Central Lancashire Online Knowledge (CLoK)

Title Path planning algorithms in the autonomous driving system: A

comprehensive review
Type Article
URL https://clok.uclan.ac.uk/50380/
DOI https://doi.org/10.1016/j.robot.2024.104630
Date 2024
Citation Reda, Mohamed, Onsy, Ahmed, Ghanbari, Ali and Haikal, Amira Y. (2024)
Path planning algorithms in the autonomous driving system: A
comprehensive review. Robotics and Autonomous Systems, 174. ISSN 0921-
8890
Creators Reda, Mohamed, Onsy, Ahmed, Ghanbari, Ali and Haikal, Amira Y.

It is advisable to refer to the publisher’s version if you intend to cite from the work.
https://doi.org/10.1016/j.robot.2024.104630

For information about Research at UCLan please go to http://www.uclan.ac.uk/research/

All outputs in CLoK are protected by Intellectual Property Rights law, including Copyright law.
Copyright, IPR and Moral Rights for the works on this site are retained by the individual authors
and/or other copyright owners. Terms and conditions for use of this material are defined in the
http://clok.uclan.ac.uk/policies/
 Robotics and Autonomous Systems 174 (2024) 104630

Contents lists available at ScienceDirect

Robotics and Autonomous Systems

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

Path planning algorithms in the autonomous driving system: A

comprehensive review
Mohamed Reda a,b ,∗,1 , Ahmed Onsy a ,2 , Ali Ghanbari a , Amira Y. Haikal b ,3
a
School of Engineering, University of Central Lancashire, Preston, PR1 2HE, UK
b
Computers and Control Systems Engineering Department, Faculty of Engineering, Mansoura University, Mansoura, 35516, Egypt

ARTICLE INFO ABSTRACT

Keywords: This comprehensive review focuses on the Autonomous Driving System (ADS), which aims to reduce human
Autonomous driving system errors that are the reason for about 95% of car accidents. The ADS consists of six stages: sensors, perception,
Path planning localization, assessment, path planning, and control. We explain the main state-of-the-art techniques used
Driverless cars
in each stage, analyzing 275 papers, with 162 specifically on path planning due to its complexity, NP-hard
Local planning
optimization nature, and pivotal role in ADS. This paper categorizes path planning techniques into three
Global planning
primary groups: traditional (graph-based, sampling-based, gradient-based, optimization-based, interpolation
curve algorithms), machine and deep learning, and meta-heuristic optimization, detailing their advantages
and drawbacks. Findings show that meta-heuristic optimization methods, representing 23% of our study, are
preferred for being general problem solvers capable of handling complex problems. In addition, they have
faster convergence and reduced risk of local minima. Machine and deep learning techniques, accounting for
25%, are favored for their learning capabilities and fast responses to known scenarios. The trend towards
hybrid algorithms (27%) combines various methods, merging each algorithm’s benefits and overcoming the
other’s drawbacks. Moreover, adaptive parameter tuning is crucial to enhance efficiency, applicability, and
balancing the search capability. This review sheds light on the future of path planning in autonomous driving
systems, helping to tackle current challenges and unlock the full capabilities of autonomous vehicles.

1. Introduction
there has been a surge in research and development in ADS, with
many companies and academic institutions dedicating resources to this
1.1. Background of ADS
area [2].
The history of autonomous driving systems (ADS) dates back nearly Modern autonomous vehicles use a combination of technologies
a hundred years. The early concepts of self-driving cars have evolved like radar, GPS, cameras, and lidar to navigate their surroundings
into today’s sophisticated ADS due to technological advancements. safely. These technologies allow vehicles to perceive and interact with
Significant developments in computing, artificial intelligence, and sen- their environment, making real-time decisions essential for autonomous
sor technology have marked this evolution. Initially, the progress in operation. Over the years, the improvements in these technologies have
ADS could have been faster, constrained by the limitations of early made autonomous cars more reliable and safer [3].
technology. However, the last few decades have seen a rapid transfor- Autonomous vehicles are now seen as a transformative innovation
mation in this field, bringing us closer to the goal of fully autonomous in transportation, potentially greatly enhancing road safety and ef-
vehicles [1]. ficiency. While fully autonomous vehicles are still being developed,
One of the critical turning points in the history of ADS was the the advancements made thus far point to a future where vehicles can
DARPA Challenges in the mid-2000s. These challenges encouraged the operate independently, shaping the future of transportation [4].
collaboration of experts from diverse backgrounds to address complex Around 95% of car accidents are because of human errors based
problems in vehicle automation. This event significantly boosted in- on a recent report by Kent County Council in the UK [5]. Therefore,
terest and investment in autonomous vehicle technology. Since then, intensive research is behind the improvement of the Automated Driving

∗ Corresponding author at: School of Engineering, University of Central Lancashire, Preston, PR1 2HE, UK.
E-mail addresses: mohamed.reda.mu@gmail.com, mohamed.reda@mans.edu.eg, mramohamed@uclan.ac.uk (M. Reda).
1
https://www.scopus.com/authid/detail.uri?authorId=57220204540
2
https://www.scopus.com/authid/detail.uri?authorId=36110045700
3
https://www.scopus.com/authid/detail.uri?authorId=37022897600

https://doi.org/10.1016/j.robot.2024.104630
Received 13 March 2023; Received in revised form 7 December 2023; Accepted 2 January 2024
Available online 12 January 2024
0921-8890/© 2024 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

Fig. 1. Phases AD System.

System (ADS) to reduce accidents and harmful emissions by reducing The real challenges start from level three, where the ADS stages
the path length of trip [6]. should be considered a whole system rather than sub-systems. In level
Recently, the improvements in the artificial intelligence research three, the car can drive itself, but the human driver can only interfere to
and availability of sensors greatly participated in the improvement respond to emergency alerts. It can only work in limited situations, such
in the ADS [7]. Before going through the challenges of the current as motorways. Audi claims to be the first manufacturer to design a level
algorithms, the phases of the ADS will be explained in the following 3 autonomous car on the highways. However, the transition between
section. manual and autonomous driving can increase the risk of collision [11].
Level four can only work in limited environments in which detailed
1.2. Stages of AD systems maps exist but without the need for human driver attention. The car can
start the trip, drive safely to the destination, stop, and park without
The AD system has six stages, as shown in Fig. 1. It starts with the interference from a human driver. Level five is a fully automated
sensors’ hardware layer, which gathers the data from the environment. driving system that can operate under any weather conditions and in
The second stage is the perception stage, where object tracking, object any area without a human driver. According to Toyota research, no
detection, and lane detection tasks are performed. The third stage is company is close to the level five autonomous driving system [7].
called the localization and mapping stage, followed by the assessment Levels four and five are open-challenging problems that still need
stage. The fifth stage is concerned with planning and decision-making.
to be achieved. Moreover, levels two and three have not been safely
The final stage is the hardware control layer in which the control
achieved as well. From July 2021 to May 2022, around 400 level 2
actions, such as steering angles, are considered [7].
ADS car accidents have been reported by the National Highway Traffic
The ADS starts with the hardware sensors layer. The sensors are
Safety Administration (NHTSA), involving 273 T cars. Therefore, op-
the main factors in interacting with the environment in the ADS.
timal performance is essential to achieve these levels, beginning with
The information about the surroundings, including static and dynamic
the perception layer and continuing to the path planning layer. Failure
objects, is gathered to be fed to the next perception layer. The primary
in one layer can lead to an accident. The rain was the reason for the
purpose of the perception stage is to process the data gathered from
Hyundai ADS crash. The ADS of Google crashed into a bus because of
the sensors and extract information that will be useful in the following
a failure to measure the bus speed. A driver was killed in a Tesla car
stages for the ADS [8].
The third stage is the localization and mapping layer. The main accident because the car could not recognize a white truck.
objective of the mapping and localization is to get the vehicle’s position The research is still working on the challenges in each stage of the
described in the reference frame in the environment [9]. The fourth ADS. Most stages in the AD system are well-researched. However, the
layer is the assessment layer. Assessment is concerned with the overall most challenging problem in the AD system is the path planning stage
risk estimation and predicting the intentions of the surrounding human with avoiding obstacles [7]. Therefore, the focus of this study will be
drivers to avoid accidents. on path planning.
The fifth layer is the path planning and decision-making layer. It is The contributions of this review paper can be listed as follows:
concerned with getting the shortest path, free of collisions, in real-time
for the vehicle between the start point and the endpoint [8]. The last 1. New and comprehensive taxonomy of the path planning algo-
layer concerns the vehicle control actions used to perform the path in rithms.
real such as torque, acceleration, steering wheel angle, etc. [10]. 2. The stages of the ADS will be discussed in detail, starting from
sensors to the decision-making layer.
1.3. Challenges and levels of AD systems 3. The recent state-of-the-art techniques to address the path plan-
ning problem in the last five years have been discussed inten-
There are five levels of the ADS, as shown in Fig. 2. Level zero is a sively.
fully manual driving system where human drivers make all decisions. 4. The main concept of each path planning algorithm is explained,
Level one uses basic systems to assist the human driver, such as the including the pros and cons of recent research; they are also
Anti-lock Braking System (ABS). Level two is a hands-off system, called compared in a tabular manner.
a partially automated system, in which advanced assistance is provided 5. Statistics are conducted on the references used in the path
to help the human driver, such as collision avoidance and emergency planning algorithms, which can help to give insight for future
braking systems. research.

2
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

Fig. 2. Levels of Automated driving.

The paper is organized as follows: Section 2 discusses related review Tyagi and Aswathy proposed a paper that explores the open chal-
papers. Section 3 covers sensor technologies in autonomous driving lenges in the smart autonomous vehicle. They did not focus on a specific
systems. Section 4 explores perception techniques, followed by Sec- stage or specific category of algorithms. However, they described the
tion 5, which addresses localization and mapping technologies. Sec- tasks and challenges in the ADS [20]. Vagale et al. introduced a review
tion 6 discusses assessment methods. The path planning phase, a crucial paper that explores the path planning problem for surface vehicles, not
component in ADS, is divided into three main categories: Traditional autonomous cars [21]. Ye et al. proposed a review paper about deep
Techniques, Machine Learning Techniques, and Meta-heuristic Opti- reinforcement learning algorithms that deal with path planning in the
mization Techniques. These categories are detailed, starting with tra- ADS until 2020. It did not explore the other categories of the algorithm,
ditional techniques in Section 7, which includes graph-based methods, and it did not present the other phases of the ADS [22].
sampling-based methods, gradient-based methods, optimization-based Xiao et al. presented a review paper exploring machine learning
methods, and interpolation curve methods. Section 8 discusses machine techniques for the path-planning problem in the ADS. They focused
learning techniques for path planning, and Section 9 delves into meta- only on the machine learning category [23]. Ignatious et al. proposed a
heuristic optimization methods. Section 10 analyzes the path planning review of the sensors in the ADS. It did not focus on the other stages of
literature from a statistical point of view. The paper concludes with a the ADS or the path planning phase [24]. This review paper discusses
summary of the pros and cons of these algorithms. Appendix (Table 11) all the stages of ADS. Moreover, a detailed taxonomy of the recent
lists all the abbreviations used in the paper. techniques that solve the path planning problem in ADS until 2022 is
explored.
2. Related work
3. Layer 1: Sensors
The literature has multiple review papers on path planning and the
autonomous driving system. Gonzalez et al. introduced a review paper The sensors are the first layer of the ADS, which is responsible for
on the motion planning algorithm in the ADS in 2015, but it did not interacting with the environment in the ADS. The information about the
cover the machine learning and meta-heuristic optimization categories. surroundings, including static and dynamic objects, is gathered to be
Moreover, it did not explore the other phases in the ADS [12]. Clauss- fed to the next perception layer. Multiple types of sensors have recently
mann et al. proposed a review of the path-planning algorithms in the been used in the ADS [25].
motorway driving scenario until 2018. Furthermore, They did not cover
the other phases of the ADS [13]. 3.1. Monocular cameras
Aradi proposed a review paper that focuses only on the deep re-
inforcement learning algorithms that deal with path planning in the Monocular Cameras are used to sense color. They are passive sen-
ADS until the year 2019. This paper did not explore the other stages sors; no signals are emitted to obtain the measurements. The cameras
of the ADS [14]. Yoganandhan et al. introduced a paper that explored have different characteristics, such as resolution, frame rate, and lens
ADS’s fundamentals and stages. This paper focused more on the method size. One of the challenges in the cameras is night visibility and cap-
than the related work behind each method. They cited only 42 ref- turing dark shadow spots [26]. This problem is addressed via multiple
erences until 2019 [15]. Hadi et al. proposed a review paper that solutions such as infrared cameras [27]. High Dynamic Range (HDR) is
presents the recent path-planning techniques for underwater vehicles, another challenging problem in cameras. This problem happens when
not autonomous cars [16]. dark and highly illuminated regions exist in the same frame. Recently,
Zhou et al. proposed a review paper exploring the ADS’s path the cameras have HDR mode that can deal with this problem [28].
planning algorithms until 2020. However, this review did not cover the The main drawback of Monocular Cameras is the difficulty of getting
meta-heuristic optimization algorithms, and they did not explore the the depth information. 3D cameras and event cameras can solve this
other stages of the ADS [17]. Puente et al. introduced a review paper problem.
on the algorithms applied to address the motion planning problem in
Unmanned Aerial Vehicles (UAVs), not autonomous cars [18]. Huang 3.2. 3D cameras
and Chen introduced a review paper that explores the path planning
algorithms in the ADS until 2020, but it did not cover machine learning The 3D camera can obtain depth information by three principles.
and meta-heuristic optimization algorithms. Moreover, the other stages The first principle is stereo vision. In the stereo cameras, the image is
of the ADS were not explained [19]. taken by two monocular cameras separated by baseline distance and

3
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

pointing in the same direction [29]. The stereo cameras suffer from detection. Black cars can be invisible to LiDAR because the black color
low-textured pattern problems. absorbs most of the radiation with little reflection. Moreover, weather
The second principle of the 3D camera is structured light cameras. conditions like rain and fog can negatively affect LiDAR. The water
A monocular camera is integrated into a structured light camera with drops of rain and fog can scatter the light beams of LiDAR, which can
an infrared emitting device. It overcomes the low-textured and has fast lead to false readings [40]. LiDAR performance is better than that of
computational time. However, they need calibration with high accuracy cameras but worse than that of radar. Furthermore, the LiDAR was
restricted with a small operative range (less than 20 m) [30]. bigger than radar and cameras.
The third principle of the 3D camera is the time-of-flight camera. A From the above discussion of sensors, each sensor has advantages
set of infrared LEDs emits modulated light; then, the sensor captures the and disadvantages. The state-of-the-art techniques are based on sensor
reflected light. The round-trip time is estimated for each pixel. Then, fusion between multiple sensors instead of using a single sensor because
the distance can be obtained. These cameras have a high refresh rate this helps to eliminate the single point of failure problem, [41]. The
(50 Hz) with depth map creation. However, they have a short operative sensors’ characteristics can be summarized in Table 1.
range (10–20 m) [31].
4. Layer 2: Perception
3.3. Event camera
The main objective of the perception layer is to receive data from
In the event cameras, each pixel in the frame is triggered indepen- the surrounding environment and extract information that will be
dently and asynchronously when there is a light density change (the useful in the following stages of the ADS. The primary sensor used in
event). A stream of events is grouped to construct a frame-like image. the perception layer is the camera with the aid of computer vision.
This camera has a high dynamic range (up to 120 dB) and fast response Detection is the main task of the perception layer. The tasks of the
time (within microseconds). This type of camera is suitable for dynamic perception layer are discussed in the following subsections.
object detection. Nevertheless, the event cameras have limited image
resolution and pixel size [32].
4.1. Image-based 2D object detection

3.4. Omnidirectional cameras

Object detection recognizes static and dynamic objects like road
signs, traffic lights, and other moving vehicles. Image-based object
These cameras generate a 360-degree panorama vision. This vision
detection is one of the standard methods to detect objects. The main
can be done through camera arrays; each has wide-angle lenses (fish-
idea is to detect if specific objects exist in the captured image. Next,
eye). This array of cameras requires precise calibration between them
the size is identified using a rectangular bounding box. Cameras are
to construct the image. This type of camera is suitable for 3-dimensional
the primary sensors used for the Image-based object detection process.
vision [33] and is used in SLAM applications [34]. However, the com-
Deep Convolutional Neural Network (DCNN) is the most well-
plex calibration process can generate highly distorted spherical images.
known object detection method. It can be done as a single stage to
There are other emerging types of cameras, such as flash cameras [35]
detect a specific object, or it can be done in two stages: one to detect
and thermal cameras [36].
general regions of interest and the second to detect a specific object
in the region. The most common single-stage detector is the You Only
3.5. Radar
Look Once (YOLO) CNN network [42]. It is improved to multiple ver-
sions such as YOLO9000 [43] and YOLOv3 [44]. Single Shot Detector
The main concept of Radar sensors relies on the Frequency-
Modulated Continuous Wave (FMCW) technique. The radar emits high- (SSD) is another single-stage detector based on DCNN and is faster and
frequency electromagnetic radio waves and receives the reflected waves can detect small objects better than YOLO [45]. However, the Region
from the object. The time of the go and return trip of the waves is called Proposal Networks (RPN) is a popular two-stage detector [46]. Single-
round-trip time. The distance is calculated based on the frequency stage detection is faster and consumes less memory, so that it can be
shift between the reflected and emitted waves. The main advantage used in real-time more than two-stage detection.
of radar is good object detection in dark, rainy, foggy, and dusty
weather conditions. Furthermore, it has a high operative range (up to 4.2. Image segmentation
250 m) [37]. Moreover, it considers depth.
Radar has some drawbacks, such as false positive and false negative Image segmentation is classified into two types: Semantic Segmen-
detection. In false positive detection, the radar can amplify the signal tation and instance segmentation. In Semantic Segmentation, a class
of small obstacles that can be ignored, such as cans, and make its size label is assigned to each pixel in the image. In instance segmentation,
bigger than reality. In false negative detection, the radar cannot detect a class label is assigned to each instance in the image instead of
massive objects made from non-reflecting materials such as wood. The classifying pixels based on box boundaries. Instance segmentation is
horizontal resolution and accuracy are considered limitations of the more suitable for autonomous driving tasks to separate objects with
radar. The emitted signals are material-dependent, so it will not be easy different trajectories. However, Semantic Segmentation is better for
to separate between objects with similar reflection characteristics [38]. detecting objects that are difficult to recognize by box bounties, such as
sidewalks, traffic lines, and buildings. The primary sensor used in the
3.6. Light detection and ranging (LiDAR) segmentation task is the camera.
Neural networks are the most commonly used techniques in the
The LiDAR sensor is based on the same working principle as the segmentation process, such as Recurrent Neural Networks (RNN) [46],
radar sensor, but the only difference is that the LiDAR uses low-power Pyramids Networks [47], and sparse convolutions networks [48]. More-
infrared laser light instead of radio waves. A 360-degree rotating mirror over, Transposed Convolutions Networks, based on convolution to
is operated to change the direction of the laser pulse. This mirror helps extract features from images followed by the deconvolutions process,
to construct a 3D point cloud of the surrounding environment. LiDAR are used to get the labels for Pixels [49]. The main merit of segmen-
considers the depth information and has higher accuracy than radar tation networks is that they can learn as a universal feature extractor
and can measure objects within 0.1 meters [39]. that can be generalized for all perception tasks in the ADS. However,
LiDAR sensor has some disadvantages. It has a low vertical reso- the segmentation networks are slow and require high computational
lution. LiDAR has poor dark color detection, leading to false negative time.

4
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

Table 1
Characteristics of sensors in AD system.
Sensor Idea advantages drawbacks
Passive sensors, Can sense color. No depth , dynamic range.
Monocular Cameras
collect lights via the lens. low night visibility
2 monocular cameras separated low-textured
3D stereo camera by distance and point measure depth. patterns problem.
in the same direction.
3D structured monocular camera integrated with fast, measures depth, need high accuracy calibration,
light cameras infrared emitting devices. no low-textured problem. small range (< 20 m).
3D time-of-flight camera integrated with high refresh rate (50 Hz), short range (10–20 m).
camera set of infrared LEDs. generates depth map.
each pixel is triggered independently high dynamic range limited image resolution,
when a light density changes (event), (up to 120 dB), fast
Event Camera
then, A stream of events are grouped response time (within microseconds), limited pixel size.
to construct a frame-like image. good for dynamic object detection.
360 generates 360 panorama vision good for 3-dimensional vision, the calibration process is
Omnidirectional via camera arrays, each camera has used in SLAM applications. difficult. Generate highly distorted
cameras wide angle lenses (fish-eye), spherical images.
which require precise calibration.
Emits radio waves, high range (up to 250 m), false positive and negative detection,
Radar measure frequency shift between works under any weather conditions. low accuracy and horizontal resolution,
reflected and emitted waves. considers depth information. material-dependent reflection.
the same as radar but uses Depth info, higher accuracy than radar, low vertical resolution, big size.
LiDAR
low power infrared Laser light. measure objects within 0.1 meters. affected by weather conditions.

4.3. 3D object detection for real-time object tracking using images obtained from monocular
cameras [64,65].
In 3D object detection, depth information should be considered to
convert 2D objects into 3D objects. This information can be obtained
using multiple techniques like 3D stereo cameras, LiDAR, and Radar 4.5. Road and lane detection
sensors.
Grouping sparse 3D data points obtained from 3D sensors into a 3D The previously discussed object detection techniques are insufficient
object point cloud is challenging for 3D object detection. There are mul- to detect objects with continuous surfaces such as roads and lanes. Road
tiple algorithms and neural networks that can be used in this process detection aims to obtain a drivable region for the autonomous vehicle.
such as euclidean clustering [50], region-growing methods [51], Vox- Semantic segmentation can be used to find the drivable area, but more
elNet grids based on RGB-D for colored point cloud [52,53], VeloFCN is needed to understand road intersections and branches. Therefore, this
network [54]. These algorithms are tested on scenes dataset [55] and
task needs to be discussed separately [66].
KITTI dataset [56] because they provide labeled 3D scenes.
The lane and road detection techniques are performed via several
4.4. Object tracking steps. The first step is the pre-processing of the raw data obtained from
sensors. The most common pre-processing tasks are color correction
The estimation of object location is not sufficient in the ADS. It is for camera images [67], map-based filtering [60] and ground extrac-
essential to detect the dynamic object trajectory and velocity to expect tion [68] for LiDAR data and identifying the dynamic objects to isolate
the object’s future location. This data about moving objects can be gath- static objects such as the road lanes.
ered from cameras, LiDAR, and radar sensors. Using only one sensor for The second step is lane feature extraction. Road estimation and
object tracking is insufficient, but sensor fusion is required for accurate lane markings can be done based on filters, gradient information,
object tracking [57]. Object tracker algorithms are based on three and intensity information. The 3D road structure can be identified
steps: first, generate an occupancy map for all sensors; second, data based on lidars and 3D stereo cameras, machine learning methods, and
association techniques are used to find the association between objects; appearance-based segmentation.
finally, filtering methods are used for smooth dynamic detection. The third step is to construct the continuity of the lanes and
As mentioned, multiple sensors gather accurate data for the object
roads based on geometric interpolation parametric techniques such as
tracking task, called sensor fusion. Therefore, occupancy maps are used
splines [69], non-parametric continuous models [70].
as a single frame for all sensors that gather all data about surrounding
The last step is temporal integration, in which the road tracking
objects [58]. The next step is to use the data association technique
on the occupancy frame to find a relationship between objects. The system is integrated with vehicle dynamics to achieve smooth results.
most common Data association techniques that establish an association Kalman filtering [71], or particle filtering [72] can be used to make the
between objects are nearest neighbor methods, Image-based methods results smoother. The perception tasks can be summarized in Table 2.
combined with gradients and KLT features [59], Point cloud-based
methods integrated with point density and Hausdorff distance [60].
5. Layer 3: Localization and mapping
The third step is to use filtering methods for smooth dynamic detection
such as traditional Bayes filters, Kalman filters [61], particle filters as
non-parametric filter [62], Rao-Blackwellized particle filters [63]. The primary purpose of the localization and mapping is to find
Recently, deep neural networks have been used as a one-step ob- the position of the vehicle relative to the reference frame in the
ject tracker. These networks need to be trained before using it in environment [9]. There are three standard techniques for localization:
real-time applications. Convolution Neural Networks (CNN) are used GPS-IMU fusion, SLAM algorithm, and priori map-based localization.

5
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

Table 2
Perception tasks.
Task Idea Sensors Steps and algorithms
Detect if specific objects exit Single stage:
Image-based 2D in the image, determine Cameras YOLO, SDD,
Object detection its location and size using two-stage: faster.
a rectangular bounding box. RPN. Single stage faster.
Semantic Seg. classifies pixels RNN,
Image in images, instance seg. cameras Pyramids Networks,
segmentation classifies objects based Transposed Convolutions
on box boundaries. Networks, Slow.
depth information should 3D stereo Euclidean clustering ,
3D object be considered to convert the cameras, VeloFCN,
detection 2D objects into LiDAR, VoxelNet network.
3D objects. Radar
Detect the dynamic object 1- occupancy map as a single frame
Object trajectory velocity to predict cameras, for all sensors. 2- data association
tracking the future location of LiDAR, (nearest neighbor, Image, Point cloud based)
the object Radar 3- filtering for smoothing
(Bayes, Kalman, particle filters).
Road Find the drivable 1- data pre-processing (color correction,
and lane region for the autonomous cameras, map-based filtering).
Detection vehicle. lidar, 2-lane feature extraction. 3- construct the
radar continuity of lanes (geometric parametric
and non-parametric alg.). 4- temporal integration
for smoothing (Kalman, particle filters).

5.1. GPS-IMU fusion In the landmark search, signs and road markers are distributed in
the environment and used as landmarks. At the same time, the vehicle
The Inertial Measurement Unit (IMU) sensor consists of some sen- moves, sensors, and road marking detection algorithms are used to
sors such as a gyroscope, magnetometer, and accelerometer. These identify the landmarks and compare them to a predefined map to
sensors can identify the angle and the direction of the vehicle. However, identify the location. This method has low computational time, but it is
each sensor has its errors, so using multiple sensors leads to error landmark-dependent. If the landmarks are insufficient, the ADS is prone
accumulation that can lead to failure over time. Global Positioning to fail [75].
System (GPS) sensor provides the coordinates values: the latitude and The second map-based method is point cloud matching. An online
longitude, which can be represented in the X and Y axes of the earth. small point cloud map obtained from a sensor, such as LiDAR, will be
The fusion between GPS and IMU sensors is used to correct the accumu- compared to an offline large priori point cloud map. The location is de-
lated error in the IMU sensor. Therefore, it can define the location and termined based on the best match between two points on the two maps.
orientation of the object. Nevertheless, the precision of the GPS-IMU is The algorithms used for point cloud matching are normal distributions
not high enough for the ADS, but it can be used as an initial estimation transform (NDT) and iterative closest point (ICP) algorithms [76]. This
for the ADS [73]. method can give a more accurate location than landmark search, but it
has a high computational time.
5.2. Simultaneous Localization and Mapping (SLAM) The main demerit of the map-based technique is the need to build a
map before using the algorithm. Moreover, any environmental change
SLAM technique is based on building an online map for an unknown will lead to misleading localization, and the maps should be updated to
environment and obtaining the vehicle’s location simultaneously. There adapt to the environmental change [77]. The localization and mapping
are two main types of SLAMs: visual SLAM (vSLAM) and LiDAR SLAM. techniques are summarized in Table 3.
Visual SLAM builds the map based only on images collected by cameras
(monocular, stereo, omnidirectional). It is cheap and can represent 6. Layer 4: Assessment
landmarks, but no depth information exists. There are two types of
vSLAM algorithms: Dense algorithms, which rely on the overall image Assessment is concerned with the overall risk estimation and pre-
brightness such as DTAM, LSD-SLAM, DSO, and SVO; Sparse algo- dicting the intentions of the surrounding human drivers to avoid acci-
rithms, which are based on some feature points in images such as PTAM dents. The assessment has three main types as follows:
and ORB-SLAM [74].
LiDAR SLAM collects information from the LiDAR sensor to con- • uncertainty and risk assessment.
struct point cloud maps. It has high-accuracy distance measurements. • assessment of the behavior of surrounding human driving.
However, the matching process needs to be detailed with more infor- • identifying driving style.
mation. In general, SLAM does not need any information about the
environment. Therefore, it can be deployed anywhere. However, it has Uncertainty and risk assessment is concerned with monitoring the
high computational time, especially in outdoor applications [74]. risk of the overall driving scene. It can detect unsafe lane change events
and changes in the road. This process can be done using sensors such
5.3. Localization using priori map as cameras and detection algorithms such as neural networks.
Surrounding human driving behavior assessment is concerned with
Localization is obtained by comparing online information, such as understanding the intention of the surrounding human drivers by pre-
location, from a GPS sensor and a pre-built map. The exact location is dicting the future behavior of their vehicle to avoid accidents. Pre-
determined based on the best match between them. Map-based methods dicting the dangerous cuts and overtaking performed by surrounding
consist of two main types: point cloud matching methods and landmark drivers are examples of this task. Human traits should be taken into
search. consideration to judge the overall driving scenes. The main challenge

6
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

Table 3
Localization methods.
Method Idea Sensors Pros. Cons.
GPS for location, GPS sensor, Low computational time, Low accuracy.
IMU for direction, IMU (gyroscope, small size, low cost,
GPS-IMU
fusion to reduce magnetometer give quick
accumulated error. accelerometer) initial position.
Visual Build online map using camera images Camera Cheap, Lack of depth,
SLAM and determine the location at the same time. represents landmarks, high computational time.
can work anywhere.
LiDAR SLAM Build online map using lidar and LiDAR High accuracy distance, Not a fine map,
SLAM determine location at the same time. can work anywhere. large size, very slow.
Landmark Sensor detect landmark (signs, markers), Marker-based Moderate speed Landmark dependent,
mab compared to pre-defended map i.e. camera high accuracy prone to fail,
based to match the location. need a priori updated map.
Point cloud Online small cloud map is compared to GPS, Best stability Time consuming
matching offline large priori cloud map, LiDAR best accuracy largest size,
mab-based location is obtained by ICP or NDT. need a priori updated map.

