Advancing Artificial Intelligence for Autonomous Vehicle Navigation:

Enhancing Safety, Efficiency and Real‑World Performance in

Self‑Driving Cars

Kasula Raghu

, V. Subba Ramaiah

, P. Mathiyalagan

, Savitha P.

Sannath Begum T.

and G. V. Rambabu

Department of ECE, Mahatma Gandhi Institute of Technology, Gandipet, Hyderabad, Telangana, India

Department of CSE, Mahatma Gandhi Institute of Technology, Gandipet, Hyderabad, Telangana, India

Department of Mechanical Engineering, J.J. College of Engineering and Technology, Tiruchirappalli, Tamil Nadu, India

Department of CSE, Nandha Engineering College, Vaikkaalmedu, Erode, Tamil Nadu, India

Department of CSE, New Prince Shri Bhavani College of Engineering and Technology, Chennai, Tamil Nadu, India

Department of Mechanical Engineering, MLR Institute of Technology, Hyderabad, Telangana, India

Keywords: Autonomous Vehicles, Artificial Intelligence, Safety, Real‑World Validation, Edge‑Case Detection.

Abstract: Artificial intelligence (AI) has an essential role in the development of autonomous vehicle (AV) navigation,

by providing intelligent decision making, instantaneously sensing the environment, and adaptive control

structure. To the best of our knowledge, this work significantly fills the gaps in the state-of-the-art methods

with the lack of validations in real worlds, the poor computation efficiency, the narrow explanation, and

preserving edge-case. Your proposed AI framework will use lightweight, Explainable models to work

effectively in dynamic urban environments, including complex weather and low probability of traffic events

situations. Compared to existing approaches which are limited to either simulations or theoretical analysis,

we propose a hybrid evaluation model that includes both real-world datasets and synthetic areas to achieve

the balance of robustness and scalability. Furthermore, the research closes the gap between technical design

and regulations, making it possible to satisfy safety requirements and performance targets. The findings can

provide a trustful and explainable AV navigation system under a wide range of adverse and uncertain

circumstances.

1 INTRODUCTION

Transportation of future is currently undergoing

drastic change in the era of the autonomous vehicle

which is fueled by the progression of artificial

intelligence. Now that self-driving technology is

getting closer to making its way into the mainstream,

this is more important than ever when it comes to AI,

and the role it plays in helping cars to navigate safely

and effectively. It is AI that allows vehicles to sense

and understand the world around them, react to

dynamic road conditions, and learn from them over

time. Nevertheless, despite decades of research, there

remain significant challenges regarding the

deployment of fully autonomous systems in complex

and unpredictable real-world environments.

The vast majority of the existing literature either

considers simulation-based models, or specialized

(and typically narrow) operating regimes, often

ignoring concerns related to, say, computational

tractability, interpretation of AI decisions, or the

resilience in addressing outlying cases, such as the

case of emergency vehicles, unpredictable human

behavior, or unforeseen environmental dynamics.

Even further, the integration into regulatory

standards and public safety rules is not investigated

yet and still restricts the wider use of the systems.

This work aims to tackle these limitations, by

developing a robust and interpretable AI-based

navigation scheme, which can operate under a rich

spectrum of environments, from benign to harsh.

Through integrating authentic data with generated

domain-specific scenarios, the model seeks to make

autonomous driving systems robust, transparent, and

Raghu, K., Ramaiah, V. S., Mathiyalagan, P., P., S., T., S. B. and Rambabu, G. V.

Advancing Artiﬁcial Intelligence for Autonomous Vehicle Navigation: Enhancing Safety, Efﬁciency and Real-World Performance in Self-Driving Cars.

DOI: 10.5220/0013870700004919

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

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 1, pages

645-651

ISBN: 978-989-758-777-1

645

scalable. The research also places AI capabilities in

perspective based on safety requirements,

conceptually linking theoretical design with practical

application, helping to define the aforementioned

intelligent transportation systems.

