Progress of Dynamic Path Optimization Strategies for Racing Cars

Based on Machine Learning

Zihan An

Mathematics, Monica Collage, Santa Monica, California, U.S.A.

Keywords: Autonomous Racing, Path Optimization, Reinforcement Learning, Multi-Sensor Fusion, Simulation.

Abstract: Motor racing has always been a globally popular sport that pursues the ultimate and the extreme. Among

them, the racing route is undoubtedly a very important aspect. This article will elaborate in detail on how to

use machine learning technology to improve the research progress of future racing in dynamic optimization

of routes. Next, computer dynamic vision algorithms such as You Only Look Once (YOLO) and Faster R-

CNN will be introduced in detail. Meanwhile, the hardware that assists in obtaining information will also be

introduced one by one. And combine this with the simulators commonly used by drivers to create the

possibility of updated simulator driving, and at the same time, this can also be utilized to cultivate better

training plans. At the same time, the challenges that this plan may face will also be taken into account. The

first one is the lack of precise and large amounts of data. At the same time, the ethical decision-making of AI

and its temporary response ability on the track also need to be considered.

1 INTRODUCTION

On January 29, 1886, with the birth of the world's first

car, "Benz Patent-Motorwagen", humanity's yearning

for speed gave rise to a passionate and challenging

sport - racing. Throughout the history of automobiles,

motor racing has always been a test of automotive

industry technology and human driving skills. As

motor racing progresses, the competition has evolved

from a simple comparison of speed to a sport that

integrates aerodynamics, materials science,

intelligent control, team strategy, and human

physiological limits, and pursues precision. In this

sport where the gap between drivers is one-

thousandth of a second, drivers not only have to

endure extremely strong physical acceleration but

also face significant psychological pressure.

However, to get closer to this limit, the "best lane"

undoubtedly holds a core position. The "racing lane"

is an ideal track that allows the racing car to pass

through the entire track at its highest average speed.

This route requires consideration of the curvature of

the late arrival, the condition and temperature of the

road surface, the physical limits of the vehicle, as well

as the connection between entering and exiting the

corner and the center of the corner in the curve. It also

https://orcid.org/0009-0001-3299-3735

greatly tests the driver's understanding and control of

speed and stability. At the same time, the racing line

is also one of the key factors for winning the race

(González et al., 2015).

However, traditional route optimization

extremely tests the personal talent and track

experience of drivers, which undoubtedly leads to

very high costs and great difficulties. Throughout the

history of motorsport, with the advancement of

human technology, racing cars have been upgraded

time and again at the technical level. The currently

emerging artificial intelligence possesses extremely

powerful data analysis and pattern recognition

capabilities, which undoubtedly make it highly

suitable for application in the racing field. It can

reduce the operating costs of teams and, through the

self-learning of machines, more efficiently find the

best lane for racing car tuning.

This article systematically reviews and

summarizes the research progress of AI in the optimal

route design and planning of racing cars, elaborates

on how to solve the problems mentioned above

through the application of object detection, and

specifically describes how artificial intelligence can

help teams formulate better route planning and

strategies.

An, Z.

Progress of Dynamic Path Optimization Strategies for Racing Cars Based on Machine Learning.

DOI: 10.5220/0014362600004718

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 2nd International Conference on Engineering Management, Information Technology and Intelligence (EMITI 2025), pages 545-550

ISBN: 978-989-758-792-4

545

2 RACING CAR ENVIRONMENT

PERCEPTION AND TARGET

DETECTION

Before applying artificial intelligence to fleet

planning and route planning, accurately perceiving

the environment and collecting information are

undoubtedly the primary prerequisites for making

precise judgments. Take Formula One as an example.