in this task is the short time available to sense the surrounding vehicles The main demerit of the conventional 𝐴 ∗ method is the high com-
in real-time, which needs to be improved to predict long-term behavior. plexity. Nevertheless, Song et al. [85] introduced a hybrid algorithm of
Driving style recognition is the most common task in the assessment the 𝐴 ∗ and PSO algorithms. The new algorithm managed to enhance
phase. The driving style can be defined based on the human driver’s ag- the complexity of the paths. However, it has a high computational time.
gressiveness or the surrounding vehicle’s relative speed [78]. Clustering Yeong et al. proposed a hybrid algorithm between the Predictive-
unsupervised learning algorithms, such as PCA algorithm and K-means, Dynamic Window Approach (P-DWA) algorithm and the A* algorithm,
and supervised learning methods, such as neural networks and SVM, called A*-PDWA. The A* algorithm is used to obtain an approximate
are the most common methods used to classify the driving styles [79]. path. Then, a cardinal spline interpolation technique is applied to
smooth the obtained path. After that, the P-DWA algorithm is used to
7. Layer 5 - part 1: Path planning (traditional techniques) avoid obstacles in the local path and follow the points obtained from
the global trajectory. The algorithm managed to find an acceptable path
The path planning stage aims to obtain a safe trajectory for the without collision. However, the algorithm is not compared to any other
vehicle. This process can be done through two phases. First, the global algorithm [86].
path is generated between the start point and the destination based on Zhong et al. designed a novel hybrid algorithm between the A* and
GPS localization and offline map. The next step is to get an obstacle-free Adaptive Window Approach (A*-DWA) algorithms. The A* algorithm
local path that executes the global path without collision. The challenge generates the rough path. Next, the DWA algorithm is deployed to
achieve real-time trajectory planning with obstacle avoidance. The
of finding the shortest path can be considered as an NP-hard opti-
practical and simulation results indicated that the A*-DWA algorithm
mization problem. Therefore, finding the optimal path exponentially
can achieve a feasible path in complex dynamic environments. Nev-
grows with the number of available nodes. The research is still trying
ertheless, the algorithm’s parameters must be optimized as they were
to find an optimal solution for the NP-hard problem via enhancing
assumed by experience. Moreover, the algorithm is not compared to
the performance of the current algorithms [80]. Obtaining the optimal
other techniques [87].
shortest collision-free path is still challenging in the ADS [7].
Maw et al. introduced the improved Anytime Dynamic A*(iADA*)
Most stages in the AD system are well-researched. However, the
algorithm for a dynamic environment. The concept of the iADA* algo-
most challenging problem in the AD system is the path planning stage
rithm is to find an initial path to allow the vehicle to begin movement.
with avoiding obstacles [7]. Therefore, the focus of this research is on
Then, the path is optimized for a short path during the vehicle’s
path planning.
movement. If the vehicle faces an obstacle, the algorithm updates
the path to get a new collision-free path. The results proved that the
7.1. Graph-based methods for path planning in ADS iADA* algorithm is faster than ADA*, D* Lite, MPGAA*, and D* Extra
Lite algorithms. However, the algorithm gave longer paths in terms of
Dijkstra [81] and A* [82] are the most well-known techniques distance, and more scenarios need to be validated [88].
for the graph-based path-planning problem. These algorithms always Thoresen et al. proposed a Traversability Hybrid A* (THA*) algo-
give discontinuous paths instead of continuous ones, leading to jerky rithm, which uses estimated traversability to optimize the distance of
paths [12]. the path for the car. The THA* algorithm outperformed the original
Yijing et al. proposed a novel A* algorithm with an Equal-Step Sam- hybrid A* algorithm over distances of up to 270 m in the experiment.
pling (A*ESS) algorithm to address local path planning. They designed Nevertheless, the computational time of the algorithm is high, which is
an enhanced reward function based on the kinematics model of the car sometimes too short for a planning horizon [89].
to improve the path comfort. Results proved that the A*ESS algorithm Zhu et al. developed a reverse labeling Dijkstra algorithm (RLDA),
outperformed the traditional A* algorithm with the central search node which relies on the basic Dijkstra algorithm but with reverse labeling.
of each lattice cell. However, the statistical analysis is poor, and the The RLDA algorithm is compared with ACO, GA, PSO, NNa, and
algorithm is not evaluated on benchmark problems [83]. OPABRL algorithms. The results indicated that the RLDA algorithm has
Udomsil et al. applied the A* algorithm in solving motion planning a faster convergence rate than ACO, NNA, and GA. The main drawback
in a static environment. A* is used to generate the path. Moreover, is that the algorithm has a high computational time when the problem
the A* is also used to avoid obstacles by generating a new trajectory nodes are more significant than 350 [90].
through the Unity 3D collision detection system. The algorithm man- Liu et al. applied a two-level algorithm to solve the path-planning
aged to obtain a path in a simple environment. Nevertheless, it is not problem in a smart car. The DWA algorithm is deployed for local
compared to any other algorithm to measure its relative performance. path planning, whereas the Dijkstra algorithm is used for global path
Furthermore, the obstacles were too simple and static [84]. planning. This algorithm managed successfully to avoid obstacles from

7
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

Table 4
Literature Review Summary for graph-based techniques.
Algorithm Idea Pros Cons
enhanced cost function based outperformed traditional poor statistics,
A*ESS [83] on vehicle kinematics model, A* algorithm not tested on
Equal-Step Sampling. with search node. benchmark functions.
A* finds path, and avoids obstacles managed to find path in No comparison,
A*-based [84]
in Unity 3D collision detection system. simple static environment. too simple obstacles.
A*-PSO [85] Hybrid A* with PSO better than A* high computational time
Hybrid; A* gets global path, cardinal spline managed to find safe path. algorithm is not compared
A*-PDWA [86]
smooths path, P-DWA avoids obstacles. managed to find a safe path. with any other algorithm.
A* finds the global path, achieved feasible path in complex parameters are not optimized,
A*-DWA [87]
DWA avoids the obstacle. dynamic environments, applied in real-time. no comparison.
finds initial path to start movement, faster than ADA*, D* Lite, longer paths,
iADA* [88]
path is optimized to be shorter during movement. MPGAA*, D* Extra Lite. needs more scenarios.
Hybrid A* that estimates outperformed the original High computational time that
THA* [89] terrain traversability. hybrid A* algorithm, 270 m leads sometimes to
real experiment distance. a short planning horizon.
RLDA [90] Dijkstra algorithm with reverse labeling Fast solutions for simple search space Slow in complex search space
Dijkstra algorithm with Successful No comparative and
D-DWA [91]
dynamic window approach implementation statistical analysis
A* evaluation function outperformed the traditional High complexity and
IA*FC [92] includes fuel consumption A* by 16.949%. computational time, not
in idle state. accurate in actual traffic.
Long-term: A* finds path, JPS utilizes path, obtained safe path, avg. time 31s, no comparison,
LTSTP [93]
Long-term: PSO optimize path. worst-case time 94s. needs more complex scenarios.
hybrid; A* finds path, Cubic outperformed A*, needs testing on more
IA*-DWA [94] Bezier curves smooths path, DWA algorithms. complex environment and
DWA avoids obstacles. benchmark functions.

the starting position to reach the final position. However, this study autonomous driving systems could focus on hybridizing graph search
focused on hardware implementation, but no statistical analysis is algorithms with interpolation techniques to make the path smoother.
used to compare the performance of the algorithm with the current On the other hand, dealing with continuous and complex search spaces
literature [91] requires hybridizing them with other path-planning methods, such as
Liu and Zhang proposed an improved A* based on the fuel con- the PSO algorithm, which adds the ability to deal with continuous
sumption (IA*FC) algorithm for the idle states in the ADS. The objective search spaces. Another strategy is to hybridize them with sampling
function of the A* algorithm includes fuel consumption, especially in methods that facilitate the representation of the continuous search
idle states in red traffic. Experimental results showed that the IA*FC al- space for the graph-based algorithms.
gorithm outperformed the traditional A* algorithm by 16.949% during
red light traffic. However, the complexity and computational time need 7.2. Sampling-based methods
to be improved. Moreover, the algorithm needs to be more accurate in
the actual traffic environment [92].
The concept of the Sampling-Based Planning (SBP) algorithms is to
Kim et al. designed a Hierarchical Long-Term and Short-Term Plan-
construct connections in the C-space by generating random paths inside
ner (LTSTP) algorithm for highway driving scenarios. This algorithm
it [95]. Rapidly-exploring Random Tree (RRT) [96] is the most common
has two stages: short-term and long-term planning. In the long-term
SBP algorithm. The main idea of the RRT algorithm is to incrementally
planning, the A* algorithm finds a rough 2D path, and then the Jump
build the path between the start and endpoints with random tree-like
Point Search (JPS) algorithm is applied to optimize the path obtained
branches. The paths between two points are guaranteed to be found
from the A* algorithm. In short-term planning, the PSO algorithm is
if given enough run-time. Therefore, it is called a probabilistically
deployed to enhance the path and generate parameters, such as steering
complete algorithm [96]. The main drawback of the SBP are the jerky
angle. The simulation results proved that the LTSTP algorithm obtained
solutions like the Graph-based algorithms [12].
a collision-free path with 31 s average computation time and 94 s
worst-case computation time. However, the algorithm is not evalu- Lim et al. introduced a hybrid algorithm between the Sequential
ated against any other state-of-the-art algorithm to judge the relative Quadratic Programming (SQP) numerical optimization algorithm and
performance. More complex scenarios need to be considered [93]. a sampling-based algorithm named SB-SQB. The sampling-based al-
Li et al. developed a hybrid algorithm between the DWA algo- gorithm is used to get the path using environmental models. The
rithm, the Improved A* algorithm, and Bezier Curves, named IA*-DWA. SQP numerical optimization technique is deployed to obtain the tra-
The purpose of the A* algorithm is to find the path, which is im- jectory based on the path obtained from the sampling-based plan-
proved through the Cubic Bezier interpolation curves to make the path ner. The hybrid SB-SQB algorithm managed to find reasonable paths
smoother. The use of the DWA algorithm avoids moving obstacles. within 50 ms. Nevertheless, the SB-SQB algorithm is not compared
The results showed that the IA*-DWA algorithm outperformed the A* to other algorithms. Furthermore, it must be tested on more complex
and DWA algorithms. Nevertheless, the algorithm needs to be tested scenarios [97].
on a more complex environment and benchmark functions to test the Wang et al. proposed a hybrid algorithm between the Rapidly-
performance [94]. Graph-based methods are summarized in Table 4. exploring Random Tree algorithm and the DWA algorithm (RRT-DWA).
Graph-based methods are mainly used for discrete search spaces The RRT algorithm finds the global path, while the DWA algorithm
that a graph can represent. Therefore, it cannot deal with complex estimates rotational and translational velocity. Results indicated that
scenarios with continuous search spaces. Moreover, the path generated the RRT-DWA method is faster and smoother than the Dijkstra-DWA
from them is jerky. Future research in graph-based algorithms for algorithm and A*-DWA algorithm. However, the dynamic objects were

8
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

not taken into consideration. Furthermore, the statistical analysis was Moreover, inaccuracy was not considered; the statistical analysis was
insufficient, and more scenarios must be tested [98]. insufficient [107].
Varghese and Jisha proposed an Improved Rapidly-exploring Ran- Huang et al. proposed an Improved RRT (i-RRT) algorithm in which
dom Tree associated with a PI controller (IRRT-PI). The IRRT algorithm the i-RRT algorithm finds the path, and the B-spline interpolation
is integrated with the PI controller. Using the control theory, the PI algorithm makes the path smoother. Results proved that the i-RRT algo-
controller eliminates the position and acceleration errors between the rithm can find a personalized collision-free trajectory. Nevertheless, the
actual and generated paths. The algorithm managed to find a path with algorithm is not compared to any other algorithm in literature [108].
acceptable error. However, the algorithm is not compared to any other Jin et al. proposed a Modified RRT* (MRRT*) algorithm, in which
algorithm to estimate its performance [99]. the RRT* algorithm generates the path, and the cubic B-spline in-
Lim et al. designed a hybrid algorithm between the sampling and terpolation algorithm makes the path smoother. The MRRT* algo-
numerical optimization algorithms (S-NO). The sampling-based tech- rithm succeeded in finding safe paths in real. However, no statistical
nique in [100] is used to find the lateral movement for a dynamic analysis is performed with no comparison to recent state-of-the-art
environment. Quadratic programming (QP) is deployed to address the algorithms [109].
convex optimization problem to generate longitudinal movements that Rakita et al. introduced the SPRINT algorithm, a novel sampling-
ensure diversity with no restrictions. The algorithm managed to find a based algorithm. This algorithm is based on minimizing the number
comfortable and safe path practically. However, the algorithm is not of collision checks using three heuristics. The first heuristic is giving
compared to other algorithms in literature [101]. high priority to the promising search regions. The second heuristic
Li et al. introduced an Adaptive Sampling-based Motion Planning is getting samples from regions with a local minimum. The third
with a Non-Conservatively Defensive Strategy (ASMP-NCSS) algorithm. heuristic is to direct the search away from states with previous colli-
The sampling method in the algorithm is based on time-varying dis- sions. The results showed that the algorithm outperformed RRT, RRT*,
tribution combined with a non-conservatively defensive strategy to RRT-Connect, and other sampling-based algorithms. Nevertheless, the
generate safe paths. Results proved that the ASMP-NCSS algorithm is heuristics were constructed by observation. Moreover, the algorithm
better than the traditional sampling-based Motion Planning (SMP) algo- does not guarantee optimal solutions. Furthermore, results need more
rithm with uniform sampling in a dynamic environment. Nevertheless, improvements [110].
more scenarios need to be tested. Furthermore, all the statistics were Wang et al. proposed the Kinematic Constrained Bi-directional
based on the graphical methods, not the numerical ones [102]. Rapidly-exploring Random Tree with Efficient Branch Pruning (KB-
Feraco et al. implemented the basic RRT algorithm to solve the local
RRT*) algorithm. The concept of the algorithm is to avoid unnecessary
trajectory planning for racing ADS. This algorithm managed to find
tree growth by integrating the kinematic constraints. Furthermore,
the paths in an unknown environment that contains obstacles, such as
the branch pruning technique assists the obtained state in getting
traffic cones. A LIDAR sensor is used in the perception layer to sense the
a better parent state and removing edges at high costs. The sim-
surrounding environment. The algorithm is implemented successfully.
ulation results showed that the KB-RRT* algorithm performs better
However, this algorithm is not compared to any other algorithm [103].
than the conventional K-RRT and K-RRT* algorithms. Nevertheless,
Chen et al. developed a two-level algorithm to address complicated
only graphical statistical analysis had been demonstrated to judge
environments with multiple obstacles. The improved Bi-RRT algorithm
the performance. Moreover, the algorithm is model-driven and cannot
implements the high-level part to give an approximate initial path,
predict the promising sampling region [111].
achieving the non-holonomic constraints of the car. The low-level part
Zhang and Wang designed an innovative Path Planning with a
of the algorithm is based on the polynomial Vector Field Histogram-
Line Segment algorithm (LSPP). The LSPP consists of three parts. The
(VFH-) algorithm to improve the trajectory generated from the Bi-
first part is the modified RRT algorithm to find the path. Second, the
RRT algorithm. The algorithm managed to find collision-free paths in
Artificial Potential Field with Azimuth and Distance (ADAPF) algorithm
different driving scenarios. However, it took a long time. Moreover, It
finds the distance to the obstacle and can fully utilize all information
is not compared to any other algorithm [104].
Wang et al. designed a hybrid algorithm between the convolu- about the obstacles. Third, the Modified Dogleg technique integrated
tional neural network (CNN), the A* algorithm, and the RRT algorithm with the Symmetric Rank-1 algorithm guarantees that the car is far
called the neural RRT* (NRRT*). A* algorithm generates the training from obstacles. Real-world test results proved that the LSPP algorithm
dataset that forms the map information. The CNN network estimates outperformed RRT, APF, and PRM algorithms. Nevertheless, the results
the probability distribution of the optimal trajectory obtained by the were only based on graphical statistical analysis without quantities
A* algorithm. Then, this distribution is used to generate the path using criteria [112].
the RRT algorithm via the sampling process. The results showed that Ganesan et al. proposed a Directional RRT* (D-RRT*) algorithm to
the NRRT* algorithm could find better solutions than the ones obtained decrease the sampling space by forming a simple elliptical heuristic,
by the other RRT variants. However, there is no online interaction with focusing on the direction between the start and endpoints. The results
the environment [105]. showed that the D-RRT* algorithm outperforms the traditional RRT*
Zhang et al. implemented an improved Sampling-Based Motion algorithm. Moreover, the TurtleBot3 robot validated the algorithm
Planning algorithm. A novel bias sampling technique speeds up the tra- in real-time. However, the heuristic technique should be enhanced
ditional SBMP algorithm. This technique selects only the most strategi- to minimize the number of visited nodes in a narrow environment.
cally necessary sample points to construct a smooth and safe path. The Furthermore, the algorithm is model-driven, not data-driven [113].
algorithm gave smoother and faster solutions than the ones from tra- Flores et al. developed an improved RRT algorithm with a Custom
ditional sampling-based algorithms. Moreover, the solutions are even Probability Density Function (RRT-CPDF) algorithm. A Custom Proba-
better than the ones from human drivers. However, this algorithm bility Density Function (C-PDF) is used in the RRT algorithm instead
is compared only to sampling-based algorithms (uniform sampling of the Uniform Probability Density Function (U-PDF). The simulation
and bias Gaussian sampling) without considering other path planning results indicated that the RRT-CPDF technique outperformed the tra-
types [106]. ditional RRT algorithm with a Uniform Probability Density Function
Rong et al. introduced an improved RRT* algorithm based on ma- (U-PDF). However, the dynamic obstacles should be taken into consid-
chine learning named Attention-RRT*. The sampling distribution in the eration. Moreover, more complex scenarios should be tested [114].
Attention-RRT* algorithm is generated using a Conditional Variational Huang et al. addressed the path planning problem by proposing an
Encoder based on 3D CNN. The results showed that the Attention- improved RRT (IRRT) algorithm, in which random points are obtained
RRT* outperformed the traditional RRT* algorithm with uniform sam- based on the circular sampling technique to ensure randomness and
pling. However, the construction of the 3D-CNN needs to be enhanced. enhance sampling efficiency. Moreover, they designed an extended

9
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

random point rule to filter the points based on the cost function. potential, target potential, velocity potential, and vehicle potential. The
The B-spline curve interpolation method has been applied to make road and lane potential maps the road structure. Velocity potential
the path smoother. Simulation results proved that the IRRT algorithm prevents unnecessary change in the lane. Vehicle potential ensures
outperformed the traditional RRT and Bi-RRT algorithms. However, a safe distance between obstacle and vehicle. The path is obtained
the Improved RRT algorithm has some limitations in its model-driven through the integration of potential components. The results showed
scheme that need improvements. Furthermore, more scenarios and that the IAPF algorithm has a stable and robust performance. However,
statistical analysis must be done [115]. the algorithm is not validated in real life. Moreover, the algorithm is
Yang and Yao introduced a hybrid algorithm between the Pruning not compared to any other algorithm [121].
Rapidly-exploring Random Tree (PRRT) and the B-Spline Curve, named Lu et al. introduced the Adaptive Potential Field for Path Planning
the PRRT-BSC algorithm. The RRT algorithm finds the path using a (ADPF-PP) algorithm. The dynamic characteristics of the ADPF algo-
pruning method based on obstacle distribution. Then, the B-Spline rithm are used to map the surrounding environment. Then, the ADPF-
interpolation technique is applied to make the path smoother. The PP produces a trajectory free of collisions by minimizing the Potential
PRRT-BSC algorithm found acceptable paths but is not compared to any Field that considers the surrounding obstacles. The results proved that
algorithm in literature [116]. the ADPF-PP outperformed the traditional CPF-PP and Improved Poten-
Zhang et al. introduced an Adaptive Improved RRT algorithm, tial Field-based Path Planning algorithms. Nevertheless, the algorithm
where an adaptive directional sampling technique has been applied falls in local optima in a complex unknown environment [122].
to reduce the random sampling points. A node selection technique Lin et al. designed a Model Predictive Path Planning based on the
has been used for further smoothness of the trajectory. The results APF algorithm, named the MPPP-APF algorithm. They proposed an
proved that the path quality had been improved compared to the other innovative curve-fitting technique integrated with the APF algorithm
RRT variants. However, this algorithm has some disadvantages. The to generate the trajectory. The results showed that the MPPP-APF al-
algorithm takes a long time, especially on curved roads. The algorithm gorithm outperformed the standard APF algorithm. Nevertheless, more
is not applied in real situations. The algorithm is only compared to RRT scenarios need to be validated. They only used graphical statistical
variants.[117] analysis [123].
Lu proposed an Enhanced Rapidly-exploring Random Tree (ERRT) Li et al. developed a hybrid algorithm between Model Predictive
algorithm by expanding the random tree based on the target random Control (MPC) and the APF algorithm considering Driving Style, named
point. The results showed that ERRT is better than traditional RRT. APF-MPC-DS. The APF algorithm is used to model the environment and
Nevertheless, the statistical analysis needs to be more comprehensive. the driving style. Then, the MPC is integrated with APF to optimize
Moreover, the dynamic obstacles are not considered [118]. the path and the motion control outputs. The results demonstrated
Spanogiannopoulos et al. applied the traditional RRT method to that the APF-MPC-DS algorithm is more stable than the APF-MPC
the path planning of self-driving cars in real-time—the proposed algo- algorithm with no driving style proposed in [124]. However, the signif-
rithm used only the point cloud data generated from a local sensor. icance level needs to be discussed; they only used graphical statistical
The algorithm is applied and validated in a well-known benchmark. analysis [125].
Moreover, the algorithm generated safe real-time trajectories of non- Huang et al. proposed an integrated algorithm between the Con-
holonomic autonomous cars in unknown static environments. The paths strained Delaunay Triangulation (CDT) and Artificial Potential Field
are practically generated in roundabouts with obstacles for cars at (APF) Algorithm, named the CDT-APF algorithm. First, the CDT method
speeds of up to 45 km. However, it is not compared to any other has been applied to determine a safe region from the recent location of
algorithm to test its relative speed and accuracy [119]. Sampling-based the car to the endpoint. Then, the APF algorithm defines the boundaries
methods are summarized in Table 5. of this safe region. Then, the Model Predictive Controller (MPC) is
Future directions for improving Sampling-Based Planning (SBP) deployed to obtain the path and utilize the control parameters. The
methods in autonomous driving should focus on enhancing efficiency, results proved that the CDT-APF algorithm outperformed the tradi-
smoothness, and adaptability. Developing hybrids that blend SBP with tional APF algorithm. However, the algorithm has yet to be tested
interpolation approaches, like spline methods, eradicates the sharpness on real-time applications. Moreover, only graphical statistical analy-
of the jerky paths, making them smoother and more feasible, especially sis is performed. Error measurements were not considered [126]. As
in complex driving scenarios. Regarding complex search spaces, ap- previously mentioned before, in the sampling-based algorithms, Zhang
plying bias sampling techniques and machine learning models within and Wang introduced the LSPP algorithm, which is a hybrid algorithm
SBP algorithms can accelerate path planning and refine trajectories for between the modified RRT algorithm and the Artificial Potential Field
complex and dynamic environments. algorithm [112].
Zhang et al. developed a hybrid algorithm between the Improved
7.3. Gradient-based (Artificial Potential Field) methods for path planning Artificial Potential Field (IAPF) and Gradient Descent Method (GDM)
in ADS algorithm, called the IAPF-GDM algorithm. The IAPF algorithm, in-
spired by distance optimization, is used to utilize information about
The Artificial Potential Field (APF) algorithm is a gradient-based obstacles. The GDM algorithm generates the path based on the data
technique in which the vehicle is displayed as a point in a potential from the IAPF algorithm. The IAPF-GDM algorithm outperformed the
field. This point is attracted to the destination endpoint and diverges algorithm introduced in [127]. However, the algorithm assumed an
from the obstacles. The resultant trajectory in the potential field rep- idealistic environment and did not consider future and dynamic obsta-
resents the final path. The main merit of this technique is that it can cles. Moreover, the study was based only on graphical statistics, not
produce collision-free trajectories within a small computation time. numerical ones, to judge the significance level [128].
Nevertheless, the main demerit of the algorithm is that it could get Li et al. proposed the Dynamic Enhanced Firework Algorithm with
stuck into a local minimum, which means there is a better solution, but APF algorithm (DynEFWA-APF), where dynamic environments, such
the algorithm cannot reach it. The reason behind this is the gradient as vehicle dynamics and road structure, are considered. Then, the
behavior of the algorithm. The APF algorithm relies on making the optimization problem is formulated considering all environmental con-
gradient reach a zero slope. When the zero-slope state is reached, straints. Finally, the DynEFWA-APF optimizes the problem to obtain a
the algorithm terminates. Therefore, the algorithm cannot explore new safe path. The algorithm gave an acceptable performance in a static
search space [120]. and dynamic environment. However, the algorithm did not outperform
Hongyu proposed an Improved Artificial Potential Field (IAPF) al- the A*, APF, and GA-APF algorithms. Moreover, the driving style is not
gorithm that consists of five potential components: road potential, lane considered [129].

10
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

Table 5
Literature Review Summary for sampling-based Path Planning techniques.
Algorithm Idea Pros Cons
hybrid; Upper SB: finds the paths based on obtained reasonable paths no comparison to other algorithms,
SB-SQB [97]
environment, Lower SQP optimizes path. within 50 ms. needs more complex scenarios.
Hybrid; RRT finds global path, smoother and faster no dynamic objects,
RRT-DWA [98] DWA calculates velocity. than Dijkstra-DWA , no statistical analysis,
and A*-DWA algorithms. needs more scenarios.
hybrid RRT with PI controller to eliminate managed to find a path No comparison with
IRRT-PI [99]
error between the generated and actual path. with acceptable error. any other algorithm.
hybrid; sampling alg. [100] for lateral path, managed to find a comfortable, No comparison
S-NO [101]
QP for longitudinal path and diversity. safe path in real-time. with other algorithms.
based on time-varying distribution with better than SMP with needs more scenarios,
ASMP-NCSS [102]
non-conservatively defensive strategy. uniform sampling in dynamic. only graphical statistics.
Basic RRT successful in No comparison with
RRT [103]
unknown environment any other algorithm
Two levels: 1-improved Bidir. RRT Managed to find Took long time,
Bi-RRT [104] for Approx. Path. 2- polynomial Vector safe paths. , no comparison with
Field Histogram- planning for acc. path any other algorithm
Hybrid: A* generates training data, Better than RRT No relative comparison
NRRT* [105] CNN predicts probability and RRT* with the other
distribution, RRT generates path. state-of-the-art
improve SBMP by selecting only the Faster than uniform sampling does not consider other
ISBMP [106]
most strategic sample points. and bias Gaussian sampling. AI techniques.
generated sampling distribution outperformed the traditional 3D-CNN needs improvement,
Att-RRT* [107] using CVAE attention mechanism RRT* with poor statistical
based on 3D CNN. uniform sampling. analysis.
i-RRT find the path, B-spline managed to find personalized no comparison with the
i-RRT [108]
interpolation smooths path. safe trajectory. state-of-the-art algorithms.
MRRT* incrementally finds managed to find no statistical
MRRT* [109] the path, cubic B-spline collision-free analysis, no comparison
smooths the path. paths in real. with other algorithms.
minimize no. of collision outperformed the other heuristics based on
SPRINT [110]
checks using three sampling-based observation, no guarantee
heuristics. algorithms. optimal paths, results
need improvements.
kinematic constraints avoids better than model-driven,
tree overgrowth, branch pruning K-RRT and cannot predict good
KB-RRT* [111]
replaces bad edges with K-RRT* algorithms. sampling area, only
better ones. graph statistics.
Hybrid; MRRT finds path, ADAPF utilizes outperformed RRT, only graphical
LSPP [112] obstacles data, MDL-SR1 oAPF, and PRM statistics.
guarantees far obstacles distance. algorithms in real.
reduces sampling space by elliptical outperforms traditional RRT*, model-driven, high number of
D-RRT* [113]
heuristic from start to end. real tested. visited nodes in a narrow environment.
Custom Probability Density Function outperformed RRT with no dynamic obstacles,
RRT-CPDF [114]
is used as a sampling method. Uniform sampling distribution. needs more complex scenarios.
based on circular sampling, random point outperformed RRT model-driven with limitations,
IRRT [115]
filter, B-spline smooths path. and B-RRT algorithms. needs more scenarios and statistics.
hybrid; RRT finds path by pruning fn. based managed to find not compared to any other
PRRT-BSC [116]
on obstacles, B-Spline smooths the path. acceptable paths. state-of-the-art algorithm.
adaptive improved RRT with adaptive Best performance Limitation in planning,
adaptive directional sampling and node compared with RRT slow performance,
AIRRT [117]
node selection to make variants. no on-site data, Only
the path smoother. compared with RRT variants
RRT is enhanced by using the target point as better than traditional RRT. statistical analysis is not
ERRT [118]
a random point for random tree expansion. sufficient, no dynamic obstacles.
basic RRT with only point generate safe No relative comparison
RRT [119] cloud data of nonholonomic cars, trajectories with state-of-the-art
static unknown environments. algorithms.

