AI-BASED DECISION-MAKING MODELS FOR AUTONOMOUS DRIVING IN URBAN ENVIRONMENTS

Volume 15 Issue 08, August 2025

Impact factor: 2019: 4.679 2020: 5.015 2021: 5.436, 2022: 5.242, 2023:

6.995, 2024 7.75

http://www.internationaljournal.co.in/index.php/jasass

402

AI-BASED DECISION-MAKING MODELS FOR AUTONOMOUS DRIVING IN

URBAN ENVIRONMENTS

Khujamov Fazliddin

Former Student of Vocational school No/1 of Dehkanabad district

Abstract:

Autonomous driving in urban environments presents a uniquely complex challenge

due to dynamic traffic patterns, dense infrastructure, and the unpredictability of human agents

such as pedestrians and cyclists. This review explores the growing role of artificial intelligence

(AI) in addressing the decision-making demands of urban autonomous vehicles (AVs). We

categorize key AI-based approaches—deep learning, reinforcement learning, imitation learning,

multi-agent models, and hybrid systems—and analyse their applications, strengths, and

limitations in real-world scenarios. The paper also examines foundational tools, including

simulation platforms (CARLA, SUMO, AirSim), benchmark datasets (KITTI, nuScenes,

Waymo), and insights from industry leaders like Waymo, Baidu Apollo, and Tesla. Core

challenges such as safety validation, data scarcity for rare events, interpretability, and ethical

considerations are critically discussed. Finally, we outline future directions involving 5G/6G

integration, digital twins, and human-centered AI to support scalable, reliable, and transparent

decision-making in urban autonomy. This review serves as a comprehensive foundation for

researchers and practitioners aiming to advance the next generation of intelligent urban mobility

systems.

Keywords:

Autonomous vehicles · Urban driving · AI-based decision-making · Deep

learning · Reinforcement learning · Imitation learning

1. Introduction

The rapid advancement in autonomous vehicle (AV) technologies has positioned intelligent

mobility at the core of future transportation systems. Urban environments—dense, dynamic, and

multifaceted—demand sophisticated systems that go beyond rule-based driving. In such contexts,

decision-making plays a pivotal role in ensuring safety, efficiency, and adaptability. As

urbanization accelerates, the integration of artificial intelligence (AI) in AVs becomes essential

for handling the high variability of traffic, infrastructure, and human behavior in cities [1].

Urban driving is uniquely challenging due to the inherent unpredictability of pedestrian behavior,

unstructured road scenarios, occlusions, frequent intersections, and inconsistent signage. Unlike

controlled highway environments, cities involve dense traffic, vulnerable road users, and

complex legal constraints. Traditional methods, such as finite state machines or rule-based logic,

fail to generalize well in these scenarios [2-3].

To navigate such complexity, AI-based decision-making models—including reinforcement

learning, neuro-symbolic reasoning, and deep imitation learning—have emerged as powerful

tools for enabling context-aware, real-time decisions in AVs. These models can learn optimal

driving strategies through simulations, adapt to unfamiliar environments, and incorporate ethical

and safety constraints [4-5]. Furthermore, the integration of IoT and edge computing allows

decision-making modules to be decentralized, responsive, and adaptive to hyper-local changes

[6].

This review aims to provide a comprehensive overview of the state-of-the-art AI-based decision-

making approaches for autonomous driving in urban contexts. It highlights foundational models,

Volume 15 Issue 08, August 2025

Impact factor: 2019: 4.679 2020: 5.015 2021: 5.436, 2022: 5.242, 2023:

6.995, 2024 7.75

http://www.internationaljournal.co.in/index.php/jasass

403

emerging hybrid methods, benchmarking environments, and open research challenges, with the

goal of guiding future research and deployment strategies for intelligent urban mobility systems.

2. Decision-Making in Urban Autonomous Driving

Autonomous driving systems rely on multi-layered decision-making frameworks, typically

categorized into three levels: strategic, tactical, and operational. The strategic level focuses on

high-level planning such as destination selection and route optimization across a city. The

tactical level handles short- to mid-range planning tasks like overtaking, lane changes, and

intersection behavior. Lastly, the operational level involves real-time control, such as

maintaining lane position, obeying traffic signals, and reacting to nearby vehicles or pedestrians

[7-8].