Usually, for the convenience of collecting

information, a standard ECU system is installed to use

mobile phone information and transmit it to the

telemetry center. Team staff can understand vehicle

performance in real time and intuitively, and check

engine health, tire wear and fuel consumption. Many

current research reports have mentioned that a

perception system of Lupin is designed to achieve

end-to-end autonomous driving and is suitable for

rapid judgment and planning under extreme working

conditions. As mentioned above, if artificial

intelligence is to be applied to motorsports, it is

necessary to meet the requirements of real-time and

accurate identification of track conditions and the

positions of competing vehicles on the same track, as

well as the judgment and use of special terrains such

as shoulders and buffer zones. Therefore, in order to

obtain information, a series of commonly used

sensors, such as cameras, may need to be equipped

Devices such as LiDAR and millimeter-wave radar,

in combination with advanced algorithms, interpret

this sensor data(Williams et al., 2017).

2.1 Racing Car Target Detection Based

on Computer Vision

Unlike traditional racing car sensors, in terms of

object detection in computer vision, especially the

object detection algorithms based on deep learning,

they play a significant role in processing the captured

image data. Currently, there are two mainstream

algorithms They are respectively "One-Stage

Detectors" and "Two-Stage Detectors" (Kapania et

al., 2016; Kendall et al., 2019).

Firstly, regarding "One-Stage Detectors", the

characteristic of this type of algorithm is that it has an

extremely high recognition speed and is more

efficient. It can form a Bounding Box at the edge of

an object just by looking at a complete image, just like

the human eye, and analyze the category. This fast

response and processing algorithm is highly suitable

for real-time scenarios that require reactions at the

level of one-thousandth of a second. Moreover, in the

F1TENTH Challenge, the researchers precisely

utilized the advantage of You Only Look Once

(YOLO) among One-Stage Detectors (Redmon et al.,

2016). Ultra-high-frequency detection of track drop

pain has been achieved.

Secondly, there are "Two-Stage Detectors",

among which the most representative one is Faster R-

CNN. The characteristics of this type of algorithm are

that the analysis results are more accurate and the

detection accuracy is also very high. It will first scan

the image and propose Region Proposals that may

contain objects. Then, conduct a detailed analysis and

classification of each of these areas, one by one.

Although it is more accurate compared to "one-stage

Detectors", it can also greatly enhance the feasibility

and safety of the strategy.

The reason why these algorithms can recognize

objects on the track is through "learning".

Researchers will prepare a dataset with many track

maps in advance. Then, they will manually process

these images to standards, such as drawing racing

lines on the track, marking reference points,

perfecting track details, etc. They will label each

opponent's car and convert it into structured and

semantic information that machines can understand

(Geiger et al., 2012). This can also help the machine

learning stage obtain more direct input (Ren et al.,

2015).

2.2 Racing Perception Technology and

Multi-Sensor Fusion

Considering the uncontrollability of track conditions,

such as weather, humidity, and light intensity, it is

crucial to introduce sensors based on other principles

for information complementarity in order to stably

ensure that the sensors receive information. First of

all, there is LiDAR. It accurately predicts the distance

to objects by emitting laser beams that are invisible to

the human eye nearby and calculating the time it takes

for the beam to reflect and return. It can instantly

construct an accurate "three-dimensional point cloud

map" composed of a vast number of data points. This

map is like a virtual 3D model, clearly depicting the

geometric outline of the track, the undulations of the

road surface, and even the precise shapes and

positions of the obstacles.

Another feasible option is millimeter-wave radar.

Compared with LiDAR, it uses radar wave detection

and can directly measure the relative speed of the

target, which can largely avoid collisions.

However, when fusing these sensory sensors,

Early Fusion and Late Fusion are usually adopted.

Firstly, the Early Fusion strategy is typically used to

blend the data at the most primitive stage. A typical

EMITI 2025 - International Conference on Engineering Management, Information Technology and Intelligence

546

example is projecting the 3D point cloud data of

LiDAR onto the 2D images captured by the camera.