Lin et al. introduced a hybrid algorithm between Potential Field Szczepanski et al. introduced the predictive APF (PAPF), where a
and Sigmoid-based Safe Passage in the Model Predictive Controller novel local minimum avoidance method, called the top quark-based
algorithm, named the PF-SPMPC algorithm. The PF algorithm is used mechanism, is used to predict and bypass obstacles in advance. The
to model the environment. Then, the MPC controller is used to op-
Husarion ROSbot 2.0 PRO robot has been used to validate the al-
timize the path with sigmoid-based safe passage constraints. The re-
sults proved that the algorithm outperformed the PF-MPC algorithm. gorithm. The results proved that the PAPF algorithm can provide a
However, more complex scenarios need to be tested. Moreover, Only shorter path by up to 8.73% than the ones obtained from the tradi-
graphical statistical analysis is performed [130]. tional APF algorithm. However, the statistical analysis should be more

11
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

Table 6
Literature Review Summary for the gradient-based Path Planning techniques.
Algorithm Idea Pros Cons
5 potential parts(road, lane, target, velocity, achieved stable and not validated in real-life,
IAPF [121]
vehicle), the gradient method finds a path. robust performance. no comparison with other algorithms.
PF maps the environment, then path is outperformed IPF-PP fall in local optima in
ADPF-PP [122]
generated by minimizing the PF. and CPF-PP algorithms. complex unknown environment.
novel curve-fitting integrated outperformed the standard more scenarios need to
MPPP-APF [123] with the APF and MPC controller APF algorithm. to be validated,
to generate a path. only graphical statistics.
hybrid; APF maps environment more stable than only graphical
APF-MPC-DS [125] and driving style, MPC APF-MPC with no statistical
optimizes the path. driving style [124]. analysis.
hybrid; CDT finds safe outperformed traditional only graphical statistics,
CDT-APF [126] path region, APF defines path APF Error was not considered,
boundaries, MPC finds a path. not tested in real-time.
Hybrid; IAPF utilizes outperformed algorithm no future dynamic obstacles,
IAPF-GDM [128] obstacles data, GDM finds in [127]. only graphical statistics,
path based on IAFP data. idealistic environment.
consider dynamic constraints in the acceptable path in static did not outperform A*, APF, GA-APF,
DynEFWA-APF [129]
optimization problem solved by APF. and dynamic environment. driving style not considered.
Hybrid; PF maps the environment, outperformed PF-MPC need more complex
PF-SPMPC [130] MPC optimizes path with algorithm. scenarios, only graphical
sigmoid passage constraints. statistics.
Hybrid; MRRT finds path, ADAPF outperformed RRT, APF, only graphical
LSPP [112] utilizes obstacles data, MDL-SR1 APF, and PRM statistics.
guarantees far obstacles distance. algorithms in real.
top quark-based mechanism enables Real-time test, outperformed insufficient statistical analysis,
PAPF [131]
APF to bypass future obstacles. traditional APF by 8.73%. no dynamic obstacles.
the problem is a quintic polynomial gave acceptable paths. no comparison
PCAPF [132] curve, AFP implements constraints, solved other path planning with any
by fmincon MATLAB toolbox. algorithms.

comprehensive to judge the significance level. Furthermore, the dy- The stochastic optimization problem is an optimization problem
namic obstacles are not considered [131]. that has a random term in the objective function or has a random
Wang et al. designed a hybrid algorithm between the APF algorithm input in the search process. Model Predictive Control (MPC) model
and the Polynomial Curve, named (PCAPF) algorithm. In the PCAPG is an example of a model that deals with random measurements and
algorithm, the path planning problem is formulated as a quintic polyno- inputs and is formulated as a stochastic optimization problem. Dynamic
mial curve optimization problem. Then, the APF algorithm is applied to programming is an approach that is used to address the stochastic
formulate the objective function considering the obstacles. Finally, the optimization problem with unknown model parameters. The path plan-
Fmincon MATLAB toolbox is used to solve the optimization problem. ning problem in the literature can be formulated as a Model Predictive
The results showed that the algorithm gave acceptable paths, but the al-
Control (MPC) stochastic optimization problem, which can be solved
gorithm is not compared with any other path planning algorithm [132].
by dynamic programming [134]. Generally, optimization methods can
Gradient-based methods are summarized in Table 6.
deal with complex search spaces and give high-quality solutions, but
The future direction for Gradient-Based (Artificial Potential Field,
they have high computational times.
APF) methods in autonomous driving systems should focus on overcom-
ing the challenge of local minima problems due to the gradient behav- Viana et al. proposed a Distributed Model Predictive Control (MPC)
ior. Integrating APF with global search and optimization techniques is that implements the Human Driver Model integrated with Mixed-
a suitable way to provide more diversity to the search space and reduce integer quadratic programming (MIQP), named the HDM-MIQP algo-
the local minimum problem. Developing hybrid systems that combine rithm. The concept of the HDM-MIQP algorithm is to integrate the
APF with other path planning methods, like Rapidly-exploring Random Human Driver Model (HDM) into the MPC model. Then, the problem
Trees (RRT), can offer a more comprehensive solution where the other can be assumed to be a mixed-integer quadratic programming (MIQP),
algorithm finds the global path. At the same time, the APF can avoid which can be easily solved by the optimization tool of CPLEX IBM.
the dynamic obstacles in the path. The algorithm managed to find acceptable paths in multiple scenarios.
However, the algorithm is not compared to any other algorithm.
7.4. Optimization-based methods for path planning in ADS In addition, the HDM errors and uncertainties were assumed to be
constant values [135].
Optimization algorithms, or mathematical programming, are meth-
Kanchwala applied the HDM-MIQP algorithm introduced in [135]
ods to find numerical solutions for optimization problems. Optimization
under the same project supported by Innovate UK. He applied different
problems can be formulated as the maximization or minimization
vehicle dynamics to the bicycle model, which is widely used. He
of the objective function. It is called convex programming, where
implemented the Human Driver Model (HDM) using a CarSim driving
the objective function should be minimized, while it is called con-
cave programming, where the objective function should be maximized. simulator. He obtained acceptable results, but HDM uncertainties and
Quadratic programming is convex programming in which the objective errors were assumed to be fixed values. Moreover, the algorithm is not
function contains a quadratic term. Quadratic programming has been compared to any other algorithm [136].
applied in the literature to address the path planning problem in the There are some hybrid algorithms between the sampling-based algo-
ADS as a convex (minimization) optimization problem that needs to rithms and the quadratic programming algorithms mentioned before in
minimize the path length [133]. the sections of the sampling-based algorithm. The SB-SQB algorithm is a

12
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

hybrid algorithm between the sampling-based method and the Sequen- Then, the Dynamic Programming technique is deployed to generate an
tial Quadratic Programming (SQP) [97]. Furthermore, the S-NO algo- initial path. Finally, the quadratic programming optimization problem
rithm is based on the sampling-based algorithm in [100] and quadratic is solved to get the optimal trajectory considering all the driving con-
programming to solve the convex optimization problem [101]. straints. The results proved that the SBDP-QP algorithm can generate
Oliveira et al. applied the Sequential Quadratic Programming (SQP) a smooth collision-free path. However, the algorithm is not compared
optimization algorithm, where the motion planning is implemented as with other algorithms in the literature, and more scenarios need to be
Sequential Quadratic Programming (SQP) optimization solved by the tested [145].
CVX package. The SQP problem is based on the road-aligned car model Typaldos et al. used the Feasible Direction Algorithm integrated
that takes the distortion of the car and obstacles. Moreover, a new with the Dynamic Programming algorithm (FDA-DP). The DP algo-
approximation method is used to capture this distortion. The simulation rithm is used to give an initial guess path. Then, the FDA algorithm
results proved that the SQP-based algorithm performs better than the is used to find a numerical solution to the path planning problem.
one obtained from the algorithm proposed in [137]. However, the Moreover, connected vehicles can exchange information about the last
approximation error was not taken into consideration. The statistical obtained trajectory so far in real-time. The FDA-DP algorithm is used
analysis is insufficient to judge the performance [138]. within the MPC controller framework. The FDA-DP algorithm gave
Zhu et al. proposed a Parameterized Curvature Control (PCC) al-
better results than the one obtained from manually driven vehicles
gorithm where the cubic spline interpolation technique is used as
and the Forward Dynamic Programming (FDP) algorithm. Nevertheless,
a parameterization method. The cubic spline is better than a single
synergistic effects are not considered, and more scenarios need to be
polynomial because the cubic spline can represent all possible curves
validated [146].
with curvature. Then, the splined optimization problem is solved by
Jiang et al. introduced a quadratic programming-based path plan-
the SQP algorithm. The results proved that the PCC algorithm obtained
ner, the QP-SQP algorithm, used in a static, cluttered environment.
suitable safe paths in a dynamic environment. However, more com-
The algorithm has two main stages: first, a collision-free guideline
plex scenarios need to be validated, and the whole algorithm is not
is generated by Quadratic Programming (QP); second, the optimal
compared to any other algorithm in literature [139].
Zhang et al. designed a novel Two-Step Quadratic Programming path is generated with the help of the guideline using the SQP tech-
(TSQP) optimization algorithm to improve the path planning in the nique. The QP and SQP optimization problems were solved by the
ADS. This method consists of two stages: first, a smooth line is obtained Open Source Quadratic Programming (OSQP) solver. The simulation
as a driving guide by a smoothing procedure; second, the optimal results indicated that the QP-SQP algorithm is better than Piecewise
trajectory is obtained by a path optimizer based on piecewise-jerk Jerk Method (PJM) and Minimum Snap Method (MSM). However,
formulation. The TSQP problem is resolved by the Operator Splitting more scenarios should be tested with deep statistical analysis [147].
Quadratic Program (OSQP) tool. The algorithm has been validated in Optimization-based methods are summarized in Table 7.
the Baidu Apollo Platform and is used for a road test on hardware. Optimization-based methods are mainly time-consuming, with high
Nevertheless, more scenarios need to be tested. The algorithm is not computation time methods based on using optimizers to solve the
compared to any other algorithm in literature [140]. optimization problem. Therefore, they are problem dependents, as one
Changhao et al. developed a Dynamic Programming-based algo- optimizer can work with a driving scenario while it fails with another
rithm integrated with Clothoid Curve (DPCC) algorithm. The concept one. Therefore, the future direction for utilizing this method is via
of the algorithm is to obtain future 𝑇 samples and then fit them into a hybridization with other techniques to reduce computational complex-
clothoid curve path by solving it as a dynamic programming problem. ity while improving real-world adaptability. In practice, developing
This algorithm managed to find an acceptable path with acceptable hybrids that combine optimization techniques like Quadratic Program-
performance. However, more scenarios need to be tested. Moreover, ming and Model Predictive Control with sampling-based methods can
the parameters of the algorithm were not optimized. Furthermore, the offer more efficient and practical solutions.
algorithm is not compared to any other algorithm [141].
Li et al. innovated a hybrid algorithm between the Quadratic Pro- 7.5. Interpolation curve methods for path planning in ADS
gramming Optimization-based algorithm and the Search-based algo-
rithm, named the SQPO algorithm. First, Search-based algorithms are The interpolating curve methods are techniques used to generate a
used to find a rough path based on two cases: If the environment is
new path from an existing one to avoid collisions [12]. Initially, we
known, the Dijkstra algorithm generates the path. If the environment
start with a path that already exists from a global planner or another
is unknown, the Space Exploration Guided Heuristic Search (SEHS)
local planner. This path can lead to a collision with an object. Then,
algorithm will be applied to find the path. Second, path planning is
the interpolating curve methods modify the path to avoid this object.
represented as a Quadratic Programming Optimization (QPO) problem,
A new sub-path will be created by fitting a new set of collision-free
which can be resolved by the OSQP tool. The QP algorithm is applied
points between a re-entry point and an exit point on the original path.
to obtain a trajectory with high quality. The results showed that the
The obtained path is collision-free and smooth. Nevertheless, the com-
SQPO algorithm performs better than the algorithm proposed in [142].
putational time is high in comparison with the other methods. The most
Nevertheless, the interaction uncertainty is not taken into considera-
commonly-used curve families are Bezier curves [148], clothoid [149],
tion. Moreover, only graphical statistical analysis was used to judge the
performance [143]. splines [150], and polynomials [151].
Hu et al. introduced the Event-Triggered Model Predictive Adaptive Wang et al. developed a hybrid algorithm between Stackelberg
Dynamic Programming (ET-MPADP) algorithm for road intersection Differential Game Theory and Polynomial Curve (SDGT-PC) algorithm.
driving scenarios. The ET-MPADP algorithm is based on the integration The Stackelberg Differential Game Theory algorithm is deployed to get
between the control policy generation of the MPC controller and the the global path. The polynomial curve method is applied to remove
mismatch of cost function approximation using the critic–actor scheme. the errors in the global trajectory and make the path smoother. The
The ET-MPADP algorithm outperformed the time-triggered MPADP experiment and simulation results proved that the SDGT-PC algorithm
(TT-MPADP) algorithm. However, the structure of regression needs to could obtain feasible paths. Nevertheless, it is not compared to the
be considered. Moreover, the unknown motion of the obstacles is not other algorithms in the literature; only the control algorithms were
considered as well [144]. compared [152].
Wang et al. designed a hybrid algorithm among sampling-based Hu et al. designed a dynamic path planning algorithm using the
algorithm, Dynamic Programming, and Quadratic Programming (SBDP- cubic spline (DPP-CS) interpolation method. The algorithm considered
QP) algorithms. The sampling-based algorithm generates the graph. the dynamic obstacles. However, it considered only one obstacle, and

13
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

Table 7
Literature Review Summary for the optimization-based (quadratic/ dynamic programming) Path Planning techniques.
Algorithm Idea Pros Cons
Human Driver Model is integrated managed to find HDM errors and
HDM-MIQP [135] with MPC, MIQP is acceptable paths in uncertainties are not
solved by CPLEX IBM. multiple scenarios. considered, no comparison.
hybrid; Upper SB: finds the paths based on obtained reasonable paths no comparison to other algorithms,
SB-SQB [97]
environment, Lower SQP optimizes path. within 50 ms. needs more complex scenarios.
hybrid; sampling alg. [100] for lateral path, managed to find a comfortable, No comparison
S-NO [101]
QP for longitudinal path and diversity. safe path in real-time. with other algorithms.
same alg. in [135], HDM simulated obtained acceptable results. HDM uncertainties and errors
HDM-MIQP [136]
by CarSim, vehicle dynamics is bicycle model. were constants, no comparison.
SQP solved by CVX, road-aligned vehicle model better than the approx. error not considered,
SQP-based [138]
with approximator for vehicle–obstacle distortion. algorithm in [137]. poor statistics.
cubic spline is used managed to find more complex scenarios,
PCC [139] as parameterization optimization suitable paths in no comparison.
method, then solved by SQP. a dynamic environment.
1-smooth line is generated as driving guide, released in Baidu Apollo more scenarios need to be tested,
TSQP [140]
2- generates optimal path, solved by OSQP. Open Platform. no comparison with other algorithms.
get future samples, fit them into clothoid path managed to find needs more scenarios, parameters
DPCC [141]
by solving a dynamic programming problem. acceptable path. were not optimized, no comparison.
Dijkstra and SEHS are used to better than the interaction uncertainty
SQPO [143] find rough path, QP solved algorithm in [142]. not considered,
by OSQP optimizes the path. only graph statistics.
hybrid; control policy generation better than TT-MPADP regression structure
ET-MPADP [144] (MPC) and cost function approx. and SPSO2011-GM and obstacles unknown
using a critic–actor scheme of ADP. algorithms. movements are not considered.
Sampling-based gets search graph, DP finds generate smooth not compared with state-of-the-art
SBDP-QP [145]
initial path, QP finds the optimal constrained path. collision-free path. algorithms, need more scenarios.
DP generates initial path, better results than no synergistic effect,
FDA-DP [146] FDA finds numerical optimal path, manually driven cars more scenarios need
data exchange among cars. and FDP algorithm. to be validated.
guide line is generated by QP, then path is faster, better than more scenarios need to be
QP-SQP [147]
generated by SQP, solved by OSQP solver. PJM and MSM. tested with deep statistics.

the algorithm needs to be validated against other algorithms in the Piscini et al. proposed a Clothoid-Based Algorithm (CBA) based on
literature. Moreover, the algorithm is not validated in real life [153]. finding the trajectory given two directions and two points as a 𝐺1 inter-
Klanvcar et al. introduced a hybrid algorithm between the contin- polation problem solved by the clothoid curves. The results indicated
uous Bernstein–Bézier (BB) interpolation technique and the discrete that the CBA algorithm outperformed the Preview-based Algorithm and
grid-based E* search algorithm, named the HE* algorithm. The HE* Pursuit-based algorithm. However, More scenarios should be tested,
algorithm has two stages: first, the E* algorithm is used to calculate and dynamic obstacles are not considered [157].
the direction-guiding heuristics; second, the quality of the search is im- Moreau et al. proposed the Bézier Curve Optimization (BCO) al-
proved in the second stage by a continuous expansion of the fifth-order gorithm as an interpolation approach for the path planning problem.
Bernstein–Bézier (BB) curve. Finally, the Complete Node Mechanism The BCO algorithm considers the traffic and obstacles constraints as a
(CNM) guarantees the complete bounded path. The HE* algorithm constraint Bézier curves optimization problem. The Newton–Raphson
outperformed the HA* algorithm and SST* algorithm. However, there method and the Lagrangian algorithm have been applied to resolve
is no path of re-planning in case of a dynamic environment [154]. this optimization problem. The BCO algorithm managed to obtain a
You et al. introduced two algorithms based on the Bézier curves safe path free of collision. However, the BCO algorithm has not been
compared to the state-of-the-art algorithms, and more scenarios need
algorithm: the first one is called the Joint Quadratic Bézier Curves
to be tested [158].
(JQBC) algorithm; the second one is called the Fourth-Order Bézier
Zambon et al. introduced the B-Spline Curve Optimization (BSCO)
Curves (FOBC) algorithm. The JQBC algorithm uses a smaller search
algorithm to deal with trajectory planning in the ADS. The concept
space to find the path, but it is not easy to track because the path
of the BSCO algorithm is to obtain smooth paths that a combination
is 𝐶 1 continuous. The FOBC algorithm makes it easier to track its
of B-spline linear functions can describe. Furthermore, based on the
paths because it is 𝐶 2 continuous and can generate smoother paths.
minimum distance to obstacles, a smooth penalty function is used
The dynamic cell (DC) concept is introduced to modify the traffic state to convert a constrained optimization problem to an unconstrained
dynamically. The results showed that the algorithms can find smooth, one. The algorithm managed to find smooth paths. Nevertheless, the
acceptable collision-free paths. However, the JQBC algorithm is not algorithm is not compared to other algorithms, and the sub-optimal
compared to the other algorithms in the literature, and more scenarios solution depends on the tuning parameter [159].
need to be validated [155]. Mu et al. developed an improved PSO (IPSO) algorithm to address
Sedighi et al. proposed the Direct Visibility Diagram Algorithm with the three-dimensional route planning problem. A penalty term has been
Clothoid Curves (DVD-CC) algorithm as a path planning algorithm. added to the fitness function, and They made the paths smoother using
The Direct Visibility Diagram (DVD) algorithm is deployed to obtain cubic spline interpolation. Results showed that the IPSO algorithm
the optimal holonomic path. The Clothoid Curves technique has been gave shorter paths by 90% of those obtained from the PSO algorithm.
applied to get a smoother non-holonomic path. The DVD-CC algorithm However, the statistical analysis needed to be more comprehensive to
is faster than standard Visibility Diagram (VD), RRT, and Hybrid A* judge the performance in a complex environment. [160].
(HA*) algorithms by (60%). Nevertheless, more scenarios should be Kamil et al. designed the Adaptive Dimension Limit-Artificial Bee
validated, and the dynamic obstacles were not considered [156]. Colony (ADL-ABC) Algorithm. The concept of the ADL-ABC algorithm

14
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

is to apply an adaptive limit parameter instead of a fixed, stable limit. tentacles. Then, the B-spline algorithm generates a path based on the
The cubic polynomial interpolation technique has been applied to make sampling area of the best tentacle path. Finally, the final trajectory is
the path smoother. The ADL-ABC algorithm outperforms the traditional formed by segments of the best tentacle and the B-spline curve. The
ABC algorithm. However, more complex scenarios should be tested to results proved that the TCSC algorithm outperformed the sampling-
judge the significance level [161]. based path algorithm introduced in [168]. Nevertheless, the algorithm
Luo et al. innovated the Gradient Descent and Bézier Curve (GDBC) does not consider the uncertainty of the environment [169].
algorithm as a post-optimization method. The GDBC Algorithm consists Horvath and Pozna introduced the Trajectory Following Approach
of two steps: first, the gradient descent algorithm (GDA) is deployed to (TFA) algorithm as an interpolation approach, in which the path is
utilize the path away from obstacles; second, the Bézier Curve inter- generated by a set of segments joined by clothoid curves. The results
polation technique has been applied to make the path smoother. The showed that the TFA algorithm can provide realistic approximation
results showed that the GDBC algorithm outperformed the traditional paths. Nevertheless, obstacles were not taken into consideration at all.
Gradient Descent Algorithm (GDA). However, more scenarios need to Moreover, the algorithm was not compared with other algorithms in
be validated, and only graphical statistical analysis is used, which is the literature [170].
insufficient to judge the significance level. Moreover, this is not an Shentu et al. applied the Bezier Curve (BC-based) algorithm in a
independent path planning algorithm, but it can be applied to enhance hybrid navigation ADS. The concept of using the Bezier curve is that
the path obtained from a traditional path planning algorithm [162]. the Bezier curve does not bypass all the points that define the path. The
Jin et al. proposed a coupled Longitudinal and Lateral planning BC-based algorithm managed to find reasonable paths. However, the
with Clothoid interpolation (LLC) algorithm. The algorithm starts by algorithm is not compared to any other algorithm, and more complex
designing the speed profile of the longitudinal horizon. Then, the scenarios need to be tested [171].
Clothoid interpolation algorithm has been deployed to obtain a set of Some hybrid algorithms based on interpolation methods are men-
local paths. After that, the best path is selected based on the objective tioned in previous sections. The DPCC algorithm is a hybrid algorithm
function, but if all candidate paths fail, a re-planning algorithm is used that combines dynamic programming as an optimization method with
for higher planning success. The results showed that the LLC algorithm Clothoid Curve as an interpolation method [141]. The PCC algorithm
outperformed the traditional discrete methods. Nevertheless, the path- combines the Sequential Quadratic Programming with the cubic spline
tracking method is simple, with no adaptive planning. Moreover, only curve for a smooth path [139]. The A*-PDWA algorithm is a hybrid
graphical statistical analysis was used [163]. algorithm that integrates the A* algorithm as a graph-based algorithm
Dai et al. applied the cubic spline interpolation technique to gener- and the cardinal spline interpolation method [86].
ate a smooth path in the ADS. This algorithm gave a smoother path The i-RRT algorithm is another hybrid algorithm between the RRT
than the one obtained from the cubic polynomial fitting algorithm. algorithm as a sampling-based method and the B-spline interpolation
Moreover, the Frenet coordinate frame in the cubic spline algorithm algorithm [108]. MRRT* algorithm is a hybrid algorithm between the
simplified the solution compared to the Cartesian coordinate frame RRT* and the cubic B-spline interpolation algorithm [109]. IA*-DWA
in the cubic polynomial algorithm. However, it is slower than the is a hybrid algorithm that integrates the DWA algorithm, the Improved
algorithms in the state-of-the-art [164]. A* algorithm as a graph-based algorithm, and the Bezier Curves as an
Feher et al. created a hybrid algorithm called HRL-MPC, a hier- interpolation algorithm [94].
archical RL algorithm, to obtain the optimal path. The twin Delayed Moreover, the IRRT algorithm is a hybrid between the RRT algo-
DDPG reinforcement learning algorithm has been applied to gener- rithm and the B-spline curve interpolation method [115]. The PCAPF
ate the path. Next, the clothoid curves interpolation technique is de- algorithm is another hybrid algorithm between the APF algorithm
ployed to make the path smoother. The CARLA results showed that and the polynomial curve interpolation method [132]. The PRRT-BSC
the HRL-MPCC algorithm outperformed human drivers. However, no algorithm integrates the Pruning RRT algorithm with the B-Spline
real-life tests are performed, and no comparisons are held against other Curve interpolation algorithm [116]. Curve Interpolation methods are
state-of-the-art algorithms [165] summarized in Table 8.
Lambert et al. implemented a Clothoid Curves Optimization (CCO) As future research guidance, interpolation-based algorithms are too
algorithm to generate an optimal path. The concept of the CCO al- sluggish and fail to find a path if they are used as a stand-alone
gorithm is to formulate the path planning problem as a clothoid op- algorithm. However, their great value lies in smoothing out paths
timization problem with 𝐺2 continuity constrained within the convex generated by other methods, such as graph-based and sampling-based
region. Then, a generic constrained non-linear solver has been applied algorithms. Their integration can significantly enhance path smooth-
to resolve the optimization problem and get a smooth path. Moreover, a ness, particularly in routes involving sharp turns or complex maneu-
weighting parameter 𝑏 controls the peak’s sharpness. The results proved vers. By combining interpolation techniques like Bezier curves, splines,
that the CCO gave smoother paths than the cubic spline interpolation and clothoids with other path-planning approaches, the resulting hy-
method. However, the CCO takes longer time than the cubic spline brid algorithms can offer optimally smooth paths while maintaining
method. When 𝑏 is more than 100, the convergence becomes poor. computational efficiency and adaptability in dynamic environments.
Furthermore, more scenarios should be tested in a broader range of This approach can balance practicality and path quality in autonomous
environments [166]. driving systems.
Bulut developed the Quintic Trigonometric Bézier Curve (QTBC)
algorithm, which is used to construct a predetermined trajectory based 8. Layer 5 - part 2: Path planning (machine and deep learning
on the Quintic Trigonometric Bézier Curve with 𝐶3 continuity. The techniques)
two shape parameters adjust the predetermined path without altering
any obstacle. The QTBC algorithm is compared with itself by chang- Machine Learning is the science in which a computer program can
ing the Bézier shape parameters: cubic Bézier, quintic Bézier, cubic learn from the experience and can make predictions and future deci-
trigonometric Bézier, and the proposed quintic trigonometric Bézier. sions based on this experience. In machine learning, a model is assumed
The QTBC algorithm gave the best results. However, the algorithm with unknown parameters. Then, a machine learning algorithm is used
is not compared to the other algorithms in the literature, and more to obtain these parameters. The machine learning process has three ma-
scenarios need to be validated [167]. jor phases: training phases, validation phases, and application phases.
Li et al. created a hybrid algorithm between the Tentacle and the A machine learning algorithm gets the unknown model parameters in
B-Spline Curve algorithms, named (TCSC) algorithm. This algorithm the training phase. In the validation phases, the model is tested to
uses the Tentacle algorithm to find pre-calculated paths using virtual make sure that the model can obtain optimal results with acceptable

15
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

Table 8
Literature Review Summary for the Curve interpolation Path Planning techniques.
Algorithm Idea Pros Cons
Game Theory finds global path, polynomial can find feasible and not compared to the other state-of-the-art
SDGT-PC [152]
curve removes errors and smooths the path. effective paths. path planning algorithms.
dynamic path planning method based considered dynamic obstacles. only one obstacle, no comparison,
DPP-CS [153]
on the cubic spline interpolation no real-life validation.
hybrid; E* computes heuristics, Bézier curves outperformed HA* and no path re-planning in
HE* [154]
improves path, CNM finds the final path. SST* algorithms. the dynamic environment.
2 alg. based on Bézier curves, FOBC is can find smooth need more scenarios, not
JQBC/FOBC [155]
smoother, DC modifies traffic state. acceptable paths. compared to other algorithms.
DVD finds holonomic path, faster than VD by 60%, RRT, more scenarios needed,
DVD-CC [156]
Clothoid finds a smooth non-holonomic path. and HA* algorithms. no dynamic obstacles.
clothoid solves 𝐺1 interpolation outperformed Preview- needs more scenarios,
CBA [157] problem given 2 directions based and no dynamic obstacles.
and 2 points. Pursuit-based algorithm.
constraint Bézier curves optimization is solved managed to find no comparison to other algorithms,
BCO [158]
by Newton–Raphson with Lagrangian Method. collision-free paths. needs more scenarios.
linear combination of B-spline basis managed to find sub-optimal path
BSCO [159] functions, smooth penalty fn. smooth paths. depends on tuning
for the unconstrained problem. parameter, no comparison.
Improved BSO; added penalty term better than PSO poor statistical
IBSO [160] to fitness fun., cubic spline algorithm by 90% analysis.
interpolation for a smooth path.
improve ABC, used adaptive limit parameter, outperform the traditional more complex scenarios
ADL-ABC [161]
Cubic polynomial smooths path. ABC algorithm. need to be tested.
get future samples, fit them into clothoid path managed to find needs more scenarios, parameters
DPCC [141]
by solving a dynamic programming problem. acceptable path. were not optimized, no comparison.
GDA optimizes the path, Bézier outperformed traditional more scenarios, only graphical statistics,
GDBC [162]
Curve smooths the path. GDA Algorithm. only post-algorithm to improve paths.
Clothoid generates multiple paths, cost fn. outperformed traditional simple path tracking with no adaptive
LLC [163]
chooses the best path, re-plan if it fails. discrete methods. planning, only graphical statistics.
apply the basic cubic spline smoother and simpler than slower than the other
Cubic spline [164]
interpolation algorithm cubic polynomial algorithm state-of-the-art algorithms.
cubic spline is used as parameterization managed to find suitable paths more complex scenarios,
PCC [139]
optimization method, then solved by SQP. in a dynamic environment. no comparison.
DDPG algorithm generates the path, outperformed human drivers. no real tests, no comparison
HRL-MPC [165]
clothoid algorithm makes the path smoother. against other algorithms.
Hybrid; A* gets global path, cardinal spline managed to find safe path. algorithm is not compared
A*-PDWA [86]
smooths path, P-DWA avoids obstacles. managed to find a safe path. with any other algorithm.
i-RRT find the path, B-spline managed to find personalized no comparison with the
i-RRT [108]
interpolation smooths path. safe trajectory. state-of-the-art algorithms.
MRRT* incrementally finds the path, managed to find collision-free no statistical analysis,
MRRT* [109]
cubic B-spline smooths the path. paths in real. no comparison with other algorithms.
clothoid with 𝐺2 continuity, sharpness smoother than slower,
CCO [166] control, obstacles constraints, solved by cubic spline. poor convergence when
a generic non-linear solver. 𝑏 > 100, more scenarios.
gets path based on 2 shape compared by changing shape no comparison,
QTBC [167] Bézier parameters that modify path parameters, QTBC is better need more scenarios.
without obstacle change. than QB, CB, and CTB.
Tentacle Alg. finds pre-calculated paths, better performance than environment uncertainty
TCSC [169]
B-spline finds path based on best Tentacle path. the algorithm in [168]. not considered.
generated path by set of straight segments can provide realistic obstacles are not considered,
TFA [170]
connected by clothoid curves. approximated paths. no comparison to other algorithms.
the problem is a quintic polynomial gave acceptable paths. no comparison with
PCAPF [132] curve, AFP implements constraints, any other path
solved by fmincon MATLAB toolbox. planning algorithm.
applied Bezier curve to find managed to find reasonable need more complex scenarios,
BC-based [171]
the optimal smooth path. paths, real application. no comparison to other algorithms.
hybrid; A* finds path, Cubic Bezier curves outperformed A*, needs testing on more complex
IA*-DWA [94]
smooths path, DWA avoids obstacles. DWA algorithms. environment and benchmark functions.
based on circular sampling, random point outperformed RRT model-driven with limitations,
IRRT [115]
filter, B-spline smooths path. and B-RRT algorithms. needs more scenarios and statistics.
hybrid; RRT finds path by pruning fn. based managed to find not compared to any other
PRRT-BSC [116]
on obstacles, B-Spline smooths the path. acceptable paths. state-of-the-art algorithm.

accuracy. In the application phase, the real-life model is used to make Machine Learning algorithms are categorized into three main
decisions. classes: supervised learning, unsupervised learning, and reinforcement

16
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

• Deep neural networks such as CNN and RNN can be used in the
perception layer to deal with sensory data such as images.
• Deep neural networks could be deployed as an End-to-End driving
scheme starting with perception and ending with path planning.

