient adaptively adjusted according to the
positions by convergence analysis and test functions. Jarray et al. (2022)
population size was introduced, which can control the degree of infea­
proposed a new parallel cooperative coevolutionary grey wolf optimizer
sible paths approximating the feasible region. A new hybrid algorithm
to improve the accuracy of UAV routes. Liu et al. (2021) endowed α
based on GA and FA was proposed by Zhang et al. (2022a). The core idea
wolf, β wolf, and δ wolf active search ability and added dynamic weight
of this new algorithm is that when FA falls into a local optimal solution,
in the grey wolf position update to prevent wolves from losing diversity
the local optimal fireflies are regarded as a population. Selection,
and falling into local optimum. Liu & Li (2022) proposed a mixed path
crossover and mutation of population perform in GA, and the optimal
planning method combining modified gray wolf optimization with a
individual fireflies can be obtained by genetic operations.
situation assessment mechanism. Kiani et al. (2021) proposed two UAV
3D path planning methods based on incremental GWO and Extended-
3.2.8. GWO (grey wolf optimizer)
GWO to determine the UAVs optimal path with minimum cost and
The GWO proposed by Australian scholars Mirjalili et al. (2014) is an
minimum execution time. Fei et al. (2021) used a chaotic random
intelligent optimization search method. Grey wolves are the canine
method to generate the initial population, introduced a velocity sorting
carnivore animals with high environmental adaptability, which are
system into the traditional particle swarm algorithm to propose mixed
good at long-distance running at high speed. GWO encompasses their
algorithm of GWO-PSO. Panwar & Deep (2021) proposed the discrete
stratification of social hierarchy, the process of besieging and attacking

19
L. Liu et al. Expert Systems With Applications 227 (2023) 120254

Table 25
Analysis of various path planning techniques.
Authors Path planning Dimension Basic Algorithm Path Test of Environmental modeling
of Algorithm smooth method

Global Local Optimization Fusion Simulation Actual

(Abdi et al., 2020) Y N 2 BFO Y N N Y N GM

(Albina & Lee, 2019) N Y 2 GWO Y N N Y N GM
(AlKhlidi et al., 2021) Y N 2 FL/PSO Y Y N Y N –
(Alyasin et al., 2019) Y N 2 Dijkstra N N N N Y –
(Balado et al., 2019) Y Y 3 Dijkstra Y N N Y Y –
(Bao et al., 2022) Y Y 2 APF/A* Y Y N Y Y GM
(Brunner et al., 2013) Y N 3 RRT Y N N Y N GM
(Chai et al., 2021) Y Y 2 PSO Y N N Y N –
(Challita et al., 2018) N Y 2 CD Y N N Y N GM
(Chen et al., 2021a) Y N 3 PRM Y N N Y Y GCM
(Chen et al., 2021b) Y Y 2 ACO/APF Y Y N Y Y GM
Chen & Zhou (2021) Y N 2 ACO Y N N N N –
(Chen et al., 2021c) Y N 2 GA Y N Y Y N GM
(Chen & Zou, 2021) N Y 2 BNN Y N N Y N GM
(Chengli et al., 2018) Y Y – ABC Y N N Y N –
(Cheng et al., 2021) Y N 2 GWO Y Y N Y N GM
/PSO
(Dai et al., 2019) Y N 2 ACO/A* Y Y N Y N GM
(Daniel et al., 2010) Y N 2 A* Y N Y Y N GM
(Ding et al., 2016) Y Y 2 APF Y N Y Y N GM
(Du et al., 2021) N Y 2 APF Y N N Y N –
(Esposito & Wright, 2019) Y N 2 PRM Y N N Y N GCM
(Fang et al., 2021) N Y 2 CSA Y N N Y N –
(Fei et al., 2021) Y N 2 GWO Y Y N Y N –
/PSO
(Feng et al., 2021) N Y 2 APF/FL Y Y N Y N –
(Ferguson & Stentz, 2006) Y Y 2 A* Y N Y Y Y GM
(Fink et al., 2019) Y N 3 Dijkstra Y N Y Y N –
(Fox et al., 2002) Y Y 3 DWA N N N Y N GCM
(Gao et al., 2021) Y N 2 SSA/ACO Y Y N Y N –
(Germi et al., 2018) Y N 3 GA/APF Y Y N Y N –
(Gharajeh & Jond, 2022) Y N 2 FL/NN Y Y Y Y N GM
(Godio et al., 2021) N Y 2 BNN Y N N Y N GM
(Gonzalez et al., 2017) Y N 2 CD Y N N Y N GM
(Guo et al., 2019) N Y 2 Dijkstra Y N N Y N –
(Guo et al., 2021) Y Y 2 FL/NN Y Y N Y N GM
(Guruprasad & Ranjitha, Y N 2 CD Y N N Y Y GM
2021)
(Hao et al., 2021) Y Y 2 ABC Y N N Y N –
(He et al., 2022) Y Y 2 A* Y N N Y N GM
(Huang, 2018a) Y Y 2 DWA Y N N Y N GM
(Huang, 2018b) Y N 2 Dijkstra Y N N N N –
(Islam et al., 2012) Y N 2 RRT Y N N Y N GM
(Jafarzadeh et al., 2018) Y N 2 – – – Y Y N GCM
(Jamshidi et al., 2021) N Y 3 GWO Y N N Y N –
(Jarray et al., 2022) Y N 3 GWO Y N N Y N –
(Ji & Liu, 2022) Y N 2 ACO Y N N Y N –
(Jiao et al., 2018) Y Y 2 ACO Y N N Y N GM
(Jung et al., 2019) Y N 2 CD Y N N Y N GCM
(Karaman & Frazzoli, 2011) Y N 2 RRT Y N N Y N GM
(Khan et al., 2021) N Y 3 BNN Y N N Y N GM
(Kiani et al., 2021) Y N 3 GWO Y N N Y N GM
(Lamini et al., 2018) Y N 2 GA Y N N Y N GM
(Li et al., 2020) Y N 2 FA Y N N Y N GM
(Li, 2015) N Y 2 FL Y N N Y Y –
(Li, 2021) Y N 2 SSA Y N N Y N GM
(Li et al., 2022a) Y N 2 RRT Y N Y Y N GM
(Li et al., 2021) Y N 2 ACO Y N Y Y N GM
(Li et al., 2022b) Y Y 2 A*/DWA Y Y N Y Y GM
(Liang et al., 2018) Y Y 2/3 – – – N Y Y GCM
(Liu et al., 2021a) Y N 3 SSA Y N N Y N –
(Liu et al., 2021b) Y N 2 GA/ABC Y Y Y Y N –
(Liu et al., 2022) Y N 2 APF/FL N Y N Y N GM
(Liu et al., 2021) Y Y 2 GWO Y N N Y N GM
(Liu, 2018) Y Y 2 APF Y N N Y Y GM
(Liu et al., 2021c) N Y 3 SSA/BNN Y Y Y Y N GM
(Liu & Li, 2022) Y Y 2 GWO Y N N Y N GM
(Liu et al., 2021d) Y Y 2 APF Y N N Y N –
(Long et al., 2020) Y N 2 BFO/A* Y Y N Y N GM
(Long et al., 2021) Y N 2 BFO Y N N Y N GM
(Luo et al., 2016) N Y 2 BNN Y N N Y Y GM
(Luo et al., 2019b) Y N 2 APF Y N Y Y N –
(continued on next page)

20
L. Liu et al. Expert Systems With Applications 227 (2023) 120254

Table 25 (continued )
Authors Path planning Dimension Basic Algorithm Path Test of Environmental modeling
of Algorithm smooth method