This means that the halo of three-dimensional

coordinates uses RGB color information, which

implies that a virtual track environment can be better

constructed, and the track situation can be analyzed

more accurately through condition clues such as color

and texture. However, the drawback is that the

computational load is large, and the spatio-temporal

synchronization requirements for each sensor are

extremely high, which poses certain obstacles in real-

time performance.

Late Fusion is just the opposite. Late Fusion

enables the sensor system to independently complete

the parts of perception and target detection, and draw

preliminary conclusions, which are then passed to the

central module for summarization and arbitration of

these conclusions. If it is reflected on the track, the

visual system should identify the distance between

the car in front and the car behind, and then output the

distance to the next corner center through the LiDAR

system. After that, the relative speed with the car

behind should be output through the radar system to

help analyze the difference in corner speed and obtain

a better corner exit Angle. The fusion center will

aggregate these data together and conduct a

comprehensive analysis, ultimately reaching a

comprehensive judgment. This logic is clearer.

Moreover, due to the modularity, if an accident or

collision occurs in the racing car and a certain sensor

malfunctions, other parts can still operate normally,

thus providing stronger stability for the system

(Kaufmann et al., 2018).

3 BEST PATH PLANNING FOR

RACING CARS

In the control system of artificial intelligence racing

cars, if the environmental perception system is the

"eyes", then the programmed autonomy that regulates

driving decisions is the "brain" that makes decisions.

The definition of racing path planning is to find a

driving path that allows the racing car to pass through

the track safely and stably, while adhering to various

constraints, such as the vehicle's dynamic response,

the maximum adhesion use of tires, the track edge,

and the best and fastest path within one lap or a

section of the track. This means converting

environmental information from perception

"processing" (i.e., track edge position, obstacle

position, etc.) into specific and actionable driving

instructions (i.e., steering wheel Angle, accelerator

and brake load), thereby providing a crucial bridge in

the connection and integration of perception and

control processes(Pfaff et al., 2021).

3.1 Racing Car Target Detection Based

on Computer Vision

When familiar with the track conditions and layout

and having sufficient data, traditional optimization

methods are very effective in finding the optimal

solution. One of the most prominent examples is

Model Predictive Control (MPC).

For instance, one can think about how the MPC

works with the mindset of a highly skilled F1 driver.

Drivers have a very good understanding of the

performance of their cars (which is equivalent to an

accurate vehicle dynamics model). When driving a

series of curves at a speed of 300 kilometers per hour,

their gaze is not on the center of the curve they are

about to pass, but looks towards the next center and

predicts the best combination route for passing the

next two or three curves to achieve the best total time

(this is predicted within a short "time range").

The driver can quickly calculate the "optimal

sequence of actions" (steering Angle, accelerator and

brake), thereby passing through the corner at the

highest possible speed while respecting the car and

the track. But importantly, they do not immediately

apply the entire optimal solution. They will take the

actions necessary to accurately enter the first corner,

but when they enter the corner, they will recalculate

the plan again based on the real-time status of the car

(such as the size of the tire grip or the change in the

car's posture) in order to modify the order of jumping

over the next corner.

This rolling optimization approach, which

involves planning, predicting and correcting, makes

the MPC's work very much like that of an experienced

F1 driver, continuing to maximize the vehicle's

performance within its physical limits. As a result,

they create a very smooth and positive "optimal

racing line"(Pavel et al., 2022).

3.2 Racing Path Planning Based on

Reinforcement Learning

Although model-based optimization often works

well, model-based optimization methods rely on

accurate vehicle dynamics models. Reinforcement

Learning (RL) offers an interesting option. The

principle of RL is to learn the best strategic

interactions in the environment through "trial and

error", and it usually does not require a detailed

physical model, which can also provide a large

tolerance for errors. In the field of motorsport, SONY

Progress of Dynamic Path Optimization Strategies for Racing Cars Based on Machine Learning

547

AI's "Gran Turismo Sophy" has achieved significant

success. This game uses deep reinforcement learning

technology to defeat professional drivers with world-

class driving skills in the simulator. Overall, RL is

developing rapidly.