Deep supervised learning techniques can give fast and accurate

solutions in familiar scenarios. However, they require a considerable
training dataset to build reliable models. Therefore, the main demerit
of NN is the need for offline training. Moreover, the need to re-train the
whole model to make any updates to the model is too time-consuming.
Song et al. developed a deep cascaded neural network as an End-
to-End planning scheme. The proposed architecture combines an im-
proved VGG neural network, a type of CNN network, and long short-
term memory (LSTM). The proposed network is called IVGG-LSTM.
In the IVGG-LSTM network, the IVGG layer has been used to obtain
spatial features from the input images, whereas the LSTM layer is
deployed to get temporary features of the input images. Then, the
IVGG-LSTM model uses this data to map between the output parameters
for the motion and the input images. The practical results proved
that the IVGG-LSTM scheme outperformed the network introduced
in [173] and the traditional VGG-16 neural network. Moreover, the
IVGG-LSTM network can learn from the human driver and adapt to
different roads. Nevertheless, this network is not applied to real roads,
Fig. 3. Machine Learning algorithms classification. and the algorithm’s speed needs further improvements [174].
Kiki et al. designed a novel technique based on a neural network
to address path planning in the ADS. This method can predict feasible
paths based on a gradient-based self-supervised learning algorithm. The
learning. In supervised learning, labeled training data containing the
algorithm can exploit the experience gathered in the past via the learn-
input and the corresponding output is used to train the model and
ing algorithm. It achieved solutions that are 14 times faster than the
obtain the model parameters. It is called supervised because there is
ones obtained from RRT* and State Lattices. However, the algorithm
an output for each input used during the training phase. In supervised
achieved only 74% precision on the benchmark validation [175].
learning, unlabeled training data is used. The primary purpose of unsu-
Moraes et al. developed a real-time image-based method using
pervised learning is to cluster the input data or reduce its dimensions
the CNN neural network, named DeepPath, as an End-to-End driving
as a pre-processing stage for the data. In reinforcement learning, no
training data is used at all. The agent model reacts to the environment scheme for the IARA self-driving car. The DeepPath receives the pose
via sensors to learn by trial-an-error technique. of the current car and an image as input and obtains the trajectory as
Deep learning is a sub-branch of machine learning in which the an output. A CNN is used to infer the path model. Results proved that
assumed model is a neural network that consists of multiple layers DeepPath can generate paths for IARA that are slightly different from
between the input and the output. The powerful merit of neural net- the desired path. However, No comparison had been made to any other
works is that they can act as a model approximator for a complex model algorithm made [176].
that deals with massive data with multiple dimensions, such as images. Markolf et al. introduced a supervised learning approach based on
The deep learning and neural network techniques can be applied in Nonlinear Optimal Control (NOC) theory to get approximate solutions
supervised learning as a model and can be applied in reinforcement for the path planning problem, named NN-NOC. This approach solves
learning as a Q-function approximator. nonlinear optimal control problems to generate training data. Then,
As shown in Fig. 3, machine learning techniques can be classified these data have been deployed to train a neural network. Simulation
into three main categories: supervised, unsupervised, and reinforce- results showed that this model performed well. Nevertheless, the NN-
ment learning. This section focuses on deep-supervised, reinforcement, NOC algorithm is not feasible for real-time applications and has high
and deep-reinforcement learning because they are the most commonly computational time. Furthermore, no comparison is made to judge the
used techniques in autonomous driving systems. relative performance [177].
Guo et al. designed an LSTM neural network as a supervised learn-
8.1. Deep supervised learning techniques ing approach to find the shortest path. The LSTM neural network is
trained by training data from a fuzzy control algorithm. The LSTM-
Deep learning techniques are mainly based on deep neural net- based architecture outperformed the BP neural network and the fuzzy
works consisting of layers, each containing a set of neurons. There control algorithm. However, the dynamic environment is not con-
are three main types of neural networks: Convolutional Neural Net- sidered, and the network needs further improvements to be adap-
works (CNN), Recurrent Neural Networks (RNN), and Forward Neural tive [178].
Networks (FFNN). There are other neural network structures, but we Sakurai et al. innovated a Spiking Neural Network (SNN) algorithm
mainly speak of the basic types. FFNN networks are the primary consisting of two agents; one is to approach the goal, and the other is
architecture of multi-layer neural networks [172]. They are usually to avoid obstacles. The agents can learn and deal with dynamic objects
deployed for classification and recognition problems as supervised through dynamic graphs using SNN properties. The SNN algorithm
learning applications so that they can be applied in lane detection and outperformed the threshold-based agent algorithm. Nevertheless, the
traffic road sign recognition in the ADS. algorithm needs more validation using more complex obstacles [179].
CNN networks are usually used to deal with images and recognize Wang et al. created a trajectory learning algorithm using deep
details inside images such that they can play an important role in the neural networks named CNN-Raw-RNN. This scheme is similar to End-
perception layer in the ADS. RNN networks usually deal with sequential to-End driving, but the difference is that the output of the proposed
data such as voice and time series. Deep supervised learning techniques scheme is the trajectory instead of vehicle control actions. The system
can be used in two forms in the ADS: comprises two main parts; the first part is a hybrid neural network

17
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

between CNN and RNN, named the pilot, which receives the frames
of the video input and constructs the future path. The second part is
called the Copilot, which receives the pilot’s future path, verifies the
path’s safety, and generates the results of the semantic segmentation
perception to update the pilot mechanism. Then, the trajectory can
be easily mapped to the control actions of the car. The algorithm
outperformed the CNN and CNN-LSTM architectures. However, the
algorithm needs to be verified on more scenarios and compared with
more AI algorithms [180].
Kalathi et al. deployed neural networks to address the path-planning
problem of road sign recognition. They implemented an autonomous
car prototype that can drive on different tracks. This paper claimed Fig. 4. Concept of Reinforcement Learning.
that the neural network can determine the path. However, the neural
network in this proposed structure is only used for perception and
Yao et al. introduced a hybrid algorithm called the Reinforce-
recognition. Therefore, more work is needed on the path planning
ment Learning Multi-Objective Hyper-Heuristic algorithm (RL-MOHH)
algorithm [181].
to obtain the safest best trajectory in a smart city. A parallel version
Lee et al. proposed a Depth-wise Separable UNet for Path Prediction
of the algorithm is also proposed called (RL-PMOHH). The Heuris-
(DSUNet-PP) algorithm for End-to-End learning in the ADS. In this
tic algorithm finds a set of new routes, and then the Reinforcement
algorithm, a CNN network, UNet, is combined with a path prediction
Learning (RL) algorithm selects the best route. Results showed that
algorithm based on a quadratic polynomial to maintain lane centering. the RL-PMOHH algorithm is faster than the NSGA-II and the PNSGAII
Results showed that the DSUNet-PP outperforms UNet-PP in a dynamic algorithms. However, both the RL-MOHH/RL-PMOHH algorithms need
environment. However, the DSUNet-PP takes about five images to more online training because they could only get a maximum of 80%
generate an acceptable path within a range of 10 m. Therefore, image optimal solutions [186].
processing has high computational time [182]. Liu et al. proposed a new hybrid algorithm between Reinforcement
Learning (RL) and A* algorithms, called OPABRL, specially designed
8.2. Reinforcement learning techniques for congestion and accident scenarios. The RL algorithm is used to
choose the best trajectory with length priority. At the same time, the A*
The recent reinforcement learning techniques in the path planning algorithm has been applied to improve the trajectory generated by the
of the AD system are discussed in this section. In reinforcement learn- RL algorithm. Results proved that the algorithm can give the shortest
ing, no training data is used at all. The agent model communicates with path in a short run time. However, the algorithm is not tested on a
the environment via sensors to learn by trial-and-error technique [183]. significant complex problem where the maximum number of nodes is
The agent’s performance is evaluated during the training process based only 30. Therefore, the performance of complex problems needs to be
on the reward function. In each state, the agent takes an action that tested [187].
will make the agent move to another state. If the new state makes Chen et al. designed a hybrid algorithm between the greedy se-
the agent move towards the goal, the agent receives a good reward. lection algorithm and the Q-learning algorithm to find the optimal
Otherwise, the agent will not receive a good reward. The objective of path in the ADS in a static environment and reduce fuel consumption.
the RL process is to maximize the cumulative reward as an optimization The algorithm has been compared with the K-shortest and Dijkstra
problem [184]. algorithms, outperforming them in terms of fuel consumption and
The main merit of the RL methods is that they can deal with shortest path. However, they did not deal with the situation of an
complex problems and obtain reasonable solutions in the long term. unknown environment [188].
The model can be learned from error and modified during run time. Wang et al. implemented a hybrid technique, named (QLOR), be-
tween the Q-learning algorithm and the Linear Output Regulation algo-
However, the main challenge in the RL is the compromise between ex-
rithm (LOR) in which a Q-learning algorithm has been applied to obtain
ploitation and exploration. In exploitation, the agent needs to take the
the optimal trajectory. At the same time, the LOR algorithm has been
best action from a set of known actions to exploit and choose the best
deployed to design the tracking controller. The algorithm succeeded
action with the best reward. On the other hand, the agent sometimes
in two scenarios, but the dynamic obstacles were not considered.
needs to take new actions, which can lead to better reward than the one
Moreover, no comparison had been made to other algorithms [189].
obtained from the known actions, which is called exploration [185].
Wang et al. developed the Augmented Adversarial Inverse Rein-
Therefore, parameter tuning is a challenging problem. Moreover, gen-
forcement Learning (AAIRL) algorithm, where the traditional AIRL
erating a state–action table will be too massive in complex problems,
algorithm is improved by augmenting it with semantic rewards in
which require more memory and high computational time to search for
an interactive reinforcement learning framework. The AAIRL algo-
suitable action. rithm is compared with four baseline RL algorithms. Simulation re-
The reinforcement learning process can be formulated as Markov sults showed that the AAIRL algorithm outperformed all other meth-
decision processes (MDPs). The MDP agent is composed of a set of ods. However, the AAIRL algorithm has yet to be compared to other
actions A, a set of states S, and a reward function R. 𝑅(𝑠, 𝑎) indicates powerful algorithms in state-of-the-art [190].
the value of the reward when the agent commits an action 𝑎 ∈ 𝐴 Liu et al. innovated a Hybrid algorithm between Reinforcement
at a specific state 𝑠 ∈ 𝑆. A policy 𝜋 is a set of concussive actions. Learning and the A* Algorithm (HARL) algorithm. The A* algorithm
The objective is to obtain a policy that maximizes cumulative rewards generates a reference path optimized based on the Markov process in
for its actions. The Q-function is an action-value function representing the reinforcement learning algorithm using prior knowledge. Results
the cumulative rewards for actions taken in specific states. The target proved that the HARL algorithm outperformed the PSO, ACO, GA,
is to maximize the Q-function to obtain the best policy. The concept and Artificial Neural Networks (ANN) algorithms. Nevertheless, the
of the RL is shown in Fig. 4. In RL, the 𝑄(𝑎, 𝑠) indicates the reward results of the number of turns obtained from the HARL algorithm
for a specific action 𝑎 taken at a specific state 𝑠. The Q values for could be better. Moreover, the storage space of the prior knowledge
specific actions taken at specific states are stored in a Q table. This table for reinforcement learning is massive [191].
represents the experience gained by the agent. The RL methods have Liu et al. introduced a hybrid algorithm between the PSO and
been applied and enhanced to improve the path planning solutions. Reinforcement Learning (RL) algorithms. The PSO algorithm has been

18
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

used to optimize the hyper-parameters of the RL algorithm to make using the Bellman operator, and so on. The new policy is guaranteed
the algorithm faster. Results proved that the hybrid algorithm can to be a strict enhancement of the previous one [197].
provide better solutions. The main drawback of this algorithm is that In value-based algorithms, the algorithm keeps an eye on the value
the parameters of the PSO optimizer are static. Moreover, the statistical function. It starts with a random value function (Q-function). Then,
analysis is based only on graphical methods, so more numerical analysis the algorithm gradually maximizes the value function until it reaches
is required to judge the confidence level [192]. the optimal value of the Q-function. Then, the optimal policy can be
Kim et al. used a real-time Q-learning (RTQL) technique to address generated from the obtained optimal value function using the opti-
ADS’s real-time path planning problem. The simulation results showed mality Bellman operator, which contains a max nonlinear operator.
no difficulty in reaching the destination point in various scenarios. Therefore, value-based and policy-based algorithms have the same
However, no comparison is made to judge the algorithm’s performance working principle to get the optimal policy with the optimal value. The
among the recent state-of-the-art algorithms [193]. policy-based technique can obtain the optimal policy and then evaluate
Chang et al. created an Improved Dynamic Window Approach using its value, while the value-based technique can obtain the optimal value
the Q-Learning (IDWAQ) algorithm. The Q-learning algorithm is used and then reduce the optimal policy. The value-based algorithms are the
for motion planning. Then, the Q-learning algorithm learns the DWA most commonly used in literature to address the path planning in the
parameters to adapt to the dynamic environment. Simulation results ADS using the Q-learning algorithm [198].
showed that the improved DWAQ is better than the original DWA. Bernhard et al. designed the Experience-Based Heuristic-Search
However, dynamic obstacles are not considered, and more complex (EBHS) algorithm, a hybrid heuristic search, and the Deep Q-Network.
obstacles need to be tested [194]. The DQN represents experiences and is used as a heuristic function
Low et al. developed an Improved Q-learning (IQL) to address in a heuristic search algorithm. The results proved that the EBHS
the trajectory planning in the ADS. The original Q-learning algorithm algorithm is better than the DQN and Hybrid A* algorithms. However,
is improved by adding a distance metric, a virtual target concept, the algorithm lacks generalization, and it is not used in dynamic
and a Q-function that overcomes dead-ends. Results showed that IQL environment [199]
outperformed QL in all maps, which is better than RRT and VG in most You et al. introduced the Maximum Entropy Deep Inverse Rein-
cases. The drawback of IQL is that it is time-consuming during escape forcement Learning algorithm (MEP-DIRL). The algorithm relies on
mode. Moreover, dynamic obstacles are not considered [195]. a stochastic Markov decision process (MDP). The reward function is
Rousseas et al. presented a hybrid algorithm between the Integral designed using a Deep Neural Network (DNN) trained via the Maximum
Reinforcement Learning (IRL) algorithm and the Artificial Harmonic Entropy Principle (MEP). Simulated results demonstrated desired driv-
Potential Fields (AHPFs) algorithm. The AHPFs method generates the ing behaviors. However, it is not applied in real life and is not compared
path, while the IRL algorithm enhances the path via interaction with to other algorithms to compare its performance [200].
the environment. The algorithm outperformed the RRT* algorithm. Liao et al. applied the Double Deep Q-network (DDQN) algorithm
However, it neglected the 3D and dynamic environment [196]. to find the best path when overtaking vehicles on the highways. The
DDQN algorithm is compared to the traditional Deep Q-Network (DQN)
8.3. Deep reinforcement learning techniques Algorithm. Simulation experiments showed the merits of the DDQN
algorithm in terms of the convergence rate and safe overtaking. How-
As described in traditional reinforcement learning, the Q-table rep- ever, online hardware-in-loop (HIL) experiments are not considered.
resents the reward corresponding to the action taken at a specific state. Furthermore, no comparisons are made to the other state-of-the-art
This table represents the (state, action) space and its corresponding optimizers [201].
reward (Q-value). However, getting the optimal policy will be com- Zhao et al. just applied the Double Deep Q-learning Network
plicated when the (state, action) space is vast. Moreover, the Q-table (DDQN) to train a path planning policy under an interactive highway
will consume more memory space. Therefore, the benefit of the neural environment. The SUMO simulator is used to simulate an environment
network will arise. The neural network can act as a Q-function approx- in which the traffic distribution and the speed can be manipulated
imator to predict the (state, action) pair, called the Q-value. Due to the easily. The DDQN algorithm achieved good paths without collisions.
flexibility and the multi-input multi-output characteristics of the neural However, the algorithm is not compared to any other algorithm [202].
network, it can fit any complicated state–action space by adding more Liao et al. introduced a hybrid technique between the Potential
hidden layers as required, called Deep Reinforcement Learning (DRL). Field (PF) algorithm and the Deep Reinforcement Q-Learning Algorithm
Nevertheless, this adds more parameter tuning complexity to train the deployed in unknown environments. This algorithm can give better
neural network and balance the exploration and exploitation process. solutions than the ones obtained from the original RL algorithm. How-
DRL comprises three sub-categories: policy-based DRL, value-based ever, the best success rate of the hybrid algorithm reached only 88.62%
DRL, and actor–critics DRL. for 10000 comparisons [203].
Chen et al. innovated a hybrid algorithm between the Fuzzy control
8.3.1. Value-based/policy-based methods concept and the Deep Q-network (DQN) Algorithm, named the Con-
Reinforcement learning is concerned with obtaining the optimal ditional Deep Q-network (CDQN) algorithm. This algorithm has been
policy with the best reward. To understand the difference between used to resolve the directional path planning problem in an End-to-
policy-based and value-based methods, the difference between value End driving scheme of the ADS. The global path is generated from
and policy should be clear. The policy is the set of actions that should the DQN algorithm. Then, a defuzzification output layer in the DQN is
be taken to reach the goal. The value represents how good the policy used to improve the motion commands. The algorithm is simulated via
is. The conversion between the policy and the value is a two-way the CARLA simulator and compared to DDQN and A3C algorithms. Re-
relationship. Given the policy, we can get its value using the Bellman sults demonstrated that the CDQN algorithm outperformed the others.
operator (Q-function). On the other hand, given the value, we can However, the obstacle avoidance task needs to be addressed [204].
get the policy using the optimality Bellman operator. This process is Wen et al. presented a hybrid algorithm called Dueling Deep Q-
explained in Fig. 5. Network with a Fast active SLAM (DDQN-FSLAM) algorithm operated
In policy-based algorithms, the algorithm keeps an eye on improv- in an unknown environment. The DDQN algorithm is used to construct
ing the policy itself and then evaluates its value. It starts with a a path using the DRL technique. At the same time, the FSLAM method
random policy and then finds its corresponding value using the Bellman has been applied to construct a 2D map of the surrounding environment
operator. During iterations, the policy will be improved by using the and recognize the obstacles. The natural and virtual results showed
previous policy, then the value of the improved policy will be evaluated the algorithm’s success in a different environment. Nevertheless, the

19
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

Fig. 5. Relationship between policy and value.

algorithm is not compared to any other algorithm. More improvements 8.3.2. Actor-critic methods
are required for deep reinforcement learning to reduce the time to The Actor-critic algorithms are a hybrid architecture between the
finish the task [205]. value-based and policy-based schemes. There are two networks con-
Li et al. implemented a Deep Q-Network algorithm to get a suitable nected in series: the first network is called the actor, which generates
path through intersections. In this study, an End-to-End AD scheme the policy (action) based on the current state and observations; the sec-
is assumed. The first stage of the neural network is a CNN to collect ond network is called critics, which receives the actions and optimizes
images and implement the perception process. Then, a Deep Q-Network the reward value based on the policy obtained from the actor.
is followed to obtain the optimal driving policy via environmental The HRL-MPC algorithm is a hybrid algorithm mentioned in the
interaction. This scheme is implemented successfully. However, this interpolation-based algorithms section. It combines the clothoid curves
algorithm is not statistically compared with the well-known algorithms interpolation algorithm with the hierarchical actor–critics RL algo-
to judge its performance [206]. rithm [165]. Wang et al. designed a combined reinforcement learning
Sainath et al. implemented a hybrid neural network architecture to algorithm called Monte Carlo Search and actor-critic (MCS/AC) algo-
simulate the ADS system as an End-to-End driving scheme. The part of rithm to solve path planning in the ADS. The simulation results from
the neural network responsible for path planning is the DQN network, the Simulation of Urban Mobility (SUMO) simulator and the SMARTS
which is learned by the Deep Reinforcement Learning algorithm. This simulator proved that they showed that the MCS/AC algorithm out-
study implemented deep learning and AI techniques for developing performed the Soft actor–critics (SAC) algorithm and Monte Carlo
autonomous driving systems. However, statistical analysis has yet to Q-learning (MCQ) algorithm. Nevertheless, the algorithm should be
be performed to judge the algorithm’s performance [207]. tested on more complex scenarios and interactions with the environ-
Naveed et al. designed a Robust Hierarchical Reinforcement Learn- ment [212].
ing (RHRL) architecture. The traditional Hierarchical Double Deep
Zhang et al. proposed the Lyapunov-based Soft actor–critics with
Q-learning (hDDQN) is improved by deploying an LSTM network layer
Collision Probability Prediction (LSAC-CPP) algorithm to find safe and
in the proposed network to interact with the environment and discover
stable paths for the ADS. They integrated two concepts into the well-
noise. Moreover, Hybrid Reward Mechanism and Reward-Driven Explo-
known SAC algorithm; the first is the Collision Probability Prediction
ration are used to improve efficiency. CARLA simulation results showed
to identify the risk of collisions, and the second is the Lyapunov control
that the RHRL algorithm outperformed the Vanilla DQN (VDQN) algo-
theory to generate stable paths. The results indicated that the LSAC-CPP
rithm and Hierarchical Double Deep Q-Learning (hDDQN) algorithm.
algorithm achieved a higher success rate than the SAC algorithm. The
However, more scenarios need to be tested [208].
algorithm’s limitations are as follows: no real-world test is performed,
Li et al. applied Deep Reinforcement Learning in an End-To-End
and dynamic environments are not considered [213].
driving scheme (IDQNPER-ETE). The End-to-End architecture begins
Xu et al. developed a novel Actor-Critic Reinforcement Learning
with a hybrid deep neural network consisting of CNN and LSTM and
(ACRL) algorithm for multi-vehicle complex traffic scenarios. The pol-
is used to gather and process multi-sensor data. Next, the DQN net-
icy network in the ACRL algorithm is divided into two sub-networks:
work is applied to generate the path using the IDQNPER algorithm.
The algorithm managed to achieve acceptable performance in a static the lane-keeping and the lane-changing networks. The ACRL updates its
environment. Nevertheless, the algorithm is not compared to any other policy based on the gradient of the current advantage of the policy. The
algorithm. Moreover, dynamic environments are not considered [209]. reward function is designed for efficiency, safety, and driving comfort.
Peng et al. proposed the DRL-GAT-SA algorithm to find an optimal The real-time results showed that the ACRL had smoother paths than
and safe path in autonomous driving systems. This algorithm com- the rule-based algorithm. However, the algorithm is based on the
bined deep reinforcement learning with a graph attention network into pre-trained network, and it is not able to deal with new situations [214]
a controller called Graph Attention Reinforcement Learning (GARL), Choi et al. proposed the mobile robot collision avoidance learning
which relies on the safety field model. The algorithm outperformed with path (MCAL-P) algorithm based on the soft Actor-Critic (SAC)
the DDQN algorithm and DGN algorithm under two different scenarios. as the reinforcement learning algorithm. The virtual and real results
However, no real-life tests are performed, and more complex scenarios showed that the (MCAL-P) algorithm outperformed the Timed Elastic
are required [210]. Band (TEB) algorithm and the Dynamic Window Approach (DWA).
Perez et al. implemented and compared two Deep Reinforcement The disadvantage of the algorithm is that it cannot drive in a straight
Learning algorithms as a machine learning approach to find the shortest line [215].
path for the car. The two DRL methods are the traditional DQN network Tang et al. deployed a Soft Actor-Critic (SAC) scheme as a deep
and the Deep Deterministic Policy Gradient (DDPG) algorithm. CARLA reinforcement learning technique to address highway driving scenarios
Simulator is used to simulate the training and validation processes. with continuous action space. The simulation results showed that the
Both the DDPG and the DQN algorithms can safely reach the destina- SAC algorithm can balance safety, efficiency, and comfort more than
tion, but DDPG performs better and is more similar to human driver DQN and DDPG algorithms. However, no real-world experiments are
performance. The main drawback is that the algorithms were not tested executed, and the kinematic model of the vehicle is not feasible [216].
on real-life maps in CARLA [211]. Machine learning algorithms are summarized in Table 9.

20
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

Table 9
Literature Review Summary for deep supervised learning Path Planning techniques.
Deep Supervised Learning algorithms
Algorithm Idea Pros Cons
combines CNN (IVGG) for spatial outperformed no real application,
IVGG-LSTM [174] data with LSTM for temporary data, VGG-16 , and speed needs
then End-to-End planning. NN in [173]. improvement.
novel NN trained by gradient-based faster than RRT* Only 74% accuracy,
NN [175]
self-supervised Algorithm. and SL algorithm need more statistical analysis.
CNN End-to-end scheme, CNN receives input can generate acceptable No comparison was made
DeepPath [176]
image and car pose, then output the path. paths for IARA car. with any other algorithm.
solve nonlinear optimal control problems good performance. high computational time, not suitable
NN-NOC [177]
to generate training data for NN. for real-time, no comparison.
Fuzzy control generates data outperformed BP no dynamic environment,
LSTM-based [178] to train LSTM, neural network and needs further
LSTM generates the path. fuzzy control algorithm. improvements to be adaptive.
two agents; one to reach goal, outperformed greedy requires more obstacles,
SNN-based [179] the other to avoid obstacles. and the threshold-based complex scenarios validation.
Input as a dynamic graph. agent algorithms. and validation.
Pilot: CNN-RNN receives input outperformed the CNN , needs to be verified
CNN-Raw-RNN [180] image and generates path, CNN-LSTM on more scenarios
Copilot: verifies path safety. architectures in real. with more AI algorithms.
Just used a traditional NN Good perception accuracy NN only limited for perception,
NN [181]
wrong claim about path planning
integrated UNet with path prediction DSUNet-PP outperformed DSUNet-PP takes 5 images to
DSUNet-PP [182]
algorithm based on quadratic polynomial. UNet-PP. generate path, high computational time.
Reinforcement Learning algorithms
Algorithm Idea Pros Cons
Hybrid; Heuristic algorithm faster than PNSGAII needs more
RL-MOHH [186] generates paths, and NSGA-II algorithm online training.
RL chooses the best path. to find safe path (only 80% accuracy).
Hybrid; RL generates path, High performance for Not tested on high
OPABRL [187]
A* improves the path. low complexity problems. dimensions problems.
hybrid between Q-learning outperformed Dijkstra neglected unknown
DQN-GS [188] algorithm and greedy and K-shortest algorithm dynamic
selection GS. in fuel consumption environment.
hybrid Q-learning with Linear Output successful in two scenarios. no dynamic obstacles, no
QLOR [189]
Regulation to design tracking controller. comparison to other algorithms.
improve AIRL by augmenting it Good results only compared Not compared to the other
AAIRL [190]
with semantic rewards. with RL variants state-of-the-art algorithms.
Hybrid A* with RL; A* find the path, outperformed ACO, PSO, High number of turns,
HARL [191]
RL optimizes path based on prior knowledge. GA, and ANN algorithms. large storage space for RL.
Optimize parameters of RL reliable solutions PSO parameters are static,
HPHA [192]
using PSO algorithm. only graphical statistical analysis.
Used real-time Q-learning Managed to find path No comparison or
RTQL [193]
(RTQL) algorithm. in various scenarios. statistics are made.
Hybrid Q-Learning with DWA; better than dynamic and complex
IDWAQ [194] QL to find a path, original DWA. obstacles are not
DWA for unknown environment. considered.
improved QL by adding distance outperformed QL, RRT too slow in
IQL [195] metric, virtual target, and VG algorithms. escape mode,
dead-ends-free Q-function. no dynamic obstacles.
Hybrid Artificial Harmonic PF outperformed neglect 3D and
AHPF-IRL [196]
with Integrated RL. outperformed Dynamic environment.
Deep Reinforcement Learning algorithms (Value/Policy based)
Algorithm Idea Pros Cons
Hybrid; DQN is used as a heuristic outperformed DQN and no generalization,
EBHS [199]
function in a heuristic search algorithm. Hybrid A* algorithms. no dynamic environment.
based on stochastic MDP, NN reward function obtained desired no real application, no comparison
MEP-DIRL [200]
trained via Maximum Entropy Principle. driving behaviors. with other algorithms.
applied Double DQN Better than No online interaction,
DDQN [201]
DQN algorithm no comparison.
DDQN [202] Just applied Double DQN Simulate highway driving in SUMO. No comparison made.
Hybrid DRL with Potential Better than success rate only
DQN-PF [203]
Field algorithm traditional RL 88.62% for 10000 trials.

(continued on next page)

21
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

Table 9 (continued).
hybrid DQN with Fuzzy control; DQN finds outperformed DDQN and No obstacle avoidance.
CDQN [204]
the path, fuzzy improves motion commands. A3C algorithms.
Hybrid DDQN with FSLAM; succeeded in Slow with high
DDQN to find path, different computational time,
DDQN-FSLAM [205]
FSLAM to environments. No comparison with.
map environment. other algorithms.
Applied Deep Q-Network (DQN) Consider the environment not compared with the
ETE-DRL [206]
in End-to-End scheme state-of-the-art algorithms.
DRL [207] implemented DRL in End-to-End scheme Managed to implement DRL in ADS No statistical analysis
improve the Hierarchical DDQN by using LSTM outperformed VDQN and more scenarios needs
RHRL [208]
layer and a hybrid reward mechanism. hDDQN algorithms. to be tested.
End-to-End; CNN-LSTM processes sensor data, achieved acceptable performance no comparison,
IDQNPER-ETE [209]
DQN generates path using IDQNPER algorithm. in a static environment. no dynamic environment.
combined graph attention network and outperformed DDQN no real tests, more complex
DRL-GAT-SA [210]
DRL based on the safety model. and DGN algorithms. scenarios are required.
Implemented DQN and Deep Deterministic Both reach the goal, human-like Lack of real maps.
DDPG [211]
Policy Gradient (DDPG) performance in CARLA simulator.
Deep Reinforcement Learning algorithms (actor–critics)
Algorithm Idea Pros Cons
DDPG algorithm generates the path, outperformed human drivers. no real tests, no comparison
HRL-MPC [165]
clothoid algorithm makes the path smoother. against other algorithms.
combined RL techniques; faster than need more complex
MCS/AC [212] Monte Carlo Search (MCS), SAC and scenarios and interactions
and Actor-Critic (AC). MCQ algorithms. with the environment.
improved SAC by Collision Probability Prediction, more success rate no real-world test,
LSAC-CPP [213]
and Lyapunov theory for stability. than traditional SAC. no dynamic environment.
Actor-Critic based with two sub-networks, smoother paths than based on pre-trained network,
ACRL [214]
gradient-based policy update. Rule-based Algorithm. cannot deal with new situations.
based on Soft Actor-Critic outperformed DWA and Timed cannot drive in
MCAL-P [215]
(SAC) algorithm. Elastic Band (TEB) algorithm. a straight line.
implemented Soft better than no real-world
SAC [216] Actor-Critic (SAC) DQN and DDPG experiments, car kinematic
algorithms. model is not feasible.