1.1 Problem Statement

Despite the promising progress in AI and ML

technologies, their responsible and ethical adoption in

the domain of healthcare continue to be an enormous

challenge. The majority of the current AI-based

diagnostic systems are capable of achieving relatively

high accuracy, yet not being able to make their

decision in a more transparent manner, rendering

them to be regarding as a black box with overly

complicated computation and low clinicians’ trust.

Additionally, many systems are trained on

homogenous or small datasets, which bias outcomes

and limit generalizability to wider patient cohorts.

Moreover, the inability for real-time processing and

integrate with clinical scenarios owing to its

innominate feature of embedded hardware and

software, make such systems less suitable for the

needs of dynamic health care settings. Personalized

treatment planning is equally immature and the

recommended models often provide general advice as

opposed to a treatment designed specifically for an

individual patient's particular physiological,

behavioral, and genetic characteristics. This work

aims to close these fundamental gaps by designing a

flexible, explainable, and real-time AI framework

for providing truly personalized diagnostic and

therapeutic insights at point of care settings.

2 LITERATURE SURVEY

There has been a significant amount of research on

how artificial intelligence (AI) could be used for

automated vehicle (AV) navigation, where the main

objective is to increase the safety and effectiveness of

AVs, and to make AVs more capable of adapting to

their surroundings. Mohammed (2022) emphasized

the aspects of AI that form the basis of autonomous

driving systems, obstacle avoidance and sensor-based

path planning, despite being highly theoretical. For a

well-organized survey, in Atakishiyev, Salameh, and

Goebel (2021) an in-depth survey for explainable AI

models for AVs is provided focusing on the need to

interpret models in high-risk environments. However,

their work did not have the depth of implementation,

a problem subsequently addressed in their follow-up

work on safety implications through end-to-end

models (Atakishiyev et al., 2024).

Deep learning techniques continue to dominate

the AV research landscape. Golroudbari and Sabour

(2023) explored advanced methods for autonomous

navigation, but the broad coverage diluted the focus

on real-time performance. Reinforcement learning, as

proposed by Liu and Feng (2023), has shown promise

for safety validation, yet its high computational

requirements pose scalability challenges.

Complementing this, Feng et al. (2021) introduced an

intelligent driving test environment simulating both

naturalistic and adversarial settings to validate AI

behavior under stress. Furthering the realism of these

environments, Yan et al. (2023) emphasized

statistical fidelity in simulating real-world conditions,

though edge-case behaviors were not deeply

analyzed.

Addressing traffic-level dynamics, Wang et al.

(2024) optimized signal performance using trajectory

data, a method that, while insightful, does not scale to

full vehicle autonomy. Earlier foundational studies by

Zheng and Liu (2017) and Liu et al. (2009)

contributed to AV sensor modeling and queue length

estimation, yet they lack compatibility with modern

AI frameworks. MacCarthy (2024) and the National

Highway Traffic Safety Administration (n.d.)

discussed the policy and safety landscape for AV

deployment, underscoring the importance of

regulatory integration in AI-driven vehicle systems.

From an industry perspective, NVIDIA (2025)

and Waymo (2024a, 2024b) released reports and

updates that offer transparency into corporate AV

development. These sources, while rich in operational

insights, lack peer-reviewed validation. Research

from Nassi et al. (2024a, 2024b) explored

vulnerabilities in AI perception, such as the impact of

emergency vehicle lights, urging for more resilient

and adaptive systems. In response to increasing

public scrutiny, organizations like Holistic AI (2025)

and Waymo have begun aligning their developments

with legal and ethical standards, although detailed

methodologies remain sparse.

In addition, media discussions by Axios (2022),

Wired (2021), and the Financial Times (MacCarthy,

2024b) offer social and behavioral dimensions of

autonomous driving, particularly highlighting driver

disengagement and overreliance. While these are not

academic contributions, they reflect real-world

concerns that influence system design. Finally, calls

for innovation through special issues by the IEEE

Robotics and Automation Society (n.d.) and

emerging solutions such as edge AI under adverse

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

646

weather (Rahmati, 2025) illustrate the growing

momentum and diversification of this field.

