Evaluating Rule-Based and AI-based approaches for

real-time decision-making in Autonomous Vehicles

1st Ngo Minh Tien Dat, 2nd Dang Bach Tung, 3rd Pham Tuong Han,
4th Vo Nguyen Kieus My
FPT University, Ho Chi Minh Campus, Vietnam
{TienDatNM, TungDB, HanPT, MyVNK}@[Link]

Abstract

In autonomous vehicle (AV) operations, real-time decision-making is a key factor to ensure safety during travel.
There are two popular approaches to decision-making systems: rule-based, which is interpretable and accurate but
limited in generalization; and AI-based, which generalizes better and provides flexibility in complex environments
but requires high computational resources. Current research largely focuses on advancing one approach in isolation,
with limited direct comparisons between the two, and hybrid solutions remain underexplored without comprehensive
benchmarks. To address this gap, this study employs a Systematic Literature Review (SLR) to synthesize and analyze
recent academic work on AV decision-making. The selected studies come from reputable academic sources and are
classified into two main groups: (1) real-time decision-making mechanisms, and (2) practical problems where both
approaches are used. This research aims to clarify the fundamental differences between rule-based and AI-based
methods, highlight their respective strengths and limitations, and provide insights for selecting appropriate strategies
in real-world AV deployment.
Keywords—rule-based, AI-based, autonomous vehicles, decision-making, real-time

I. Introduction
Real-time decision-making is a critical component of autonomous vehicle (AV) operations to ensure
safety and efficiency. Traditional rule-based approaches have been widely used in the behavioral planning
layer, offering transparency and alignment with traffic laws. However, these systems lack flexibility
in adapting to dynamic traffic environments [1]–[3]. To address this limitation, AI-based approaches
such as Deep Reinforcement Learning (DRL), Large Language Models (LLMs), and Edge AI have
been introduced, enabling self-learning, adaptability, and improved performance under adverse weather
conditions [4]–[6].
In Vietnam, research has focused on operational challenges in local traffic contexts characterized by
high mixed-vehicle density, dominance of motorbikes, complex signposts, and uneven lanes. Hybrid
architectures that combine deep learning models with traditional control algorithms are emerging as
a promising direction. However, systematic publications and real-world experiments on such hybrid
approaches remain limited [2], [5], [7].
In practice, rule-based algorithms have been applied in Advanced Driver Assistance Systems (ADAS),
such as Automated Emergency Braking (AEB) at Level 2 and Adaptive Cruise Control (ACC) at Level 1
[1]. At higher autonomy levels, decision-making systems increasingly rely on AI-based models. A notable
example is Tesla’s Full Self-Driving (FSD v12), which eliminates hard-coded rules and instead employs
an end-to-end neural network [8].
These observations highlight a critical research gap: although rule-based and AI-based approaches have
been widely investigated, direct comparisons between them in real-time autonomous vehicle decision-
making remain limited. In particular, existing studies often evaluate different methods, scenarios, and
performance criteria, making it difficult to determine the conditions in which each approach is most
appropriate. Therefore, this study aims to provide a structured comparison of rule-based and AI-based
approaches based on their responsiveness, safety, adaptability, transparency, and computational require-
ments.
Accordingly, this study addresses the following research question:
RQ: How do rule-based and AI-based approaches compare in terms of responsiveness, safety, adapt-
ability, transparency, and computational requirements in real-time autonomous vehicle decision-making?

II. Related Works

Fig. 1. Decision-making architecture of an autonomous vehicle with rule-based and AI-based planning.

A. Rule-based Decision-making
Rule-based approaches rely on finite sets of traffic situations and corresponding actions, often modeled
as state machines. They are transparent, easy to implement, and closely follow traffic laws [1]. In Vietnam,
Nguyen et al. introduced Vi-DiSC, a framework that uses rule-based logic to extract and interpret
information from Vietnamese signposts [3]. Similarly, Tiến and Hương applied deep learning for lane
detection in self-driving systems [5], while Quang and Hiền evaluated scheduling algorithms such as
A*, Dijkstra, and EBAND for autonomous navigation [7]. Pham et al. also proposed multi-sensor fusion
combined with rule-based decision-making for navigation in mixed urban traffic [2]. The main limitation
of rule-based systems is their inability to handle situations outside predefined rules, reducing adaptability
in dynamic environments.

