1

Autonomous Vehicles: Evolution of Artificial

Intelligence and Learning Algorithms
Divya Garikapati, Senior Member, IEEE, Sneha Sudhir Shetiya, Senior Member, IEEE

Abstract—The advent of autonomous vehicles has heralded a the promise of reshaping traditional development processes,
transformative era in transportation, reshaping the landscape of enhancing efficiency, and accelerating innovation. AI tech-
mobility through cutting-edge technologies. Central to this evolu- nologies are becoming integral in numerous facets of software
tion is the integration of Artificial Intelligence (AI) and learning
arXiv:2402.17690v2 [cs.LG] 28 Feb 2024

algorithms, propelling vehicles into realms of unprecedented development within autonomous vehicles making a paradigm
autonomy. This paper provides a comprehensive exploration of shift towards Software-Defined Vehicles (SDVs) [[1]][[2]].
the evolutionary trajectory of AI within autonomous vehicles, The success of autonomous vehicles hinges on balancing
tracing the journey from foundational principles to the most their potential benefits with addressing the challenges through
recent advancements. collaborative efforts in technology development, regulation,
Commencing with a current landscape overview, the paper
delves into the fundamental role of AI in shaping the autonomous and public communication. Some of the challenges include:
decision-making capabilities of vehicles. It elucidates the steps • Safety and Reliability: Ensuring flawless AI performance
involved in the AI-powered development life cycle in vehicles, in all scenarios is paramount.
addressing ethical considerations and bias in AI-driven software • Regulations and Law: Clear standards for safety, insur-
development for autonomous vehicles. The study presents statis-
tical insights into the usage and types of AI/learning algorithms
ance, and liability are needed.
over the years, showcasing the evolving research landscape within • Public Trust and Acceptance:Addressing concerns about
the automotive industry. Furthermore, the paper highlights the safety, data privacy, and ethical dilemmas is crucial.
pivotal role of parameters in refining algorithms for both trucks • Cybersecurity: Protecting against hacking and unautho-
and cars, facilitating vehicles to adapt, learn, and improve rized access is essential.
performance over time. It concludes by outlining different levels
• Ethical Dilemmas: Defining AI decision-making in am-
of autonomy, elucidating the nuanced usage of AI and learning
algorithms, and automating key tasks at each level. Additionally, biguous situations raises moral questions.
the document discusses the variation in software package sizes • Addressing Edge cases: Being able to handle unforeseen
across different autonomy levels. scenarios is challenging as those scenarios are rare and
Index Terms—Artificial Intelligence (AI), Machine Learning could be hard to imagine in some cases.
(ML), Deep Neural Networks (DNNs), Natural Language Process-
ing (NLP), Autonomous Vehicles (AVs), Safety, Security, Ethics, A. Benefits of AI/Learning Algorithms for Autonomous
Emerging Trends, Trucks vs.Cars, Autonomy Levels, Operational
Design Domain (ODD), Software-Defined Vehicles (SDVs), Con- Vehicles
nected and Automated Vehicles (CAVs), In-Vehicle AI Assistant, AI/Learning Algorithms are currently influencing various
Internet Of Things (IOT), Natural Language Processing (NLP), stages from initial coding to post-deployment maintenance in
Generative AI (GenAI).
autonomous vehicles. Some of the benefits include:
• Safety: AI can significantly reduce accidents by eliminat-
I. I NTRODUCTION ing human error, leading to safer roads.
• Traffic Flow: Platooning and efficient routing can ease
RTIFICIAL Intelligence (AI) and learning algorithms
A such as Machine Learning (ML), Deep Learning using
Deep Neural Networks (DNNs) and Natural Language Pro-
congestion and improve efficiency.
• Accessibility: People with physical impairments or dif-

cessing (NLP) currently play a crucial role in the develop- ferent abilities, the elderly, and the young can gain
ment and operation of autonomous vehicles. The integration independent mobility.
• Energy Savings: Optimized driving reduces fuel con-
of AI and learning algorithms enable autonomous vehicles
to navigate, perceive, and adapt to dynamic environments, sumption and emissions.
• Productivity and Convenience: Passengers use travel
making them safer and more efficient. Continuous advance-
ments in AI technologies are expected to further enhance the time productively while delivery services become more
capabilities and safety of autonomous vehicles in the future. efficient.
Autonomous system development has been experiencing a AI in autonomous vehicles is poised for a bright future,
transformational evolution through the integration of Artifi- shaping everyday life and creating exciting opportunities.
cial Intelligence (AI). This revolutionary combination holds Here’s a glimpse of the possibilities:
1) Technological Advancements:
Corresponding Author: Divya Garikapati is a Senior IEEE Member. E-mail: • Sharper perception and decision-making: AI algorithms
([email protected], [email protected]),
Sneha.S.Shetiya is a Senior IEEE Member. Email:([email protected]) are more adept at understanding environments with ad-
Manuscript created January 20, 2024. vanced sensors and robust machine learning.
 2