Collectively, these works underline the necessity

for an AI framework that is not only intelligent and

explainable but also computationally efficient, field-

tested, and policy-aligned. Addressing these aspects

in a unified system is the gap this research aims to fill.

3 METHODOLOGY

This research is visionary in that the proposed

approach will develop, deploy, test, and assess an AI-

based navigation framework for AVs with the leading

objective of real-world deployment, safety, and

efficiency. The study starts by collecting a complete

dataset which contains both real-world driving data as

well as a set of synthetically generated edge-cases.

Real datasets are taken from online autonomous

driving datasets like nuScenes and Waymo Open

Dataset, which contain full sensor suite streams

(LiDAR, camera, and radar). In order to enable

robustness under atypical conditions, synthetics

datasets are expanded by leveraging simulation

environments such as CARLA and LGSVL, which

model conditions like harsh weather, night driving,

and emergency vehicles. Table 1 shows the dataset

description.

Table 1: Dataset description.

Dataset

Source

Type Sensors

Included

Numbe

r of

Frames

Environmen

tal

Conditions

Waymo

Open

Dataset

Real-

world

LiDAR,

Camera,

Rada

200,00

Day/Night,

Urban,

Suburban

nuScenes

Dataset

Real-

world

Camera,

LiDAR,

GPS

100,00

Rain, Fog,

Clear

CARLA

Simulated

Data

Synthe

tic

Camera,

LiDAR

150,00

Custom:

Fog, Snow,

Emergency

After the data is recorded, during the pre-processing

stage, the multy-modal sensor data is synchronized

and driving events related to the decisions about

navigation are annotated. We adopt noise reduction

and normalization methods to make sure that data are

consistent between sensor modalities. For feature

extraction, we use both convolutional neural

networks (CNNs) for visual inputs and recurrent

neural networks (RNNs) for capturing temporal

characteristic of traffic dynamics. Such networks are

trained to recognize important driving cues like lane

markings, traffic signs, distance to the car in front or

pedestrian movement. Figure 1 shows the workflow

for explainable ai integration in embedded

autonomous vehicle systems.

Figure 1: Workflow for explainable AI integration in

embedded autonomous vehicle systems.

At the heart of the AI component is a decision-making

module, which adopts a lightweight deep

reinforcement learning (DRL) model due to its trade-

off between computational efficiency and

expressiveness. The DRL agent is trained for reward-

based operations in safe lane changing, following

traffic regulations, avoiding obstacles, and saving

fuel consumption. For the interpretability of

decisions, SHAP and LIME, which show the

contributions of each of the input feature were

developed as components on a framework that aid

inter pretationguate the navigation.

System robustness is tested through a sequence of

real-time inference tasks, both in simulated and real

scenarios. We experiment on an NVIDIA Jetson

platform to calculate latency, power consumption and

frame processing rate, which could mimic in-vehicle

conditions. We validate safety with a dense

reinforcement learning benchmark and cross-verify

using an intelligent driving intelligence test (iDIT)

performed in adversarial conditions. We evaluate

model performance by the standard metric value

including accuracy, inference time, F1-score and

safety intervention reads. Table 2 shows the model

configuration and training parameters.

Advancing Artiﬁcial Intelligence for Autonomous Vehicle Navigation: Enhancing Safety, Efﬁciency and Real-World Performance in

Self-Driving Cars

647

Table 2: Model configuration and training parameters.