Global Local Optimization Fusion Simulation Actual

(Mahmud et al., 2019) Y N 2 GA Y N N Y N –

(Martins et al., 2022) Y N 2 A* Y N Y Y N –
(Miao et al., 2021) Y N 2 ACO Y N N Y N GM
(Muni et al., 2021) Y N 2 BFO Y N N Y Y –
(Muntasha et al., 2021) Y Y 2 ABC Y N N Y N –
(Nazarahari et al., 2019) Y N 2 APF/GA Y Y Y Y N –
(Ni & Yang, 2011) Y Y 2 BNN Y N N Y N –
(Ni et al., 2017) N Y 3 BNN Y N N Y N GM
(Nian, 2021) N Y 2 A* Y N Y N Y GM
(Ntakolia & Lyridis, 2021) N Y 2 FL/RS Y Y N Y N GM
(Oršulić et al., 2019) Y N 2 CD Y N Y N Y GM
(Palmieri et al., 2016) Y N 2 RRT Y N Y Y Y GM
(Pang et al., 2021) Y N 2 APF Y N N Y N TM
(Panwar & Deep, 2021) Y N 2 GWO Y N N Y N –
(Park et al., 2015) Y Y 2 CD Y N Y Y N GM
(Patle et al., 2018) Y Y 2 FA Y N N Y Y –
(Phung & Ha, 2021) Y N 3 PSO Y N N Y Y –
(Qi et al., 2020) N Y 2 RRT Y N Y Y Y GM
(Qian et al., 2020) Y N 2 RRT Y N N Y N GM
(Qiuyun et al., 2021) Y N 2 PSO/GA Y Y N Y N –
(Rahmaniar & (Rakhmania, N Y 2 GA Y N Y Y N –
2022)
(Ravankar et al., 2020) Y Y 2 PRM/APF Y Y N Y Y GCM
(Ryu & Park, 2019) Y N 2 RRT Y N N Y N GM
(Sadhu et al., 2018) Y N 3 FA Y N N Y Y GM
(Samaniego et al., 2019) N Y 3 CD Y N N Y N GM
(Sang et al., 2021) N Y 2 APF Y N N Y N –
(Sangeetha et al., 2021a) N Y 3 ACO Y N N Y N GM
(Sangeetha et al., 2021b) N Y 2 FL/ACO Y Y Y Y N –
(Sepehri & Moghaddam, Y N 3 APF/RRT Y Y N Y Y –
2021)
(Shao et al., 2021) Y N 2 ACO Y N Y Y N GM
(Sharma et al., 2021) Y N 2 CSA Y N N Y N –
(Shin et al., 2021) N Y 3 APF/A* Y Y Y Y N GM
(Shivgan & Dong, 2020) Y N 2 GA Y N N Y N –
(Song, 2021) Y Y 2 FA Y N N Y N –
(Song et al., 2020) Y N 3 CSA Y N N Y N –
(Souza et al., 2019) Y N 2 Dijkstra Y N N Y N GM
(Stentz, 1997) Y Y 2 A* Y N N Y N GM
(Stenzel et al., 2021) Y N 2 RS Y N N Y N GM
(Sun, 2017a) Y Y 2 RRT Y N N Y N –
(Sun, 2017b) N Y 2 BNN Y N Y Y N GM
(Szczepanski & Tarczewski, Y Y 2 ABC Y Y N Y N GM
2021) /Dijkstra
(Szczepanski et al., 2022) Y Y 2 APF Y N Y N Y –
(Tang, 2019) Y Y 2 PSO Y Y Y Y N GM
/GWO
(Tao et al., 2021) Y N 2 FL/ACO Y Y N Y N GM
(Tao et al., 2022) Y N 2 GA Y N Y Y N GM
(Wang et al., 2021a) N Y 2 APF/NN N Y N Y Y GM
/FL
(Wang et al., 2022a) Y N 2 RRT Y N Y Y N GM
(Wang et al., 2021b) Y Y 2 FL/APF Y Y Y Y N –
(Wang et al., 2021c) Y N 2 ACO Y N Y Y N GM
(Wang et al., 2021d) Y Y 3 BFO Y N N Y N –
(Wang et al., 2022b) Y N 2 A* Y N N Y N GM
(Wang et al., 2022c) Y N 2 A* Y N Y Y N GM
(Wang et al., 2022d) Y N 3 BFO Y N Y Y N –
(Wei et al., 2019) Y N 2 Dijkstra Y N N N N –
(Wu, 2021) Y N 2 CSA Y N N Y N –
(Wu et al., 2022) Y Y 2 A*/DWA Y Y Y Y N GM
(Wu et al., 2021a) Y N 2 SSA/GA Y Y N Y N –
(Wu et al., 2021b) Y Y 3 PSO Y N N Y N GM
(Xie, 2016) Y N 2 FL/APF Y Y N Y Y –
(Xue & Shen, 2020) Y N 2 SSA N N N Y N –
(Xu et al., 2021) Y N 2 GA Y N N N N –
(Xue et al., 2021) Y N 2 ACO Y N Y Y N GM
(Xu et al., 2020) Y N 2 RRT Y N Y Y N GM
(Yang et al., 2018) N Y 2 BNN Y N N Y N GM
(Yang et al., 2021) N Y 3 RS Y N N Y Y TM
(Yang et al., 2022) Y Y 2 ACO Y N Y Y N GM
(Yan, 2021) Y N 2 ACO Y N N Y N GM
(Yan & Zhao, 2021) N Y 3 BNN Y N N Y N GM
(continued on next page)

21
L. Liu et al. Expert Systems With Applications 227 (2023) 120254

Table 25 (continued )
Authors Path planning Dimension Basic Algorithm Path Test of Environmental modeling
of Algorithm smooth method

Global Local Optimization Fusion Simulation Actual

(Yi et al., 2016) N Y 3 BNN Y N N Y N –

(Yu & Ge, 2017) Y N 2 Dijkstra Y N N Y N –
(Yu et al., 2018) Y Y 2 Dijkstra N N Y Y Y –
(Yu et al., 2021) Y N 3 SSA/PSO Y Y Y Y N –
(Zagradjanin et al., 2021) N Y 2 FL/A* Y Y N Y N GM
(Zhang et al., 2015) Y Y 2 DWA/FL Y Y Y Y Y GCM
(Zhang et al., 2020a) Y N 2 PSO Y N Y Y Y GM
(Zhang et al., 2021a) Y Y 3 CSA Y N Y Y Y –
(Zhang, 2021b) N Y 2 APF Y N N Y N –
(Zhang et al., 2016) Y N 2 GWO Y N N Y N –
(Zhang et al., 2020b) Y N 2 ACO Y N Y Y N GM
(Zhang et al., 2021c) Y N 2 ACO Y N N Y N GM
(Zhang et al., 2021d) Y N 2 APF Y N N Y N GM
(Zhang et al., 2022a) Y N 3 FA/GA N Y N Y N GM
(Zhang et al., 2021e) Y N 3 GWO Y N N Y N –
(Zhang et al., 2018a) Y Y 2 A* Y N Y N Y GM
(Zhang, 2020) N Y 2 FL/APF Y Y N Y Y GM
(Zhang, 2021) Y Y 3 RS Y N N Y Y MR
(Zhang et al., 2018b) Y Y 2 Dijkstra Y N N Y Y GM
(Zhang et al., 2022b) Y N 2 SSA Y N Y Y N GM
(Zhenyang & Wei, 2022) Y Y 2 A*/DWA Y Y Y Y Y GM
(Zheyi & Bing, 2021) Y N 2 APF/FL Y Y N Y N –
(Zhou et al., 2018) Y Y 3 CD Y N N Y N GM
(Zhu & Yang, 2021) Y N 3 BNN Y N N Y N GM
(Zhu et al., 2021) Y N 2 ACO Y Y N Y N GM
/GWO
(Ziegler et al., 2008) Y Y 2 A* Y N Y Y Y TM
Zou et al. (2019) Y N 2 PRM Y N N Y N GCM
(Zou, 2005) Y Y 3 A*/GA Y Y N Y N MR

Here Y stands for “Yes”, N stands for “No” and “-” stands unknown.

3.2.9. SSA (sparrow search algorithm)

SSA is a swarm intelligence optimization algorithm proposed by Xue
& Shen (2020) in 2020 based on sparrows’ behavior of foraging and
escaping predators (pseudo-code as shown in Table 22).
SSA is a new algorithm with the qualities of better search capability
and fast convergence. In the last two years, the optimization for SSA has
been mainly focused on:
(1) Smoothed paths. Zhang et al. (2022b) performed path smoothing
for turning points on the basis of the traditional algorithm, a new
neighborhood search strategy was used to improve the fitness value of
the global optimal individual and a new position update function was
used to speed up the convergence rate. Liu et al. (2021c) used SSA to find
a series of nodes, and then used b-sample curves to fit these nodes, thus
achieving path smoothing. Yu et al. (2021) proposed a sparrow particle
swarm algorithm for UAV path planning. Three spline interpolation was
used to the path nodes, which improved the smoothness of the path and
the smoothness of the UAV flight trajectory, making it more applicable
to the actual UAV flight trajectory.
(2) Optimization search methods. In order to improve the search
capability of SSA in the field of UAV 3D path planning, Liu et al. (2021a)
transformed the route planning problem into a multidimensional func­
tion optimization problem, introduced chaos strategy to enhance the
diversity of the algorithm population, and used adaptive inertia weights
Fig. 14. Data statistics of environment modeling methods for mobile robot. to balance the convergence speed and search capability of the algorithm.
Finally, a Cauchy-Gaussian mutation strategy was used to improve the
ability of the algorithm to solve the stagnation problem. Wu et al.
GWO for solving complex discrete traveler problems. And the 2-optimi­
(2021a) proposed a greedy genetic SSA based on a sine and cosine search
zation algorithm was incorporated in it to improve the performance of
strategy and applied to the traveler problem. Li (2021) introduced a K-
the algorithm. Cheng et al. (2021) combined the traditional particle
medoids clustering method to dynamically classify populations based on
swarm optimization algorithm with GWO, chaos, and adaptive inertia
SSA, speeding up the information exchange between populations. In
weights to solve the problems of premature convergence and poor global
addition, an adaptive weighting factor was introduced to make the al­
search capability of the particle swarm optimization algorithm,
gorithm search more detailed and extensive, which improved its opti­
improving the convergence speed of the algorithm and that of path
mization capabilities. Gao et al. (2021) proposed the ant-sparrow
search.
algorithm based on the same number of constraint sets of ants and
sparrows, which combined the qualities of a better initial solution of the

22
L. Liu et al. Expert Systems With Applications 227 (2023) 120254

Fig. 15. Data statistics of classical path planning algorithms based on literatures in this paper.

Fig. 16. Data statistics of bionic path planning algorithms based on literatures
in this paper.

Fig. 18. Data statistics of the application environment of path planning algo­
rithms based on literatures in this paper.

on the path search, and significantly reduce the blind search.

Table 23 clearly shows the difference between the above-mentioned
bionics-based path planning algorithms (Tan et al., 2021).

3.3. AI (artificial intelligence) algorithm

Fig. 17. Data statistics of artificial intelligence path planning based on litera­ AI is a study of making computers to simulate thought process and
tures in this paper.
intelligent behaviors of human beings (e.g., learning, reasoning,
thinking, planning, etc.). The three cornerstones of artificial intelligence
ant colony optimization algorithm and the fast convergence of SSA. are algorithms, data, and computing capability. In this paper, we focus
After limited initial iterations, the ant colony algorithm was transferred on two popular AI algorithms for path planning of mobile robots,
into SSA. Yu et al. (2021) proposed that SPSA algorithm can change the including bioinspired neural network algorithm and fuzzy control
position update of the finders, enhance the influence of the starting line algorithm.

23
L. Liu et al. Expert Systems With Applications 227 (2023) 120254

Fig. 19. Data statistics of optimization methods and experimental methods of path planning algorithms based on literatures in this paper.