• Faster, more autonomous operation: Edge computing good pavement quality, Daytime and nighttime operation,
enables on-board AI processing for quicker decisions and moderate traffic density, dynamic route planning, Traffic
greater independence. light/stop sign recognition, intersection navigation, ma-
• Enhanced safety and reliability: Redundant systems and neuvering in yards/warehouses etc.,
rigorous fail-safe mechanisms prioritizes safety above all • TuSimple and Embark Trucks: Sunny, dry weather, clear
else. visibility. Temperature range -10°C to 40°C, Limited-
2) Education and Career Boom: access highways with clearly marked lanes, Daytime
• Surging demand for AI expertise: Specialized courses and
operation only, maximum speed 70 mph, limited traf-
degrees in autonomous vehicle technology will cater to fic density, pre-mapped routes, Lane changes, highway
a growing need for AI, robotics, and self-driving car merging/exiting, platooning with other AV trucks etc.,
professionals. • Pony.ai and Einride: Diverse weather conditions, in-
• Interdisciplinary skills will be key: Professionals with
cluding heavy rain/snow. Variable lighting and complex
cross-functional skills bridging AI, robotics, and trans- urban environments, Narrow city streets, residential areas,
portation will be highly sought after. and parking lots, Low speeds (20-30 mph), high traffic
• New career paths in safety and ethics: Expertise in
density, frequent stops and turns, geo-fenced delivery
ethical considerations, safety audits, and regulatory [[3]] zones, Pedestrian and cyclist detection/avoidance, obsta-
compliance will be crucial as self-driving cars become cle avoidance in tight spaces, dynamic rerouting due to
widespread. congestion etc.,
• Komatsu Autonomous Haul Trucks, Caterpillar MineS-
3) Regulatory Landscape: tar Command for Haul Trucks: Harsh weather con-
• Standardized safety guidelines: Governments will estab- ditions (dust, heat, extreme temperatures). Limited or
lish common frameworks for performance and safety, no network connectivity, Unpaved roads, uneven terrain,
building public trust and ensuring industry coherence. steep inclines/declines, Autonomous operation with re-
• Stringent testing and validation: Autonomous systems mote monitoring, pre-programmed routes, high ground
will undergo rigorous testing before deployment, guar- clearance, Obstacle detection in unstructured environ-
anteeing reliability and safety standards. ments, path planning around natural hazards, dust/fog
• Data privacy and security safeguards: Laws and regula- mitigation, etc.,
tions will address data privacy and cybersecurity con- • Baidu Apollo: Highways and city streets in specific zones
cerns, protecting personal information and mitigating like Beijing and Shenzhen. Operates in daytime and
cyberattacks. nighttime, under clear weather conditions, and limited
• Ethical and liability frameworks: Clearly defined legal traffic density. Designed for passenger transportation and
frameworks will address ethical decision-making and de- robotaxis. Specific scenarios include Lane changes, high-
termine liability in situations involving self-driving cars. way merging/exiting, traffic light/stop sign recognition,
This future holds immense potential for revolutionizing trans- intersection navigation, low-speed maneuvering in urban
portation, creating new jobs, and improving safety. However, areas.
navigating ethical dilemmas, ensuring robust regulations, and • WeRide: Limited-access highways and urban streets in
building public trust will be crucial for harnessing this tech- Guangzhou and Nanjing. Operates in daytime and night-
nology responsibly and sustainably. time, under clear weather conditions. Targeted for rob-
otaxi services and last-mile delivery. Specific scenarios
B. Operational Design Domains (ODDs) Expansions into include Lane changes, highway merging/exiting, traffic
new areas and Diversity - The Current Industry Landscape light/stop sign recognition, intersection navigation, auto-
mated pick-up and drop-off for passengers/packages.
These examples illustrate the diverse evolution of Opera-
tional Design Domains (ODDs) [[4]] across various vehicle • Bosch & Daimler: Motorways and specific highways
types, including trucks and cars, and within different geograph- in Germany. Operates in daytime and nighttime, under
ical locations such as the United States, China, and Europe. good weather conditions. Focused on highway trucking
applications. Specific scenarios include Platooning with
• Waymo Driver: Can handle a wider range of weather
other AV trucks, automated lane changes and overtak-
conditions, city streets, and highway driving, but speed
ing, emergency stopping procedures, communication with
limitations and geo-fencing restrictions apply.
traffic management systems.
• Tesla Autopilot: Primarily for highway driving with lane
markings, under driver supervision, and within specific • Volvo Trucks: Defined sections of Swedish highways.
speed ranges. Operates in daytime and nighttime, under varying weather
• Mobileye Cruise AV: Operates in sunny and dry weather,
conditions. Tailored for autonomous mining and quarry
on highways with clearly marked lanes, and at speeds operations. Specific scenarios include Obstacle detection
below 45 mph. and avoidance in unstructured environments, path plan-
• Aurora and Waymo via: Wider range of weather condi-
ning around natural hazards, pre-programmed routes with
tions, including light rain/snow. Variable lighting (sun- high precision, remote monitoring and control.
rise/sunset), Multi-lane highways and rural roads with In this paper, we discuss the AI-powered software develop-
 3