Initially, Q-learning was a method of filling the

value "lookup table" with all possible states or

actions, which was impossible in the complex racing

environment. The next significant advancement is

Deep Q-Network (DQN), as AI can use deep neural

networks to estimate all these action values and learn

directly from the data and information provided by

complex observations, such as camera lenses.

The next-generation reinforcement learning, namely

the latest policy gradient method, is more

straightforward because artificial intelligence can

learn a "policy network" to select the next action.

Among the Policy gradients, Proximal Policy

Optimization (PPO) is the most commonly used and

accepted algorithm for complex autonomous control

tasks, such as racing, due to its training consistency

and speed, which is an evolution of the algorithm that

allows AI to adopt complex racing strategies (Chen et

al., 2018).

3.3 Racing Path Smoothing and

Tracking Control

Whether generated through optimization theory or

reinforcement learning, the "optimal path" produced

by advanced planning modules is typically composed

of a series of discrete road points. Although this path

might be the most optimal on a macroscopic level,

simply connecting these points would form a zigzag

line with many sharp turns, which is impossible for a

real vehicle to travel smoothly at high speeds.

Therefore, before sending the path to the vehicle's

actuator, two key subsequent steps must be taken,

namely path smoothing and path tracking control.

The goal of achieving the path smoothing step is

to convert the discrete sequence of path points into a

geometrically continuous smooth curve. This curve

must be physically feasible for the vehicle to follow,

which means its curvature must be continuous to

allow for a smooth steering input. Common

techniques include the use of mathematical tools,

such as Bezier curves or mathematical curves similar

to B-Splines, to fit new trajectories to the original

points and create a path that is both close to the

optimal route and inherently smooth.

Realizing path tracking control is the final step to

putting a smooth and ideal trajectory into practice. Its

task is to design a controller that continuously

compares the actual position of the vehicle with the

ideal trajectory and calculates the necessary steering

angles and accelerator/brake commands in real time

to minimize tracking errors. In the field of

autonomous driving, there are several classic path

tracking control methods:

First of all, there is the Pure Pursuit Controller,

which is a very intuitive geometry-based control

method. The core idea is to select a target point on the

ideal path in front of the vehicle with a fixed "Look-

ahead Distance". Then, the controller calculates the

curvature of a perfect arc, which starts from the

current position of the vehicle and intersects with this

forward-looking point. This curvature is directly

converted into the required steering Angle, guiding

the vehicle towards the target.

Secondly, there is the Stanley Controller. This is

the famous controller used by the Stanford University

team to win the DARPA Grand Challenge, and it

shows greater stability at high speeds. It operates by

simultaneously correcting two core errors: one is the

"Cross-track Error", that is, the distance from the

front axle of the vehicle to the nearest point on the

path; The other one is "Heading Error", which is the

Angle difference between the direction of the vehicle

and the path tangent at that point. By correcting these

two types of errors, the Stanley controller enables the

aircraft to converge to the target trajectory quickly

and smoothly.

These path tracking controllers form the lowest

level of the "brain" of an AI racing driver. They are

the "nerve endings" that faithfully execute advanced

strategic intentions, ensuring that every maneuver

precisely follows the planned "optimal racing route".

4 CONSTRUCTION AND

SIMULATION TRAINING OF

VIRTUAL RACING SCENES

In the sport of racing, which is highly dangerous and

costly, it is unrealistic to directly apply the

development, training and debugging of AI

algorithms to real cars. Take the current Formula One

as an example. The annual research and development

budget cap is 175 million US dollars. Therefore,

every test error may compress the overall budget of

the team and also pose a great risk. Thus, the virtual

simulation environment plays a crucial role in the

training and learning of AI.