3.3.1. BNN (bioinspired neural network) (regions with convolutional neural networks) approach to discover
Meng & Yang (Luo et al., 2019a) firstly introduced the BNN method scattered nails and screws in real time, so that the robot can automati­
to solve the path planning problem of mobile robots in 1998. Compared cally recover nails and screws. A shortest path deduction algorithm
with other NN algorithms, BNN algorithms do not need to be learned (Ni based on neural network was designed for underwater robots by Zhu &
& Yang, 2011). The dynamic activation value function of each neuron Yang (2021). Yan & Zhao (2021) constructed the GBNN (glasius bio-
cell can be solved through information transfer between neurons. And inspired neural network) model with a topological structure, which
by solving the dynamic activation potential function of neurons in real represented the search environment. Gaussian mixture model was used
time, the feasible path from the initial position to the target position to extract high-value subregions and quantify future detection rewards.
along the direction of increasing potential value can be obtained as the The external excitatory input to the neurons of the GBNN can be directly
planning path (Luo, 2021). At present, the structure of BNN contains introduced as a way to improve the search capability. Godio et al. (2021)
input layer, hidden layer, and output layer (Zhao, 2020).. described an approach based on bio-inspired neural network, which can
A BNN-based path planning method for multi-robot capture was solve the coverage planning problem for unmanned aerial vehicle for­
proposed by Chen & Zou (2021). They constructed a dynamic alliance mations exploring critical areas. Khan et al. (2021) used the ZNNBAS
strategy to achieve the interaction of multiple robots, constructed a (zeroing neural network with beetle antennae search) to solve for mobile
target tracking strategy based on bio-inspired neural networks to home robots, further improving the computing ability and efficiency of
dynamically guide all robots in the coalition for tracking, and used the the system. ZNNBAS is a nature-inspired meta heuristic algorithm
formation strategy to capture the target. Yi et al. (2016) integrated the inspired by the food-seeking nature of beetles.
advantages and features of BNN with the self-organizing feature map to
handle the task allocation of large batches of robots in 3D dynamic 3.3.2. FL (fuzzy logic)
environments. Based on the improved BNN algorithm, a real-time Zadeth (1965) proposed the concept of FL in 1965. Based on real-
collision-free path planning method for descaling robots in a naval time information from sensors, FL regards current environmental
environment was proposed by Sun (2017b). Ni et al. (2017) proposed an obstacle information as input of the fuzzy controller, analyzes the
improved dynamic BNN algorithm to handle real-time path planning for desired speed by simulating human driving experience, constructs a
autonomous underwater vehicle in various 3D underwater environ­ planning information table, and finally controls the mobile robot to
ments. Yang et al. (2018) proposed an improved BNN-based “active complete local path planning in an unstructured environment. With
obstacle avoidance strategy” pedestrian position prediction method to better real-time performance, FL is less influenced by the external
plan collision-free complete coverage path planning trajectories for environment and is suitable for path planning in unknown environments
mobile robots. Luo et al. (2016) proposed a BNN-based dynamics (Li, 2015).
approach to handle the complete coverage path planning of multiple However, FL also has some inherent defects. For example, fuzzy
robots. Wang et al. (2019) used BNN techniques to assist the robot in inference rules should be formulated in advance based on experts’
patrolling the unknown working environments, and used the fast R-CNN experience, which is closely related to the advantages and disadvantages

24
L. Liu et al. Expert Systems With Applications 227 (2023) 120254

of path; the optimal path depends on fuzzy inference rules; as the optimizing the CD algorithm itself to deal with its shortcomings. SBM,
obstacle information increases, the inference rules or fuzzy tables will be GSA and APF are mainly used to improve the algorithm itself, but
sharply dilated, which affects the real-time performance of the algo­ compared with GSA, APF and DWA is the easiest to fuse with other al­
rithm; when the environment changes, the existing rules may not be gorithms. Among the bionics-based path planning algorithms, there are
applicable to the new environment, resulting in the failure to choose the fewer studies for local path planning than global path planning. How­
correct path; the obstacle avoidance strategy is not intelligent and ever, compared with classical path planning, these algorithms response
cannot flexibly cope with dynamic obstacles. Many scholars proposed to quickly and are more suitable to be optimized for local path planning.
improve the online learning capability and accuracy of FL by combining We also find that there are a large proportion of research papers on ACO,
the FL algorithm with others (Xie, 2016). Zagradjanin et al. (2021), used GWO, GA and PSO which have shorter search time and stronger search
D*lite algorithm in global path planning and used FL in local path capability. For these algorithms, researchers are keen on optimizing
planning, and introduced this method into collaborative multi-robot them for a particular defect rather than combining them with other al­
operations in complex environments. Ntakolia & Lyridis (2021) pro­ gorithms. But PSO, SSA, and GA are more easily fused with other al­
posed swarm intelligence graph-based pathfinding algorithm (SIGPA), gorithms. In the field of AI algorithms, the proportion of researches on
which aims to improve the performance of the SIGPA algorithm, cope NN and FL is basically equal. What’s more, AI algorithm is more often
with multiple objectives and produce high-quality solutions by inte­ used in local path planning due to its higher environment exploration
grating fuzzy logic. Gharajeh & Jond (2022) proposed a collision-free capability and more flexible operation capability. FL is rarely used for
autonomous navigation path planning method for wheeled mobile ro­ path planning in three-dimension environments because it is more
bots based on adaptive neuro-fuzzy inference system (ANFIS). And complex to design and requires larger memory. But FL is often fused
experimental results showed that the paths found by this method were with other algorithms, or it is used to optimize the certain defect of other
about 30% shorter than those found by other methods. Guo et al. (2021) algorithms. In Fig. 19, it can be seen that bionics-based and artificial
proposed a new algorithm based on a combination of long and short- intelligence algorithms contribute more to local path planning than
term memory neural networks, fuzzy logic control and reinforcement classical algorithms. Some intrinsic defects of classical algorithm, such
learning control. And low-dimensional input fuzzy logic control (FL) as local optimal solutions, lack of intelligence, and low computational
algorithm was applied to collect training data and the transfer learning intensity, cause the failure of performing real-time path planning.
method was applied to learn the basic capabilities. Wang et al. (2021b), This paper reviews the path planning techniques of mobile robots
proposed hierarchical path planning (HPP) scheme combining global based on a large amount of literature and categories the research-
and local tasks for unmanned surface vehicles by considering fuzzy oriented literature. The conclusions are as follows.
artificial potential field (FAPF). Wang et al. (2021a) introduced NN al­ (1) The environment modeling for mobile robots usually uses gird
gorithm into FAPF. The training samples of the neural network consists method that is easy to implement and extend to three-dimension
of the coordinates of the target and the obstacle, as well as the motion environment.
direction of the robot corresponding to this position relationship, which (2) Classical path planning algorithms are often used to perform
can further improve the path optimization capability of the mobile ro­ global path planning. They need small memory occupation without
bots. AlKhlidi et al. (2021) proposed FPSO algorithm combining PSO expensive environmental monitoring systems, becoming very popular in
with FL that can improve the problem of PSO algorithm falling into local low-cost mobile robots.
optimal solutions. Sangeetha et al. (2021b) proposed fuzzy gain-based (3) Artificial intelligence algorithms are most preferred in local path
dynamic ant colony optimization (FGDACO) for dynamic path plan­ planning because they are trained to learn with higher intelligence and
ning, which effectively plans collision-free and smooth paths with the better to handle uncertainty in the environment.
shortest length and time. Tao et al. (2021) introduced a pheromone gain- (4) There is less research on path planning in three-dimension en­
based function to the path planning process in order to effectively utilize vironments than two-dimension environments.
the pheromone update mechanism of the ant colony system. A new (5) Researchers are keen on improving a particular algorithm rather
pheromone update rule was proposed, which used fuzzy control to than fusing different algorithms.
change the magnitude of the pheromone heuristic factor and the desired (6) The experiment verification method used in most of the papers is
heuristic factor, as well as adjust the evaporation rate in stages. This simulating on a computer rather than applying it in real environments of
method converged quickly, ensuring the global search capability. mobile robots.
Table 24 clearly shows the difference between BNN and FL. (7) One third of the researchers considered path smoothing
techniques.
4. Conclusion and prospection
4.2. Prospection
4.1. Conclusion
With the continuous development of computer science, mobile ro­
Based on the statistics of the research-oriented literature covered in bots are widely used in industry, agriculture, service industry and other
this paper, we made Table 25. In order to explore the development trend fields, which requires that mobile robot systems have good real-time and
of path planning algorithms, we performed a data-based analysis of robustness. Therefore, in the future path planning technology, in addi­
Table 3 and obtained Figs. 14–19. tion to research and discovery of new path planning algorithms, the
Among the 105 papers regarding environment modeling, the per­ following aspects deserve attention:
centages of models using grid method, topology method, geometrical
measurers, and hybrid methods are 85.71%, 3.81%, 8.57%, and 1.90%, Environment modeling technology is combined with path planning
respectively. In Fig. 14, it can be intuitively found that the grid method, algorithm. When a mobile robot is in a complex 2D or even 3D
as the mainstream modeling method in path planning of mobile robots, continuous dynamic environment, the function of the algorithm is
has been irreplaceable in the past and even in the future. limited. The combination of good modeling technology and path
In recent years, many classical path planning algorithms have been planning algorithm will become a method to solve this problem.
proposed. CD, SBM and GSA have contributed more to global path The problem of combining global path planning with local path
planning than local path planning, while APF’s contribution to global planning. Global path planning refers to the condition that all the
path planning is almost equal to that to local path planning. DWA is a data in the environmental map are known, and the safety node and
path planning algorithm for dynamic environment. CD is seldom fused safety area, as well as the optimal driving path, are judged by
with other algorithms, and researchers usually adopt the approach of combining the real-time information obtained with the feedback