B. AI-based Decision-Making
AI-based approaches leverage machine learning models to improve adaptability. Deep Reinforcement
Learning (DRL) has been applied to freeway decision-making, showing the ability to learn safe and
efficient strategies but requiring large computational resources [1]. To enhance learning efficiency, Pang
et al. proposed combining DRL with Large Language Models (LLMs), reducing reliance on human expert
guidance but facing high inference costs [4]. Edge AI has also been explored to process data locally,
reducing latency and improving performance in adverse weather conditions [6].

C. Comparison and Research Gap

Surveys such as Paden et al. [8] show the evolution from modular rule-based systems to end-to-end
AI-based architectures. In summary:
• Rule-based: transparent, easy to verify, suitable for traffic laws, but lacks adaptability.
• AI-based: strong in learning and handling complex environments, but costly and less transparent.
• Hybrid approaches: combining both methods has been suggested, especially in Vietnam, but sys-
tematic studies and benchmarks for real-time decision-making are still missing.
To provide a clearer overview of the approaches, Table I summarizes the eight selected studies,
categorized by method, focus, result, and limitation.

TABLE I
Summary of selected studies on rule-based and AI-based decision-making in autonomous vehicles

Reference Approach Focus Result Limitation

Liu, 2024 [1] AI-based DRL for freeway Achieved safe and efficient High computation
driving driving strategies in freeway cost
scenarios
Pang et al., AI-based DRL + LLM for Reduced reliance on human Slow inference,
2024 [4] decision-making expert guidance during costly
training
Tiến & AI-based Deep learning Detected lanes effectively in Accuracy depends on
Hương, 2025 for lane detection mixed traffic conditions training data
[5]
Quang & Rule- Scheduling Provided fast and Limited in dynamic
Hiền, 2023 based algorithms (A*, transparent path planning traffic
[7] Dijkstra,
EBAND)
Pham et al., Hybrid Multi-sensor Demonstrated promising Real-world testing
2025 [2] fusion in adaptability in mixed traffic limited
Vietnam traffic
Nguyen et Rule- Vi-DiSC traffic Achieved high accuracy in Struggles with
al., 2023 [3] based sign recognition Vietnamese traffic sign unexpected cases
recognition
Rahmati, AI-based Edge AI under Lowered latency and Still slower than
2025 [6] adverse weather improved safety under rule-based
adverse weather
Paden et al., Rule- Survey of Confirmed fast, Poor adaptability in
2016 [8] based motion planning interpretable, and verifiable complex scenarios
(FSM/HFSM) decision-making
 III. Methodology
This study adopts a scoped, PRISMA-informed literature review to compare rule-based and AI-based
approaches in real-time autonomous vehicle decision-making. The review process began with the defini-
tion of the research question, followed by keyword-based searches in selected academic databases. The
identified studies were screened based on their titles, abstracts, and full-text relevance to the research
topic. Duplicate records and studies that did not directly address autonomous vehicle decision-making
were excluded.
Studies were included if they examined rule-based, AI-based, or hybrid decision-making approaches
and provided an applied method, evaluation, or relevant findings. After the screening and eligibility
assessment, eight studies were selected for the final synthesis. These studies were categorized into rule-
based, AI-based, and hybrid approaches rather than by domestic or overseas origin.
Data were extracted from each study, including the proposed method, application scenario, main
findings, and reported limitations. The selected studies were then compared according to five criteria:
responsiveness, safety, adaptability, transparency, and computational requirements. Because the studies
used different datasets, scenarios, and evaluation metrics, the findings were synthesized mainly through
qualitative comparison rather than direct statistical analysis. The overall study-selection process is illus-
trated in Fig. 2.

IV. Comparative Analysis

A. Comparison of Decision-Making Mechanisms
Rule-based decision-making converts sensor data into predefined traffic states and then matches them
with a set of rules to generate actions such as lane keeping, lane changing, or stopping. Models like
FSM and HFSM allow the system to respond quickly, remain transparent, and be easy to verify [8]. For
example, algorithms such as A*, Dijkstra, and EBAND have been applied to find the shortest route [7],
while the Vi-DiSC study shows that rule-based captioning can recognize Vietnamese traffic signs with
high accuracy [3]. However, rule-based methods struggle to handle complex or unexpected situations.
AI-based decision-making replaces fixed rules with machine learning models, representing the environ-
ment state and inferring control policies. Methods such as DRL and LLM-assisted policy enable vehicles
to learn driving strategies from data and adapt better in dynamic environments [1], [4]. Nevertheless, they
require significant computational resources, have higher latency, and are less transparent. For instance,
DRL freeway achieves strong performance but at high cost [1], while DRL+LLM reduces the need for
human experts but still suffers from slow inference [4].
In the Vietnamese context, studies on lane detection using deep learning [5] and multi-sensor fusion
[2] demonstrate the potential of AI-based approaches in mixed traffic, though real-world testing remains
limited. Conversely, rule-based methods continue to be effective in tasks with clear rules, such as traffic
sign recognition [3].
 Fig. 2. Research methodology based on keyword search and relevance to the research question (RQ).