The future of Deep Supervised Learning Techniques in autonomous of meta-heuristic optimization algorithms is that they are problem-
driving systems should focus on enhancing real-time processing capa- independent and are considered general problem-solvers. They can
bilities and reducing reliance on extensive training datasets. Integrating give an optimal solution for any complex search space. Therefore, it
these techniques more effectively with other path-planning strategies, can be applied to any problem with few modifications [219]. Meta-
like optimization or interpolation methods, could yield more efficient heuristic optimization algorithms are less likely to get stuck into a local
and adaptive solutions. In addition, using pre-trained models can in- minimum, compared to the gradient-based algorithms, because of the
crease training speed. Additionally, improving the algorithms’ ability to exploration phase that can help to explore different search spaces.
learn from limited data and incorporating self-improving mechanisms Nevertheless, the main challenge of the meta-heuristic optimization
through continuous learning will be vital in advancing these techniques algorithms is the parameter tuning to achieve the balance between
for practical autonomous driving applications. exploitation and exploration. In exploitation, the algorithm seeks the
The Traditional Reinforcement Learning (RL) techniques do not best solution in the current search space region. While in exploration,
need a training data set to be trained; they can interact with the the algorithm explores different search spaces to ensure diversity in the
environment and update the model in real time. However, storing the candidate solutions [220]. Moreover, they can provide different paths
Q table requires a massive memory for complex driving scenarios. The for the same problem if run multiple times. The reason behind this
research direction solved this problem by using deep reinforcement problem is the random behavior in specific stages of the algorithms.
The meta-heuristic optimization techniques can be classified into
learning (DRL), which incorporates advanced neural network architec-
four categories: evolutionary algorithms, trajectory-based algorithms,
tures that can replace the need to store substantial state–action tables.
swarm-based algorithms, and nature-based algorithms. Evolutionary
However, these networks generate the need for parameter tuning for
algorithms imitate the concept of the survival of the fittest. Genetic
the weights of the neural networks, which makes the training process
Algorithm (GA), Differential Evolution (DE), Harmony Search (HS) al-
too time-consuming. The best way to tackle this problem is to use pre-
gorithm, and Genetic Programming are the most common evolutionary
trained deep Q networks to accelerate the learning process for practical
algorithms. Trajectory-based algorithms rely on updating solutions by
and reliable autonomous vehicle navigation.
moving through neighboring candidates. The most used algorithms are
Tabu Search (TS), Simulated Annealing (SA), hill climbing, and Iterated
9. Layer 5 - part 3: Path planning (meta-heuristic optimization Local Search (ILS) algorithm.
techniques) Swarm-based algorithms mimic the nature of animals and humans,
such as animal behavior in searching for food in groups. The most fa-
Nowadays, meta-heuristic optimization algorithms have been in- mous Swarm-based algorithms are the Grey Wolf Optimization (GWO)
tensely applied to a wide range of optimization problems [217]. As algorithm, Artificial Bee Colony (ABC) algorithm, Cat Swarm Optimiza-
long as the path planning problem in the AD system can be modeled as tion Algorithm (CSOA), Particle Swarm Optimization (PSO) algorithm,
an optimization problem, meta-heuristic optimization algorithms can Bacterial Foraging Optimization (BFO), Ant Colony Optimization (ACO)
be easily applied to resolve this problem [218]. The main advantage algorithm, and Chicken Swarm Optimization (CSO) Algorithm. The

22
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

Fig. 6. Meta-heuristic optimization types.

last category is the nature-based algorithms, which are inspired by 9.2. Differential Evolution (DE)
natural phenomena that exist in animals and humans, such as the
Moth–Flame Optimization (MFO) Algorithm, Firefly Algorithm (FA), Differential Evolution (DE) is another evolutionary meta-heuristic
Teaching-Learning-Based Optimization (TLBO) Algorithm, Beetle An- optimization algorithm, like the GA algorithm. It starts with a set of
tennae Search (BAS) Algorithm, and Cuckoo Search Algorithm (CSA). candidate solutions called agents. These agents explore the search space
The categories of the meta-heuristic algorithms are shown in Fig. 6. based on a simple mathematical rule to combine the position of agents
in the population. If the new position is improved, the new position
will be accepted to form part of the population. Otherwise, it will be
9.1. Genetic Algorithm (GA) neglected. This algorithm was first introduced by Storn and Price in
1997 [225]. It has recently been improved to solve the path planning
problem.
The genetic Algorithm (GA) is the most popular evolutionary al- Guo et al. applied the Differential Evolution (DE) algorithm to the
gorithm, which imitates the principle of survival of the fittest. This Hongqi autonomous car HQ430. A model predictive control (MPC)
principle was first discovered in 1958 by Bremermann [221]. The GA in scheme has been implemented to design the controller; then, the DE
its recent form as an optimization algorithm was proposed by Holland algorithm is deployed to obtain the optimal path. The experimental
in 1975 [222]. GA and its variants have recently been applied in various results showed that the implementation can achieve suitable control
applications, including path planning. performance for path planning in autonomous cars. However, no the-
GA begins with an initial randomly generated population of can- oretical statistical analysis or comparison has been made to any other
didate solutions. Each candidate is called an individual and has a algorithm [226].
fitness value, which describes how good the solution is. Each solution
9.3. Simulated Annealing (SA)
is selected based on its fitness and passes its genetics to two newly
generated solutions. The previous operation is called crossover. To
Simulated Annealing (SA) is a trajectory-based meta-heuristic al-
ensure the diversity of the new generation, another operation, called
gorithm recently used to solve optimization problems. The concept
a mutation, is applied by a low percentage. The previous steps are
of the SA algorithm relies on the idea of annealing in metallurgy,
repeated until they are solved with the desirable accuracy. which involves cooling and heating the materials to change their phys-
Utami et al. proposed the Modified Crossover Genetic Algorithm ical properties of the material. In the SA algorithm, the temperature
(MCGA), in which the crossover operator was modified by comparing smoothly decreases from a starting value until it reaches zero. The slow
the newly generated solutions’ fitness with the original solutions’ fit- cooling process is interpreted in the SA algorithm as the slow decay
ness. The best fitness will survive, and the poor fitness will be ignored. in the probability of accepting poor solutions while exploring new so-
Simulation results showed that the MCGA algorithm can generate opti- lutions. During the cooling process, the thermodynamic characteristics
mal paths based on the desired conditions. Nevertheless, the algorithm will be changed, representing the improvement of solutions [227]. This
is not compared to any other algorithm, and the dynamic obstacles are algorithm has recently been adapted to enhance trajectory planning
not considered [223]. solutions.
Magdy et al. implemented the SA algorithm for motion planning
Receveur et al. designed a hybrid algorithm called Genetic
in the Autonomous Transportation Operating Modules (ATOM) system.
Algorithm-Potential Field (GA-PF), which combines the GA with the
The SA algorithm is compared to ACO and PSO algorithms. Concerning
Potential Field Algorithm. The global path is generated and optimized
the computational time, the results of the SA outperformed the PSO
by GA. Then, the PF algorithm generates the local path based on the algorithm and ACO Algorithm. However, PSO gave solutions with
global path obtained from the GA algorithm. This algorithm overcame better cost than ACO and SA. The algorithms are not tested in a
the drawbacks of the PF algorithm, such as getting stuck in a local real-life experiment. Moreover, it is compared only to ACO and PSO
minimum. However, the final trajectory is not periodically recalculated. algorithms, neglecting the other promising meta-heuristic optimization
Moreover, the non-holonomic car model and the dynamic constraints algorithms [228].
are not considered. Therefore, the solutions are not feasible for real-life Yin et al. designed a hybrid algorithm between the SA algorithm
ADS [224]. and the PSO algorithm called the SAPSO algorithm, in which the

23
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

PSO algorithm was kept as simple as possible. The hybrid algorithm Fusic et al. [237] implemented Satellite PSO (SPSO) and five other
showed better results than those obtained from the Simulated Anneal- PSO variants to address the path optimization problem based on images
ing Teaching Learning Based Optimization (SA-TLBO) and the original from the satellite. Satellite map data are used to map the unknown en-
PSO algorithm. However, the dynamic environment is not considered, vironment. SPSO achieves the best among PSO variants. However, this
and no numerical statistical analysis has been made [229]. study focused only on the PSO algorithm and its variants, neglecting
Shang et al. developed an Improved Simulated Annealing (ISA) algo- other prominent meta-heuristic optimization algorithms.
rithm, where the original SA algorithm is improved by designing neigh- Chai et al. proposed Enhanced Multi-Objective PSO (EMOPSO),
borhood transformation techniques and better solution acceptance rules where the original path planning problem is reformulated to a multi-
to improve the global search. The results proved the robustness of the criterion unconstrained problem by adding a normalized objective
ISA. However, the execution time of the algorithm is relatively high, function that reflects the total amount of constraint violation. More-
and the algorithm was not tested on high dimensions problems, more over, two local exploration operations evolution RS and 𝜖 -bias se-
than 200 nodes [230]. lection method were added to improve the progress. The algorithm
outperformed the Potential Field PSO (PFPSO) and Potential Field
9.4. Particle Swarm Optimization (PSO) ABC (PFABC) algorithms. However, dynamic obstacles are not con-
sidered, and the algorithm needs to be tested on a more complex
Particle Swarm Optimization (PSO) is a widely used swarm-based environment [238].
meta-heuristic optimization algorithm. It imitates the organized move- Zhang et al. designed a hybrid algorithm between Improved PSO,
ments of bird flocks to find food without a leader; the birds go with the Improved Artificial Potential Field (IAPF), and Simulated Annealing
one nearest to the food source. It was first proposed by Kennedy and (SA) algorithm, called the IPSO-APF/SA algorithm. The improved PSO
Eberhart in 1995 [231]. The PSO algorithm represents the candidate algorithm, with adaptive weights, is used for global path planning. The
solutions as a group of particles. Recently, the PSO algorithm has been improved APF algorithm has been deployed for local path planning and
modified and customized to the path planning. avoiding collisions. The parameters of the APF algorithm are optimized
Li et al. [232] introduced a hybrid algorithm called SLPSO, which using the SA algorithm. The practical and simulation results showed
combined Particle Swarm Optimization (PSO) and a Self-adaptive that the IPSO outperformed the original and Mutation PSO (MPSO)
Learning algorithm. The PSO algorithm finds the optimal path under algorithms. However, the dynamic obstacles are not considered, and
multiple constraints, such as collision risk, path length, and smooth- more statistical analysis is required [239]
ness. Then, the Self-adaptive Learning algorithm is deployed to enhance
the capability of the PSO algorithm. The results proved that the SLPSO
9.5. Ant Colony Optimization(ACO)
is better and more effective than PSO and GA. However, SLPSO is
not used for the complex environment for path planning of the ADS.
The ACO algorithm is a swarm-based meta-heuristic optimization
Furthermore, it is not applied in the 3D environment.
algorithm introduced by Dorigo in 1997 [240]. ACO algorithm mimics
Zhang et al. innovated a hybrid algorithm called the Evolutionary
the ants’ behavior in organized groups while searching for the food
Scatter Search Particle Swarm Optimization (ESS-PSO) algorithm. This
source. It is required that the ants find the shortest path to the source of
algorithm uses the ESS algorithm to generate paths in a scatter search
food. Recent research adopted the ACO algorithm to the path-planning
template. The PSO algorithm improves the paths generated from the
problem.
ESS algorithm, where the velocity and the position of the car are
implemented using the ‘‘ruin and recreate’’ concept. The algorithm Chen et al. improved the ACO algorithm by combining the A* and
gave excellent results concerning accuracy and speed. However, the ACO algorithms. The results proved that the new hybrid algorithm
algorithm has yet to be used for local planning [233]. gave more efficient and accurate solutions than the ones from the
Zhang et al. developed an Improved Localized Particle Swarm Op- ACO algorithm. However, the algorithm still has high computational
timization (ILPSO) algorithm for trajectory planning. Improvements in time [241].
PSO parameters are made to increase convergence speed and prevent Ali et al. developed a hybrid algorithm among the A* Multi-
the ILPSO algorithm from falling into a local minimum. Furthermore, directional algorithm, the ACO algorithm, and the Markov Decision
extended Gaussian distribution has been used to increase the diversity Process (MDP) model. The A* Multi-directional algorithm searches and
of particles. Compared to the basic PSO algorithm and A* algorithm, stores the best nodes between the start and the endpoints. Next, the
the ILPSO algorithm outperformed them. However, the ILPSO is not ACO algorithm optimizes the generated path to avoid blind searches.
applied in complex,3D, and large-scale environments [234]. Then, the MDP scheme is implemented to reduce the sharpness of
Wahab et al. presented a hybrid algorithm between the PSO algo- the global path. The algorithm gave better results than the other
rithm and the Fringe Search Algorithm, named PSOFS, for the indoor algorithms. However, the algorithm did not take into consideration the
environment. Generally, the PSO algorithm generates the path as a moving obstacles. Moreover, it has high computational time [242].
sequence of points. The role of the FS algorithm arises by generating Wu et al. presented an Improved ACO algorithm optimized by the
a safe point if the generated point by the PSO algorithm is less than Particle Swarm Algorithm (IACO-PSO). Not only is the path length
a one-meter distance from the obstacles. The PSOFS algorithm gener- taken into consideration, but also the flow of the incoming traffic and
ated safer, shorter, and smoother paths than the ones obtained from the number of lanes are considered. The ACO algorithm is deployed
PSOA and PSOD algorithms. Nevertheless, the algorithm needs more for dynamic path planning, whereas the PSO algorithm is used to
investigations against more state-of-the-art algorithms [235]. optimize the parameters of the ACO algorithm. The practical results
Qiuyun et al. introduced an Improved Particle Swarm Optimization showed that the IACO algorithm reduced the average rate of congestion
(IPSO) algorithm for the material transportation problem in the indoor from 13.63% to 9.73% compared to the ACO algorithm. However, no
static environment. Two main modifications had been made to the comparison had been held with other algorithms [243].
original PSO; a crossover operator had been added to the particle Pohan et al. designed a hybrid algorithm between the RRT algo-
position, and a mutation operator had been added to prevent falling rithm and the Ant Colony System (ACS) algorithm, named the RRT-ACS
into a local minimum. The IPSO algorithm outperformed the original algorithm. The RRT algorithm generates a quick sub-optimal path. After
PSO, GA, and ACO algorithms in material transportation. Neverthe- that, the optimal path is generated from the sub-optimal paths using the
less, dynamic obstacles are not considered, and the algorithm is not ACS algorithm. The RRT-ACS algorithm outperformed the original RRT,
applied in the outdoor environment; it is problem-specific with high RRT*, and the other RRT variants. However, the algorithm became too
constraints [236]. slow to find the optimal path, especially when the effect of the ACS

24
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

algorithm was minimal. Furthermore, the dynamic environment is not 9.7. Grey Wolf Optimization (GWO) algorithm
considered [244].
Zhang et al. introduced the Adaptive Improved ACS algorithm GWO algorithm is a recent swarm-based meta-heuristic optimiza-
using the population information Entropy (AIACSE) algorithm. The tion algorithm that was first introduced by Mirjalili et al. in 2014.
information entropy concept is added to the ACO algorithm to increase The GWO algorithm is inspired by the hierarchy of the wolves and the
the diversity of the population. The population is initialized based hunting group’s behavior. The GWO algorithm is widely used because
on non-uniform distribution to avoid blind searches at the beginning. of its simplicity, but it has a slow convergence rate and can be stuck
Moreover, the adaptive parameters and pheromone diffusion model into a local minimum for some problems [252].
have been applied to ensure an acceptable balance between exploita- Liu et al. proposed the Improved Grey Wolf Optimization (IGWO)
tion and exploration. The AIACSE algorithm outperformed the original algorithm, where further improvements had been made to the origi-
ACS, the Rank-based Ant System (RAS), and PS-ACO algorithms. Never- nal GWO algorithm; the lion optimizer optimizes the parameters to
theless, the dynamic environment is not considered, and the algorithm improve search capability, and the weights became dynamic to add
needs parameter improvements [245]. more diversities to the wolves. The IGWO algorithm outperformed
Jiang et al. implemented a hybrid algorithm between ACO and the Chicken Swarm Optimizer (CSO) algorithm, the original GWO
PSO based on feedback from the interaction with the environment to algorithm, the Butterfly Optimization Algorithm (BOA), and the Whale
make the algorithm adaptive. The algorithm gave better and faster Optimizer Algorithm (WOA) on a maximum of 30 dimensions prob-
solutions than the traditional ACO algorithm. However, this algorithm lems. However, dynamic obstacles are not considered, and a more
is compared to the original ACO for low dimensions problem [246]. complex environment needs to be validated [253].
Zhou et al. developed a hybrid algorithm between the Dijkstra Al- Kiani et al. proposed three hybrid algorithms between the Grey Wolf
gorithm and the ACO algorithm, named ACO-DA, for the indoor airport Optimization (GWO) and the Adaptive Rapidly-exploring Random Tree
environment. The ACO algorithm can find a global path using the (ARRT) algorithm. The ARRT algorithm can be hybridized with one
feedback information. Dijkstra’s algorithm can avoid obstacles while of the following GWO variants: GWO (ARRT-GWO), Incremental GWO
selecting the best path. The ACO-DA algorithm outperformed the A*, (ARRT-IGWO), and the Expanded GWO (ARRT-ExGWO) algorithms.
Dijkstra, and ACO-A* algorithms. Nevertheless, the algorithm is not The GWO algorithm constructs the path, while the RRT algorithm
used in the outdoor environment, and it needs to be tested on more avoids possible obstacles. The ARRT-ExGWO algorithm outperformed
scenarios and benchmark functions with deep statistical analysis [247]. the other GWO variants, PSO, Whale Optimization Algorithm (WOA),
and the improved BA algorithm under four scenarios. However, these
algorithms are not tested on complex benchmark functions. Moreover,
9.6. Artificial Bee Colony (ABC) algorithm
dynamic and 3D environments are not considered [254].
Zafar et al. implemented a hybrid algorithm between the GWO
The ABC algorithm is a popular swarm-based meta-heuristic opti- and APF algorithms, namely the GWO-APF algorithm. This algorithm
mization algorithm based on honey bees’ natural activity while search- has been applied to on-time path planning of ADS in a static envi-
ing for a food source. Three types of bees play an essential role in ronment. The APF algorithm finds obstacle-free locations. Then, the
the ABC algorithm. The scout bees search for food sources. Then, the GWO algorithm minimizes the path and locations generated by the APF
employed bees test and evaluate the quality of nectar food sources algorithm. The results revealed that the GWO-APF algorithm outper-
obtained from the scout bees. The onlooker bees obtain the data from formed the Surrounding Point Set (SPS) [255]. However, the algorithm
the employed bees and choose the best food source. This algorithm was was not applied in a dynamic environment and complex benchmark
first proposed by Kharaboga in 2005 [248]. functions [256].
Nayyar et al. designed the Arrhenius ABC (aABC) algorithm, im-
proving the balance between exploration and exploitation using the 9.8. Chicken Swarm Optimization (CSO) algorithm
Arrhenius equation. The results showed that the aABC algorithm out-
performed the traditional ABC, DE, and PSO algorithms. However, the The CSO algorithm is a recent and effective swarm-based meta-
algorithm needs more driving scenarios to be tested in the ADS [249]. heuristic optimization algorithm introduced by Meng et al. in 2014.
Kamil et al. introduced the Adaptive Dimension Limit-ABC (ADL- This algorithm mimics the natural behavior of chicken groups. How-
ABC) Algorithm, in which an adaptive limit parameter has been used ever, the CSO algorithm can fall in a local minimum, especially in
instead of a fixed, stable limit. The cubic polynomial interpolation complex optimization problems [257].
technique is used to make the path smoother. The ADL-ABC algorithm Liang et al. designed the Improved CSO (ICSO) algorithm as a meta-
outperforms the traditional ABC algorithm. However, more complex heuristic optimizer to get the shortest path for the autonomous car.
scenarios should be tested to judge the level of significance [161]. They improved the original CSO algorithm by adding Levy flight to the
Xu et al. developed the Coevolution framework with the Global best update rule of the hen’s location to increase the population’s diversity
Leading ABC (Co-GLABC) algorithm. In this algorithm, the Differential and exploration. Moreover, a nonlinear weight reduction strategy is
Evolution (DE) search equation is introduced to the ABC algorithm added to the update rule of the chicken’s position to improve the
to speed up the convergence speed. Moreover, a crossover technique ability of self-learning. The results showed that the ICSO is faster and
is integrated with the global best position to handle the dimension- more accurate than the PSO and original CSO algorithms. However, the
dependent part. Furthermore, an innovative setting for the food source algorithms are only tested on a few benchmark functions and only one
has been applied to boost performance. The Co-GLABC outperformed path planning scenario [258].
the state-of-the-art algorithms. However, dynamic obstacles are not
considered, and more complex scenarios need to be tested [250]. 9.9. Cat Swarm Optimization Algorithm (CSOA)
Kumar and Sikande proposed a hybrid improved Artificial Bee
Colony with Evolutionary Programming (EP), named iABC–EP algo- The classical Cat Swarm Optimization (CSO) algorithm is a recent
rithm. The improved ABC algorithm generates the path. Then, evolu- swarm-based meta-heuristic optimization algorithm that mimics the
tionary programming (EP) is applied to improve the path achieved by cats’ behavior. However, the algorithm does not fully consider the other
the iABC algorithm. The simulation results proved that the iABC–EP optimal cat position in the update process. Therefore, the classical
algorithm can give shorter paths than the ABC–EP algorithm by 5.75%. CSO lacks population diversity, which may lead to falling into local
However, more complex scenarios should be tested [251]. minimum [259].

25
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

Zhao et al. introduced the MOCMCSO-APFM, a hybrid algorithm Reda et al. proposed a new discrete CSA variant named the Dis-
between the APF Method (APFM) and the Multi-Objective Cauchy Mu- crete Damped Cuckoo Search (DDCS) algorithm to solve the order-
tation Cat Swarm (MOCMCSO) algorithm. In the MOCMCSO algorithm, picking routing problem inside warehouses. The proposed algorithm
they introduced the Cauchy mutation operators to improve the search improved the traditional CSA algorithm by combining 2-opt move,
patterns to obtain the shortest path. The APFM algorithm is used to crossover operators, and random key encoding. The DDCS algorithm
avoid collisions in the indoor environment. The results proved that was customized to address the OPR problem inside a warehouse en-
the MOCMCSO algorithm outperformed the Multi-Objective Particle, vironment. Results indicated the significant performance of the DDCS
Swarm Optimization (MOPSO) and the original Multi-Objective Cat algorithm over GA, PSO, and ACO algorithms, especially in complex
Swarm Optimization (MOCSO) algorithms. However, the path points problems. However, this algorithm is not tested for local or global path
are repeated in the inspection case, and anti-collision efficiency needs planning [80].
to be increased [260].
9.12. Beetle Antennae Search (BAS) algorithm
9.10. Firefly Algorithm (FA)
BAS Algorithm is a new nature-based meta-heuristic optimization
FA algorithm is a nature-based meta-heuristic optimization algo- algorithm proposed in 2017 by Jiang and Li [268]. The navigation of
rithm used to solve optimization problems. The FA algorithm mimics longhorn beetles inspires the algorithm and mimics the random walking
the fireflies’ natural flashing behavior. Fireflies belong to the winged mechanism of beetles and the function of antennae in nature. However,
beetle family, which is called the lightning bug, because of their ability the convergent BAS algorithm is based heavily on the direction of a ran-
to generate light. Fireflies use that light to select their mate without dom beetle in every iteration. Therefore, this algorithm was improved
wasting heat energy. Sometimes, they use the light to make their by Wang and Chen in 2018, who proposed the Beetle swarm antennae
enemies scary to avoid them. The goal is to find the proper mate and search (BSAS) algorithm. The BSAS algorithm overcomes the draw-
avoid the enemy. Yang first proposed the FA algorithm in 2008 [261]. backs of the BAS algorithm. It combines the feedback-based step-size
It has been used recently as a meta-heuristic approach in trajectory update technique with the swarm intelligence algorithm [269].
planning. Towards the improvement of the BAS algorithm, the Beetle Swarm
Zhou et al. introduced the Modified Firefly Algorithm (MFA), in Optimization (BSO) algorithm was introduced by Wang and Yang in
which the step factor 𝛼 is adaptive, and the convergence speed is 2018 [270]. The BSO algorithm is a hybrid algorithm between Beetle
controlled by the parameter 𝛽0 . The results indicated that the MFA Antennae Search (BAS) and the PSO algorithm (BAS-PSO). The BSO
algorithm outperformed the Genetic Algorithm (GA) and the tradi- algorithm combines the PSO algorithm’s update strategy and the BAS
tional Firefly Algorithm (FA). However, More scenarios and benchmark algorithm’s search mechanism, thus increasing convergence speed and
functions should be applied to test the algorithm’s performance [262]. avoiding the local minimum trap.
Bisen and Kaundal applied the original FA algorithm to the path Mu et al. presented an improved version of the BSO (IBSO) algo-
planning for autonomous cars. They applied the FA algorithm to multi- rithm to address the three-dimensional route planning problem, where
ple scenarios. The algorithm managed to find an acceptable path. Nev- a penalty term has been added to the fitness function. Next, the cubic
ertheless, the algorithm is not compared to any other algorithm. The spline interpolation technique has been applied to make the paths
performance of the traditional algorithm needs to be enhanced [263]. smoother. Results showed that the IBSO algorithm gave shorter paths
Li et al. developed the self-adaptive population size firefly algorithm by 90% of those obtained from the PSO algorithm. However, the
(SPSFA), where the population size is adaptive to the collision degree statistical analysis needed to be more comprehensive to judge the
using two nonlinear functions. An adaptively adjusted parameter is performance in a complex environment. [160].
added to control the population size. The results proved that the Jiang et al. developed three variants based on the BAS algorithm for
SPSFA algorithm outperformed the traditional FA algorithm with a addressing the 2D and 3D environments. The first one is called the Local
fixed population size. Nevertheless, the SPSFA algorithm is slower than Fast Search with BAS (LFS-BAS) algorithm, which improves the explo-
the traditional FA algorithm [264]. ration and the convergence speed of path finding and can avoid missing
Abbas proposed a hybrid algorithm between the D* algorithm and early solutions. The second one is called ACO-BAS, in which ACO
the FA algorithm, named (the FA-D*) algorithm. The D* algorithm has initializes the population to obtain a fast local optimal path in real-time
been deployed to find the shortest path. The FA algorithm is applied applications. The third one is called Searching Information Orientation
to generate a trajectory of intermediate points in the free space to with BAS (SIO-BAS), which guarantees the stability between speed
achieve the path from the D* algorithm. Finally, a Quadratic parametric and accuracy of the path-finding. The results proved that the three
equation is used to make the path smooth. The algorithm managed variants are better than the original ACO and BAS algorithms. However,
to find collision-free and safe paths. However, the algorithm is not dynamic and complex obstacles are not considered. Moreover, the
compared to any other algorithm in literature [265]. statistical analysis needs to be stronger to judge the performance and
needs to be compared with more algorithms [271].
9.11. Cuckoo Search Algorithm (CSA)
9.13. Teaching-Learning-Based Optimization (TLBO) algorithm
The CSA is a recent nature-based meta-heuristic optimization algo-
rithm introduced by Yang and Deb in 2009 [266]. The CSA algorithm The TLBO Algorithm is a nature-based meta-heuristic optimization
is based on the cuckoo bird behavior, which lays eggs in another algorithm inspired by the idea of learning between teachers and learn-
host bird. The main goal of the cuckoo bird is to mimic the host ers. A population is a group of learners. The TLBO algorithm consists
bird’s eggs and avoid being discovered by the host bird. Hence, the of two phases: the teacher phase, where the learners learn from the
optimization goal is reached if the host bird cannot discover the cuckoo teacher, and the learner phase, where learners learn from each other.
egg. Researchers widely apply the CSA algorithm to solve various The TLBO algorithm was first proposed by Rao et al. in 2011 [272].
optimization problems, including path planning problems. This algorithm is used in many optimization problems.
Alireza et al. designed the Enhanced Mutated Cuckoo Optimization Sabiha et al. designed a TLBO algorithm to address the online path
Algorithm (EMCOA). The experimental results indicated that the EM- planning for the ADS. LiDAR and IMU sensors are used in the percep-
COA algorithm is better than the GA and A* algorithms. However, the tion layer. The TLBO algorithm took the static obstacles and vehicle
statistical analysis needs to be more robust to judge the performance of dynamics into consideration. The algorithm is compared with the PSO,
the algorithms. Local planning is not considered as well [267]. GA, and hybrid GA–PSO algorithms. The robot operating system (ROS)

26
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

Fig. 7. General Statistics about the state-of-the-art path planning algorithms.

results showed the effectiveness of the TLBO algorithm. Nevertheless, environment. The first improvement to the MFO algorithm is the his-
the algorithm did not consider dynamic obstacles. Moreover, it did not torical best flame concept to enhance the ability of the updated law and
deal with obstacles with irregular shapes. The algorithm did not use avoid local minimum. The second improvement is quasi-opposition-
cameras that reduce the system vision [273]. based learning (QOBL) to ensure the diversity of the population. The
results proved that the IMFO algorithm is better than the original
MFO, PSO, GWO, and DA algorithms. However, the algorithm is tested
9.14. Moth–Flame Optimization (MFO) algorithm
only on a maximum of 10 Dimensions benchmark functions. More-
over, the dynamic environment is not considered [274]. Meta-heuristic
The MFO algorithm is a new nature-based meta-heuristic optimiza- optimization algorithms are summarized in Table 10.
tion algorithm inspired by the natural movement mechanism of moths Meta-heuristic optimization Techniques have been widely used in
towards the moon at night. However, if there is any other light source, the literature to address the path planning problems in the ADS system.
this disturbs this movement by attracting the moths to it. This concept These algorithms act as general problem solvers, whatever the complex-
is used as an optimization algorithm in various applications. The main ity of the problem, making them a general template that can adapt to
disadvantage of the algorithm is the slow convergence speed in the late any driving scenario. However, the main challenge of these algorithms
iterations, which leads to a fall in local minimum [275]. is to find the balance between exploring new solutions and exploiting
Dai and Wei proposed the Improved Moth–Flame Optimization known reasonable solutions, which is the key to the efficiency of these
(IMFO) algorithm to solve path planning problems in the ADS in a static algorithms.

27
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

Table 10
Literature Summary for Meta-heuristic Optimization Path Planning techniques.
Genetic Algorithm (GA)
Algorithm Idea Pros Cons
modified Crossover by comparing can generate optimal paths no comparison at all,
MCGA [223] children to parents based on based on the desired no dynamic obstacles.
fitness, the best will survive. conditions.
Hybrid; GA for the global path, Avoid local minimum. optimal path not recalculated,
GA-PF [224]
PF for the local path. not admissible paths/
Differential Evolution (DE)
Algorithm Idea Pros Cons
DE solves the path planning achieved experimental suitable no theoretical statistical analysis,
DE-based [226]
in the MPC model. control performance. no comparison to other algorithms.
Simulated Annealing (SA) algorithm
Algorithm Idea Pros Cons
traditional SA faster solutions Gives high-cost solutions
SA [228]
than PSO and ACO compared to PSO.
Hybrid SA with PSO better than PSO and no dynamic environment,
SAPSO [229]
for better convergence. SA-TLBO algorithms. no numerical statistical analysis.
added neighborhood verified solutions High computational time,
ISA [230] transformation techniques in multi-constraints not tested on
to SA path. high dimensions problems.
Particle Swarm Optimization (PSO) algorithm
Algorithm Idea Pros Cons
Hybrid; PSO generates path, Better than GA PSO. Not applied to complex
SLPSO [232]
Self Learning Alg. improves PSO. 3D environment
Hybrid; ESS generate initial path, managed to solve path Not used for local
ESS-PSO [233]
PSO improves the path. planning with time window. path planning yet.
improved PSO parameters, outperformed PSO and not applied in complex,3D,
ILPSO [234] extended Gaussian distribution is A* algorithms. and large-scale
used for particles’ diversity. environment.
hybrid PSO and FS; outperformed need more investigations
PSOFS [235] if PSO generates unsafe points, PSOA and against more
FS generates safer ones. PSOD algorithms. state-of-the-art algorithms.
added crossover outperformed PSO, No dynamic obstacles,
and mutation operations GA, ACO in not generalized
IPSO [236]
in PSO material transportation to outdoor path
indoors. planning problems.
PSO with satellite unknown environment using satellite neglected other
SPSO [237]
images for mapping images, best variant (SPSO). algorithms (PSO only).
normalized objective function, local outperformed PFPSO and no dynamic obstacles,
EMOPSO [238]
exploration operations ( 𝜖-bias, RS). PFABC algorithms. need a more complex environment.
Hybrid; PSO for global planning, APF for outperformed PSO and no dynamic obstacles, need
IPSO-APF/SA [239]
local planning, SA to optimize APF. MPSO algorithms. more statistical analysis.
Ant Colony Optimization (ACO) algorithm
Algorithm Idea Pros Cons
ACO-A* [241] Hybrid ACO with A*. improved shortcomings of ACO. high computational time.
, Hybrid; A* stores best nodes, ACO optimizes More accurate paths No dynamic objects,
ACO-A*-MDP [242]
the path, MDP smooths the path. high computational time.
optimize ACO parameters IACO reduce congestion rate from No comparison to other
IACO-PSO [243]
using PSO algorithm. 13.63%. to 9.73% compared to ACO. state-of-the-art algorithms.
Hybrid ACS, RRT ; outperformed all too slow when
RRT finds sub-optimal RRT variants. ACS contribution
RRT-ACS [244]
path, ACS optimizes is minimal, no
sub-optimal path. no dynamic environment.
entropy concept for diversity, biased initial outperformed RAS, ACS, no dynamic environment ,
AIACSE [245]
population, diffusion model for balance. PS-ACO, MRCACO algorithms. needs parameter improvements.
Hybrid ACO with PSO, adaptive based fast convergence speed, only compared to ACO for
ACO-PSO [246]
on environment interaction. better than ACO low dimensions problem
hybrid ACO, Dijkstra; outperformed ACO-A* , not used in outdoor,
ACO finds global path, ,and Dijkstra more scenarios
ACO-DA [247]
DA avoids obstacles in algorithms. and statistics
selecting the best path. are required.