ment lifecycle for autonomous vehicles and discuss the details

on how to ensure software quality, security and resolve ethical
dilemmas by taking different biases into account during the
development of the AI algorithms. We explain how the AI
algorithms have been emerging and evolving over time to have
more and more decision making capabilities without human
involvement using IOT as a future direction of expansion for
autonomous vehicles to being more connected to other actors
in the driving environment. In-cabin experience enhancements
and Driver Assistant Systems were also discussed as part of the
emerging trends. A literature survey of how the AI algorithms Fig. 1. AI-Powered Development Life-Cycle
are being used within autonomous vehicles has been provided
in Section V. In Section VI, we have also provided certain
statistics on how the use of AI and Learning algorithms have Integration and Deployment: Integrate the model into the
been evolving over time, how the research in these areas has broader system under development for autonomy implemen-
been trending over time, different AI model parameters being tation. Deploy and test the software application incorporating
considered for autonomous trucks vs. passenger cars etc.,. the new model using multiple methods like software-in-the-
Another interesting statistic on how the use of AI and Learning loop, hardware-in-the-loop, human-in-the-loop etc., using sim-
algorithms change based on the levels of autonomy was also ulation, closed course and limited public road environments.
provided. Some models are trained to improve their learning even after
deployment. These models need to be tested for future direc-
tions of learning to ensure compliance to ethical considerations
II. T HE AI- POWERED D EVELOPMENT L IFE -C YCLE IN as explained in Section III and other requirements.
AUTONOMOUS V EHICLES Using a systematic process like this would help build
This section describes about the key aspects involved with the confidence levels on each model being developed and
the AI-powered development life cycles within autonomous deployed in various subsystems of autonomous vehicles like
vehicles and these could be applicable to other fields as well perception, planning, controls and Human-Machine Interface
in general. (HMI) applications.

B. Ensuring Software Quality and Security

A. Model Training and Deployment
In autonomous vehicles, the integration of AI in various
AI model training and deployment in autonomous vehicles aspects of software development and maintenance plays a
involves a systematic process and typically includes several crucial role in ensuring the robustness and security of the
stages: overall system. Automated testing, powered by AI-based tools,
Data Collection and Pre-processing: Gathering a vast emerges as a key component in the testing process. These tools
amount of data from real-world sensors, pre-existing datasets efficiently identify bugs, vulnerabilities, and ensure that the
and other sources such as synthetic datasets. Cleaning and pre- software functions as intended, contributing to the reliability
processing the data to make it suitable for machine learning of autonomous vehicle software. Additionally, AI extends its
models. capabilities to code analysis and review, providing a thorough
Model Training: Employ learning models such as neural examination of the codebase for quality and highlighting po-
networks, deep learning [[5]], or natural language processing tential issues or vulnerabilities. Predictive maintenance, facili-
(NLP) to understand patterns and structures based on the data. tated by AI, becomes essential for anticipating and addressing
Training the models to a desired level of accuracy based on potential software failures, ultimately reducing downtime and
each scenario or in generic abstract cases like being able to enhancing the overall operational efficiency of autonomous
extract the patterns during the live operation of the vehicles. vehicles. Moreover, AI-driven anomaly detection and security
Model Generation: Train models to perform a certain de- monitoring contribute significantly to the safety of autonomous
cision making task, functions, or modules based on learned vehicles. By continuously monitoring the software environ-
patterns. These models can use various architectures, such as ment, AI systems can identify abnormal patterns or behaviors,
decision trees, random forests, regression trees, deep layers, promptly responding to potential security threats in real time.
ensemble learning etc., Vulnerability assessment, another application of AI tools, con-
Code Refinement and Optimization[[6]]: Refine the gener- ducts in-depth evaluations to pinpoint weaknesses in software
ated code to improve its quality, readability, and functionality. systems, providing valuable insights to mitigate risks effec-
Post-generation processing ensures the code adheres to coding tively. Behavioral analysis powered by AI proves instrumental
standards, conventions [[7]] and requirements. in understanding user interactions within the software. This
Quality Assessment: Evaluate the generated code for cor- capability aids in detecting and preventing suspicious or ma-
rectness, efficiency, and adherence to the intended functional- licious activities, fostering a secure and reliable autonomous
ities. This involves testing, debugging, and validation proce- vehicle ecosystem. Finally, AI’s role in fraud detection within
dures. software applications adds an extra layer of security, ensuring
 4