25
L. Liu et al. Expert Systems With Applications 227 (2023) 120254

control. The global path can only be roughly planned, but mobile Alyasin, A., Abbas, E. I., & Hasan, S. D. (2019). An efficient optimal path finding for
mobile robot based on dijkstra method. In 2019 4th Scientific International Conference
robot driving is a very complex problem. Especially after the mobile
Najaf (SICN) (pp. 11–14). IEEE.
robot is applied to military industry, agriculture, service industry An, Y. Y., & Zeng, X. D. (2001). Optical sensing and measurement. Beijing Electronic
and other occasions with high real-time requirements, not only the Industry Press, 293, 294.
initial environmental data processing, but also the dynamic factors Balado, J., Díaz-Vilariño, L., Arias, P., et al. (2019). Point clouds for direct pedestrian
pathfinding in urban environments. ISPRS Journal of Photogrammetry and Remote
such as the state of the robot when moving and the characteristics of Sensing, 148, 184–196.
obstacles should be considered. Of course, the width and curvature of Bao J.S., Zhang M.Y., Ge S.R., et al., Underground Driverless Path Planning of Trackless
the path, Static factors such as the particularity of intersections Rubber Tyred Vehicle Based on Improved A * and Artificial Potential Field
Algorithm[J/OL]. Journal of China Coal Society: 1-14[2022-03-22]. DOI:10.13225/
should also be included. Therefore, the mobile robot constantly in­ j.cnki.jccs.XR21.1717.
teracts with the environment to obtain local paths, and enough local Brunner M., Brüggemann B., Schulz D., Hierarchical rough terrain motion planning using
paths constitute the global path. an optimal sampling-based method. 2013 IEEE International Conference on Robotics
and Automation. IEEE, 2013: 5539-5544.
Improvement of advanced path planning algorithm. When the Bullnheimer B., Hartl R.F., Strauss C. A new rank based version of the Ant System. A
environment changes, traditional algorithms will expose their computational study[J]. 1997.
shortcomings. For example, mobile robots are widely used to deliver Chai, R., Tsourdos, A., Savvaris, A., et al. (2021). Solving constrained trajectory planning
problems using biased particle swarm optimization. IEEE Transactions on Aerospace
meals in many countries, and the scene is densely populated. Re­ and Electronic Systems, 57(3), 1685–1701.
searchers can consider adding learning function to the existing al­ Chang, Z., Wu, H., Sun, Y., et al. (2022). RGB-D Visual SLAM Based on Yolov4-Tiny in
gorithm. After continuous learning, the probability of collision of the Indoor Dynamic Environment. Micromachines, 13(2), 230.
Challita, U., Saad, W., & Bettstetter, C. (2018). Deep reinforcement learning for
delivery robot will become smaller and smaller.
interference-aware path planning of cellular-connected UAVs. In 2018 IEEE
Effective combination of path planning algorithms. Each algorithm international conference on communications (ICC) (pp. 1–7). IEEE.
has inherent defects due to its own limitations in practical applica­ Chen, G., Luo, N., Liu, D., et al. (2021a). Path planning for manipulators based on an
tion. Through the fusion of two or more algorithms, we can learn improved probabilistic roadmap method. Robotics and Computer-Integrated
Manufacturing, 72, Article 102196.
from each other’s strengths to offset each other’s weaknesses. To Chen, Y. G. (2020). Local Tracking and Global Localization of Mobile Robot based on
some extent, it can avoid the defects of a single algorithm. Sensor Fusion of Vision and LiDAR[D]. Guangdong University of Technology. https://
Design of path planning algorithm for multi machine collaborative doi.org/10.27029/d.cnki.ggdgu.2020.000295
Chen, Y., Bai, G., Zhan, Y., et al. (2021b). Path planning and obstacle avoiding of the USV
work. With the continuous development of science and technology, based on improved ACO-APF hybrid algorithm with adaptive early-warning. IEEE
the problem of multi machine collaborative parallel operation and Access, 9, 40728–40742.
path conflict has been gradually concerned. How to realize the path Chen, Y., & Zhou, X. (2021). Path planning of robot based on improved ant colony
algorithm in computer technology. Journal of Physics: Conference Series IOP
planning of multi robot collision free operation will become one of Publishing, 1744(4), Article 042092.
the research focuses in the future. Chen, Z., Zhou, J., Sun, R., et al. (2021c). A new evolving mechanism of genetic
algorithm for multi-constraint intelligent camera path planning. Soft Computing, 25
(7), 5073–5092.
CRediT authorship contribution statement Chen, Z., & Zou, A. C. (2021). Research on multi-robot cooperative roundup based on
biological heuristic neural network. Electronic Measurement Technology, 44(10),
Lixing Liu: Conceptualization, Methodology, Validation, Writing – 82–90. https://doi.org/10.19651/j.cnki.emt.2106018
Chengli, F. A. N., Qiang, F. U., Guang, Z. L., et al. (2018). Hybrid artificial bee colony
original draft. Xu Wang: Software, Conceptualization, Supervision. Xin algorithm with variable neighborhood search and memory mechanism. Journal of
Yang: Writing – review & editing, Formal analysis, Supervision, Visu­ Systems Engineering and Electronics, 29(2), 405–414.
alization. Hongjie Liu: Investigation, Resources, Data curation. Jianp­ Cheng, X., Li, J., Zheng, C., et al. (2021). An improved PSO-GWO algorithm with chaos
and adaptive inertial weight for robot path planning. Frontiers in Neurorobotics, 15.
ing Li: Formal analysis, Funding acquisition, Project administration. Chien, R. T., Zhang, L., & Zhang, B. (1984). Planning collision-free paths for robotic arm
Pengfei Wang: Software, Data curation, Validation. among obstacles. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1,
91–96.
Chou, J. S., & Ngo, N. T. (2017). Modified firefly algorithm for multidimensional
optimization in structural design problems. Structural and Multidisciplinary
Declaration of Competing Interest Optimization, 55(6), 2013–2028.
Chu, K. X., Chang, T. Q., Wang, Q. D., et al. (2021). Path Planning Method of Ant Colony
The authors declare that they have no known competing financial Algorithm Based on Effective Turning Point and Shortest-minimum Path.
Transactions of the Chinese Society for Agricultural Machinery, 52(12), 400–407.
interests or personal relationships that could have appeared to influence Chu, X., Hu, G., Niu, B., et al. (2016). An superior tracking artificial bee colony for global
the work reported in this paper. optimization problems. In 2016 IEEE Congress on Evolutionary Computation (CEC)
(pp. 2712–2717). IEEE.
Dai, X., Long, S., Zhang, Z., et al. (2019). Mobile robot path planning based on ant colony
Data availability algorithm with A* heuristic method. Frontiers in Neurorobotics, 13, 15.
Daniel, K., Nash, A., Koenig, S., et al. (2010). Theta*: Any-angle path planning on grids.
Data will be made available on request. Journal of Artificial Intelligence Research, 39, 533–579.
Daraei, M. H. (2018). Tightly-coupled lidar and camera for autonomous vehicles[M]. Santa
Cruz: University of California.
Acknowledgements Debnath, S. K., Omar, R., Bagchi, S., et al. (2021). Different cell decomposition path
planning methods for unmanned air vehicles-A review. In Proceedings of the 11th
National Technical Seminar on Unmanned System Technology 2019 (pp. 99–111).
This work was supported by the earmarked fund for CARS (CARS-27) Singapore: Springer.
and supported by the Hebei Province Graduate Innovation Funding Detone D., Malisiewicz T., Rabinovich A., Toward Geometric Deep SLAM[J]. 2017.
Project (CXZZBS2023080). Dhal, K. G., Iqbal, Q. M., & Das, S. (2016). Development of firefly algorithm via chaotic
sequence and population diversity to enhance the image contrast. Natural Computing,
15(2), 307–318.
References Ding, X. Q., Du, Z. Y., Lu, Y. Q., et al. (2016). Visual trajectory planning for mobile robots
based on hybrid artificial potentialfield. Journal of Zhejiang University(Engineering
Science), 50(07), 1298–1306.
Abdallaoui, S., Aglzim, E. H., Chaibet, A., et al. (2022). Thorough review analysis of safe
Dorigo M., Maniezzo V., Colorni A., The ant system: An autocatalytic optimizing process
control of autonomous vehicles: Path planning and navigation techniques. Energies,
[J]. 1991.
15(4), 1358.
Dorigo, M., & Gambardella, L. M. (1997). Ant colonies for the travelling salesman
Abdi, M. I. I., Khan, M. U., Güneş, A., et al. (2020). Escaping Local Minima in Path
problem. Bio Systems, 43(2), 73–81.
Planning Using a Robust Bacterial Foraging Algorithm. Applied Sciences, 10(21),
Du, W. R., Wang, X. Y., Tian, T., et al. (2021). Artificial potential field path planning
7905.
algorithm for unknown environment and dynamic obstacles. Computer Science, 48
Albina, K., & Lee, S. G. (2019). Hybrid stochastic exploration using grey wolf optimizer
(02), 250–256.
and coordinated multi-robot exploration algorithms. IEEE Access, 7, 14246–14255.
Alkhlidi, B., Abdulsadda, A. T., & Al, B. A. (2021). Optimal Robotic Path Planning Using
Intlligents Search Algorithms. Journal of Robotics and Control (JRC), 2(6), 519–526.