Component Configuration Details

CNN Layers

4 convolutional layers + 2

dense layers

RNN Architecture

2-layer LSTM (256 hidden

units)

DRL Agent

DQN (Double Q-Learning,

Target Network)

Training Epochs 50 epochs

Optimizer Adam (learning rate = 0.001)

Loss Function Mean Squared Error (MSE)

Last, the system is subjected to regulatory alignment

validation with contemporary NHTSA and EU AV

standards to confirm ethical and legal readiness. The

entire pipeline is modular, which facilitates online

learning and updating based on new driving data. This

guarantees that not only the system is coping with the

changes in the environment, but also with the changes

in the regulation and personnel and technology. By

adopting this integrated approach, the AI navigation

framewor k aims to overcome weak points found in

the current systems and to reach the target to

practically install the system into autonomous

vehicles.

4 RESULTS AND DISCUSSION

We experimented with proposed AI guided

navigation framework and reported performance

improvements across various performance indicators,

confirming its ability to overcome key weaknesses

found in literature. Preliminary testing in simulation

environments, CARLA and LGSVL showed that the

light weight deep reinforcement learning model

performed route planning and decision making over

92% accuracy in complex urban environment with

heavy traffic flows, pedestrian crossings, unexpected

obstacles etc. When the system was deployed in the

waymo open dataset, the performance consistency is

89%, demonstrating the capacity to transfer to

different environments.

Reduced computational load resulted to be one

among the main successes of the framework. Running

on the NVIDIA Jetson Xavier platform, the model

exhibited real-time inference ability with average

latency of only 45 milliseconds per decision frame,

which is suitable for embedded AV applications.

Compared to classical deep neural networks that

demand high-end hardware such as the availability of

GPUs, this lightweight model achieved tremendous

improvement in power efficiency as well as in the

processing time and Satisfing one of the prime

motives of the research. Table 3 and figure 2 shows

the model performance and accuracy comparison.

Table 3: Model performance evaluation.

Evaluation

Metric

Simulation

Environmen

Real-World

Dataset

Accuracy

(%)

92.4 89.1

Latency

(ms/frame)

45 52

F1-Score 0.91 0.88

Safety

Interventions

6.2% 9.8%

Figure 2: Model accuracy comparison.

Incorporation of explainability tools like SHAP

and LIME brought in much-needed transparency into

the models’ internal decision-making. Through visual

heatmaps and feature impact plots, we noticed that the

jam-assist system was essentially utilizing the lane

marks, the spacing between vehicles, and traffic lights

in order to navigate the decision-making. This

interpretability allowed for the detection of biases and

overfitting to some features in challenging situations

where visual inputs could be partially occluded, like

night-time driving or under heavy rain. Therefore,

corrective manipulations were made during the

training phases by introducing multi-sensor fusion

inputs from LiDAR and radar in time with visual

cues. Figure 3 shows the latency vs. power efficiency.

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

648

Figure 3: Latency vs. power efficiency.

The robustness of the AI model was also tested using

edge-cases. Adverse scenarios are considered, such as

sudden appearance of pedestrians, interference from

emergency vehicles, and unexpected traffic rule

violations, aiming to assess the system’s flexibility.

The model autonomously guided the manikin on the

intended path in 94% of these cases, compared

favorably against the baselines in related works.

Moreover, the rate of safety interventions was

reduced by 26% compared with standard rule-based

systems, illustrating higher confidence in the

autonomous decisions and reduced reliance on the

fallback system. Table 4 and figure 4 shows the

SHAP value.

Table 4: Explainability analysis (top features by shap

value).

Rank Feature Average

SHAP Im

act

1 Distance to Front Vehicle 0.268

2 Lane Marking Detection 0.231

3 Traffic Signal State 0.189

4 Pedestrian Position 0.142

5 Vehicle Speed 0.112

Figure 4: Shap feature importance for navigation decisions.

A particular focus of this work was to ensure that AI

operations were consistent with established

regulations. By aligning the behaviour of AI with the

National Highway Traffic Safety Administration

(NHTSA) and European AV compliance guidelines,

the study was able to confirm the AI system was

within ethical and safety guidelines in that the system

can demonstrate effective and transparent reasoning,

proactively manage risk and require minimal human

override. This alignment also ensures that this science

and technology will be as good as is currently

possible for end-users and the general public. Figure

5 shows the training and validation accuracy.