B. Comparison of Operational Performance

Latency: Rule-based systems show lower latency because they only match the current state with
predefined rules. Their structure is simple and requires fewer processing steps. In contrast, AI-based
systems need to perform several operations such as feature extraction and policy inference, which makes
them slower. Edge AI can reduce latency, but it still cannot reach the speed of rule-based approaches [1],
[4], [6].
Safety: Rule-based systems are safe in environments with clear traffic rules (for example, lane keeping
or traffic sign recognition). However, they may fail when facing unexpected situations or cases not included
in the rule set [3]. AI-based systems can learn to avoid collisions in dynamic environments and adapt
better under adverse weather conditions, but they are less transparent and harder to certify legally [1],
[6].
Accuracy: Rule-based systems achieve high accuracy in scenarios with well-defined rules, such as
Vi-DiSC’s recognition of Vietnamese traffic signs [3]. Yet they are limited when applied to more complex
situations. AI-based systems have stronger generalization ability and can handle diverse scenarios, but
their accuracy depends on training data and may drop in new or noisy environments [2], [4], [5].

C. Comparison in Practical Driving Scenarios

TABLE II
Comparison in practical driving scenarios

Scenario Rule-based AI-based

Freeway FSM/HFSM allow fast DRL learns safe,
driving reactions, keeping or efficient strategies in
changing lanes according high-speed environments,
to predefined rules. adapts better but needs
Effective in moderate significant computation
traffic density [8]. [1].
Intersections Difficult to handle DRL and LLM-assisted
/ Round- complex situations, often policy manage complex
abouts fails when many vehicles scenarios, but latency is
compete for space or high and transparency is
behave unpredictably. limited [4].
Mixed Effective in traffic sign Lane detection [5] and
traffic in recognition and route multi-sensor fusion [2]
Vietnam planning [3], [7], but show potential, though
struggles with real-world testing is
motorbike-dominated limited.
traffic.
Adverse Easily affected by noisy Edge AI processes data
weather sensor data, reducing locally, lowering latency
accuracy. and collision rates [6].

V. Results and Discussion

A. Prototype Implementation
The studies analyzed include two approaches:
Rule-based: FSM/HFSM [8], path planning algorithms such as A*, Dijkstra, EBAND [7], and the
Vi-DiSC traffic sign recognition system [3].
AI-based: DRL freeway [1], DRL+LLM [4], lane detection using deep learning [5], multi-sensor fusion
in mixed traffic [2], and Edge AI under adverse weather conditions [6].

B. Results
Overall, the synthesis of the eight selected studies reveals a clear trend: rule-based approaches con-
sistently outperform AI-based methods in latency and transparency, while AI-based approaches excel in
adaptability and generalization. Quantitatively, rule-based algorithms, such as Dijkstra’s and A*, demon-
strate consistent low-latency performance, maintaining response times below 50 ms [7]. By contrast, AI-
based approaches—particularly DRL-LLM hybrid models like LGDRL—significantly enhance efficiency,
achieving decision-making within roughly 0.01 s [8]. Furthermore, deployment architecture critically
impacts speed: Cloud AI exhibits a latency of 240–310 ms, whereas Edge AI operates at just 45–55 ms,
representing a nearly fivefold improvement [6].
In simulation environments, the A*, Dijkstra, and EBAND algorithms all exhibit good collision avoid-
ance, with EBAND additionally calculating the vehicle’s occupied space [7]. For scenarios with clear
regulations, the rule-based approach ensures good safety but is prone to failure when encountering
unexpected situations outside its rule set [3]. Conversely, the AI-based approach handles complex traffic
and adverse weather conditions well. Specifically, the model combining DRL with LLM achieves a
collision rate of only 10%, significantly lower than that of conventional DRL algorithms (over 26%) [4].
Notably, shifting from Cloud AI to Edge AI reduces the collision rate from 12.3% to 5.7% under extreme
weather conditions [6].
The two approaches exhibit a sharp contrast in their accuracy strengths, highly depending on the specific
nature of the task [3], [4]. While AI-based methods dominate in dynamic decision-making—demonstrated
by the LGDRL model achieving a success rate of up to 90% in complex lane-changing scenarios [4]—
rule-based approaches excel in processing rule-rich tasks with fixed logic [3]. Specifically, for signpost
recognition, the rule-based model achieves an accuracy of 89.25%, which is 2.5 times higher than that of
deep-learning architectures such as CNN-Transformer (34.10%) and Transformer-Transformer (38.35%)
[3].