26
L. Liu et al. Expert Systems With Applications 227 (2023) 120254

Esposito, J. M., & Wright, J. N. (2019). Matrix completion as a post-processing technique Karaman, S., & Frazzoli, E. (2011). Sampling-based algorithms for optimal motion
for probabilistic roadmaps. The International Journal of Robotics Research, 38(2/3), planning. The International Journal of Robotics Research, 30(7), 846–894.
388–400. Karaboga, D., & Basturk, B. (2007). A powerful and efficient algorithm for numerical
Fan, Y., Sun, X., Wang, G., et al. (2021). Collision avoidance controller for unmanned function optimization: Artificial bee colony (ABC) algorithm. Journal of Global
surface vehicle based on improved cuckoo search algorithm. Applied Sciences, 11(20), Optimization, 39(3), 459–471.
9741. Khan, A. T., Li, S., & Cao, X. (2021). Control framework for cooperative robots in smart
Fei W., Ziwei W., Meijin L., Robot Path Planning Based on Improved Particle Swarm home using bio-inspired neural network. Measurement, 167, Article 108253.
Optimization. 2021 IEEE 2nd International Conference on Big Data, Artificial Kiani, F., Seyyedabbasi, A., Aliyev, R., et al. (2021). 3D path planning method for multi-
Intelligence and Internet of Things Engineering (ICBAIE). IEEE, 2021: 887-891. UAVs inspired by grey wolf algorithms. Journal of Internet Technology, 22(4),
Feng, S., Qian, Y., & Wang, Y. (2021). Collision avoidance method of autonomous vehicle 743–755.
based on improved artificial potential field algorithm. Proceedings of the Institution of Lamini, C., Benhlima, S., & Elbekri, A. (2018). Genetic algorithm based approach for
Mechanical Engineers, Part D: Journal of Automobile Engineering, 235(14), 3416–3430. autonomous mobile robot path planning. Procedia Computer Science, 127, 180–189.
Ferguson, D., & Stentz, A. (2006). Using interpolation to improve path planning: The Latombe, J. C. (1991). Approximate cell decomposition. In Robot Motion Planning (pp.
Field D* algorithm. Journal of Field Robotics, 23(2), 79–101. 248–294). Boston, MA: Springer.
Fink, W., Baker, V. R., Brooks, A. J. W., et al. (2019). Globally optimal rover traverse Li, F., Fan, X., & Hou, Z. (2020). A Firefly Algorithm With Self-Adaptive Population Size
planning in 3D using Dijkstra’s algorithm for multi-objective deployment scenarios. for Global Path Planning of Mobile Robot. IEEE Access, 8, 168951–168964.
Planetary and Space Science, 179, Article 104707. Li, H. L. (2020). Improved artificial bee colony algorithm and related application
Fox, D., Burgard, W., & Thrun, S. (2002). The Dynamic Window Approach to Collision research[D]. Nanjing University of Posts and Telecommunications. https://doi.org/
Avoidance. IEEE Robotics & Automation Magazine, 4(1), 23–33. 10.27251/d.cnki.gnjdc.2020.001231
Gao, Q., Zheng, J., & Zhang, W. (2021). Research on optimization of manned robot Li, M. (2014). Visual-inertial odometry on resource-constrained systems[M]. Riverside:
swarm scheduling based on ant-sparrow algorithm. Journal of Physics: Conference University of California.
Series. IOP Publishing, 2078(1), Article 012002. Li, M. (2015). Mobile robot path planning based on fuzzycontrol. Hebei University of
Germi S.B., Khosravi M.A., FesharakiFard R., Adaptive GA-based potential field Technology.
algorithm for collision-free path planning of mobile robots in dynamic Li, J. (2021). Robot path planning based on improved sparrow algorithm. Journal of
environments. 2018 6th RSI International Conference on Robotics and Mechatronics Physics: Conference Series. IOP Publishing, 1861(1), Article 012017.
(IcRoM). IEEE, 2018: 28-33. Li, Q. Q., Bu, S. Q., & Yang, J. F. (2021). Research review on biological swarm
Gharajeh, M. S., & Jond, H. B. (2022). An intelligent approach for autonomous mobile intelligence algorithm in mobile robot path planning. World Sci-Tech R&D, 43(05),
robots path planning based on adaptive neuro-fuzzy inference system. Ain Shams 535–546. https://doi.org/10.16507/j.issn.1006-6055.2021.04.004
Engineering Journal, 13(1), Article 101491. Li, Q., Wang, J., Li, H., et al. (2022a). Fast-RRT*: An improved motion planner for mobile
Godio, S., Primatesta, S., Guglieri, G., et al. (2021). A Bioinspired neural network-based robot in two-dimensional space. IEEJ Transactions on Electrical and Electronic
approach for cooperative coverage planning of UAVs. Information, 12(2), 51. Engineering, 17(2), 200–208.
Gomez C., Fehr M., Millane A., et al., Hybrid topological and 3d dense mapping through Li, Y., Ming, Y., Zhang, Z., et al. (2021). An adaptive ant colony algorithm for
autonomous exploration for large indoor environments. 2020 IEEE International autonomous vehicles global path planning. In 2021 IEEE 24th International
Conference on Robotics and Automation (ICRA). IEEE, 2020: 9673-9679. Conference on Computer Supported Cooperative Work in Design (CSCWD) (pp.
Gonzalez R., Kloetzer M., Mahulea C., Comparative study of trajectories resulted from 1117–1122).
cell decomposition path planning approaches. 2017 21st International Conference Li, Y., Li, J., Zhou, W., et al. (2022b). Robot path planning navigation for dense planting
on System Theory, Control and Computing (ICSTCC). IEEE, 2017: 49-54. red jujube orchards based on the joint improved A* and DWA algorithms under laser
Graeter, J., Wilczynski, A., & Lauer, M. (2019). LIMO: Lidar-monocular visual odometry. SLAM. Agriculture, 12(9), 1445.
2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE. Liang, X. (2016). Indoor slam for robots based on laser and mono-vision fusion. Harbin
Gregorio, D. D., & Stefano, L. D. (2017). SkiMap: An Efficient Mapping Framework for Institute of Technology.
Robot Navigation. IEEE. IEEE. Liang, X., Meng, G., Xu, Y., et al. (2018). A geometrical path planning method for
Guivant, J., Nebot, E., & Durrant-Whyte, H. (2000). Simultaneous localization and map unmanned aerial vehicle in 2D/3D complex environment. Intelligent Service Robotics,
building using natural features in outdoor environments. Intelligent Autonomous 11(3), 301–312.
Systems., 6(1), 581–586. Lin, H. X., Xiang, D., & Ou, Y. J. (2021). Review of Path Planning Algorithms for Mobile
Guo, D., Wang, J., Zhao, J. B., et al. (2019). A vehicle path planning method based on a Robots. Computer Engineering and Applications, 57(18), 38–48.
dynamic traffic network that considers fuel consumption and emissions. Science of Liu, G., Shu, C., Liang, Z., et al. (2021a). A modified sparrow search algorithm with
the Total Environment, 663, 935–943. application in 3d route planning for UAV. Sensors, 21(4), 1224.
Guo, N., Li, C., Gao, T., et al. (2021). A fusion method of local path planning for mobile Liu, H., Ge, J., Wang, Y., et al. (2021b). Multi-UAV optimal mission assignment and path
robots based on LSTM neural network and reinforcement learning. Mathematical planning for disaster rescue using adaptive genetic algorithm and improved artificial
Problems in Engineering, 2021. bee colony method. Actuators MDPI, 11(1), 4.
Guruprasad, K. R., & Ranjitha, T. D. (2021). CPC algorithm: Exact area coverage by a Liu, H. P., Wang, D. S., Wang, Y. X., et al. (2022). Research of path planning for mobile
mobile robot using approximate cellular decomposition. Robotica, 39(7), robots based on fuzzy artificial potential field method. Control Engineering of China,
1141–1162. 29(01), 33–38. https://doi.org/10.14107/j.cnki.kzgc.20200187
Hao, W., Luo, B., & Zhang, Z. (2021). Application study of UAV path planning based on Liu, J. (2019). Research and application of clustering recommendation algorithm based
the balanced search factor artificial bee colony algorithm. Journal of Physics: on cuckoo search. Software Guide, 18(04), 91–94.
Conference Series. IOP Publishing, 2083(3), Article 032064. Liu, J., Wei, X., & Huang, H. (2021). An improved grey wolf optimization algorithm and
He, Z., Liu, C., Chu, X., et al. (2022). Dynamic anti-collision A-star algorithm for multi- its application in path planning. IEEE Access, 9, 121944–121956.
ship encounter situations. Applied Ocean Research, 118, Article 102995. Liu, K. (2018). Object tracking and path planning of indoor mobile robot based on visual
Howden, W. E. (1968). The sofa problem. The Computer Journal, 11(3), 299–301. information. South China University of Technology.
Huang, C. (2018a). Study on Path Planning and Location of Mobile Robot Based on Intelligent Liu, Q., Zhang, Y., Li, M., et al. (2021c). Multi-UAV path planning based on fusion of
Optimization Algorithm [2018]. Dalian Jiaotong University. sparrow search algorithm and improved bioinspired neural network. IEEE Access, 9,
Huang Y. Research on the improvement of Dijkstra algorithm in the shortest path 124670–124681.
calculation. Proceedings of the 2017 4th International Conference on Machinery, Liu, Y., & Li, X. (2022). A Hybrid mobile robot path planning scheme based on modified
Materials and Computer (MACMC 2017), Atlantis Press. 2018b: 745-749. gray wolf optimization and situation assessment. Journal of Robotics, 2022.
Islam, F., Nasir, J., Malik, U., et al. (2012). Rrt*-smart: Rapid convergence Liu, Z., Yuan, X., Huang, G., et al. (2021d). Two potential fields fused adaptive path
implementation of rrt* towards optimal solution. In 2012 IEEE international planning system for autonomous vehicle under different velocities. ISA Transactions,
conference on mechatronics and automation (pp. 1651–1656). IEEE. 112, 176–185.
Jafarzadeh, H., & Fleming, C. H. (2018). An exact geometry-based algorithm for path Long, Y., Zuo, Z., Su, Y., et al. (2020). An A*-based bacterial foraging optimisation
planning. International Journal of Applied Mathematics and Computer Science, 28(3). algorithm for global path planning of unmanned surface vehicles. The Journal of
Jain, M., Saihjpal, V., Singh, N., et al. (2022). An Overview of Variants and Navigation, 73(6), 1247–1262.
Advancements of PSO Algorithm. Applied Sciences, 12(17), 8392. Long, Y., Su, Y., Shi, B., et al. (2021). A multi-subpopulation bacterial foraging
Jamshidi, V., Nekoukar, V., & Refan, M. H. (2021). Real time UAV path planning by optimisation algorithm with deletion and immigration strategies for unmanned
parallel grey wolf optimization with align coefficient on CAN bus. Cluster Computing, surface vehicle path planning. Intelligent Service Robotics, 14(2), 303–312.
24(3), 2495–2509. Luo, C., Yang, S. X., Li, X., et al. (2016). Neural-dynamics-driven complete area coverage
Jarray, R., Al-Dhaifallah, M., Rezk, H., et al. (2022). Parallel cooperative coevolutionary navigation through cooperation of multiple mobile robots. IEEE Transactions on
grey wolf optimizer for path planning problem of unmanned aerial vehicles. Sensors, Industrial Electronics, 64(1), 750–760.
22(5), 1826. Lu, C. (2017). Study on theory and method of multi-objective scheduling problems with
Ji, Y., & Liu, B. (2022). Research and Implementation of Robot Path Planning Based on controllable processing times. Huazhong University of Science and Technology.
Ant Colony Algorithm. Journal of Physics: Conference Series. IOP Publishing, 2171(1), Luo, M. (2021). Research on path planning and trajectory tracking of mobile robot in
Article 012074. complex environment. University of Electronic Science and Technology of China.
Jiao, Z., Ma, K., Rong, Y., et al. (2018). A path planning method using adaptive https://doi.org/10.27005/d.cnki.gdzku.2021.000212
polymorphic ant colony algorithm for smart wheelchairs. Journal of Computational Luo M, Hou X, Yang J. Multi-robot one-target 3d path planning based on improved
Science, 25, 50–57. bioinspired neural network[C]//2019 16th International Computer Conference on
Jung, J. W., So, B. C., Kang, J. G., et al. (2019). Expanded douglas–peucker polygonal Wavelet Active Media Technology and Information Processing. IEEE, 2019a: 410-
approximation and opposite angle-based exact cell decomposition for path planning 413.
with curvilinear obstacles. Applied Sciences, 9(4), 638.