Virtual scenes can provide a safe, low-cost,

repeatable and efficient testing condition, which can

give AI sufficient space for trial and error. More

importantly, it can simulate various extreme weather

EMITI 2025 - International Conference on Engineering Management, Information Technology and Intelligence

548

and dangerous working conditions that are difficult to

encounter in the real world, and also enhance the

robustness of AI algorithms.

First of all, open-source simulators, such as

CARLA, are an important part of autonomous driving

research (Dosovitskiy et al., 2017). As an open-source

simulator, CARLA has a rich API that enables

researchers to configure sensors (such as cameras,

lidars, etc.), traffic flow, and various environmental

conditions (such as weather, lighting, etc.). This

makes CARLA a useful research simulator for testing

perception and planning algorithms running in

software. Microsoft's AirSim is another open-source

emulator that offers excellent physics and realistic

features. AirSim features SIL/HIL functionality,

which means developers can test their algorithms in

an environment similar to hardware deployment.

Secondly, some commercial racing games, such

as "Assetto Corsa", "irracing" and "Gran Turismo

Sport", based on minimal physical modeling, provide

excellent realism in vehicle dynamics and also offer

track data from laser scans. This level of fidelity

significantly enhances the realism of driving

feedback, which builds more accurate scenarios for

the teaching of reinforcement learning agents that

require precise physical interaction. For example,

Project (Pfaff et al., 2021) of "Gran Turismo Sophy"

utilized the high fidelity of the Gran Turismo

commercial racing simulator to train AI systems to

compete with human drivers.

These simulation platforms all provide

researchers with an interface (such as a Python API),

allowing control commands to be sent to the

simulation platform and vehicle status and related

sensor information to be read. Basically, each

platform provides a complete definition of robots or

agents that learn in a closed-loop "AI simulator"

training paradigm.

5 CURRENT LIMITATIONS AND

FUTURE PROSPECTS

Although the dynamic path optimization and

planning strategies for racing cars through machine

learning have great potential, when truly

implemented in the padtrack, there are still many

hidden dangers and limitations, which also point out

the ultimate research direction of this study in the

future.

5.1 Limitations of the Dataset

One major issue in this research is that it is difficult

to obtain high-precision, diverse and genuine racing

car data without the confidential information from the

event officials and some teams. However, compared

with the general autonomous driving field where

there is a lot of public information, because racing car

data has high commercial value and confidentiality,

the leakage of racing car data will cause some teams

to lose their competitiveness So this is also the main

protected information for many teams, which makes

it very difficult for researchers to obtain real and

reliable information. Even if the simulator can

produce some data, there will still be a significant

difference from the real data (Sakai et al., 2022).

5.2 Ethical Challenges

Firstly, in the complex and highly dangerous

environment of the track, when facing unavoidable

accidents, the decisions made by AI are also very

important. One is to preserve the vehicle and reduce

damage, and the other is to protect other drivers on

the track. Therefore, such moral and ethical decisions

go beyond the technical scope and can be a hidden

danger for AI.

So Wiesmüller think if AI can only be used for

human-machine collaboration, race engineers or team

managers might be more in line with the situation and

be able to intervene at the moment of mistakes.

Perhaps this is more likely to be the future application

direction (Wiesmüller, 2023).

6 CONCLUSIONS

This article systematically studies and reviews how

machine learning can be applied to the planning and

optimization techniques of dynamic racing paths.

Research shows that if AI is to be applied to racing

cars, it requires the support and integration of key

technologies such as perception, planning, and

simulation.

Although there are current issues such as data

scarcity, insufficient hardware computing power, and

ethical challenges in AI, it is already clear that this

sector will have a clear application direction and

development blueprint in the future. AI can be deeply

integrated with vehicle engineering, incorporating

digital twin technology to optimize the engineering

details of vehicles in real time, endowing racing cars

with self-regulating and self-summarizing functions.