Urban driving presents unique challenges for decision-making systems. Common tasks include

lane changing in dense traffic, handling intersections with ambiguous right-of-way rules, and

ensuring pedestrian and cyclist safety in dynamic environments. These scenarios often involve

partial observability, delayed feedback, and multiple agents interacting concurrently [9]. As such,

AVs must operate with both predictive foresight and real-time adaptability.

Traditional rule-based systems, while interpretable and easier to verify, struggle in urban

contexts due to their rigidity and inability to generalize across novel situations. They are heavily

reliant on pre-programmed logic and often fail in scenarios involving uncertainty or incomplete

information. In contrast, AI-based approaches, including deep reinforcement learning, imitation

learning, and neuro-symbolic reasoning, offer more robust adaptability by learning from

experience and simulating human-like reasoning [4, 3].

AI systems can infer intentions of surrounding agents, prioritize ethical decisions in ambiguous

cases, and improve continuously via data-driven feedback. These capabilities make them

particularly suitable for complex urban environments, where flexibility and contextual awareness

are critical for safety and efficiency.

3. AI-Based Decision-Making Approaches

3.1 Deep Learning

Deep learning (DL) methods, particularly convolutional and recurrent neural networks, have

been applied to end-to-end autonomous driving, where raw sensory inputs (like camera images)

are mapped directly to driving actions. These models eliminate the need for modular pipelines

and perform well in structured environments. A notable example is Nvidia’s PilotNet, which

demonstrated real-time steering prediction using only camera data. More recently, CARLA-

Apollo integrations have allowed researchers to simulate DL-based driving under urban

conditions [10].

However, the black-box nature of deep learning raises concerns about interpretability and safety

in critical decisions, particularly in ambiguous urban scenarios like jaywalking pedestrians or

unclear lane markings. Deep networks also require large-scale annotated datasets for robust

performance, making real-world deployment costly and data-intensive. Furthermore, DL systems

often struggle with generalization to unseen conditions such as rare weather events or unusual

road geometries.

3.2 Reinforcement Learning

Reinforcement learning (RL) offers a powerful framework for decision-making by enabling

autonomous agents to learn optimal policies through interaction with the environment. In urban

driving, RL has shown promise in managing intersections, adaptive lane changes, and multi-

Volume 15 Issue 08, August 2025

Impact factor: 2019: 4.679 2020: 5.015 2021: 5.436, 2022: 5.242, 2023:

6.995, 2024 7.75

http://www.internationaljournal.co.in/index.php/jasass

404

scenario path planning. Techniques such as Deep Q-Networks (DQN) and Proximal Policy

Optimization (PPO) are used in simulators like CARLA to train vehicles for safe, reactive

behavior [4].

One strength of RL is its ability to handle delayed rewards and long-term planning, crucial for

anticipating complex events in urban settings. However, the need for extensive exploration and

simulation episodes limits its scalability. In the real world, deploying RL-trained policies safely

is still a major challenge due to the sim-to-real gap and the potential for unsafe exploration.

3.3 Imitation Learning

Imitation learning (IL) focuses on mimicking expert human behavior using supervised learning.

Behavioral cloning, a popular IL approach, uses recorded driving data to train models that

reproduce driver actions in similar situations. This method is intuitive and data-efficient,

allowing quick deployment in familiar traffic environments [2].

While IL enables rapid prototyping and produces human-like behavior, it suffers from

compounding errors—small deviations from expected inputs can snowball over time, causing

unsafe behavior. Moreover, models trained via IL can lack robustness to novel or out-of-

distribution scenarios unless augmented with safety mechanisms or hybrid learning techniques.

The quality and diversity of the human dataset directly influence system reliability.