27
L. Liu et al. Expert Systems With Applications 227 (2023) 120254

Luo, Q., Wang, H. B., Cui, X. J., et al. (2019b). Autonomous mobile robot path planning Ryu, H., & Park, Y. (2019). Improved informed RRT* using gridmap skeletonization for
based on improved artificial potential method. Control Engineering of China, 26(06), mobile robot path planning. International Journal of Precision Engineering and
1091–1098. https://doi.org/10.14107/j.cnki.kzgc.170907 Manufacturing, 20(11), 2033–2039.
Mahmud, M. S. A., Abidin, M. S. Z., Mohamed, Z., et al. (2019). Multi-objective path Sadhu, A. K., Konar, A., Bhattacharjee, T., et al. (2018). Synergism of firefly algorithm
planner for an agricultural mobile robot in a virtual greenhouse environment. and Q-learning for robot arm path planning. Swarm and Evolutionary Computation,
Computers and Electronics in Agriculture, 157, 488–499. 43, 50–68.
Ma, F., Shi, J., Yang, Y., et al. (2019). ACK-MSCKF: Tightly-coupled Ackermann multi- Samaniego, F., Sanchis, J., García-Nieto, S., et al. (2019). Recursive rewarding modified
state constraint Kalman filter for autonomous vehicle localization. Sensors, 19(21), adaptive cell decomposition (RR-MACD): A dynamic path planning algorithm for
4816. UAVs. Electronics, 8(3), 306.
Martins, O. O., Adekunle, A. A., Olaniyan, O. M., et al. (2022). An Improved multi- Sang, H., You, Y., Sun, X., et al. (2021). The hybrid path planning algorithm based on
objective a-star algorithm for path planning in a large workspace: Design, improved A* and artificial potential field for unmanned surface vehicle formations.
Implementation, and Evaluation. Scientific African, 15, e01068. Ocean Engineering, 223, Article 108709.
Miao, C., Chen, G., Yan, C., et al. (2021). Path planning optimization of indoor mobile Sangeetha, V., Krishankumar, R., Ravichandran, K. S., et al. (2021a). Energy-efficient
robot based on adaptive ant colony algorithm. Computers & Industrial Engineering, green ant colony optimization for path planning in dynamic 3D environments. Soft
156, Article 107230. Computing, 25(6), 4749–4769.
Mirjalili S., Mirjalili S.M., Lewis A., Grey wolf optimizer. Advances in Engineering Sangeetha, V., Krishankumar, R., Ravichandran, K. S., et al. (2021b). A fuzzy gain-based
Software 69, 46-61[J]. 2014. dynamic ant colony optimization for path planning in dynamic environments.
Mohanty, P. K., Singh, A. K., Kumar, A., et al. (2021). Path planning techniques for Symmetry, 13(2), 280.
mobile robots: A review. In International Conference on Soft Computing and Pattern Santos, A. H. M., de Lima, R. M., Pereira, C. R. S., et al. (2019). Optimizing routing and
Recognition (pp. 657–667). Cham: Springer. tower spotting of electricity transmission lines: An integration of geographical data
Mourikis, A. I., & Roumeliotis, S. I. (2007). A multi-state constraint kalman filter for and engineering aspects into decision-making. Electric Power Systems Research, 176,
vision-aided inertial navigation. ICRA, 2, 6. Article 105953.
Muni, M. K., Parhi, D. R., & Kumar, P. B. (2021). Improved motion planning of humanoid Šeda, M. (2007). Roadmap methods vs. cell decomposition in robot motion planning. In
robots using bacterial foraging optimization. Robotica, 39(1), 123–136. Proceedings of the 6th WSEAS international conference on signal processing, robotics and
Muntasha G., Karna N., Shin S.Y., Performance Analysis on Artificial Bee Colony automation (pp. 127–132). Athens, Greece: World Scientific and Engineering
Algorithm for Path Planning and Collision Avoidance in Swarm Unmanned Aerial Academy and Society (WSEAS).
Vehicle[C]//2021 International Conference on Artificial Intelligence and Sepehri, A., & Moghaddam, A. M. (2021). A motion planning algorithm for redundant
Mechatronics Systems (AIMS). IEEE, 2021: 1-6. manipulators using rapidly exploring randomized trees and artificial potential fields.
Nazarahari, M., Khanmirza, E., & Doostie, S. (2019). Multi-objective multi-robot path IEEE Access, 9, 26059–26070.
planning in continuous environment using an enhanced genetic algorithm. Expert Shah, H., Tairan, N., Garg, H., et al. (2018). Global gbest guided-artificial bee colony
Systems with Applications, 115, 106–120. algorithm for numerical function optimization. Computers, 7(4), 69.
Ni, J., & Yang, S. X. (2011). Bioinspired neural network for real-time cooperative hunting Shao, X., Zheng, R., Wei, J., et al. (2021). Path planning of mobile Robot based on
by multirobots in unknown environments. IEEE Transactions on Neural Networks, 22 improved ant colony algorithm based on Honeycomb grid. In 2021 IEEE 5th
(12), 2062–2077. Advanced Information Technology, Electronic and Automation Control Conference
Ni, J., Wu, L., Shi, P., et al. (2017). A dynamic bioinspired neural network based real- (IAEAC) (pp. 1358–1362). IEEE.
time path planning method for autonomous underwater vehicles. Computational K. Sharma S. Singh R. Doriya Optimized cuckoo search algorithm using tournament
intelligence and neuroscience, 2017. selection function for robot path planning[J] International Journal of Advanced
Nian, J. (2021). Indoor map construction and navigation of mobile robot based on multi- Robotic Systems 18 3 2021 1729881421996136.
sensor fusion. Beijing University of Posts and Telecommunications. https://doi.org/ Shi, Q., Li, C., Wang, C., et al. (2017). Design and implementation of an omnidirectional
10.26969/d.cnki.gbydu.2021.002501 vision system for robot perception. Mechatronics, 41, 58–66.
Ntakolia, C., & Lyridis, D. V. (2021). A swarm intelligence graph-based pathfinding Shin, Y., & Kim, E. (2021). Hybrid path planning using positioning risk and artificial
algorithm based on fuzzy logic (SIGPAF): A case study on unmanned surface vehicle potential fields. Aerospace Science and Technology, 112, Article 106640.
multi-objective path planning. Journal of Marine Science and Engineering, 9(11), 1243. Shivgan R., Dong Z., Energy-efficient drone coverage path planning using genetic
Oršulić, J., Miklić, D., & Kovačić, Z. (2019). Efficient dense frontier detection for 2- algorithm. 2020 IEEE 21st International Conference on High Performance Switching
d graph slam based on occupancy grid submaps. IEEE Robotics and Automation Letters, and Routing (HPSR). IEEE, 2020: 1-6.
4(4), 3569–3576. Song, F. Q. (2021). Improved firefly algorithm and its application and research in vehicle
Palmieri L., Koenig S., Arras K.O., RRT-based nonholonomic motion planning using any- routing problem[D]. Jiangxi University of Finance and Economics. https://doi.org/
angle path biasing[C]//2016 IEEE International Conference on Robotics and 10.27175/d.cnki.gjxcu.2021.000640
Automation (ICRA). IEEE, 2016: 2775-2781. Song, P. C., Pan, J. S., & Chu, S. C. (2020). A parallel compact cuckoo search algorithm
Pang, B., Dai, W., Hu, X., et al. (2021). Multiple air route crossing waypoints for three-dimensional path planning. Applied Soft Computing, 94, Article 106443.
optimization via artificial potential field method. Chinese Journal of Aeronautics, 34 Souza, F. R., Câmara, T. R., Torres, V. F. N., et al. (2019). Mine fleet cost evaluation-
(4), 279–292. Dijkstra’s optimized path. REM-International Engineering Journal, 72, 321–328.
Panwar, K., & Deep, K. (2021). Discrete Grey Wolf Optimizer for symmetric travelling Stentz, A. (1997). Optimal and efficient path planning for partially known environments
salesman problem. Applied Soft Computing, 105, Article 107298. [M]//Intelligent unmanned ground vehicles (pp. 203–220). Boston, MA: Springer.
Park, J., Karumanchi, S., & Iagnemma, K. (2015). Homotopy-based divide-and-conquer Stenzel, J., Lünsch, D., & Schmitz, L. (2021). Automated topology creation for global
strategy for optimal trajectory planning via mixed-integer programming. IEEE path planning of large AGV fleets. In 2021 IEEE International Intelligent Transportation
Transactions on Robotics, 31(5), 1101–1115. Systems Conference (ITSC) (pp. 3373–3380). IEEE.
Passino, K. M. (2002). Biomimicry of bacterial foraging for distributed optimization and Stützle, T., & Hoos, H. H. (2000). MAX–MIN ant system. Future generation computer
control. IEEE control systems magazine, 22(3), 52–67. systems, 16(8), 889–914.
Patle, B. K., Pandey, A., Jagadeesh, A., et al. (2018). Path planning in uncertain Sun, C. Y. (2021). Research on Trajectory Planning and Path Tracking Control Strategy
environment by using firefly algorithm. Defence technology, 14(6), 691–701. for Intelligent Vehicle Emergency Collision Avoidance[D]. Beijing Jiaotong University.
Patle, B. K., Ganesh, B. L., Pandey, A., et al. (2019). A review: On path planning strategies https://doi.org/10.26944/d.cnki.gbfju.2021.000262
for navigation of mobile robot. Defence Technology, 15(04), 582–606. Sun, H. (2017a). Studies on Trajectory Planning Considering Motion Uncertainties of Traffic
Pei, W., Huayu, G., Zheqi, Z., et al. (2019). A novel hybrid firefly algorithm for global Vehicles[D]. Jilin University.
optimization. In 2019 IEEE 4th international conference on computer and Sun, L. (2017b). A real-time collision-free path planning of a rust removal robot using an
communication systems (ICCCS) (pp. 164–168). IEEE. improved neural network. Journal of Shanghai Jiaotong University (Science), 22(5),
Phung, M. D., & Ha, Q. P. (2021). Safety-enhanced UAV path planning with spherical 633–640.
vector-based particle swarm optimization. Applied Soft Computing, 107, Article Su Z., Zhou X., Cheng T., et al., Global localization of a mobile robot using lidar and
107376. visual features. 2017 IEEE International Conference on Robotics and Biomimetics
Preparata F.P., Shamos M.I., Zhuang X.G., Introduction to computational geometry[J]. (ROBIO). IEEE, 2017c: 2377-2383.
1990. Szczepanski, R., & Tarczewski, T. (2021). Global path planning for mobile robot based on
Qi, J., Yang, H., & Sun, H. (2020). MOD-RRT*: A sampling-based algorithm for robot Artificial Bee Colony and Dijkstra’s algorithms. In 2021 IEEE 19th International Power
path planning in dynamic environment. IEEE Transactions on Industrial Electronics, 68 Electronics and Motion Control Conference (PEMC) (pp. 724–730). IEEE.
(8), 7244–7251. Szczepanski, R., Tarczewski, T., & Erwinski, K. (2022). Energy Efficient Local Path
Qian, K., Liu, Y., Tian, L., et al. (2020). Robot path planning optimization method based Planning Algorithm Based on Predictive Artificial Potential Field. IEEE Access, 10,
on heuristic multi-directional rapidly-exploring tree. Computers & Electrical 39729–39742.
Engineering, 85, Article 106688. Tan, C. S., Mohd-Mokhtar, R., & Arshad, M. R. (2021). A Comprehensive Review of
Qiuyun, T., Hongyan, S., Hengwei, G., et al. (2021). Improved Particle Swarm Coverage Path Planning in Robotics using Classical and Heuristic Algorithms. IEEE.
Optimization Algorithm for AGV Path Planning. IEEE Access, 9, 33522–33532. Access.
Rahmaniar, W., & Rakhmania, A. E. (2022). Mobile Robot Path Planning in a Trajectory Tang, H. W. (2019). Research on Multi-Robot Target Searching Based on Intelligent
with Multiple Obstacles Using Genetic Algorithms. Journal of Robotics and Control Optimization Algorithms in Unknown Environments[D]. Hunan university. https://
(JRC), 3(1), 1–7. doi.org/10.27135/d.cnki.ghudu.2019.004690
Ravankar, A. A., Ravankar, A., Emaru, T., et al. (2020). HPPRM:Hybrid potential based Tang, J., Chen, G., & Coon, J. P. (2018). Route selection based on connectivity-delay-
probabilistic roadmap algorithm for improved dynamic path planning of mobile trust in public safety networks. IEEE Systems Journal, 13(2), 1558–1567.
robots. IEEE Access, 8, 221743–221776.