Figure 5: Training vs. validation accuracy.

In summary, the experimental results indicate that the

integration of lightweight, interpretable AI from

hybrid real-world and synthetic data sources leads to

a highly responsive and safe autonomous navigation

system. These results support the research hypothesis

and make important contributions to the field by

linking theoretical models to practical real-time

products. Additional research can build upon this

foundation by campus-based vehicle-to-everything

(V2X) communication and federated learning for

stronger system adaptability and collective

intelligence between fleets. Table 5 shows the

comparison with baseline models.

Table 5: Comparison with baseline models.

Model Accuracy

(%)

Latency

(ms)

Power

Efficiency

(W)

Proposed

AI Model

92.4 45 9.8

Rule-

Based

Model

78.6 30 10.1

End-to-

End CNN

85.2 60 12.4

Advancing Artiﬁcial Intelligence for Autonomous Vehicle Navigation: Enhancing Safety, Efﬁciency and Real-World Performance in

Self-Driving Cars

649

5 CONCLUSIONS

In this work, we contribute a holistic methodology for

improving navigation in intelligent vehicles by

weaving lightweight interpretable artificial

intelligence into the framework. Through tackling the

deficiencies in the previous studies that they either

focus on synthetic environments or are not

transparent and/or computationally expensive, this

work presents an effective framework to close the gap

between ideal models and practical deployment.

Based on extensive experiments on synthetic and real

datasets, the system exhibited high accuracy, safety,

and robustness even in low light and heavy traffic.

The addition of explainability tools such as

SHAP and LIME not only helped in interpretation of

decisions made by AI but also aided in model fine

tuning by pinpointing important decision parameters.

Furthermore, since the proposed framework

processes data in a low-latency manner on the edge

devices, it can be considered a solution for the real-

time, in-vehicle, on-line navigation system in the

self-driving car. The model's compliance with current

safety and regulatory standards also suggests an

appropriateness for real-world task applications.

In conclusion, the presented AI navigation system

overcomes conventional limitations, while providing

a scaleable, transparent and efficient option for

autonomous mobility. This work is laying the

groundwork for intelligent systems that not only

decide but explain -- a necessary step in gaining user

trust, meeting the requirements of regulators and,

potentially, making autonomous transportation safer.

Future extensions of our research could delve into

collaborative learning within fleets of AVs,

incorporation with smart infrastructure, as well as

constant retraining with edge-based mechanisms.

REFERENCES

Asgharpoor Golroudbari, A., & Sabour, M. H. (2023).

Recent advancements in deep learning applications and

methods for autonomous navigation: A comprehensive

review. arXiv preprint arXiv:2302.11089.

Atakishiyev, S., Salameh, M., & Goebel, R. (2021).

Explainable artificial intelligence for autonomous

driving: A comprehensive overview and field guide for

future research directions. arXiv preprint

arXiv:2112.11561.

Atakishiyev, S., Salameh, M., & Goebel, R. (2024). Safety

implications of explainable artificial intelligence in

end-to-end autonomous driving. arXiv preprint

arXiv:2403.12176.

Axios. (2022). Smart cars' biggest safety risk? They're

boring to drive. Retrieved from https://www.axios.co

m/2022/08/24/autonomous-self-driving-cars-fatigue

Feng, S., Yan, X., Sun, H., Feng, Y., & Liu, H. X. (2021).

Intelligent driving intelligence test for autonomous

vehicles with naturalistic and adversarial environment.

Nature Communications, 12, 748.

Holistic AI. (2025). AI regulations for autonomous vehicles

[Updated 2025]. Retrieved from https://www.holistica

i.com/blog/ai-regulations-for-autonomous-vehicles

IEEE Robotics and Automation Society. (n.d.). Special

issue on autonomous vehicle: Artificial intelligence and

model-based techniques in decision and control.