3.4 Multi-Agent Models

Multi-agent reinforcement learning (MARL) has emerged as a strategy for cooperative

autonomous driving, where vehicles interact with one another (V2V) and with infrastructure

(V2I). MARL models enable coordinated lane merges, intersection negotiation, and platooning

by treating each vehicle as an agent with shared or individual rewards [2].

For example, vehicles communicating via 5G can cooperatively adjust speeds to minimize fuel

consumption or optimize intersection throughput. Despite its potential, MARL faces challenges

including scalability, reward shaping, and non-stationarity—since each agent's behavior evolves

over time, making the learning environment unstable. Ensuring safety in adversarial or noisy

communication scenarios also remains a hurdle.

3.5 Hybrid Approaches

To bridge the gap between data-driven learning and human-designed safety, hybrid systems

combine AI models with explicit rule-based constraints. This fusion allows deep models to learn

flexible behavior while being constrained by safety or legal rules—ensuring compliance even in

uncertain environments. For example, an RL agent may learn optimal gap selection for a lane

change but be bounded by speed limits or safety margins defined by rules [10].

Hybrid approaches are also used in hierarchical decision-making, where a high-level planner

based on symbolic logic governs a low-level controller driven by deep learning. This layered

structure supports modularity

,

verifiability, and adaptive decision-making, making it increasingly

popular in real-world AV deployments. Table 1 compares the strengths and limitations of AI-

Based Decision-Making Approaches.

Table 1.

Comparative Analysis of AI-Based Decision-Making Approaches for Urban

Autonomous Driving

Volume 15 Issue 08, August 2025

Impact factor: 2019: 4.679 2020: 5.015 2021: 5.436, 2022: 5.242, 2023:

6.995, 2024 7.75

http://www.internationaljournal.co.in/index.php/jasass

405

Approach

Key Feature

Strengths

Limitations

Urban

Application

Example

Key

Referenc

e

Deep

Learning

End-to-end

mapping from

sensor input to

actions

High

scalability;

learns

complex

features

automaticall

y

Black-box

nature; poor

interpretabilit

y

End-to-end

navigation in

city scenarios

(CARLA,

Apollo)

[10]

Reinforcemen

t Learning

Trial-and-error

learning

for

policy

optimization

Capable of

long-term

reward

optimization

Requires

massive

simulation;

safety risks in

real world

Intersection

handling

with traffic

signal

learning

[4]

Imitation

Learning

Learning from

expert

demonstrations

Data-

efficient;

human-like

behavior

replication

Compounding

errors;

poor

generalization

to unseen

Behavior

cloning for

lane

following or

yielding

[2]

Multi-Agent

Models

Coordination

among

multiple agents

(V2V, V2I)

Enables

cooperative

driving and

negotiation

High

complexity;

non-stationary

environment

Platooning

and

intersection

negotiation

with V2V

[7]

Hybrid

Approaches

Combining AI

with

rule-

based

constraints

Balances

learning

flexibility

with

rule

safety

Integration

complexity;

potential

performance

bottlenecks

Rule-guided

lane merging

using learned

policies

[10]

4. Applications & Tools

4. Applications & Tools

The development of AI-based decision-making systems for autonomous urban driving is

supported by a growing ecosystem of simulation platforms, curated datasets, and industrial

innovation projects. These tools enable robust training, validation, and deployment of models

under safe, controlled, and scalable conditions.

4.1 Simulation Platforms

Simulation platforms are indispensable for prototyping and testing autonomous driving policies,

particularly where safety, cost, and repeatability are critical.



CARLA (Car Learning to Act) is a widely used open-source urban driving simulator that

supports complex urban layouts, diverse traffic agents, and dynamic weather conditions.

CARLA is integrated with Python APIs and ROS, making it suitable for AI model training and

Volume 15 Issue 08, August 2025

Impact factor: 2019: 4.679 2020: 5.015 2021: 5.436, 2022: 5.242, 2023:

6.995, 2024 7.75

http://www.internationaljournal.co.in/index.php/jasass

406

benchmarking [11-12].



SUMO (Simulation of Urban MObility) enables traffic-level simulation, focusing on

vehicle flow, signal control, and routing. SUMO is often used alongside AI-driven control

models to simulate city-scale optimization and V2X coordination [13].