Secondly, more efficient simulation display

Progress of Dynamic Path Optimization Strategies for Racing Cars Based on Machine Learning

549

technologies can be studied and produced through

predictive algorithms, ultimately achieving multi-

agent collaboration in motorsports. This will also

largely change the current strategies of motorsports

competitions.

In conclusion, there is still much room for

improvement in artificial intelligence in extreme

sports such as car racing. However, as it develops, car

racing will evolve from a mere test of drivers' skills

and teams' strategies into an era that promotes data-

driven, competitive, and human-machine

interconnection and collaboration.

REFERENCES

Chen, L. C., Zhu, Y., Papandreou, G., Schroff, F., & Adam,

H. (2018). Encoder-decoder with atrous separable

convolution for semantic image segmentation. In

Proceedings of the European conference on computer

vision (ECCV) (pp. 801–818).

Dosovitskiy, A., Ros, G., Codevilla, F., Lopez, A., &

Koltun, V. (2017, October). CARLA: An open urban

driving simulator. In Conference on robot learning (pp.

1–16). PMLR.

Geiger, A., Lenz, P., & Urtasun, R. (2012, June). Are we

ready for autonomous driving? The KITTI vision

benchmark suite. In 2012 IEEE Conference on

Computer Vision and Pattern Recognition (pp. 3354–

3361). IEEE.

González, D., Pérez, J., Milanés, V., & Nashashibi, F.

(2015). A review of motion planning techniques for

automated vehicles. IEEE Transactions on Intelligent

Transportation Systems, 17(4), 1135–1145.

Kapania, N. R., Subosits, J., & Christian Gerdes, J. (2016).

A sequential two-step algorithm for fast generation of

vehicle racing trajectories. Journal of Dynamic Systems,

Measurement, and Control, 138(9), 091005.

Kaufmann, E., Loquercio, A., Ranftl, R., Dosovitskiy, A.,

Koltun, V., & Scaramuzza, D. (2018, October). Deep

drone racing: Learning agile flight in dynamic

environments. In Conference on Robot Learning (pp.

133–145). PMLR.

Kendall, A., Hawke, J., Janz, D., Mazur, P., Reda, D., Allen,

J. M., ... & Shah, A. (2019, May). Learning to drive in

a day. In 2019 International Conference on Robotics

and Automation (ICRA) (pp. 8248–8254). IEEE.

Ma, H., Sun, Y., Li, J., & Tomizuka, M. (2021). Multi-agent

driving behavior prediction across different scenarios

with self-supervised domain knowledge. In 2021 IEEE

International Intelligent Transportation Systems

Conference (ITSC) (pp. 3122–3129).

Pavel, M. I., Tan, S. Y., & Abdullah, A. (2022). Vision-

based autonomous vehicle systems based on deep

learning: A systematic literature review. Applied

Sciences, 12(14), 6831.

Redmon, J., Divvala, S., Girshick, R., & Farhadi, A. (2016).

You only look once: Unified, real-time object detection.

In Proceedings of the IEEE Conference on Computer

Vision and Pattern Recognition (pp. 779–788).

Ren, S., He, K., Girshick, R., & Sun, J. (2015). Faster R-

CNN: Towards real-time object detection with region

proposal networks. Advances in Neural Information

Processing Systems, 28.

Ugurlu, H. I., Pham, X. H., & Kayacan, E. (2022). Sim-to-

real deep reinforcement learning for safe end-to-end

planning of aerial robots. Robotics, 11(5), 109.

Wiesmüller, S. (2023). The relational governance of

artificial intelligence. Relational Economics and

Organization Governance.

Williams, G., Wagener, N., Goldfain, B., Drews, P., Rehg,

J. M., Boots, B., & Theodorou, E. A. (2017, May).

Information theoretic MPC for model-based

reinforcement learning. In 2017 IEEE International

Conference on Robotics and Automation (ICRA) (pp.

1714–1721). IEEE.

EMITI 2025 - International Conference on Engineering Management, Information Technology and Intelligence

550