(continued on next page)

28
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

Table 10 (continued).
Artificial Bee Colony Optimization (ABC) algorithm
improve balance between exploration outperformed PSO, DE more scenarios need
aABC [249] and exploitation in ABC using and basic ABC algorithms. to be tested in ADS.
the Arrhenius equation.
improve ABC, used adaptive limit parameter, outperform the traditional more complex scenarios
ADL-ABC [161]
Cubic polynomial smooths path. ABC algorithm. need to be tested.
DE search eq. speeds up convergence, Co-GLABC outperformed the no dynamic obstacles,
Co-GLABC [250] crossover handles, dimension-dependent, state-of-the-art algorithms. needs more scenarios.
novel food source setting.
improved ABC finds path, Evolutionary gave shorter path than more complex scenarios
iABC–EP [251]
Programming refines the path. ABC–EP algorithm by 5.75%. need to be tested.
Grey Wolf Optimization (WFO) algorithm
optimized parameters by the outperformed GWO, CSO, no dynamic obstacles,
IGWO [253] lion optimizer, dynamic BOA, WOA algorithms. need more complex
weights for diversity. environment (> 30 Dim.).
Hybrid RRT with ExGWO; outperformed all other no benchmark test,
ARRT-ExGWO [254] RRT avoids obstacles, GWO algorithms on no dynamic and
finds the optimal path. 4 scenarios. 3D environment.
Hybrid GWO, APF; APF finds outperformed (SPS) no dynamic environment,
GWO-APF [256] obstacle-free locations, algorithm in [255]. no complex benchmark
GWO optimizes the path. functions.
Chicken Swarm Optimization (CSO) algorithm
improved CSO by adding Levy flight and better than PSO and tested on few benchmark functions,
ICSO [258]
nonlinear weight reduction to update rules. original CSO algorithms. and on only one scenario.
Cat Swarm Optimization algorithm (CSOA)
MOCMCSO- hybrid; MOCMCSO to find path better than MOCSO and repeated points in the
APFM [260] with Cauchy Mutation, MOPSO algorithms. inspection case,
APFM to avoid obstacles. low anti-collision.
Firefly Algorithm (FA)
step factor 𝛼 becomes adaptive, outperformed GA and More scenarios and
MFA [262]
convergence parameter is used 𝛽0 . FA algorithms. benchmark is needed.
Applied traditional FA Managed to find No comparison, performance
FA [263]
to multiple scenarios. acceptable paths. needs to be improved.
population size is adaptive to collision outperformed traditional slower than the
SPSFA [264]
degree using two nonlinear functions. FA algorithm. traditional FA algorithm.
D* finds path, FA finds intermediate managed to find not compared to the
FA-D* [265]
path, Quadratic fn. smooths path. collision-free path. state-of-the-art algorithms.
Cuckoo Search Algorithm (CSA)
added mutation operator better than weak statistical analysis
EMCOA [267]
for diversity GA and A* no local planning
added 2-opt move, crossover, Solved TSP and path Not tested in path planning
DDCS [80]
random-key encoding to CSA planning in indoor environment. in outdoor environment.
Beetle Antennae Search (BAS) algorithm
Improved BSO; added penalty term, better than PSO poor statistical
IBSO [160] to fitness fun., cubic spline algorithm by 90% analysis.
interpolation for a smooth path.
hybrid BAS; LFS to find early path, better than BAS no dynamic, complex obstacles,
LFS/ACO/SIO [271]
ACO for a fast path, SIO for a stable path. and ACO algorithms. poor statistics.
Teaching Learning Based Optimization (TLBO) algorithm
applied traditional outperformed GA, PSO no dynamic
TLBO [273]
TLBO in static environment irregular obstacles
Moth Flaming Optimization (MFO) algorithm
Improved MFO; historical best flame, better than MFO, PSO, only max. of 10 Dim.
IMFO [274] quasi-opposition based GWO, and DA algorithms. functions,
learning concepts are added. no dynamic environment.

Most research direction tackles this balance problem via adaptive 10. Statistics of the review paper
rather than static parameters. This adaptive method will eliminate
the need to tune the parameters before applying the algorithm to This section explores the numerical statistics about the state-of-
new driving scenarios. Moreover, the adaptive method will add the the-art publications of the path planning algorithms in the ADS. The
proper balance to the algorithm, leading to an optimal path. Further- number of references used in the path planning problem is 162 ref-
more, hybridization with some local search techniques can improve the erence. There are five metrics used to categorize these references:
convergence speed, making them suitable choices for real-life driving publisher, publication type, year, hybridization, and category of the
scenarios. algorithm.

29
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

Fig. 7. (continued).

10.1. General statistics Fig. 8b demonstrates the algorithm hybridization percentages with
time. In 2018, 14.29% of the algorithms were hybrid, while 85.71%
The general statistics show the percentage of each metric. Fig. 7 were not hybrid. In 2019, 21.05% of the algorithms were hybrid, while
presents the general statistics about the state-of-the-art path planning 78.95% were not hybrid. In 2020, 32.65% of the algorithms are hybrid,
algorithms in the ADS discussed in this paper. Concerning the publish- while 67.35% are not hybrid. In 2021, 25% of the algorithms are
ers, 56% of the references were published in IEEE, 25% were published hybrid, while 75% are not hybrid. In 2022, 31.25% of the algorithms
in Springer, 12% were published in Elsevier, and 7% were published are hybrid, while 68.75% are not hybrid.
via other publishers such as arXiv, AAAI, Hindawi, and SAGE UK, as Fig. 8c shows the publisher percentage with time. In 2018, 78.57%
shown in Fig. 7a. Fig. 7b shows that 75% of the publications are journal of the publications are in IEEE, 14.29% are in Springer, and 7.14% are
papers, while 25% are conference papers. Concerning the publication in Elsevier. The percentage of publishers in 2019 is 73.86%, 5.26%,
year, Fig. 7c represents that 9% of the publications were published in 15.79%, and 5.09% in IEEE, Springer, Elsevier, and others, respec-
2018, 12% in 2019, 30% in 2020, 29% in 2021, and 20% in 2022. tively. The publisher percentages in 2020 are 59.18%, 16.32%, 10.20%,
Concerning algorithm hybridization, Fig. 7d shows that 27% of and 14.30% in IEEE, Springer, Elsevier, and others, respectively. In
the algorithms are hybrid with another algorithm of an elementary 2021, the publishers’ distribution was 54.17%, 29.17%, 10.42%, and
category. On the other hand, 73% of the algorithms are not hybrid with 6.24% in IEEE, Springer, Elsevier, and others, respectively. In 2022, the
another primary algorithm, but they may include some concepts from publisher distribution was 31.25%, 50%, 15.63%, and 3.12% in IEEE,
another algorithm embedded in the main steps. Fig. 7e demonstrates Springer, Elsevier, and others, respectively.
the percentages of algorithm categories as follows: 7% of the algo- Fig. 8d presents the algorithm types percentages with time. In 2018,
rithms are graph-based, 14% are sampling-based, 7% gradient-based, 7.14% of the algorithms are graph-based, 21.43% are sampling-based,
8% optimization-based, 16% interpolation-based, 6% are supervised 7.14% gradient-based, 14.29% optimization-based, 7.14%
learning-based, 19% are reinforcement learning-based, and 23% are interpolation-based, 7.14% are supervised learning-based, 14.29% are
meta-heuristic optimization-based. reinforcement learning-based, and 21.43% are meta-heuristic
optimization-based. In 2019, 10.53% of the algorithms are graph-
10.2. Chronicle order based, 5.26% are sampling-based, 15.79% optimization-based, 36.84%
interpolation-based, 10.53% are reinforcement learning-based, and
This section explores the metrics from a chronicle point-of-view as 21.05% are meta-heuristic optimization-based.
shown in Fig. 8. Fig. 8a shows the publication type percentages with In 2020, 6.12% of the algorithms are graph-based, 12.25% are
time. In 2018, 57.14% of the publications were journal papers, while sampling-based, 4.08% gradient-based, 8.16% optimization-based,
42.86% were conference papers. In 2019, 63.16% of the publications 16.33% interpolation-based, 6.12% are supervised learning-based,
were journal papers, while 36.84% were conference papers. In 2020, 18.37% are reinforcement learning-based, and 28.57% are
67.45% of the publications were journal papers, while 32.65% were meta-heuristic optimization-based. In 2021, 6.25% of the algorithms
conference papers. In 2021, 81.25% of the publications were journal are graph-based, 10.42% are sampling-based, 10.42% gradient-based,
papers, while 18.75% were conference papers. In 2022, 93.75% of the 4.17% optimization-based, 12.50% interpolation-based, 8.33% are su-
publications were journal papers, while 6.25% were conference papers. pervised learning-based, 22.91% are reinforcement learning-based, and

30
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

Fig. 8. Chronicle statistics of the state-of-the-art path planning algorithms.

25% are meta-heuristic optimization-based. In 2022, 9.38% of the algo- published in 2019, 25% were published in 2020, 25% were published
rithms are graph-based, 21.88% are sampling-based, 9.38% gradient- in 2021, and 25% were published in 2022. For sampling-based algo-
based, 6.25% optimization-based, 15.63% interpolation-based, 3.13% rithms, 13.64% were published in 2018, 4.55% were published in 2019,
are supervised learning-based, 18.75% are reinforcement learning- 27.27% were published in 2020, 22.73% were published in 2021, and
based, and 15.60% are meta-heuristic optimization-based. 31.81% were published in 2022.
For gradient-based algorithms, 9.09% were published in 2018,
10.3. Algorithm category point of view 18.18% in 2020, 45.45% in 2021, and 27.28% in 2022. For
optimization-based algorithms, 15.38% were published in 2018,
This section discusses the metrics from the algorithm category point- 23.08% in 2019, 30.77% in 2020, 15.38% in 2021, and 15.39% in
of-view as shown in Fig. 9. Fig. 9a demonstrates the publication year 2022. For interpolation-based algorithms, 3.70% of them were pub-
percentage for each algorithm type. For graph-based algorithms, 8.33% lished in 2018, 25.93% in 2019, 29.63% in 2020, 22.22% in 2021,
of the graph-based algorithms were published in 2018, 16.67% were and 18.52% in 2022. For supervised learning-based algorithms, 11.11%

31
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

Fig. 8. (continued).

were published in 2018, 33.33% in 2020, 44.44% in 2021, and 11.12% in Elsevier. For optimization-based algorithms, 76.92% are published
in 2022. For reinforcement learning-based algorithms, 6.67% of them in IEEE, 15.38% in Springer, and 7.70% in Elsevier.
were published in 2018, 6.67% in 2019, 30% in 2020, 36.67% in For interpolation-based algorithms, 48.15% are published in IEEE,
2021, and 19.99% in 2022. For meta-heuristic optimization algorithms, 33.33% in Springer, 11.11% in Elsevier, and 7.41% in other publishers.
7.89% of them were published in 2018, 10.53% in 2019, 36.84% in For supervised learning-based algorithms, 44.44% of them are pub-
2020, 31.58% in 2021, and 13.15% in 2022. lished in IEEE, 33.33% in Springer, 11.11% in Elsevier, and 11.12% in
other publishers. For reinforcement learning-based algorithms, 56.67%
Fig. 9b demonstrates the publisher percentage for each algorithm are published in IEEE, 26.67% in Springer, 13.33% in Elsevier, and
type. For graph-based algorithms, 58.33% are published in IEEE, 25% 3.33% in other publishers. For meta-heuristic optimization algorithms,
in Springer, and 16.67% in other publishers. For sampling-based al- 52.63% of them are published in IEEE, 23.68% in Springer, 15.79% in
gorithms, 50% are published in IEEE, 22.73% in Springer, 13.64% Elsevier, and 7.90% in other publishers.
in Elsevier, and 13.63% in other publishers. For gradient-based algo- Fig. 9c explores the publication type percentage for each algorithm
rithms, 72.73% are published in IEEE, 18.18% in Springer, and 9.09% type. For graph-based algorithms, 75% of the publications are journal

32
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

papers, while 25% are conference papers. For sampling-based algo- the main challenge is the need for vast, complex training data to build
rithms, 72.72% of the publications are journal papers, while 27.28% reliable models. Moreover, the need to re-train the whole model to
are conference papers. For gradient-based algorithms, 90.9% of the make any updates to the model is too time-consuming. The future
publications are journal papers, while 9.1% are conference papers. For of Deep Supervised Learning Techniques should focus on enhancing
optimization-based algorithms, 46.15% of the publications are journal real-time processing capabilities by using pre-trained models with the
papers, while 53.85% are conference papers. capability of real-time training, which reduces reliance on extensive
For interpolation-based algorithms, 77.78% of the publications are training datasets.
journal papers, while 22.22% are conference papers. For supervised Reinforcement learning techniques have recently been deployed to
learning-based algorithms, 66.67% of the publications are journal pa- provide optimal and fast solutions for complex search spaces in the
pers, while 33.33% are conference papers. For reinforcement learning- long term. Unlike traditional supervised learning, they do not need
based algorithms, 73.33% of the publications are journal papers, while training data to learn. They interact with the environment to update
26.67% are conference papers. For meta-heuristic optimization algo- the policy. The main challenge of traditional reinforcement learning is
rithms, 84.21% of the publications are journal papers, while 15.79% that generating a state–action table will be too massive for complex
are conference papers. problems, which require more memory and high computational time
Fig. 9d explores the hybridization percentage for each algorithm to search for suitable action.
type. For graph-based algorithms, 16.67% are hybrid, while 83.33% Deep Reinforcement learning can best use neural networks as a
are not hybrid. Only 4.55% of the sampling-based algorithms are function approximator instead of storing a vast state–action Q-table.
hybrid. 18.18% of gradient-based algorithms are hybrid. 46.15% of However, this adds more parameter tuning complexity to train the
optimization-based algorithms are hybrid. 44.4% of interpolation-based neural network, which is too time-consuming. The future of Deep
algorithms are hybrid. 100% of the supervised learning algorithms Reinforcement Learning techniques should emphasize using pre-trained
are not hybrid. 20% of reinforcement learning algorithms are hybrid. deep Q networks, which can make it faster; however, it is still challeng-
39.47% of meta-heuristic optimization algorithms are hybrid. ing.
Meta-heuristic optimization techniques such as GA, ABC, PSO, SA,
11. Conclusion ACO, and CSA have recently been widely used to solve path-planning
problems, 23% of the literature. The main advantage of meta-heuristic
This research has intensively explored the recent state-of-the-art optimization algorithms is that they are considered general problem-
algorithms in the AD system. The stages of the AD system have been solver and problem-independent. They can give an optimal solution for
discussed in detail, starting from sensors, perception, localization, map- any complex search space. Therefore, it can be applied to any problem
ping, risk assessment, and path planning. The stages of the AD system with few modifications. Furthermore, they rarely get stuck in the local
are well-researched. However, path planning is still one of the most minimum because of the ability to explore different search spaces and
challenging problems in the AD system [7]. utilize local search spaces. Nevertheless, they can provide different
There are three main groups of path-planning algorithms: Tradi- paths for the same problem if run multiple times. The reason behind
tional, Machine learning, and Meta-heuristic optimization techniques. this problem is the random behavior in specific stages of the algorithms.
The main challenge of these algorithms is the parameter tuning to
The most common traditional techniques are graph-based search meth-
balance exploitation and exploration. The most trending methodology
ods such as Dijkstra and A* algorithms. These methods give accurate
to address this challenge is using adaptive rather than static parameters;
solutions for limited search space. However, they are very slow for large
this adds the appropriate balance between global and local search.
search spaces and give jerky routes. Future research in graph-based
In this study, we claim that we proposed a comprehensive and
algorithms could focus on hybridizing them with sampling-based and
extensive review with high algorithm diversity in the path planning
optimization techniques for enhanced search capability for complex
problem in the ADS. This paper can mark a starting point for future
search spaces.
research on solving the path planning problem. This review can give
The sampling-based methods such as RRT have been used to solve
an intuition about the most recent techniques used in solving the
the path planning problem. These algorithms are faster than graph
path planning problem of the ADS. Based on the cons of the recent
search, but they give jerky solutions. Improving Sampling-Based Plan-
state-of-the-art algorithms, new ideas can be generated to solve the
ning methods should involve integrating optimization and interpolation
path planning problem more efficiently. Therefore, as described in this
techniques for enhanced efficiency and smoothness.
paper, the research community can start from where the literature
A gradient-based algorithm such as the APF algorithm has been used
ended.
to address the trajectory planning in the ADS. The main merit of this
technique is that it can produce collision-free trajectories within a small CRediT authorship contribution statement
computation time. Nevertheless, the main demerit of the algorithm is
the high possibility of falling into a local minimum because of the gradi- Mohamed Reda: Conceptualization, Data curation, Formal analy-
ent behavior. The future direction for Gradient-Based methods should sis, Methodology, Software, Validation, Visualization, Writing – origi-
focus on overcoming the challenge of local minima via hybridization nal draft. Ahmed Onsy: Conceptualization, Funding acquisition, Inves-
with global search methods that can add diversity to the solutions. tigation, Project administration, Resources, Supervision, Writing – re-
Interpolation-based algorithms should primarily be utilized to view & editing. Ali Ghanbari: Investigation, Project administration, Su-
smooth out paths generated by other methods. However, they are too pervision, Writing – review & editing. Amira Y. Haikal: Investigation,
slow and have high computational time if used as an independent Supervision, Writing – review & editing.
technique. Numerical optimization methods such as quadratic pro-
gramming can deal with complex search spaces and give high-quality Declaration of competing interest
solutions but have high computational time. Future developments in
optimization-based methods should reduce computational complexity The authors declare that they have no known competing finan-
using appropriate optimizers, making them problem-dependent. cial interests or personal relationships that could have appeared to
Supervised and deep learning techniques have been deployed in two influence the work reported in this paper.
manners: they can be used only in the perception layer to deal with
images and sensory data, or they can be used as an End-To-End driving Data availability
scheme to include the path-planning phase. It can give a high-speed
solution for familiar scenarios during the application phase. However, No data was used for the research described in the article.

33
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

Fig. 9. Algorithm prospective of the state-of-the-art path planning algorithms.

34
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

Fig. 9. (continued).

35
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

Table 11
List of abbreviations.
Symbol Acronym/Abbreviation
A*ESS A* algorithm with Equal-Step Sampling
A*-PDWA A* algorithm and Predictive-Dynamic Window Approach
aABC Arrhenius Artificial Bee Colony
AAIRL Augmented Adversarial Inverse Reinforcement Learning
ABC Artificial Bee Colony
ACO Ant Colony Optimization
ACRL Actor-Critic Reinforcement Learning
ACS Ant Colony System
ADAPF Azimuth and Distance Artificial Potential Field
ADL-ABC Adaptive Dimension Limit-Artificial Bee Colony
ADP Adaptive Dynamic Programming
ADPF-PP Adaptive Potential Field for Path Planning
ADS Automated driving system
AHPFs Artificial Harmonic Potential Fields
AI Artificial Intelligence
AIACSE Adaptive Improved Ant Colony System algorithm based on population information Entropy
ANN Artificial Neural Networks
APF Artificial Potential Field
ARRT-ExGWO Adapted Rapidly-exploring Random Tree with Extended Grey Wolf Optimization
ARRT-GWO Adapted Rapidly-exploring Random Tree with Grey Wolf Optimization
ASMP-NCSS Adaptive Sampling-based Motion Planning with a Non-Conservatively Defensive Strategy
BAS Beetle Antennae Search
BAS-PSO Beetle Antennae Search and Particle Swarm Optimization
BB Bernstein–Bézier
BC-based Bezier Curve-based
BCO Bézier Curve Optimization
BFO Bacterial Foraging Optimization
Bi-RRT Bidirectional Rapidly-exploring Random Tree
BOA Butterfly Optimization Algorithm
BSAS Beetle swarm antennae search
BSCO B-Spline Curve Optimization
BSO Beetle Swarm Optimization
CBA Clothoid-Based Algorithm
CCO Clothoid Curves Optimization
CDQN Conditional Deep Q-network
CDT Constrained Delaunay Triangulation
CNM Complete Node Mechanism
CNN Convolution Neural Networks
Co-GLABC Coevolution framework with the Global best Leading Artificial Bee Colony
C-PDF Custom Probability Density Function
CSA Cuckoo Search Algorithm
CSO Chicken Swarm Optimization
CSOA Cat Swarm Optimization Algorithm
CVAE Conditional Variational Encoder
DC dynamic cell
DCNN Deep Convolutional Neural Network
DDCS Discrete Damped Cuckoo Search
DDPG Deep Deterministic Policy Gradient
DDQN Double Deep Q-network
DDQN-FSLAM Dueling Deep Q-Network with a Fast active SLAM
DE Differential Evolution
DL Deep Learning
DNN Deep Neural Network
DPCC Dynamic Programming-based algorithm integrated with Clothoid Curve
DQN Deep Q-Network
DRL Deep Reinforcement Learning
D-RRT* Directional Rapidly-exploring Random Trees*
DSUNet-PP Depth-wise Separable UNet for Path Prediction
DVD-CC Direct Visibility Diagram Algorithm with Clothoid Curves
DWA Dynamic Window Approach
DynEFWA-APF Dynamic Enhanced Firework Algorithm with Artificial Potential Field
EBHS Experience-Based Heuristic-Search
EES-PSO Evolutionary Scatter Search Particle Swarm Optimization
EMCOA Enhanced Mutated Cuckoo Optimization Algorithm
EMOPSP Enhanced Multi-Objective Particle Swarm Optimization
ERRT Enhanced Rapidly-exploring Random Tree
ET-MPADP Event-Triggered Model Predictive Adaptive Dynamic Programming
FA Firefly Algorithm
FA-D* Firefly algorithm and the D* algorithm
FDA-DP Feasible Direction Algorithm integrated with Dynamic Programming (
FDP Forward Dynamic Programming
FFNN Feed Forward Neural Networks
FMCW Frequency-Modulated Continuous Wave
FOBC Fourth-Order Bézier Curves
(continued on next page)

36
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

Table 11 (continued).
Symbol Acronym/Abbreviation
GA Genetic Algorithm
GA-PF Genetic Algorithm-Potential Field
GARL Graph Attention Reinforcement Learning
GDA Gradient Descent Algorithm
GDBC Gradient Descent and B∖’ezier Curve algorithm
GDM Gradient Descent Method
GPS-IMU Global Positioning System and Inertial Measurement Unit
GWO Grey Wolf Optimization
HA* Hybrid A*
HARL Hybrid A* Algorithm and Reinforcement Learning
hDDQN Hierarchical Double Deep Q-learning
HDM Human Driver Model
HDM-MIQP Human Driver Model integrated with Mixed-integer quadratic programming
HDR High Dynamic Range
HIL hardware-in-loop
HS Harmony Search
IA*-DWA Improved A* algorithm and Dynamic Window Approach (DWA)
IA*FC Improved A* based on the Fuel Consumption
iABC–EP Improved Artificial Bee Colony with Evolutionary Programming
IACO-PSO Improved Ant Colony Algorithm optimized by Particle Swarm Algorithm
iADA* Improved Anytime Dynamic A*
IAPF Improved Artificial Potential Field
IAPF-GDM Improved Artificial Potential Field and Gradient Descent Method
IBSO Improved Beetle Swarm Optimization
ICP Iterative closest point
ICSO Improved Chicken swarm optimization
IDQNPER-ETE Deep Reinforcement Learning in an End-To-End driving scheme
IDWAQ Improved Dynamic Window Approach based on the Q-Learning
IGWO Improved Grey Wolf Optimization
ILPSO Improved Localized Particle Swarm Optimization
ILS Iterated Local Search
IMFO Improved Moth–Flame Optimization
IPF-PP Improved Potential Field-based Path Planning
IPSO Improved Particle Swarm Optimization
IPSO-APF/SA Improved PSO, Improved Artificial Potential Field , and Simulated Annealing algorithm
IQL Improved Q-learning
IRL Integral Reinforcement Learning
IRRT Improved Rapidly-exploring Random Trees
i-RRT improved Rapidly-exploring Random Trees
IRRT-PI Improved Rapidly-exploring Random Tree associated with PI controller
ISA Improved Simulated Annealing
JPS Jump Point Search
JQBC Joint Quadratic Bézier Curves
KB-RRT* Kinematic Constrained Bi-directional Rapidly-exploring Random Tree
LFS-BAS Local Fast Search with Beetle Antennae Search
LiDAR Light Detection And Ranging
LLC Longitudinal and Lateral planning with Clothoid interpolation
LSAC-CPP Lyapunov-based Soft Actor–Critics with Collision Probability Prediction
LSPP Path Planning with a Line Segment
LSTM long short-term memory
LTSTP Long-Term and Short-Term Planner
MCAL-P Mobile robot Collision Avoidance Learning with Path
MCGA Modified Crossover Genetic Algorithm
MCQ Monte Carlo Q-learning
MCS/AC Monte Carlo Search and actor-critic
MDL Modified Dogleg
MDP Markov decision process
MEP-DIRL Maximum Entropy Deep Inverse Reinforcement Learning algorithm
MFA Modified Firefly Algorithm
MFO Moth–Flame Optimization
ML Machine Learning
MOCMCSO Multi-Objective Cauchy Mutation Cat Swarm
MPC Model Predictive Control
MPPP Model Predictive Path Planning
MPSO Mutation Particle Swarm Optimization
FDA-DP Feasible Direction Algorithm integrated with Dynamic Programming (
FDP Forward Dynamic Programming
FFNN Feed Forward Neural Networks
FMCW Frequency-Modulated Continuous Wave
FOBC Fourth-Order Bézier Curves
GA Genetic Algorithm
GA-PF Genetic Algorithm-Potential Field
GARL Graph Attention Reinforcement Learning
GDA Gradient Descent Algorithm
(continued on next page)

37
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

Table 11 (continued).
Symbol Acronym/Abbreviation
GDBC Gradient Descent and B∖’ezier Curve algorithm
GDM Gradient Descent Method
GPS-IMU Global Positioning System and Inertial Measurement Unit
GWO Grey Wolf Optimization
HA* Hybrid A*
HARL Hybrid A* Algorithm and Reinforcement Learning
hDDQN Hierarchical Double Deep Q-learning
HDM Human Driver Model
HDM-MIQP Human Driver Model integrated with Mixed-integer quadratic programming
HDR High Dynamic Range
HIL hardware-in-loop
HS Harmony Search
IA*-DWA Improved A* algorithm and Dynamic Window Approach (DWA)
IA*FC Improved A* based on the Fuel Consumption
iABC–EP Improved Artificial Bee Colony with Evolutionary Programming
IACO-PSO Improved Ant Colony Algorithm optimized by Particle Swarm Algorithm
iADA* Improved Anytime Dynamic A*
IAPF Improved Artificial Potential Field
IAPF-GDM Improved Artificial Potential Field and Gradient Descent Method
IBSO Improved Beetle Swarm Optimization
ICP Iterative closest point
ICSO Improved Chicken swarm optimization
IDQNPER-ETE Deep Reinforcement Learning in an End-To-End driving scheme
IDWAQ Improved Dynamic Window Approach based on the Q-Learning
IGWO Improved Grey Wolf Optimization
ILPSO Improved Localized Particle Swarm Optimization
ILS Iterated Local Search
IMFO Improved Moth–Flame Optimization
IPF-PP Improved Potential Field-based Path Planning
IPSO Improved Particle Swarm Optimization
IPSO-APF/SA Improved PSO, Improved Artificial Potential Field , and Simulated Annealing algorithm
IQL Improved Q-learning
IRL Integral Reinforcement Learning
IRRT Improved Rapidly-exploring Random Trees
i-RRT Improved Rapidly-exploring Random Trees
IRRT-PI Improved Rapidly-exploring Random Tree associated with PI controller
ISA Improved Simulated Annealing
JPS Jump Point Search
JQBC Joint Quadratic Bézier Curves
KB-RRT* Kinematic Constrained Bi-directional Rapidly-exploring Random Tree
LFS-BAS Local Fast Search with Beetle Antennae Search
LiDAR Light Detection And Ranging
LLC Longitudinal and Lateral planning with Clothoid interpolation
LSAC-CPP Lyapunov-based Soft Actor–Critics with Collision Probability Prediction
LSPP Path Planning with a Line Segment
LSTM Long short-term memory
LTSTP Long-Term and Short-Term Planner
MCAL-P Mobile robot Collision Avoidance Learning with Path
MCGA Modified Crossover Genetic Algorithm
MCQ Monte Carlo Q-learning
MCS/AC Monte Carlo Search and actor-critic
MDL Modified Dogleg
MDP Markov decision process
MEP-DIRL Maximum Entropy Deep Inverse Reinforcement Learning algorithm
MFA Modified Firefly Algorithm
MFO Moth–Flame Optimization
ML Machine Learning
MOCMCSO Multi-Objective Cauchy Mutation Cat Swarm
MPC Model Predictive Control
MPPP Model Predictive Path Planning
MPSO Mutation Particle Swarm Optimization
MRCACO Multi-Role Adaptive Collaborative Ant Colony Optimization
MRRT* Modified Rapidly-exploring Random Trees*
MSM Minimum Snap Method
NDT Normal distributions transform
NHTSA National Highway Traffic Safety Administration
NOC Nonlinear Optimal Control
NRRT* Neural Rapidly-exploring Random Tree*
ODDs Operational design domains
OSQP Open Source Quadratic Programming
PAPF Predictive Artificial Potential Field
PCAPF Polynomial Curve and Artificial Potential Field
PCC Parameterized Curvature Control
P-DWA Predictive-Dynamic Window Approach
PF Potential Field
(continued on next page)

38
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

Table 11 (continued).
Symbol Acronym/Abbreviation
PFABC Potential Field Artificial Bee Colony
PFPSO Potential Field Particle Swarm Optimization
PJM Piecewise Jerk Method
PRM Probabilistic Roadmap method
PRRT Pruning Rabidly-exploring Random Tree
PSO particle swarm optimization
QLOR Q-learning algorithm and Linear Output Regulation algorithm
QOBL Quasi-opposition-based learning
QP Quadratic programming
QPO Quadratic Programming Optimization
QTBC Quintic Trigonometric Bézier Curve
RAS Rank-based Ant System
RHRL Robust Hierarchical Reinforcement Learning
RL Reinforcement Learning
RLDA reverse labeling Dijkstra algorithm
RL-MOHH Reinforcement Learning based Multi-Objective Hyper-Heuristic
RNN Recurrent Neural Networks
ROS Robot operating system
RPN Region Proposal Networks
RRT Rapidly-exploring Random Tree
RRT-CPDF Rapidly-exploring Random Tree with a Custom Probability Density Function
RRT-DWA Random Tree algorithm and the Dynamic Window Approach algorithm
RTQL Real-time Q-learning
SA Simulated Annealing
SAC Soft Actor–Critics
SAE Society of Automotive Engineers
SAPSO Simulated Annealing algorithm and the Particle Swarm Optimization
SA-TLBO Simulated Annealing Teaching Learning Based Optimization
SBDP-QP Sampling-based algorithm, Dynamic Programming, and Quadratic Programming
SBP Sampling-Based Planning
SDGT-PC Stackelberg Differential Game Theory and Polynomial Curve
SEHS Space Exploration Guided Heuristic Search
SIO-BAS Searching Information Orientation with Beetle Antennae Search
SLAM Simultaneous Localization And Mapping
S-NO Sampling and Numerical Optimization algorithm
SPS Surrounding Point Set
SPSO Satellite Particle Swarm Optimization
SQP Sequential Quadratic Programming
SR1 Symmetric Rank-1
SSD Single Shot Detector
SUMO Simulation of Urban Mobility
TCSC Tentacle algorithm and the B-Spline Curve algorithm
TEB Timed Elastic Band
TFA Trajectory Following Approach
THA* Traversability Hybrid A*
TLBO Teaching-Learning-Based Optimization
TS Tabu Search
TSQP Two-Step Quadratic Programming
TT-MPADP Time-triggered Model Predictive Adaptive Dynamic Programming
UAV Unmanned Aerial Vehicle
U-PDF Uniform Probability Density Function
VD Visibility Diagram
VDQN Vanilla Deep Q-Network
VFH- Vector Field Histogram-
vSLAM visual Simultaneous Localization And Mapping
WOA Whale Optimizer Algorithm
YOLO You Only Look Once

Appendix. List of abbreviations [4] J. Zhao, W. Zhao, B. Deng, Z. Wang, F. Zhang, W. Zheng, W. Cao, J. Nan, Y.
Lian, A.F. Burke, Autonomous driving system: A comprehensive survey, Expert
Syst. Appl. (2023) 122836.
See Table 11.
[5] Kent County Council, Crash and Casualty Data, Tech. Rep., Kent County
Council, 2022, https://www.kent.gov.uk/roads-and-travel/road-safety/crash-
References and-casualty-data. (Online; accessed 28 October 2022).
[6] T.J. Crayton, B.M. Meier, Autonomous vehicles: Developing a public health
[1] É. Zablocki, H. Ben-Younes, P. Pérez, M. Cord, Explainability of deep vision- research agenda to frame the future of transportation policy, J. Transp. Health
based autonomous driving systems: Review and challenges, Int. J. Comput. Vis. 6 (2017) 245–252.
130 (10) (2022) 2425–2452. [7] E. Yurtsever, J. Lambert, A. Carballo, K. Takeda, A survey of autonomous
[2] C. Badue, R. Guidolini, R.V. Carneiro, P. Azevedo, V.B. Cardoso, A. Forechi, driving: Common practices and emerging technologies, IEEE Access 8 (2020)
L. Jesus, R. Berriel, T.M. Paixao, F. Mutz, et al., Self-driving cars: A survey, 58443–58469.
Expert Syst. Appl. 165 (2021) 113816. [8] M. Maurer, J.C. Gerdes, B. Lenz, H. Winner, Autonomous Driving: Technical,
[3] A. Chougule, V. Chamola, A. Sam, F.R. Yu, B. Sikdar, A comprehensive review Legal and Social Aspects, Springer Nature, 2016.
on limitations of autonomous driving and its impact on accidents and collisions, [9] S. Kuutti, S. Fallah, K. Katsaros, M. Dianati, F. Mccullough, A. Mouzakitis,
IEEE Open J. Veh. Technol. (2023) 1–20, http://dx.doi.org/10.1109/OJVT. A survey of the state-of-the-art localization techniques and their potentials for
2023.3335180. autonomous vehicle applications, IEEE Internet Things J. 5 (2) (2018) 829–846.