AirSim, developed by Microsoft, supports both aerial and ground vehicles and offers

realistic physics and rendering via Unreal Engine. It facilitates perception and control algorithm

training through photorealistic environments [14-15].

These platforms provide synthetic but high-fidelity environments for developing deep learning,

reinforcement learning, and hybrid models in urban contexts.

4.2 Urban Driving Datasets

Large-scale annotated datasets are essential for training and validating perception and decision-

making models. Several benchmarks have emerged to serve diverse urban use cases:



KITTI: One of the earliest real-world datasets, KITTI provides stereo imagery, LiDAR

scans, and GPS data. While limited in scope, it remains a foundational dataset for urban object

detection and tracking [16].



nuScenes: This dataset contains 1,000 driving scenes from Boston and Singapore,

including 3D object annotations, LiDAR, radar, and weather conditions. It supports tasks like

trajectory prediction and behavior modeling [13].



Waymo Open Dataset: Offers over 1,000 driving sequences from urban areas across the

U.S., with high-resolution sensor data and extensive labels. It is commonly used for large-scale

AI model benchmarking [16, 12].

These datasets are instrumental in both imitation learning and supervised perception tasks, aiding

in the development of safe and adaptive AI driving agents.

4.3 Industry Applications

Several major players are driving real-world deployment of AI-based decision-making through

large-scale urban autonomy projects:



Waymo has deployed fully autonomous vehicles in selected U.S. cities and uses

advanced neural networks and reinforcement learning for complex driving tasks, including

merging and pedestrian negotiation [12].



Baidu Apollo has released an open-source AV platform with modules for planning,

control, and simulation. Apollo's Robotaxi services use deep learning and rule-based hybrids for

decision-making in Beijing and other cities [12].



Tesla’s Autopilot and Full Self-Driving (FSD) systems rely heavily on vision-based

neural networks. Tesla emphasizes end-to-end learning and leverages its fleet for real-world data

collection at scale [17].

These industry projects reflect different philosophies—Waymo and Apollo lean toward high-

definition mapping and LiDAR, while Tesla focuses on data-driven perception and minimal map

dependence.

5. Challenges & Future Directions

Despite rapid advancements in AI-based decision-making for autonomous vehicles (AVs),

multiple challenges continue to hinder their safe and scalable deployment in complex urban

environments. From reliability concerns to regulatory ambiguities, addressing these issues is

critical to achieving widespread societal trust and adoption.

5.1 Safety Validation and Reliability

Volume 15 Issue 08, August 2025

Impact factor: 2019: 4.679 2020: 5.015 2021: 5.436, 2022: 5.242, 2023:

6.995, 2024 7.75

http://www.internationaljournal.co.in/index.php/jasass

407

Safety assurance remains one of the most difficult challenges in deploying AI-driven decision-

making systems in urban environments. Unlike traditional systems, AI-based models—especially

deep neural networks and reinforcement learning agents—lack deterministic behavior, making it

difficult to validate safety across all operational design domains (ODDs). Scenarios involving

unpredictable human behavior (e.g., pedestrians darting across streets) are hard to simulate or

exhaustively test [18].

Simulation tools such as CARLA and AirSim offer synthetic environments, but these do not

always capture the full complexity of the real world. Moreover, there's a lack of standardized

formal verification methods for non-rule-based AI agents. Techniques like adversarial testing

and scenario-based validation are evolving, but regulatory frameworks are still catching up to

ensure these models are certifiably safe.

5.2 Data Scarcity for Rare and Edge Cases

AI systems learn from data—but urban driving contains countless rare but critical scenarios,

such as emergency vehicles approaching from blind spots, ambiguous traffic signals, or unusual

weather conditions. These edge cases are underrepresented in datasets like KITTI and nuScenes,

which often focus on routine traffic scenarios [16].

This data imbalance hampers generalization and robustness. Emerging approaches, such as

synthetic data augmentation, scenario generation using Generative AI, and collaborative dataset

sharing among stakeholders, are being explored to improve coverage. Still, ensuring that AVs

behave safely and ethically in the "long tail" of driving events remains an unresolved problem.