28
L. Liu et al. Expert Systems With Applications 227 (2023) 120254

Tao, Y., Gao, H., Ren, F., et al. (2021). A Mobile Service Robot Global Path Planning Yan S., Research on path planning of auv based on improved ant colony algorithm. 2021
Method Based on Ant Colony Optimization and Fuzzy Control. Applied Sciences, 11 4th International Conference on Artificial Intelligence and Big Data (ICAIBD). IEEE,
(8), 3605. 2021: 121-124.
Tao, Y., Wen, Y., Gao, H., et al. (2022). A Path-Planning Method for Wall Surface X.H. Yan Y.L. Zhu Z.C. Zhang et al. A Review of Bacterial Optimization and Its
Inspection Robot Based on Improved Genetic Algorithm. Electronics, 11(8), 1192. Applications[J] Journal of Zhengzhou University (Engineering Science) 39 05 2018
Thrun, S. (1998). Learning metric-topological maps for indoor mobile robot navigation. pp. 1–10+21 10.13705/j.issn.1671-6833.2018.05.010.
Artificial Intelligence, 99(1), 21–71. Yang, X. S., & Deb, S. (2009). Cuckoo search via Lévy flights. In 2009 World congress on
Thrun, S. (2002). Probabilistic robotics. Communications of the ACM, 45(3), 52–57. nature & biologically inspired computing (NaBIC) (pp. 210–214). IEEE.
Tsardoulias, E. G., Iliakopoulou, A., Kargakos, A., et al. (2016). A Review of Global Path Yang, X. S. (2010). Nature-inspired metaheuristic algorithms. Luniver press.
Planning Methods for Occupancy Grid Maps Regardless of Obstacle Density. Journal Yeap, W. K., & Jefferies, M. E. (1999). Computing a representation of the local
of Intelligent & Robotic Systems, 84(1–4), 829–858. environment. Artificial Intelligence, 107(2), 265–301.
Wang, D., Chen, S., Zhang, Y., et al. (2021a). Path planning of mobile robot in dynamic Yi, X., Zhu, A., Yang, S. X., et al. (2016). A bio-inspired approach to task assignment of
environment: Fuzzy artificial potential field and extensible neural network. Artificial swarm robots in 3-D dynamic environments. IEEE transactions on cybernetics, 47(4),
Life and Robotics, 26(1), 129–139. 974–983.
Wang, H., Wang, W., Sun, H., et al. (2016). Firefly algorithm with random attraction. Yu J., Ge X.X., Optimal route planning of parking lot based on Dijkstra algorithm. 2017
International Journal of Bio-Inspired Computation, 8(1), 33–41. International Conference on Robots & Intelligent System (ICRIS). IEEE, 2017: 221-
Wang, H., Cui, Z., Sun, H., et al. (2017a). Randomly attracted firefly algorithm with 224.
neighborhood search and dynamic parameter adjustment mechanism. Soft Yu, L., Kong, D., Shao, X., et al. (2018). A path planning and navigation control system
Computing, 21(18), 5325–5339. design for driverless electric bus. IEEE Access, 6, 53960–53975.
Wang, H., Zhou, X., Sun, H., et al. (2017b). Firefly algorithm with adaptive control Yu, S., Zhu, S., & Zhou, X. (2016). An Improved Firefly Algorithm Based on Nonlinear
parameters. Soft computing, 21(17), 5091–5102. Time-varying Step-size. International Journal of Hybrid Information Technology, 9(7),
Wang, H., Li, G., Hou, J., et al. (2022a). A Path Planning Method for Underground 397–410.
Intelligent Vehicles Based on an Improved RRT* Algorithm. Electronics, 11(3), 294. Yu, W., Liu, J., & Zhou, J. (2021). A Novel Sparrow Particle Swarm Algorithm (SPSA) for
Wang, N., Xu, H., Li, C., et al. (2021b). Hierarchical path planning of unmanned surface Unmanned Aerial Vehicle Path Planning. Scientific Programming, 2021.
vehicles: A fuzzy artificial potential field approach. International Journal of Fuzzy Yun, X., & Tan, K. C. (1997). A wall-following method for escaping local minima in
Systems, 23(6), 1797–1808. potential field based motion planning. In 1997 8th International Conference on
Wang, S. B. (2020). Research on Interaction-Aware Predictive Trajectory Planning for Advanced Robotics. Proceedings. ICAR’97 (pp. 421–426). IEEE.
Autonomous Vehicle in Dynamic Environments[D]. University of science and Zadeth, L. A. (1965). Fuzzy sets. Information and control, 8(3), 338–353.
technology of China. https://doi.org/10.27517/d.cnki.gzkju.2020.001752 Zagradjanin, N., Rodic, A., Pamucar, D., et al. (2021). Cloud-based multi-robot path
Wang, W., Zhao, J., Li, Z., et al. (2021c). Smooth Path Planning of Mobile Robot Based on planning in complex and crowded environment using fuzzy logic and online
Improved Ant Colony Algorithm. Journal of Robotics. learning. Information Technology and Control, 50(2), 357–374.
Wang, X., Kou, X., Huang, J., et al. (2021d). A Collision Avoidance Method for Intelligent Zhang, C. (2013). Path tracking of a mobile robot using inertial measurement unit.
Ship Based on the Improved Bacterial Foraging Optimization Algorithm. Journal of Control Theory & Applications, 30(03), 398–403.
Robotics, 2021. Zhang, H., Lin, W., & Chen, A. (2018). Path planning for the mobile robot: A review.
Wang, X., Zhang, H., Liu, S., et al. (2022b). Path planning of scenic spots based on Symmetry, 10(10), 450.
improved A* algorithm. Scientific Reports, 12(1), 1–7. Zhang, H., Sun, C. L., Zheng, Z. J., et al. (2015). A modified dynamic window approach to
Wang Y.F., Cao X.H., Guo X., Warehouse AGV path planning method based on improved obstacle avoidance combined with fuzzy logic[C]. International Symposium on
A* algorithm and system short-term state prediction[J/OL]. Computer Integrated Distributed Computing and Applications for Business. Engineering and Science,
Manufacturing Systems:1-22[2022c-03-20]. 523–526.
Wang, Z., Li, H., & Zhang, X. (2019). Construction waste recycling robot for nails and Zhang J., Singh S., Visual-lidar odometry and mapping: Low-drift, robust, and fast. 2015
screws: Computer vision technology and neural network approach. Automation in IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2015:
Construction, 97, 220–228. 2174-2181.
Wang, Z., Peng, J., & Ding, S. (2022d). A Bio-inspired trajectory planning method for Zhang, L., Zhang, Y., & Li, Y. (2020a). Mobile robot path planning based on improved
robotic manipulators based on improved bacteria foraging optimization algorithm localized particle swarm optimization. IEEE Sensors Journal, 21(5), 6962–6972.
and tau theory. Mathematical Biosciences and Engineering, 19(1), 643–662. Zhang, L., Wang, Y., Zhao, X., et al. (2021a). Time-optimal trajectory planning of serial
Wei K., Gao Y., Zhang W., et al., A modified Dijkstra’s algorithm for solving the problem manipulator based on adaptive cuckoo search algorithm. Journal of Mechanical
of finding the maximum load path. 2019 IEEE 2nd International Conference on Science and Technology, 35(7), 3171–3181.
Information and Computer Technologies (ICICT). IEEE, 2019: 10-13. R.K. Zhang Research on the Automatic Overtaking Control Method Based onthe