39
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

[10] D. Gruyer, S. Demmel, V. Magnier, R. Belaroussi, Multi-hypotheses tracking [37] J. Hasch, Driving towards 2020: Automotive radar technology trends, in: 2015
using the Dempster–Shafer theory, application to ambiguous road context, Inf. IEEE MTT-S International Conference on Microwaves for Intelligent Mobility,
Fusion 29 (2016) 40–56. ICMIM, IEEE, 2015, pp. 1–4.
[11] C. Gold, M. Körber, D. Lechner, K. Bengler, Taking over control from highly [38] D. Kissinger, Millimeter-Wave Receiver Concepts for 77 GHz Automotive Radar
automated vehicles in complex traffic situations: The role of traffic density, in Silicon-Germanium Technology, Springer Science & Business Media, 2012.
Human Factors 58 (4) (2016) 642–652. [39] M.E. Warren, Automotive LIDAR technology, in: 2019 Symposium on VLSI
[12] D. González, J. Pérez, V. Milanés, F. Nashashibi, A review of motion planning Circuits, IEEE, 2019, pp. C254–C255.
techniques for automated vehicles, IEEE Trans. Intell. Transp. Syst. 17 (4) [40] M. Wang, W. Liu, Y. Lu, X. Zhao, B. Song, Y. Zhang, Y. Wang, C. Lian, J. Chen,
(2015) 1135–1145. Y. Cheng, et al., Study on the measurement of the atmospheric extinction of
[13] L. Claussmann, M. Revilloud, D. Gruyer, S. Glaser, A review of motion planning fog and rain by forward-scattering near infrared spectroscopy, Guang Pu Xue
for highway autonomous driving, IEEE Trans. Intell. Transp. Syst. 21 (5) (2019) Yu Guang Pu Fen Xi=Guang Pu 28 (8) (2008) 1776–1780.
[41] P. Radecki, M. Campbell, K. Matzen, All weather perception: Joint data
1826–1848.
association, tracking, and classification for autonomous ground vehicles, 2016,
[14] S. Aradi, Survey of deep reinforcement learning for motion planning of
arXiv preprint arXiv:1605.02196.
autonomous vehicles, IEEE Trans. Intell. Transp. Syst. (2020).
[42] J. Redmon, S. Divvala, R. Girshick, A. Farhadi, You only look once: Unified,
[15] A. Yoganandhan, S. Subhash, J.H. Jothi, V. Mohanavel, Fundamentals and
real-time object detection, in: Proceedings of the IEEE Conference on Computer
development of self-driving cars, Mater. Today: Proc. 33 (2020) 3303–3310.
Vision and Pattern Recognition, 2016, pp. 779–788.
[16] B. Hadi, A. Khosravi, P. Sarhadi, A review of the path planning and formation [43] J. Redmon, A. Farhadi, Yolov3: An incremental improvement, 2018, arXiv
control for multiple autonomous underwater vehicles, J. Intell. Robot. Syst. 101 preprint arXiv:1804.02767.
(4) (2021) 1–26. [44] J. Redmon, A. Farhadi, YOLO9000: Better, faster, stronger, in: Proceedings of
[17] C. Zhou, B. Huang, P. Fränti, A review of motion planning algorithms for the IEEE Conference on Computer Vision and Pattern Recognition, 2017, pp.
intelligent robots, J. Intell. Manuf. (2021) 1–38. 7263–7271.
[18] A. Puente-Castro, D. Rivero, A. Pazos, E. Fernandez-Blanco, A review of artificial [45] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y. Fu, A.C. Berg, Ssd:
intelligence applied to path planning in UAV swarms, Neural Comput. Appl. Single shot multibox detector, in: European Conference on Computer Vision,
(2021) 1–18. Springer, 2016, pp. 21–37.
[19] Y. Huang, D. Chen, Research progress of automatic driving path planning, [46] R. Rampriya, R. Suganya, et al., RSNet: Rail semantic segmentation network
in: 2021 2nd International Conference on Artificial Intelligence and Computer for extracting aerial railroad images, J. Intell. Fuzzy Systems 41 (2) (2021)
Engineering, ICAICE, IEEE, 2021, pp. 95–99. 4051–4068.
[20] A.K. Tyagi, S. Aswathy, Autonomous intelligent vehicles (AIV): Research state- [47] H. Zhao, J. Shi, X. Qi, X. Wang, J. Jia, Pyramid scene parsing network,
ments, open issues, challenges and road for future, Int. J. Intell. Netw. 2 (2021) in: Proceedings of the IEEE Conference on Computer Vision and Pattern
83–102. Recognition, 2017, pp. 2881–2890.
[21] A. Vagale, R. Oucheikh, R.T. Bye, O.L. Osen, T.I. Fossen, Path planning and [48] L.-C. Chen, Y. Zhu, G. Papandreou, F. Schroff, H. Adam, Encoder-decoder with
collision avoidance for autonomous surface vehicles I: A review, J. Mar. Sci. atrous separable convolution for semantic image segmentation, in: Proceedings
Technol. (2021) 1–15. of the European Conference on Computer Vision, ECCV, 2018, pp. 801–818.
[49] H. Noh, S. Hong, B. Han, Learning deconvolution network for semantic
[22] F. Ye, S. Zhang, P. Wang, C.-Y. Chan, A survey of deep reinforcement learning
segmentation, in: Proceedings of the IEEE International Conference on Computer
algorithms for motion planning and control of autonomous vehicles, in: 2021
Vision, 2015, pp. 1520–1528.
IEEE Intelligent Vehicles Symposium, IV, IEEE, 2021, pp. 1073–1080.
[50] R.B. Rusu, Semantic 3D object maps for everyday manipulation in human living
[23] X. Xiao, B. Liu, G. Warnell, P. Stone, Motion planning and control for mobile
environments, KI-Künstliche Intell. 24 (4) (2010) 345–348.
robot navigation using machine learning: A survey, Auton. Robots (2022) 1–29.
[51] W. Wang, K. Sakurada, N. Kawaguchi, Incremental and enhanced scanline-based
[24] H.A. Ignatious, M. Khan, et al., An overview of sensors in autonomous vehicles,
segmentation method for surface reconstruction of sparse LiDAR data, Remote
Procedia Comput. Sci. 198 (2022) 736–741.
Sens. 8 (11) (2016) 967.
[25] E. Marti, M.A. De Miguel, F. Garcia, J. Perez, A review of sensor technologies [52] S. Song, J. Xiao, Sliding shapes for 3d object detection in depth images, in:
for perception in automated driving, IEEE Intell. Transp. Syst. Mag. 11 (4) European Conference on Computer Vision, Springer, 2014, pp. 634–651.
(2019) 94–108. [53] Y. Zhou, O. Tuzel, Voxelnet: End-to-end learning for point cloud based 3d
[26] N. Pinchon, O. Cassignol, A. Nicolas, F. Bernardin, P. Leduc, J.-P. Tarel, object detection, in: Proceedings of the IEEE Conference on Computer Vision
R. Brémond, E. Bercier, J. Brunet, All-weather vision for automotive safety: and Pattern Recognition, 2018, pp. 4490–4499.
Which spectral band? in: International Forum on Advanced Microsystems for [54] B. Li, T. Zhang, T. Xia, Vehicle detection from 3d lidar using fully convolutional
Automotive Applications, Springer, 2018, pp. 3–15. network, 2016, arXiv preprint arXiv:1608.07916.
[27] R. O’Malley, M. Glavin, E. Jones, A review of automotive infrared pedestrian [55] H. Caesar, V. Bankiti, A.H. Lang, S. Vora, V.E. Liong, Q. Xu, A. Krishnan, Y.
detection techniques, in: IET Irish Signals and Systems Conference, ISSC 2008, Pan, G. Baldan, O. Beijbom, Nuscenes: A multimodal dataset for autonomous
IET, 2008, pp. 168–173. driving, in: Proceedings of the IEEE/CVF Conference on Computer Vision and
[28] B. Pueo, High speed cameras for motion analysis in sports science, J. Hum. Pattern Recognition, 2020, pp. 11621–11631.
Sport Exercise 11 (1) (2016) 53–73. [56] A. Geiger, P. Lenz, R. Urtasun, Are we ready for autonomous driving? The
[29] R.A. Hamzah, H. Ibrahim, Literature survey on stereo vision disparity map Kitti vision benchmark suite, in: 2012 IEEE Conference on Computer Vision
algorithms, J. Sensors 2016 (2016). and Pattern Recognition, IEEE, 2012, pp. 3354–3361.
[30] P. Garbat, W. Skarbek, M. Tomaszewski, Structured light camera calibration, [57] C. Urmson, J. Anhalt, D. Bagnell, C. Baker, R. Bittner, M. Clark, J. Dolan,
Opto-Electron. Rev. 21 (1) (2013) 23–38. D. Duggins, T. Galatali, C. Geyer, et al., Autonomous driving in urban
environments: Boss and the urban challenge, J. Field Robot. 25 (8) (2008)
[31] M. Hansard, S. Lee, O. Choi, R.P. Horaud, Time-of-Flight Cameras: Principles,
425–466.
Methods and Applications, Springer Science & Business Media, 2012.
[58] T.-N. Nguyen, B. Michaelis, A. Al-Hamadi, M. Tornow, M.-M. Meinecke, Stereo-
[32] J. Binas, D. Neil, S.-C. Liu, T. Delbruck, DDD17: End-to-end DAVIS driving
camera-based urban environment perception using occupancy grid and object
dataset, 2017, arXiv preprint arXiv:1711.01458.
tracking, IEEE Trans. Intell. Transp. Syst. 13 (1) (2011) 154–165.
[33] M. Schönbein, A. Geiger, Omnidirectional 3d reconstruction in augmented
[59] J. Shi, et al., Good features to track, in: 1994 Proceedings of IEEE Conference
Manhattan worlds, in: 2014 IEEE/RSJ International Conference on Intelligent
on Computer Vision and Pattern Recognition, IEEE, 1994, pp. 593–600.
Robots and Systems, IEEE, 2014, pp. 716–723. [60] J. Lambert, L. Liang, L.Y. Morales, N. Akai, A. Carballo, E. Takeuchi, P. Narksri,
[34] D. Scaramuzza, R. Siegwart, Appearance-guided monocular omnidirectional S. Seiya, K. Takeda, Tsukuba challenge 2017 dynamic object tracks dataset for
visual odometry for outdoor ground vehicles, IEEE Trans. Robot. 24 (5) (2008) pedestrian behavior analysis, J. Robot. Mechatronics 30 (4) (2018) 598–612.
1015–1026. [61] J. Ziegler, P. Bender, T. Dang, C. Stiller, Trajectory planning for Bertha—A local,
[35] C. Jang, C. Kim, K. Jo, M. Sunwoo, Design factor optimization of 3D flash continuous method, in: 2014 IEEE Intelligent Vehicles Symposium Proceedings,
lidar sensor based on geometrical model for automated vehicle and advanced IEEE, 2014, pp. 450–457.
driver assistance system applications, Int. J. Automot. Technol. 18 (1) (2017) [62] A. Petrovskaya, S. Thrun, Model based vehicle detection and tracking for
147–156. autonomous urban driving, Auton. Robots 26 (2) (2009) 123–139.
[36] Q. Ha, K. Watanabe, T. Karasawa, Y. Ushiku, T. Harada, MFNet: Towards real- [63] M. He, E. Takeuchi, Y. Ninomiya, S. Kato, Precise and efficient model-based
time semantic segmentation for autonomous vehicles with multi-spectral scenes, vehicle tracking method using Rao-Blackwellized and scaling series particle
in: 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems, filters, in: 2016 IEEE/RSJ International Conference on Intelligent Robots and
IROS, IEEE, 2017, pp. 5108–5115. Systems, IROS, IEEE, 2016, pp. 117–124.

40
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

[64] D. Held, S. Thrun, S. Savarese, Learning to track at 100 fps with deep regression [88] A.A. Maw, M. Tyan, J.-W. Lee, iADA*: Improved anytime path planning and
networks, in: European Conference on Computer Vision, Springer, 2016, pp. replanning algorithm for autonomous vehicle, J. Intell. Robot. Syst. 100 (3)
749–765. (2020) 1005–1013.
[65] B. Huval, T. Wang, S. Tandon, J. Kiske, W. Song, J. Pazhayampallil, M. [89] M. Thoresen, N.H. Nielsen, K. Mathiassen, K.Y. Pettersen, Path planning for
Andriluka, P. Rajpurkar, T. Migimatsu, R. Cheng-Yue, et al., An empirical UGVs based on traversability hybrid A, IEEE Robot. Autom. Lett. 6 (2) (2021)
evaluation of deep learning on highway driving, 2015, arXiv preprint arXiv: 1216–1223.
1504.01716. [90] D.-D. Zhu, J.-Q. Sun, A new algorithm based on Dijkstra for vehicle path
[66] C. Fernandez, D. Fernandez-Llorca, M.A. Sotelo, A hybrid vision-map method planning considering intersection attribute, IEEE Access 9 (2021) 19761–19775.
for urban road detection, J. Adv. Transp. 2017 (2017). [91] L.-s. Liu, J.-f. Lin, J.-x. Yao, D.-w. He, J.-s. Zheng, J. Huang, P. Shi, Path
[67] M. Paton, K. MacTavish, C.J. Ostafew, T.D. Barfoot, It’s not easy seeing green: planning for smart car based on Dijkstra algorithm and dynamic window
Lighting-resistant stereo visual teach & repeat using color-constant images, in: approach, Wirel. Commun. Mob. Comput. 2021 (2021).
2015 IEEE International Conference on Robotics and Automation, ICRA, IEEE, [92] T. Liu, J. Zhang, An improved path planning algorithm based on fuel
2015, pp. 1519–1526. consumption, J. Supercomput. (2022) 1–31.
[68] P. Narksri, E. Takeuchi, Y. Ninomiya, Y. Morales, N. Akai, N. Kawaguchi, A [93] D. Kim, G. Kim, H. Kim, K. Huh, A hierarchical motion planning framework
slope-robust cascaded ground segmentation in 3D point cloud for autonomous for autonomous driving in structured highway environments, IEEE Access 10
vehicles, in: 2018 21st International Conference on Intelligent Transportation (2022) 20102–20117.
Systems, ITSC, IEEE, 2018, pp. 497–504. [94] Y. Li, R. Jin, X. Xu, Y. Qian, H. Wang, S. Xu, Z. Wang, A mobile robot path
[69] A.S. Huang, D. Moore, M. Antone, E. Olson, S. Teller, Finding multiple lanes planning algorithm based on improved A* algorithm and dynamic window
in urban road networks with vision and lidar, Auton. Robots 26 (2) (2009) approach, IEEE Access (2022).
103–122. [95] M. Elbanhawi, M. Simic, Sampling-based robot motion planning: A review, Ieee
[70] A.V. Nefian, G.R. Bradski, Detection of drivable corridors for off-road au- Access 2 (2014) 56–77.
tonomous navigation, in: 2006 International Conference on Image Processing, [96] S.M. LaValle, J.J. Kuffner Jr., Randomized kinodynamic planning, Int. J. Robot.
IEEE, 2006, pp. 3025–3028. Res. 20 (5) (2001) 378–400.
[71] R. Labayrade, J. Douret, J. Laneurit, R. Chapuis, A reliable and robust lane [97] W. Lim, S. Lee, M. Sunwoo, K. Jo, Hierarchical trajectory planning of an
detection system based on the parallel use of three algorithms for driving safety autonomous car based on the integration of a sampling and an optimization
assistance, IEICE Trans. Inf. Syst. 89 (7) (2006) 2092–2100. method, IEEE Trans. Intell. Transp. Syst. 19 (2) (2018) 613–626.
[72] R. Danescu, S. Nedevschi, Probabilistic lane tracking in difficult road scenarios
[98] J. Wang, S. Wu, H. Li, J. Zou, Path planning combining improved rapidly-
using stereovision, IEEE Trans. Intell. Transp. Syst. 10 (2) (2009) 272–282.
exploring random trees with dynamic window approach in ROS, in: 2018 13th
[73] J. Levinson, M. Montemerlo, S. Thrun, Map-based precision vehicle localization
IEEE Conference on Industrial Electronics and Applications, ICIEA, IEEE, 2018,
in urban environments, in: Robotics: Science and Systems, vol. 4, Citeseer, 2007,
pp. 1296–1301.
p. 1.
[99] A.M. Varghese, V. Jisha, Motion planning and control of an autonomous mobile
[74] G. Bresson, Z. Alsayed, L. Yu, S. Glaser, Simultaneous localization and mapping:
robot, in: 2018 International CET Conference on Control, Communication, and
A survey of current trends in autonomous driving, IEEE Trans. Intell. Veh. 2
Computing, IC4, IEEE, 2018, pp. 17–21.
(3) (2017) 194–220.
[100] M. Werling, S. Kammel, J. Ziegler, L. Gröll, Optimal trajectories for time-critical
[75] X. Qu, B. Soheilian, N. Paparoditis, Vehicle localization using mono-camera and
street scenarios using discretized terminal manifolds, Int. J. Robot. Res. 31 (3)
geo-referenced traffic signs, in: 2015 IEEE Intelligent Vehicles Symposium, IV,
(2012) 346–359.
IEEE, 2015, pp. 605–610.
[101] W. Lim, S. Lee, M. Sunwoo, K. Jo, Hybrid trajectory planning for autonomous
[76] M. Magnusson, The Three-Dimensional Normal-Distributions Transform: An
driving in on-road dynamic scenarios, IEEE Trans. Intell. Transp. Syst. 22 (1)
Efficient Representation for Registration, Surface Analysis, and Loop Detection
(2019) 341–355.
(Ph.D. thesis), Örebro universitet, 2009.
[102] Z. Li, W. Zhan, L. Sun, C.-Y. Chan, M. Tomizuka, Adaptive sampling-based
[77] N. Akai, L.Y. Morales, E. Takeuchi, Y. Yoshihara, Y. Ninomiya, Robust localiza-
motion planning with a non-conservatively defensive strategy for autonomous
tion using 3D NDT scan matching with experimentally determined uncertainty
driving, IFAC-PapersOnLine 53 (2) (2020) 15632–15638.
and road marker matching, in: 2017 IEEE Intelligent Vehicles Symposium, IV,
[103] S. Feraco, S. Luciani, A. Bonfitto, N. Amati, A. Tonoli, A local trajectory
IEEE, 2017, pp. 1356–1363.
planning and control method for autonomous vehicles based on the RRT
[78] E. Yurtsever, K. Takeda, C. Miyajima, Traffic trajectory history and drive path
algorithm, in: 2020 AEIT International Conference of Electrical and Electronic
generation using GPS data cloud, in: 2015 IEEE Intelligent Vehicles Symposium,
Technologies for Automotive, AEIT AUTOMOTIVE, IEEE, 2020, pp. 1–6.
IV, IEEE, 2015, pp. 229–234.
[104] J. Chen, R. Zhang, W. Han, W. Jiang, J. Hu, X. Lu, X. Liu, P. Zhao, Path
[79] C.M. Martinez, M. Heucke, F.-Y. Wang, B. Gao, D. Cao, Driving style recognition
planning for autonomous vehicle based on a two-layered planning model in
for intelligent vehicle control and advanced driver assistance: A survey, IEEE
complex environment, J. Adv. Transp. 2020 (2020).
Trans. Intell. Transp. Syst. 19 (3) (2017) 666–676.
[80] M. Reda, A. Onsy, M.A. Elhosseini, A.Y. Haikal, M. Badawy, A discrete variant [105] J. Wang, W. Chi, C. Li, C. Wang, M.Q.-H. Meng, Neural RRT*: Learning-based
of cuckoo search algorithm to solve the travelling salesman problem and path optimal path planning, IEEE Trans. Autom. Sci. Eng. 17 (4) (2020) 1748–1758.
planning for autonomous trolley inside warehouse, Knowl.-Based Syst. (2022) [106] Y. Zhang, J. Zhang, J. Zhang, J. Wang, K. Lu, J. Hong, A novel learning
109290. framework for sampling-based motion planning in autonomous driving, in:
[81] D. Delling, A.V. Goldberg, A. Nowatzyk, R.F. Werneck, PHAST: Hardware- Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 34, 2020,
accelerated shortest path trees, J. Parallel Distrib. Comput. 73 (7) (2013) pp. 1202–1209.
940–952. [107] J. Rong, S. Arrigoni, N. Luan, F. Braghin, Attention-based sampling distribution
[82] P.E. Hart, N.J. Nilsson, B. Raphael, A formal basis for the heuristic determi- for motion planning in autonomous driving, in: 2020 39th Chinese Control
nation of minimum cost paths, IEEE Trans. Syst. Sci. Cybern. 4 (2) (1968) Conference, CCC, IEEE, 2020, pp. 5671–5676.
100–107. [108] C. Huang, H. Huang, P. Hang, H. Gao, J. Wu, Z. Huang, C. Lv, Personalized
[83] W. Yijing, L. Zhengxuan, Z. Zhiqiang, L. Zheng, Local path planning of trajectory planning and control of lane-change maneuvers for autonomous
autonomous vehicles based on A* algorithm with equal-step sampling, in: 2018 driving, IEEE Trans. Veh. Technol. 70 (6) (2021) 5511–5523.
37th Chinese Control Conference, CCC, IEEE, 2018, pp. 7828–7833. [109] X. Jin, Z. Yan, H. Yang, Q. Wang, A practical sampling-based motion
[84] R. Udomsil, T. Sangpet, T. Sapsaman, Environment generation from real map to planning method for autonomous driving in unstructured environments,
investigate path planning and obstacle avoidance algorithm for electric vehicle, IFAC-PapersOnLine 54 (10) (2021) 449–453.
in: 2019 Research, Invention, and Innovation Congress, RI2C, IEEE, 2019, pp. [110] D. Rakita, B. Mutlu, M. Gleicher, Single-query path planning using sample-
1–5. efficient probability informed trees, IEEE Robot. Autom. Lett. 6 (3) (2021)
[85] Y. Song, Z. Wang, Path planning simulation based on improved A* algorithm, 4624–4631.
J. Changchun Univ. Technol. 40 (2) (2019) 138–141. [111] J. Wang, B. Li, M.Q.-H. Meng, Kinematic constrained bi-directional RRT with
[86] L. Yeong-Ho, K. Yeong-Jun, J. Da-Un, W. Ihn-Sik, Development of an integrated efficient branch pruning for robot path planning, Expert Syst. Appl. 170 (2021)
path planning algorithm for autonomous driving of unmanned surface vessel, 114541.
in: 2020 20th International Conference on Control, Automation and Systems, [112] Y. Zhang, S. Wang, LSPP: A novel path planning algorithm based on perceiving
ICCAS, IEEE, 2020, pp. 27–32. line segment feature, IEEE Sens. J. 22 (1) (2021) 720–731.
[87] X. Zhong, J. Tian, H. Hu, X. Peng, Hybrid path planning based on safe [113] S. Ganesan, S.K. Natarajan, J. Srinivasan, A global path planning algorithm for
A* algorithm and adaptive window approach for mobile robot in large-scale mobile robot in cluttered environments with an improved initial cost solution
dynamic environment, J. Intell. Robot. Syst. 99 (1) (2020) 65–77. and convergence rate, Arab. J. Sci. Eng. 47 (3) (2022) 3633–3647.

41
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

[114] G.O. Flores-Aquino, J.I. Vasquez-Gomez, O. Gutierrez-Frias, Custom distribution [139] S. Zhu, B. Aksun-Guvenc, Trajectory planning of autonomous vehicles based on
for sampling-based motion planning, J. Braz. Soc. Mech. Sci. Eng. 44 (3) (2022) parameterized control optimization in dynamic on-road environments, J. Intell.
1–15. Robot. Syst. 100 (3) (2020) 1055–1067.
[115] G. Huang, Q. Ma, Research on path planning algorithm of autonomous vehicles [140] Y. Zhang, H. Sun, J. Zhou, J. Pan, J. Hu, J. Miao, Optimal vehicle path planning
based on improved RRT algorithm, Int. J. Intell. Transp. Syst. Res. 20 (1) (2022) using quadratic optimization for baidu apollo open platform, in: 2020 IEEE
170–180. Intelligent Vehicles Symposium, IV, IEEE, 2020, pp. 978–984.
[116] G. Yang, Y. Yao, Vehicle local path planning and time consistency of unmanned [141] J. Changhao, H. Miaohua, S. Liyang, An autonomous vehicle motion planning
driving system based on convolutional neural network, Neural Comput. Appl. method based on dynamic programming, in: 2020 17th International Computer
34 (15) (2022) 12385–12398. Conference on Wavelet Active Media Technology and Information Processing,
[117] X. Zhang, T. Zhu, Y. Xu, H. Liu, F. Liu, Local path planning of the au- ICCWAMTIP, IEEE, 2020, pp. 394–398.
tonomous vehicle based on adaptive improved RRT algorithm in certain lane [142] Y. Zhang, H. Chen, S.L. Waslander, J. Gong, G. Xiong, T. Yang, K. Liu, Hybrid
environments, Actuators 11 (4) (2022) 109. trajectory planning for autonomous driving in highly constrained environments,
[118] S. Lu, Path tracking control algorithm for unmanned vehicles based on IEEE Access 6 (2018) 32800–32819.
improved RRT algorithm, in: 2022 IEEE 2nd International Conference on [143] J. Li, J. Gong, G. Kong, Y. Zhao, X. Zhang, A hierarchical trajectory planning
Electronic Technology, Communication and Information, ICETCI, IEEE, 2022, framework for autonomous driving, in: 2020 3rd International Conference on
pp. 1201–1204. Unmanned Systems, ICUS, IEEE, 2020, pp. 428–434.
[119] S. Spanogiannopoulos, Y. Zweiri, L. Seneviratne, Sampling-based non-holonomic [144] C. Hu, L. Zhao, G. Qu, Event-triggered model predictive adaptive dynamic
path generation for self-driving cars, J. Intell. Robot. Syst. 104 (1) (2022) 1–17. programming for road intersection path planning of unmanned ground vehicle,
[120] F. Bounini, D. Gingras, H. Pollart, D. Gruyer, Modified artificial potential field IEEE Trans. Veh. Technol. 70 (11) (2021) 11228–11243.
method for online path planning applications, in: 2017 IEEE Intelligent Vehicles [145] L. Wang, B. Wang, C. Wang, Collision-free path planning with kinematic
Symposium, IV, IEEE, 2017, pp. 180–185. constraints in urban scenarios, J. Shanghai Jiaotong Univ. (Science) 26 (5)
[121] H. Hongyu, Z. Chi, S. Yuhuan, Z. Bin, G. Fei, An improved artificial (2021) 731–738.
potential field model considering vehicle velocity for autonomous driving, [146] P. Typaldos, M. Papageorgiou, I. Papamichail, Optimization-based path-
IFAC-PapersOnLine 51 (31) (2018) 863–867. planning for connected and non-connected automated vehicles, Transp. Res.
[122] B. Lu, G. Li, H. Yu, H. Wang, J. Guo, D. Cao, H. He, Adaptive potential field- C 134 (2022) 103487.
based path planning for complex autonomous driving scenarios, IEEE Access 8 [147] Y. Jiang, Z. Liu, D. Qian, H. Zuo, W. He, J. Wang, Robust online path planning
(2020) 225294–225305. for autonomous vehicle using sequential quadratic programming, in: 2022 IEEE
[123] P. Lin, W.Y. Choi, S.-H. Lee, C.C. Chung, Model predictive path planning based Intelligent Vehicles Symposium, IV, IEEE, 2022, pp. 175–182.
on artificial potential field and its application to autonomous lane change, [148] J.P. Rastelli, R. Lattarulo, F. Nashashibi, Dynamic trajectory generation using
in: 2020 20th International Conference on Control, Automation and Systems, continuous-curvature algorithms for door to door assistance vehicles, in: 2014
ICCAS, IEEE, 2020, pp. 731–736. IEEE Intelligent Vehicles Symposium Proceedings, IEEE, 2014, pp. 510–515.
[124] Z. Huang, D. Chu, C. Wu, Y. He, Path planning and cooperative control for [149] H. Fuji, J. Xiang, Y. Tazaki, B. Levedahl, T. Suzuki, Trajectory planning for auto-
automated vehicle platoon using hybrid automata, IEEE Trans. Intell. Transp. mated parking using multi-resolution state roadmap considering non-holonomic
Syst. 20 (3) (2018) 959–974. constraints, in: 2014 IEEE Intelligent Vehicles Symposium Proceedings, IEEE,
[125] H. Li, C. Wu, D. Chu, L. Lu, K. Cheng, Combined trajectory planning and 2014, pp. 407–413.
tracking for autonomous vehicle considering driving styles, IEEE Access 9 [150] C. Bergenhem, S. Shladover, E. Coelingh, C. Englund, S. Tsugawa, Overview of
(2021) 9453–9463. platooning systems, in: Proceedings of the 19th ITS World Congress, Oct 22-26,
[126] C. Huang, C. Lv, P. Hang, Y. Xing, Toward safe and personalized autonomous Vienna, Austria (2012), 2012, pp. 2–7.
driving: Decision-making and motion control with DPF and CDT techniques, [151] P. Petrov, F. Nashashibi, Modeling and nonlinear adaptive control for au-
IEEE/ASME Trans. Mechatronics 26 (2) (2021) 611–620. tonomous vehicle overtaking, IEEE Trans. Intell. Transp. Syst. 15 (4) (2014)
[127] J. Ji, A. Khajepour, W.W. Melek, Y. Huang, Path planning and tracking for vehi- 1643–1656.
cle collision avoidance based on model predictive control with multiconstraints, [152] H.-w. Wang, X.-c. Yu, H.-b. Song, Z.-h. Lu, J. Lloret, F. You, A global optimal
IEEE Trans. Veh. Technol. 66 (2) (2016) 952–964. path planning and controller design algorithm for intelligent vehicles, Mob.
[128] Z. Zhang, L. Zheng, Y. Li, P. Zeng, Y. Liang, Structured road-oriented motion Netw. Appl. 23 (5) (2018) 1165–1178.
planning and tracking framework for active collision avoidance of autonomous [153] X. Hu, L. Chen, B. Tang, D. Cao, H. He, Dynamic path planning for autonomous
vehicles, Sci. China Technol. Sci. 64 (11) (2021) 2427–2440. driving on various roads with avoidance of static and moving obstacles, Mech.
[129] H. Li, W. Liu, C. Yang, W. Wang, T. Qie, C. Xiang, An optimization-based path Syst. Signal Process. 100 (2018) 482–500.
planning approach for autonomous vehicles using dynEFWA-artificial potential [154] G. Klančar, M. Seder, S. Blažič, I. Škrjanc, I. Petrović, Drivable path planning
field, IEEE Trans. Intell. Veh. (2021). using hybrid search algorithm based on E* and Bernstein–Bézier motion
[130] P. Lin, M. Tsukada, Model predictive path-planning controller with potential primitives, IEEE Trans. Syst., Man, Cybern.: Syst. 51 (8) (2019) 4868–4882.
function for emergency collision avoidance on highway driving, IEEE Robot. [155] C. You, J. Lu, D. Filev, P. Tsiotras, Autonomous planning and control for
Autom. Lett. 7 (2) (2022) 4662–4669. intelligent vehicles in traffic, IEEE Trans. Intell. Transp. Syst. 21 (6) (2019)
[131] R. Szczepanski, T. Tarczewski, K. Erwinski, Energy efficient local path planning 2339–2349.
algorithm based on predictive artificial potential field, IEEE Access 10 (2022) [156] S. Sedighi, D.-V. Nguyen, P. Kapsalas, K.-D. Kuhnert, Fusing direct visibility
39729–39742. diagram with clothoid curves for motion planning, in: 2019 IEEE Intelligent
[132] C. Wang, Z. Wang, L. Zhang, H. Yu, D. Cao, Post-impact motion planning and Transportation Systems Conference, ITSC, IEEE, 2019, pp. 3579–3586.
tracking control for autonomous vehicles, Chin. J. Mech. Eng. 35 (1) (2022) [157] D. Piscini, E. Pagot, G. Valenti, F. Biral, Experimental comparison of trajectory
1–18. control and planning algorithms for autonomous vehicles, in: IECON 2019-45th
[133] J. Nocedal, S.J. Wright, Numerical Optimization, Springer, 1999. Annual Conference of the IEEE Industrial Electronics Society, Vol. 1, IEEE, 2019,
[134] E.V. Denardo, Dynamic Programming: Models and Applications, Courier pp. 5217–5222.
Corporation, 2012. [158] J. Moreau, P. Melchior, S. Victor, L. Cassany, M. Moze, F. Aioun, F.
[135] Í.B. Viana, N. Aouf, Distributed cooperative path-planning for autonomous Guillemard, Reactive path planning in intersection for autonomous vehicle,
vehicles integrating human driver trajectories, in: 2018 International Conference IFAC-PapersOnLine 52 (5) (2019) 109–114.
on Intelligent Systems, IS, IEEE, 2018, pp. 655–661. [159] A.Z. Zambom, B. Seguin, F. Zhao, Robot path planning in a dynamic en-
[136] H. Kanchwala, Path planning and tracking of an autonomous car with high vironment with stochastic measurements, J. Global Optim. 73 (2) (2019)
fidelity vehicle dynamics model and human driver trajectories, in: 2019 389–410.
IEEE 10th International Conference on Mechanical and Aerospace Engineering, [160] Y. Mu, B. Li, D. An, Y. Wei, Three-dimensional route planning based on the
ICMAE, IEEE, 2019, pp. 306–313. beetle swarm optimization algorithm, IEEE Access 7 (2019) 117804–117813.
[137] P.F. Lima, R. Oliveira, J. Mårtensson, B. Wahlberg, Minimizing long vehicles [161] R.T. Kamil, M.J. Mohamed, B.K. Oleiwi, Path planning of mobile robot using
overhang exceeding the drivable surface via convex path optimization, in: 2017 improved artificial bee colony algorithm, Eng. Technol. J. 38 (9) (2020)
IEEE 20th International Conference on Intelligent Transportation Systems, ITSC, 1384–1395.
IEEE, 2017, pp. 1–8. [162] S. Luo, X. Li, Z. Sun, An optimization-based motion planning method for au-
[138] R. Oliveira, P.F. Lima, G.C. Pereira, J. Mårtensson, B. Wahlberg, Path planning tonomous driving vehicle, in: 2020 3rd International Conference on Unmanned
for autonomous bus driving in highly constrained environments, in: 2019 Systems, ICUS, IEEE, 2020, pp. 739–744.
IEEE Intelligent Transportation Systems Conference, ITSC, IEEE, 2019, pp. [163] X. Jin, Z. Yan, G. Yin, S. Li, C. Wei, An adaptive motion planning technique
2743–2749. for on-road autonomous driving, IEEE Access 9 (2020) 2655–2664.