5.3 Black-Box Interpretability

Many AI-based models—particularly deep learning systems—function as black boxes, offering

little insight into how decisions are made. This raises concerns in high-stakes situations where

explanations are necessary for legal accountability

,

debugging, or user trust.

Recent research has proposed explainable AI (XAI) techniques, including saliency maps,

decision trees over latent embeddings, and post-hoc rule extraction. However, these techniques

often provide only approximations, not guarantees. Interpretability becomes even more

important in multi-agent settings, where emergent behaviors can lead to unpredictable

interactions between AVs and human drivers or pedestrians [18].

5.4 Ethical and Legal Challenges

Urban environments often involve ethically ambiguous scenarios, such as choosing between two

suboptimal outcomes during an unavoidable collision. Unlike rule-based logic, AI-based

decision models must encode value-based judgments, which raise societal and legal questions.

For instance, Tesla's approach to vision-based end-to-end autonomy has led to debates about

liability and transparency in the event of system failures [17]. Furthermore, privacy concerns

arise from data-intensive driving systems that may collect sensitive pedestrian or license plate

information. The legal systems in many countries are still evolving to define the liability of

autonomous systems, especially those driven by non-deterministic AI.

5.5 Future Directions

To address these issues and expand AI’s role in AV decision-making, several promising

directions are emerging:



5G/6G integration: Low-latency communication networks can enable real-time Vehicle-

to-Everything (V2X) coordination, supporting collaborative driving and dynamic route

optimization in congested urban settings [13].

Volume 15 Issue 08, August 2025

Impact factor: 2019: 4.679 2020: 5.015 2021: 5.436, 2022: 5.242, 2023:

6.995, 2024 7.75

http://www.internationaljournal.co.in/index.php/jasass

408



Digital Twins: These virtual replicas of real urban environments allow AVs to be tested

in highly accurate, continuously updated simulations. Digital twins can enable continuous

learning, urban planning, and proactive safety assessments.



Human-Centered AI: Incorporating psychological and behavioral insights into decision-

making can improve trust and cooperation between humans and AVs. This includes predicting

pedestrian intent, emotion-aware planning, and ethical reasoning modules that align with societal

values.



Federated and Collaborative Learning: Rather than centralizing all driving data, AVs can

learn from each other using privacy-preserving distributed learning frameworks—enhancing data

efficiency while maintaining user trust.

As autonomous driving enters the next decade, bridging the gap between technical robustness

and societal alignment will determine the long-term success of AI in urban mobility.

6. Conclusion

Artificial intelligence has become an indispensable pillar in the pursuit of safe and efficient

autonomous urban mobility. By enabling data-driven reasoning and adaptive behavior, AI-based

decision-making models have shown remarkable potential in addressing the complexity of city

driving—ranging from dense traffic and unstructured road layouts to unpredictable pedestrian

behavior. From deep learning to reinforcement learning and hybrid models, AI continues to

transform how autonomous vehicles perceive, interpret, and act in dynamic urban environments.

However, progress must be met with caution and responsibility. Ensuring safety

,

transparency

,

and ethical alignment is just as important as achieving technical sophistication. The challenges

surrounding interpretability, validation under uncertainty, and societal trust demand that

innovation in AV decision-making be paired with robust safeguards and inclusive regulatory

frameworks.

Looking forward, the integration of next-generation infrastructure (such as 5G/6G and digital

twins), along with the rise of human-centered and explainable AI, promises a future of

autonomous systems that are not only intelligent but also accountable and aligned with human

values. Achieving this balance will be essential for deploying scalable and trustworthy AI-based

driving systems in the cities of tomorrow.

References:

1. Soica, A., & Gheorghe, C. (2025). Revolutionizing Urban Mobility: A Systematic Review

of AI, IoT, and Predictive Analytics in Adaptive Traffic Control Systems. Electronics, 14(4),

719.