C. Discussion
The comparison of rule-based and AI-based approaches shows that each has clear advantages and
limitations. Rule-based methods, such as FSM/HFSM [8], classical path planning algorithms [7], and Vi-
DiSC traffic sign recognition [3], provide fast responses, transparency, and high accuracy in environments
with clear rules. Yet they struggle in complex or unpredictable traffic situations. AI-based approaches,
including DRL freeway [1], DRL+LLM [4], lane detection [5], multi-sensor fusion [2], and Edge AI [6],
demonstrate stronger adaptability and generalization, allowing vehicles to learn strategies directly from
data and perform better in dynamic conditions. However, they require greater computational resources,
introduce higher latency, and remain less transparent, which complicates safety certification. In Vietnam’s
mixed traffic, hybrid solutions such as multi-sensor fusion [2] appear promising, combining rule-based
clarity with AI-based learning.

VI. Conclusion
This study, based on eight papers, highlights the differences between rule-based and AI-based decision-
making in autonomous vehicles. Rule-based systems excel in low latency, transparency, and accuracy in
structured environments [3], [7], [8], while AI-based systems are superior in adaptability and generaliza-
tion, as shown in DRL freeway [1], DRL+LLM [4], lane detection [5], and Edge AI [6]. Yet AI-based
methods face challenges of computation cost, latency, and certification. In Vietnam’s traffic context,
multi-sensor fusion [2] offers a promising path forward, combining the strengths of both approaches.
Overall, the findings confirm that rule-based and AI-based systems differ significantly in mechanism and
performance, and hybrid models may provide the most effective solution for real-time autonomous driving.

This study is limited by the small number of selected papers and the absence of real-world validation.
In the future, our group can expand the scope of the literature review with more sources and design
evaluation criteria tailored to Vietnam’s traffic conditions to strengthen the practical relevance of the
analysis.

References
[1] T. Liu, Y. Yang, W. Xiao, X. Tang, and M. Yin, “A comparative analysis of deep reinforcement learning-enabled freeway decision-making
for automated vehicles,” IEEE Access, vol. 12, pp. 24 090–24 103, 2024.
[2] L. T. Pham, H. M. Le, and T. N. Do, “Integration of multi-sensor fusion and decision-making architecture for autonomous vehicle
navigation apply in vietnam traffic conditions,” Preprints, 2025, art. no. 202408.0123.
[3] T. M. Nguyen, H. V. Tran, and Q. T. Nguyen, “Vi-disc: A novel dataset and framework for extracting information from vietnamese
signposts images,” in Proc. Int. Conf. Computer Vision and Image Processing, 2023, pp. 112–125.
[4] H. Pang, Z. Wang, and G. Li, “Large language model guided deep reinforcement learning for decision making in autonomous driving,”
arXiv preprint arXiv:2412.18511, 2024, available: [Link]
[5] L. Q. Tiến and T. T. Hương, “Research on applying deep learning networks in lane detection for self-
driving vehicle systems,” Journal of Maritime Science and Technology, no. 82, pp. 196–203, apr 2025, available:
[Link]
[6] M. Rahmati, “Edge ai-powered real-time decision-making for autonomous vehicles in adverse weather conditions,” arXiv preprint
arXiv:2503.09638, 2025.
[7] N. V. A. Quang and Đ. Q. Hiền, “Implement performance evaluation for scheduling algorithms used in autonomous driving vehicles,”
University of Danang, Tech. Rep. ĐHVH-2022-13, may 2023.
[8] B. Paden, M. Čáp, S. Z. Yong, D. Yershov, and E. Frazzoli, “A survey of motion planning and control techniques for self-driving urban
vehicles,” IEEE Transactions on Intelligent Vehicles, vol. 1, no. 1, pp. 33–55, 2016.