42
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

[164] Y. Dai, S.-G. Lee, Perception, planning and control for self-driving system based [190] P. Wang, D. Liu, J. Chen, H. Li, C.-Y. Chan, Decision making for autonomous
on on-board sensors, Adv. Mech. Eng. 12 (9) (2020) 1687814020956494. driving via augmented adversarial inverse reinforcement learning, in: 2021 IEEE
[165] Á. Fehér, S. Aradi, T. Bécsi, Hierarchical evasive path planning using re- International Conference on Robotics and Automation, ICRA, IEEE, 2021, pp.
inforcement learning and model predictive control, IEEE Access 8 (2020) 1036–1042.
187470–187482. [191] X. Liu, D. Zhang, T. Zhang, Y. Cui, L. Chen, S. Liu, Novel best path selection
[166] E.D. Lambert, R. Romano, D. Watling, Optimal smooth paths based on clothoids approach based on hybrid improved A* algorithm and reinforcement learning,
for car-like vehicles in the presence of obstacles, Int. J. Control Autom. Syst. Appl. Intell. 51 (12) (2021) 9015–9029.
19 (6) (2021) 2163–2182. [192] X. Liu, D. Zhang, T. Zhang, J. Zhang, J. Wang, A new path plan method based
[167] V. Bulut, Path planning for autonomous ground vehicles based on quintic on hybrid algorithm of reinforcement learning and particle swarm optimization,
trigonometric Bézier curve, J. Braz. Soc. Mech. Sci. Eng. 43 (2) (2021) 1–14. Eng. Comput. (2021).
[168] D. Zeng, Z. Yu, L. Xiong, J. Zhao, P. Zhang, Z. Fu, A Steerable Curvature [193] H. Kim, W. Lee, Real-time path planning through Q-learning’s exploration strat-
Approach for Efficient Executable Path Planning for On-Road Autonomous egy adjustment, in: 2021 International Conference on Electronics, Information,
Vehicle, Tech. Rep., SAE Technical Paper, 2019. and Communication, ICEIC, IEEE, 2021, pp. 1–3.
[169] Z. Li, L. Xiong, D. Zeng, Z. Fu, B. Leng, F. Shan, Real-time local path [194] L. Chang, L. Shan, C. Jiang, Y. Dai, Reinforcement based mobile robot path
planning for intelligent vehicle combining tentacle algorithm and B-spline planning with improved dynamic window approach in unknown environment,
curve, IFAC-PapersOnLine 54 (10) (2021) 51–58. Auton. Robots 45 (1) (2021) 51–76.
[170] E. Horváth, C.R. Pozna, Clothoid-based trajectory following approach for self- [195] E.S. Low, P. Ong, C.Y. Low, R. Omar, Modified Q-learning with distance metric
driving vehicles, in: 2021 IEEE 19th World Symposium on Applied Machine and virtual target on path planning of mobile robot, Expert Syst. Appl. 199
Intelligence and Informatics, SAMI, IEEE, 2021, pp. 000251–000254. (2022) 117191.
[171] S. Shentu, Z. Gong, X.-J. Liu, Q. Liu, F. Xie, Hybrid navigation system based [196] P. Rousseas, C.P. Bechlioulis, K.J. Kyriakopoulos, Optimal motion planning in
autonomous positioning and path planning for mobile robots, Chin. J. Mech. unknown workspaces using integral reinforcement learning, IEEE Robot. Autom.
Eng. 35 (1) (2022) 1–13. Lett. (2022).
[172] D.S. Drake, Using Ensemble Learning Techniques to Solve the Blind Drift [197] D. Silver, G. Lever, N. Heess, T. Degris, D. Wierstra, M. Riedmiller, Deterministic
Calibration Problem (Ph.D. thesis), Old Dominion University, 2022. policy gradient algorithms, in: International Conference on Machine Learning,
[173] M. Bojarski, D. Del Testa, D. Dworakowski, B. Firner, B. Flepp, P. Goyal, PMLR, 2014, pp. 387–395.
L.D. Jackel, M. Monfort, U. Muller, J. Zhang, et al., End to end learning for [198] R.S. Sutton, A.G. Barto, Reinforcement Learning: An Introduction, MIT Press,
self-driving cars, 2016, arXiv preprint arXiv:1604.07316. 2018.
[174] S. Song, X. Hu, J. Yu, L. Bai, L. Chen, Learning a deep motion planning model
[199] J. Bernhard, R. Gieselmann, K. Esterle, A. Knol, Experience-based heuristic
for autonomous driving, in: 2018 IEEE Intelligent Vehicles Symposium, IV, IEEE,
search: Robust motion planning with deep q-learning, in: 2018 21st Interna-
2018, pp. 1137–1142.
tional Conference on Intelligent Transportation Systems, ITSC, IEEE, 2018, pp.
[175] P. Kicki, T. Gawron, P. Skrzypczyński, A self-supervised learning approach to 3175–3182.
rapid path planning for car-like vehicles maneuvering in urban environment,
[200] C. You, J. Lu, D. Filev, P. Tsiotras, Advanced planning for autonomous vehicles
2020, arXiv preprint arXiv:2003.00946.
using reinforcement learning and deep inverse reinforcement learning, Robot.
[176] G. Moraes, A. Mozart, P. Azevedo, M. Piumbini, V.B. Cardoso, T. Oliveira-
Auton. Syst. 114 (2019) 1–18.
Santos, A.F. De Souza, C. Badue, Image-based real-time path generation using
[201] J. Liao, T. Liu, X. Tang, X. Mu, B. Huang, D. Cao, Decision-making strategy
deep neural networks, in: 2020 International Joint Conference on Neural
on highway for autonomous vehicles using deep reinforcement learning, IEEE
Networks, IJCNN, IEEE, 2020, pp. 1–8.
Access 8 (2020) 177804–177814.
[177] L. Markolf, J. Eilbrecht, O. Stursberg, Trajectory planning for autonomous
[202] J. Zhao, T. Qu, F. Xu, A deep reinforcement learning approach for autonomous
vehicles combining nonlinear optimal control and supervised learning,
highway driving, IFAC-PapersOnLine 53 (5) (2020) 542–546.
IFAC-PapersOnLine 53 (2) (2020) 15608–15614.
[203] X. Liao, Y. Wang, Y. Xuan, D. Wu, AGV path planning model based on
[178] N. Guo, C. Li, D. Wang, Y. Song, G. Liu, T. Gao, Local path planning of
reinforcement learning, in: 2020 Chinese Automation Congress, CAC, IEEE,
mobile robot based on long short-term memory neural network, Autom. Control
2020, pp. 6722–6726.
Comput. Sci. 55 (1) (2021) 53–65.
[204] L. Chen, X. Hu, B. Tang, Y. Cheng, Conditional DQN-based motion planning
[179] M. Sakurai, Y. Ueno, M. Kondo, Path planning and moving obstacle avoidance
with fuzzy logic for autonomous driving, IEEE Trans. Intell. Transp. Syst.
with neuromorphic computing, in: 2021 IEEE International Conference on
(2020).
Intelligence and Safety for Robotics, ISR, IEEE, 2021, pp. 209–215.
[205] S. Wen, Y. Zhao, X. Yuan, Z. Wang, D. Zhang, L. Manfredi, Path planning for ac-
[180] D. Wang, C. Wang, Y. Wang, H. Wang, F. Pei, An autonomous driving approach
tive SLAM based on deep reinforcement learning under unknown environments,
based on trajectory learning using deep neural networks, Int. J. Automot.
Intell. Serv. Robot. 13 (2) (2020) 263–272.
Technol. 22 (6) (2021) 1517–1528.
[181] D. Kalathil, V.K. Mandal, A. Gune, K. Talele, P. Chimurkar, M. Bansode, Self [206] G. Li, S. Li, S. Li, Y. Qin, D. Cao, X. Qu, B. Cheng, Deep reinforcement learning
driving car using neural networks, in: 2022 International Conference on Applied enabled decision-making for autonomous driving at intersections, Automot.
Artificial Intelligence and Computing, ICAAIC, IEEE, 2022, pp. 213–217. Innov. 3 (4) (2020) 374–385.
[182] D.-H. Lee, J.-L. Liu, End-to-end deep learning of lane detection and path [207] V. Sainath, S. Reddy, et al., Deep learning for autonomous driving system,
prediction for real-time autonomous driving, Signal, Image Video Process. in: 2021 Second International Conference on Electronics and Sustainable
(2022) 1–7. Communication Systems, ICESC, IEEE, 2021, pp. 1744–1749.
[183] S. Russell, P. Norvig, A modern, agent-oriented approach to introductory [208] K.B. Naveed, Z. Qiao, J.M. Dolan, Trajectory planning for autonomous ve-
artificial intelligence, Acm Sigart Bull. 6 (2) (1995) 24–26. hicles using hierarchical reinforcement learning, in: 2021 IEEE International
[184] B.R. Kiran, I. Sobh, V. Talpaert, P. Mannion, A.A. Al Sallab, S. Yogamani, P. Intelligent Transportation Systems Conference, ITSC, IEEE, 2021, pp. 601–606.
Pérez, Deep reinforcement learning for autonomous driving: A survey, IEEE [209] J. Li, Y. Chen, X. Zhao, J. Huang, An improved DQN path planning algorithm,
Trans. Intell. Transp. Syst. (2021). J. Supercomput. 78 (1) (2022) 616–639.
[185] Z.-W. Hong, T.-Y. Shann, S.-Y. Su, Y.-H. Chang, T.-J. Fu, C.-Y. Lee, Diversity- [210] Y. Peng, G. Tan, H. Si, J. Li, DRL-GAT-SA: Deep reinforcement learning for
driven exploration strategy for deep reinforcement learning, Adv. Neural Inf. autonomous driving planning based on graph attention networks and simplex
Process. Syst. 31 (2018). architecture, J. Syst. Archit. 126 (2022) 102505.
[186] Y. Yao, Z. Peng, B. Xiao, Parallel hyper-heuristic algorithm for multi-objective [211] Ó. Pérez-Gil, R. Barea, E. López-Guillén, L.M. Bergasa, C. Gómez-Huélamo,
route planning in a smart city, IEEE Trans. Veh. Technol. 67 (11) (2018) R. Gutiérrez, A. Díaz-Díaz, Deep reinforcement learning based control for
10307–10318. autonomous vehicles in carla, Multimedia Tools Appl. 81 (3) (2022) 3553–3576.
[187] X.-H. Liu, D.-G. Zhang, H.-R. Yan, Y.-y. Cui, L. Chen, A new algorithm of [212] Z. Wang, J. Tu, C. Chen, Reinforcement learning based trajectory planning for
the best path selection based on machine learning, IEEE Access 7 (2019) autonomous vehicles, in: 2021 China Automation Congress, CAC, IEEE, 2021,
126913–126928. pp. 7995–8000.
[188] C. Chen, J. Jiang, N. Lv, S. Li, An intelligent path planning scheme of [213] L. Zhang, R. Zhang, T. Wu, R. Weng, M. Han, Y. Zhao, Safe reinforcement
autonomous vehicles platoon using deep reinforcement learning on network learning with stability guarantee for motion planning of autonomous vehicles,
edge, IEEE Access 8 (2020) 99059–99069. IEEE Trans. Neural Netw. Learn. Syst. 32 (12) (2021) 5435–5444.
[189] X. Wang, X. Yu, W. Sun, A path planning and tracking control for autonomous [214] C. Xu, W. Zhao, Q. Chen, C. Wang, An actor-critic based learning method for
vehicle with obstacle avoidance, in: 2020 Chinese Automation Congress, CAC, decision-making and planning of autonomous vehicles, Sci. China Technol. Sci.
IEEE, 2020, pp. 2973–2978. 64 (5) (2021) 984–994.

43
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

[215] J. Choi, G. Lee, C. Lee, Reinforcement learning-based dynamic obstacle avoid- [242] H. Ali, D. Gong, M. Wang, X. Dai, Path planning of mobile robot with improved
ance and integration of path planning, Intell. Serv. Robot. 14 (5) (2021) ant colony algorithm and MDP to produce smooth trajectory in grid-based
663–677. environment, Front. Neurorobotics 14 (2020) 44.
[216] X. Tang, B. Huang, T. Liu, X. Lin, Highway decision-making and motion [243] C. Wu, S. Zhou, L. Xiao, Dynamic path planning based on improved ant colony
planning for autonomous driving via soft actor-critic, IEEE Trans. Veh. Technol. algorithm in traffic congestion, IEEE Access 8 (2020) 180773–180783.
71 (5) (2022) 4706–4717. [244] M.A.R. Pohan, B.R. Trilaksono, S.P. Santosa, A.S. Rohman, Path planning
[217] M. Shehab, A.T. Khader, M.A. Al-Betar, A survey on applications and variants algorithm using the hybridization of the rapidly-exploring random tree and ant
of the cuckoo search algorithm, Appl. Soft Comput. 61 (2017) 1041–1059. colony systems, IEEE Access 9 (2021) 153599–153615.
[218] B. Patle, A. Pandey, D. Parhi, A. Jagadeesh, et al., A review: On path planning [245] S. Zhang, J. Pu, Y. Si, An adaptive improved ant colony system based on
strategies for navigation of mobile robot, Def. Technol. 15 (4) (2019) 582–606. population information entropy for path planning of mobile robot, IEEE Access
[219] M. Reda, M. Elhosseini, A. Haikal, M. Badawy, A novel cuckoo search algorithm 9 (2021) 24933–24945.
with adaptive discovery probability based on double Mersenne numbers, Neural [246] C. Jiang, J. Fu, W. Liu, Research on vehicle routing planning based on adaptive
Comput. Appl. 33 (23) (2021) 16377–16402. ant colony and particle swarm optimization algorithm, Int. J. Intell. Transp.
Syst. Res. 19 (1) (2021) 83–91.
[220] M. Reda, A.Y. Haikal, M.A. Elhosseini, M. Badawy, An innovative damped
[247] Y. Zhou, N. Huang, Airport AGV path optimization model based on ant colony
cuckoo search algorithm with a comparative study against other adaptive
algorithm to optimize Dijkstra algorithm in urban systems, Sustain. Comput.:
variants, IEEE Access 7 (2019) 119272–119293.
Inf. Syst. 35 (2022) 100716.
[221] H.J. Bremermann, et al., Optimization through evolution and recombination,
[248] D. Karaboga, et al., An Idea Based on Honey Bee Swarm for Numerical
Self-Organ. Syst. 93 (1962) 106.
Optimization, Tech. Rep., Technical report-tr06, Erciyes university, engineering
[222] J.H. Holland, Adaptation in Natural and Artificial Systems: An Introductory
faculty, computer . . . , 2005.
Analysis with Applications to Biology, Control, and Artificial Intelligence, MIT
[249] A. Nayyar, N.G. Nguyen, R. Kumari, S. Kumar, Robot path planning using
Press, 1992.
modified artificial bee colony algorithm, in: Frontiers in Intelligent Computing:
[223] N.S. Utami, A. Jazidie, R.E.A. Kadier, Path planning for differential drive
Theory and Applications, Springer, 2020, pp. 25–36.
mobile robot to avoid static obstacles collision using modified crossover genetic
[250] F. Xu, H. Li, C.-M. Pun, H. Hu, Y. Li, Y. Song, H. Gao, A new global best guided
algorithm, in: 2019 International Seminar on Intelligent Technology and Its
artificial bee colony algorithm with application in robot path planning, Appl.
Applications, ISITIA, IEEE, 2019, pp. 282–287.
Soft Comput. 88 (2020) 106037.
[224] J.-B. Receveur, S. Victor, P. Melchior, Autonomous car decision making and [251] S. Kumar, A. Sikander, Optimum mobile robot path planning using improved
trajectory tracking based on genetic algorithms and fractional potential fields, artificial bee colony algorithm and evolutionary programming, Arab. J. Sci. Eng.
Intell. Serv. Robot. 13 (2) (2020) 315–330. 47 (3) (2022) 3519–3539.
[225] R. Storn, K. Price, Differential evolution–A simple and efficient heuristic for [252] S. Mirjalili, S.M. Mirjalili, A. Lewis, Grey wolf optimizer, Adv. Eng. Software
global optimization over continuous spaces, J. Global Optim. 11 (4) (1997) 69 (2014) 46–61.
341–359. [253] J. Liu, X. Wei, H. Huang, An improved grey wolf optimization algorithm and
[226] H. Guo, D. Cao, H. Chen, Z. Sun, Y. Hu, Model predictive path following control its application in path planning, IEEE Access 9 (2021) 121944–121956.
for autonomous cars considering a measurable disturbance: Implementation, [254] F. Kiani, A. Seyyedabbasi, R. Aliyev, M.U. Gulle, H. Basyildiz, M.A. Shah,
testing, and verification, Mech. Syst. Signal Process. 118 (2019) 41–60. Adapted-RRT: Novel hybrid method to solve three-dimensional path planning
[227] P.J. Van Laarhoven, E.H. Aarts, Simulated annealing, in: Simulated Annealing: problem using sampling and metaheuristic-based algorithms, Neural Comput.
Theory and Applications, Springer, 1987, pp. 7–15. Appl. 33 (22) (2021) 15569–15599.
[228] Y. Magdy, O.M. Shehata, M. AbdelAziz, M. Ghoneima, F. Tolbah, Metaheuris- [255] J. Han, Y. Seo, Mobile robot path planning with surrounding point set and path
tic optimization in path planning of autonomous vehicles under the ATOM improvement, Appl. Soft Comput. 57 (2017) 35–47.
framework, in: 2017 IEEE International Conference on Vehicular Electronics [256] M. Zafar, J. Mohanta, A. Keshari, et al., GWO-Potential field method for mobile
and Safety, ICVES, IEEE, 2017, pp. 32–37. robot path planning and navigation control, Arab. J. Sci. Eng. 46 (8) (2021)
[229] J. Yin, W. Fu, A hybrid path planning algorithm based on simulated annealing 8087–8104.
particle swarm for the self-driving car, in: 2018 International Computers, [257] X. Meng, Y. Liu, X. Gao, H. Zhang, A new bio-inspired algorithm: Chicken
Signals and Systems Conference, ICOMSSC, IEEE, 2018, pp. 696–700. swarm optimization, in: International Conference in Swarm Intelligence,
[230] Z. Shang, J. Gu, J. Wang, An improved simulated annealing algorithm for the Springer, 2014, pp. 86–94.
capacitated vehicle routing problem, Comput. Integr. Manuf. Syst. 13 (2020). [258] X. Liang, D. Kou, L. Wen, An improved chicken swarm optimization algorithm
[231] J. Kennedy, R. Eberhart, Particle swarm optimization, in: Proceedings of and its application in robot path planning, IEEE Access 8 (2020) 49543–49550.
ICNN’95-International Conference on Neural Networks, Vol. 4, IEEE, 1995, pp. [259] B. Crawford, R. Soto, N. Berríos, F. Johnson, F. Paredes, Binary cat swarm
1942–1948. optimization for the set covering problem, in: 2015 10th Iberian Conference
[232] G. Li, W. Chou, Path planning for mobile robot using self-adaptive learning on Information Systems and Technologies, CISTI, IEEE, 2015, pp. 1–4.
particle swarm optimization, Sci. China Inf. Sci. 61 (5) (2018) 1–18. [260] D. Zhao, H. Yu, X. Fang, L. Tian, P. Han, A path planning method
based on multi-objective cauchy mutation cat swarm optimization algo-
[233] J. Zhang, F. Yang, X. Weng, An evolutionary scatter search particle swarm
rithm for navigation system of intelligent patrol car, IEEE Access 8 (2020)
optimization algorithm for the vehicle routing problem with time windows,
151788–151803.
IEEE Access 6 (2018) 63468–63485.
[261] X.-S. Yang, Nature-Inspired Metaheuristic Algorithms, Luniver Press, 2010.
[234] L. Zhang, Y. Zhang, Y. Li, Mobile robot path planning based on improved
[262] J. Zhou, P. Chen, H. Liu, J. Gu, H. Zhang, H. Chen, H. Zhou, Improved
localized particle swarm optimization, IEEE Sens. J. 21 (5) (2020) 6962–6972.
path planning for mobile robot based on firefly algorithm, in: 2019 IEEE
[235] M.N. Ab Wahab, C.M. Lee, M.F. Akbar, F.H. Hassan, Path planning for
International Conference on Robotics and Biomimetics, ROBIO, IEEE, 2019, pp.
mobile robot navigation in unknown indoor environments using hybrid PSOFS
2885–2889.
algorithm, IEEE Access 8 (2020) 161805–161815.
[263] A.S. Bisen, V. Kaundal, Mobile robot for path planning using firefly algo-
[236] T. Qiuyun, S. Hongyan, G. Hengwei, W. Ping, Improved particle swarm
rithm, in: 2020 Research, Innovation, Knowledge Management and Technology
optimization algorithm for AGV path planning, IEEE Access 9 (2021)
Application for Business Sustainability, INBUSH, IEEE, 2020, pp. 232–235.
33522–33531.
[264] F. Li, X. Fan, Z. Hou, A firefly algorithm with self-adaptive population size for
[237] S.J. Fusic, G. Kanagaraj, K. Hariharan, S. Karthikeyan, Optimal path planning of global path planning of mobile robot, IEEE Access 8 (2020) 168951–168964.
autonomous navigation in outdoor environment via heuristic technique, Transp. [265] N.A.F. Abbas, Mobile robot path planning optimization based on integration
Res. Interdiscipl. Perspect. 12 (2021) 100473. of firefly algorithm and quadratic polynomial equation, in: The International
[238] R. Chai, A. Tsourdos, A. Savvaris, S. Chai, Y. Xia, Solving constrained trajectory Conference on Intelligent Systems & Networks, Springer, 2021, pp. 538–547.
planning problems using biased particle swarm optimization, IEEE Trans. [266] X.-S. Yang, S. Deb, Cuckoo search via Lévy flights, in: 2009 World Congress on
Aerosp. Electron. Syst. 57 (3) (2021) 1685–1701. Nature & Biologically Inspired Computing, NaBIC, Ieee, 2009, pp. 210–214.
[239] T. Zhang, J. Xu, B. Wu, Hybrid path planning model for multiple robots [267] M. Alireza, D. Vincent, W. Tony, Experimental study of path planning problem
considering obstacle avoidance, IEEE Access 10 (2022) 71914–71935. using EMCOA for a holonomic mobile robot, J. Syst. Eng. Electron. 32 (6)
[240] M. Dorigo, L.M. Gambardella, Ant colony system: A cooperative learning (2021) 1450–1462.
approach to the traveling salesman problem, IEEE Trans. Evol. Comput. 1 (1) [268] X. Jiang, S. Li, BAS: Beetle antennae search algorithm for optimization
(1997) 53–66. problems, 2017, CoRR arXiv: abs/1710.10724.
[241] Y. Ge, T. Chen, X. Kong, L. Gao, Application of improved ant colony algorithm [269] J. Wang, H. Chen, BSAS: Beetle swarm antennae search algorithm for
in car navigation, Control Eng. China 23 (1) (2016) 133–137. optimization problems, 2018, arXiv preprint arXiv:1807.10470.

44
M. Reda et al. Robotics and Autonomous Systems 174 (2024) 104630

[270] T. Wang, L. Yang, Beetle swarm optimization algorithm: Theory and application, Institute for Tribotechnology. Ahmed has taken different
2018, arXiv preprint arXiv:1808.00206. roles at the University of Central Lancashire School of
[271] X. Jiang, Z. Lin, T. He, X. Ma, S. Ma, S. Li, Optimal path finding with beetle Engineering since 2014: Lecturer, Senior Lecturer, Principal
antennae search algorithm by using ant colony optimization initialization and Lecturer, and Academic Lead for the Mechanical and Main-
different searching strategies, IEEE Access 8 (2020) 15459–15471. tenance Engineering area. Ahmed is currently the Deputy
[272] R.V. Rao, V.J. Savsani, D. Vakharia, Teaching–learning-based optimization: Head of the School of Engineering for Business Development
A novel method for constrained mechanical design optimization problems, and Partnerships, driving the School of Engineering UK and
International Partnerships. Ahmed is a member of the Uni-
Comput.-Aided Des. 43 (3) (2011) 303–315.
versity Research and Innovation Committee. He successfully
[273] A.D. Sabiha, M.A. Kamel, E. Said, W.M. Hussein, Real-time path planning for
placed the Mechanical Engineering Programmes in the top
autonomous vehicle based on teaching–learning-based optimization, Intell. Serv.
15 universities in the UK in 2020 NSS results in overall
Robot. (2022) 1–18.
student satisfaction.
[274] X. Dai, Y. Wei, Application of improved moth-flame optimization algorithm for
robot path planning, IEEE Access 9 (2021) 105914–105925.
[275] S. Mirjalili, Moth-flame optimization algorithm: A novel nature-inspired Amira Y. Haikal received the B.Sc., M.Sc., and Ph.D. de-
heuristic paradigm, Knowl.-Based Syst. 89 (2015) 228–249. grees from the Computers and Control Systems Engineering
Department, Mansoura University, Egypt. She is the vice
dean of environmental and community affairs and the head
of the Computers and Control Systems Engineering Depart-
Mohamed Reda Mohamed received his B.Sc. and M.Sc. ment, Faculty of Engineering, Mansoura University. Her
degrees in Computers and Control Systems Engineering major research interests include artificial intelligence fields:
from the Faculty of Engineering at Mansoura University, meta-heuristic optimization techniques, machine learning,
Egypt. He is currently pursuing his Ph.D. at the School deep learning, and smart systems engineering.
of Engineering, University of Central Lancashire (UCLan)
in the UK. Since 2016, he has been an Assistant Lecturer
at the Faculty of Engineering, Mansoura University, and Ali Ghanbari received his Ph.D. in Mechanical Engineering
an Associate Lecturer at UCLan’s School of Engineering from Amirkabir University of Technology in 2011. His
since 2022. His primary research areas include artificial research focuses on control of microrobots for biomedical
intelligence, meta-heuristic optimization techniques, evolu- applications. Dr Ghanbari is interested in how to make and
tionary computation, and control engineering. Additionally, model intelligent devices at small scale. He was a researcher
Mohamed has a strong interest in the application of AI and in Robotics Engineering Department at Daegu Gyeongbuk
deep learning to embedded systems, biomedical systems, Institute of Science and Technology (DGIST) and University
and robotics, with a particular focus on automated vehicles. of Leeds working on micro-/nanorobots and soft robotics. Dr
Ghanbari was also a guest researcher at Multi-Scale Robotics
Lab in ETH Zurich. Currently, he is a lecturer in the School
Ahmed Onsy awarded his Ph.D. from the School of Mechan- of Engineering and Computing at the University of Central
ical and Systems Engineering, Design Unit and Mechatronics Lancashire.
Group, Newcastle University, UK. His main research in-
terests are intelligent diagnostics and health management
systems, intelligent maintenance systems, advanced mecha-
tronics, and embedded systems, which can be directly
applied to Intelligent Diagnostic and Health Management
(DHM) and Predictive Health Monitoring (PHM) systems for
oil well, wind turbine, aerospace (SHM & HUM), marine
and automotive applications. He is a member of the Jost

45