2. Gondhalekar, G., & Mathiyalagan, P. (2025). Artificial Intelligence Applications in

Autonomous Vehicles: Navigating the Future of Transportation Systems. ITM Web of

Conferences.

3. Pentela, V.K., & Deepalakshmi, P. (2025). Improved Threat Intelligence and Real-Time

Protection with Next-Generation Firewall Solutions. IEEE Global Conference, 2025. IEEE

4. Muniyandy, P., El, T.D.Y.A.B., & Ebiary, D.D.N.P.D. (2025). Neuro-Symbolic

Reinforcement Learning for Context-Aware Decision Making in Safe Autonomous Vehicles.

ResearchGate.

5. Ramesh, J.V.N., Khan, H., & Chaudhari, T.D. (2025). Neuro-Symbolic Reinforcement

Learning for Context-Aware Decision Making in Safe Autonomous Vehicles. International

Journal of Advanced Research in Computer Science, 2025.

Volume 15 Issue 08, August 2025

Impact factor: 2019: 4.679 2020: 5.015 2021: 5.436, 2022: 5.242, 2023:

6.995, 2024 7.75

http://www.internationaljournal.co.in/index.php/jasass

409

6. Haider, I., & Marchant, R. (2025). AI and Edge Computing in Smart Parking: Advancing

Real-Time Space Allocation and Billing. ResearchGate.

7. Ceccarelli, A., Trapp, M., Bondavalli, A., & Schoitsch, E. (2024). Computer Safety,

Reliability, and Security: SAFECOMP Workshops. Springer.

8. Soica, A., & Gheorghe, C. (2025). Revolutionizing Urban Mobility: A Systematic Review

of AI, IoT, and Predictive Analytics in Adaptive Traffic Control Systems. Electronics, 14(4),

719.

9. Gondhalekar, G., & Mathiyalagan, P. (2025). Artificial Intelligence Applications in

Autonomous Vehicles. ITM Web of Conferences, 2025.

10. Rosas Otero, M. (2025). From rules to rewards: AI path planning for autonomous driving

within the CARLA-Apollo framework. Universitat Politècnica de Catalunya.

11. Sarker, S., Maples, B., Islam, I., Fan, M., et al. (2024). A Comprehensive Review on Traffic

Datasets

and

Simulators

for

Autonomous

Vehicles.

arXiv:2412.14207.

https://arxiv.org/abs/2412.14207

12. Liu, H., Cao, Z., Yan, X., Feng, S., & Lu, Q. (2025). Autonomous Vehicles: A Critical

Review

(2004–2024)

and

a

Vision

for

the

Future.

TechRxiv.

https://www.techrxiv.org/doi/full/10.36227/techrxiv.174857767.78237989

13. Wang, Y., Xing, S., Can, C., Li, R., et al. (2025). Generative AI for Autonomous Driving:

Frontiers and Opportunities. arXiv:2505.08854. https://arxiv.org/abs/2505.08854

14. Fadaie, J. (2019). The State of Modeling, Simulation, and Data Utilization within Industry:

An Autonomous Vehicles Perspective. arXiv:1910.06075. https://arxiv.org/abs/1910.06075

15. Niranjan, D.R., & VinayKarthik, B.C. (2021). Deep Learning Based Object Detection

Model for Autonomous Driving Using CARLA. IEEE 2nd Intl. Conf. on Electronics and

Sustainable Communication Systems. https://ieeexplore.ieee.org/document/9591747

16. Dasgupta, A., Gopi, O., & Chowdhury, A. (2023). On the Road to Autonomy: A

Comparative Analysis of Multimodal Datasets. In Springer Conference on Recent Trends.

17. Rosero, L.A.R. (2024). Leveraging Modular Architectures and End-to-End Learning for

Autonomous Driving in Unmapped Environments. USP Thesis Repository.

18. Ibrahum, A.D.M., Hussain, M., & Hong, J.E. (2024). Deep Learning Adversarial Attacks

and Defenses in Autonomous Vehicles: A Systematic Review from a Safety Perspective.

Artificial Intelligence Review. https://doi.org/10.1007/s10462-024-11014-8