B. Wu Research on obstacle avoidance of multi-agent based on bionic optimization Improved Artificial Potential Field[D].Dalian 2021 University of Technology
algorithm[D].Taiyuan 2021 University of Science and Technology 10.27721/d.cnki. 10.26991/d.cnki.gdllu.2021.002495.
gyzjc.2021.000260. Zhang, S., Zhou, Y., Li, Z., et al. (2016). Grey wolf optimizer for unmanned combat aerial
Wu, B., Chi, X., Zhao, C., et al. (2022). Dynamic Path Planning for Forklift AGV Based on vehicle path planning. Advances in Engineering Software, 99, 121–136.
Smoothing A* and Improved DWA Hybrid Algorithm. Sensors, 22(18), 7079. Zhang, S. C., Pu, J. X., Si, Y. N., et al. (2020b). Survey on Application of Ant Colony
Wu, C., Fu, X., Pei, J., et al. (2021a). A Novel Sparrow Search Algorithm for the Traveling Algorithm in Path Planning of Mobile Robot.Computer. Engineering and Applications,
Salesman Problem. IEEE Access, 9, 153456–153471. 56(08), 10–19.
Wu, J., Song, C., Ma, J., et al. (2021b). Reinforcement Learning and Particle Swarm Zhang, S., Pu, J., & Si, Y. (2021c). An adaptive improved ant colony system based on
Optimization Supporting Real-Time Rescue Assignments for Multiple Autonomous population information entropy for path planning of mobile robot. IEEE Access, 9,
Underwater Vehicles. IEEE Transactions on Intelligent Transportation Systems. 24933–24945.
Wu, Y., Wu, S., & Hu, X. (2021c). Multi-constrained cooperative path planning of S.C. Zhang Z. Ming A VNF-aware Virtualization Layer Constructing Algorithm Based on
multiple drones for persistent surveillance in urban environments. Complex & Adjustable Hop Count[J] Journal of Zhengzhou University(Engineering Science)
Intelligent Systems, 7(3), 1633–1647. 2021, 42(05), pp. 32–36+55 10.13705/j.issn.1671-6833.2021.05.008.
Xie, Y. N. (2016). The research for the mobile robot path planning algorithm[D]. Xi ’an Zhang, T. W., Xu, G. H., Zhan, X. S., et al. (2022a). A new hybrid algorithm for path
University of Architecture and Technology. planning of mobile robot. The Journal of Supercomputing, 78(3), 4158–4181.
Xue, J., & Shen, B. (2020). A novel swarm intelligence optimization approach: Sparrow Zhang, W., Zhang, S., Wu, F., et al. (2021e). Path planning of UAV based on improved
search algorithm. Systems Science & Control Engineering, 8(1), 22–34. adaptive grey wolf optimization algorithm. IEEE Access, 9, 89400–89411.
Xu J., Han F.Q., Liu Q.X., et al., Bioinformation Heuristic Genetic Algorithm for Solving Zhang, X., Wang, J., Fang, Y., et al. (2018a). Multilevel humanlike motion planning for
TSP[J/OL]. Journal of System Simulation: 1-10[2021-09-22]. mobile robots in complex indoor environments. IEEE Transactions on Automation
Xue, T., Li, L., Shuang, L., et al. (2021). Path planning of mobile robot based on improved Science and Engineering, 16(3), 1244–1258.
ant colony algorithm for logistics. Mathematical Biosciences and Engineering, 18(4), Zhang, X. F., & Wang, X. Y. (2019). Comprehensive Review of Grey Wolf Optimization
3034–3045. Algorithm. COMPUTER SCIENCE, 46(03), 30–38.
Xu, W., Yang, Y., Yu, L. T., et al. (2022). A global path planning algorithm based on Y. Zhang Research on Task Execution-Oriented Environment Modeling Problem for
improved RRT*. Control and Decision, 37(04), 829–838. https://doi.org/10.13195/j. Horae Service Robots[D].Shandong 2021 University 10.27272/d.cnki.
kzyjc.2020.1354 gshdu.2021.000319.
Yang C., Tang Y., Zhou L., et al., Complete coverage path planning based on bioinspired Zhang, Z., Guo, Q., Chen, J., et al. (2018b). Collision-free route planning for multiple
neural network and pedestrian location prediction. 2018 IEEE 8th Annual AGVs in an automated warehouse based on collision classification. IEEE Access, 6,
International Conference on CYBER Technology in Automation, Control, and 26022–26035.
Intelligent Systems (CYBER). IEEE, 2018: 528-533. Zhang, Z., He, R., & Yang, K. (2022b). A bioinspired path planning approach for mobile
Yang F., Lee D.H., Keller J., et al., Graph-based topological exploration planning in large- robots based on improved sparrow search algorithm. Advances in Manufacturing, 10
scale 3d environments. 2021 IEEE International Conference on Robotics and (1), 114–130.
Automation (ICRA). IEEE, 2021: 12730-12736. Zhao, C. W. (2020). A Survey on Artificial Neural Networks. Shanxi Electronic Technology,
Yang, G. F., & Cai, Y. G. (2021). A Survey of Hybrid Particle Swarm Optimization 03, 94–96.
Algorithm for Solving Vehicle Routing Problem and Its Variants. Automation & Zhao, Y., Liu, G. L., Tian, G. H., et al. (2017). A Survey of Visual SLAM Based on Deep
Information Engineering, 42(02), 7–13. Learning. ROBOT, 39(06), 889–896. https://doi.org/10.13973/j.cnki.
Yang, L., Fu, L., Li, P., et al. (2022). An Effective Dynamic Path Planning Approach for robot.2017.0889
Mobile Robots based on Ant Colony Fusion Dynamic Windows. Machines, 10(1), 50.

29
L. Liu et al. Expert Systems With Applications 227 (2023) 120254

Zhao, Y., Liu, H., & Gao, K. (2021). An evacuation simulation method based on an Zhu, Y., Zhu, L., Long, Y., et al. (2021). The Wolf pack assignment rule based on ant
improved artificial bee colony algorithm and a social force model. Applied colony algorithm and the path planning of scout ants in complex raster diagram.
Intelligence, 51(1), 100–123. Journal of Physics: Conference Series IOP Publishing, 1732(1), Article 012084.
Zhenyang, X., & Wei, Y. (2022). Mobile robot path planning based on fusion of improved Ziegler, J., Werling, M., & Schroder, J. (2008). Navigating car-like robots in unstructured
A* algorithm and adaptive DWA algorithm. Journal of Physics: Conference Series IOP environments using an obstacle sensitive cost function. In 2008 IEEE intelligent
Publishing, 2330(1), Article 012003. vehicles symposium (pp. 787–791). IEEE.
Zheyi C., Bing X., AGV path planning based on improved artificial potential field method. Zou, S. X., Wang, P., & Han, X. (2019). Improved path planning algorithm based on PRM.
2021 IEEE International Conference on Power Electronics, Computer Applications Modular Machine Tool & Automatic Manufacturing Technique, 1, 1–3.
(ICPECA). IEEE, 2021: 32-37. Zou, X. B. (2005). Researeh on design of control system and environment modeling for a
Zhou, X., Yu, X., & Peng, X. (2018). UAV collision avoidance based on varying cells prototype of mobile robot. Central South University.
strategy. IEEE Transactions on Aerospace and Electronic Systems, 55(4), 1743–1755. Zuo X., Geneva P., Lee W., et al., Lic-fusion: Lidar-inertial-camera odometry. 2019 IEEE/
Zhu, D., & Yang, S. X. (2021). Bio-inspired Neural Network-based Optimal Path Planning RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2019:
for UUVs under the Effect of Ocean Currents. IEEE Transactions on Intelligent Vehicles. 5848-5854.

30