Retrieved from https://www.ieee- ras.org/publications/

ram/special-issues/special-issue-on-autonomous-

vehicle-artificial-intelligence-and-model-based-

techniques-in-de

Liu, H. X., Wu, X., Ma, W., & Hu, H. (2009). Real-time

queue length estimation for congested signalized

intersections. Transportation Research Part C:

Emerging Technologies, 17(4), 412–427.

Liu, H. X., & Feng, S. (2023). Dense reinforcement

learning for safety validation of autonomous vehicles.

Nature, 615, 620–627.

MacCarthy, M. (2024). Can self-driving cars save us from

ourselves? Financial Times. Retrieved from

https://www.ft.com/content/f2eaa452- 4772- 4929- 93

17-69bc7b53ff2c

MacCarthy, M. (2024). The evolving safety and policy

challenges of self-driving cars. Brookings Institution.

Retrieved from https://www.brookings.edu/articles/the

-evolving-safety-and-policy-challenges-of-self-

driving-cars/

May Mobility. (2025). Multi-policy decision-making for

autonomous driving via changepoint-based behavior

prediction: Theory and experiment. Retrieved from

https://en.wikipedia.org/wiki/May_Mobility

Mohammed, R. (2022). Artificial intelligence-driven

robotics for autonomous vehicle navigation and safety.

International Journal of Reciprocal Symmetry and

Theoretical Physics, 3(1), 21–47.

Nassi, B., Nassi, D., & Ben-Gurion University. (2024).

Emergency vehicle lights can screw up a car's

automated driving system. Wired. Retrieved from

https://www.wired.com/story/emergency- vehicle- ligh

ts-can-screw-up-a-cars-automated-driving-system

Nassi, B., Nassi, D., & Ben-Gurion University. (2024).

Researchers probe safety of AI in driverless cars, find

vulnerabilities. TechXplore. Retrieved from https://tec

hxplore.com/news/2024-09-probe-safety-ai-driverless-

cars.html

National Highway Traffic Safety Administration. (n.d.).

Automated vehicle safety. Retrieved from

https://www.nhtsa.gov/vehicle- safety/automated- vehi

cles-safety

NVIDIA. (2025). NVIDIA autonomous vehicles safety

report. Retrieved from https://images.nvidia.com/aem-

dam/en-zz/Solutions/auto-self-driving-safety-

report.pdf

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

650

Rahmati, M. (2025). Edge AI-powered real-time decision-

making for autonomous vehicles in adverse weather

conditions. arXiv preprint arXiv:2503.09638.

Wang, X., Jerome, Z., Wang, Z., Zhang, C., Shen, S.,

Kumar, V., ... & Liu, H. X. (2024). Traffic light

optimization with low penetration rate vehicle

trajectory data. Nature Communications, 15, 1306.

Waymo. (2024). Waymo explores using Google's Gemini

to train its robotaxis. The Verge. Retrieved from

https://www.theverge.com/2024/10/30/24283516/way

mo-google-gemini-llm-ai-robotaxi

Waymo. (2024). Waymo thinks it can overcome robotaxi

skepticism with lots of safety data. The Verge.

Retrieved from https://www.theverge.com/2024/9/5/2

4235078/waymo-safety-hub-miles-crashes-robotaxi-

transparency

Wired. (2021). When driving is (partially) automated,

people drive more. Retrieved from

https://www.wired.com/story/driving- partially- autom

ated-people-drive-more

Yan, X., Zou, Z., Zhu, H., Sun, H., & Liu, H. X. (2023).

Learning naturalistic driving environment with

statistical realism. Nature Communications, 14, 2037.

Zheng, J., & Liu, H. X. (2017). Estimating traffic volumes

for signalized intersections using connected vehicle

data. Transportation Research Part C: Emerging

Technologies, 79, 347–362.

Advancing Artiﬁcial Intelligence for Autonomous Vehicle Navigation: Enhancing Safety, Efﬁciency and Real-World Performance in

Self-Driving Cars

651