the integrity of autonomous vehicle systems and safeguarding • Mitigation: Robust testing against adversarial sce-
against potential security breaches. In summary, the integration narios, incorporating security measures, and regular
of AI in these diverse areas significantly enhances the overall updates to address emerging threats.
safety, security, and efficiency of autonomous vehicles. 7) Inclusivity and Accessibility:
• Challenge: Biases in AI can result in excluding
III. E THICAL C ONSIDERATIONS AND B IAS IN AI-D RIVEN certain demographics, reinforcing digital divides.
S OFTWARE D EVELOPMENT FOR AUTONOMOUS VEHICLES • Mitigation: Prioritizing diversity in development
To address the challenges related to bias, understanding and teams, actively seeking user feedback, and conduct-
addressing these concerns are crucial for building responsible ing accessibility assessments to ensure inclusivity.
and fair AI-driven software for autonomous vehicles. Here are 8) Social Impact:
key points highlighting ethical considerations and bias in AI-
• Challenge: The deployment of biased AI systems
driven software development:
can have negative social implications, affecting
1) Data Bias: marginalized communities disproportionately.
• Challenge: AI models learn from historical data, • Mitigation: Conducting thorough impact assess-
and if the training data is biased, the model can ments, involving diverse stakeholders in the de-
perpetuate and amplify existing biases. velopment process, and considering societal conse-
• Mitigation: Rigorous data pre-processing, diversity quences during AI development.
in training data, and continuous monitoring for bias 9) Continuous Monitoring and Adaptation:
are essential. Ethical data collection practices must
• Challenge: AI models may encounter new biases
be upheld.
or ethical challenges as they operate in dynamic
2) Algorithmic Bias: environments.
• Challenge: Algorithms may inadvertently encode • Mitigation: Establishing mechanisms for ongoing
biases present in the training data, leading to dis- monitoring, feedback loops, and model adaptation
criminatory outcomes. to address evolving ethical considerations.
• Mitigation: Regular audits of algorithms for bias,
10) Ethical Frameworks and Guidelines:
transparency in algorithmic decision-making, and
• Challenge: The absence of standardized ethical
the incorporation of fairness metrics during model
frameworks can lead to inconsistent practices in AI
evaluation.
development.
3) Fairness and Accountability: • Mitigation: Adhering to established ethical guide-
• Challenge: Ensuring fair outcomes and establishing lines, such as those provided by organizations like
accountability for AI decisions is complex, espe- the ISO, IEEE, SAE, Government regulatory boards
cially when models are opaque. etc., and actively participating in the development of
• Mitigation: Implementing explainable AI (XAI) industry-wide standards.
techniques, defining clear decision boundaries, and
Addressing ethical considerations and bias in AI-driven soft-
establishing accountability frameworks for AI-
ware development in autonomous vehicles requires a holistic
generated decisions.
and proactive approach[[10]]. It involves a commitment to
4) Explainability and Transparency: fairness, transparency, user privacy, and social responsibility
• Challenge: Many AI models operate as ”black throughout the AI development lifecycle. As the field evolves,
boxes,” making it challenging to understand how continuous efforts are needed to refine ethical practices and
decisions are reached. AI safety is another challenge promote responsible AI deployment.
that needs to be made sure is safety-critical appli-
cations like autonomous vehicles[9] IV. AI’ S ROLE IN THE E MERGING TREND OF I NTERNET OF
• Mitigation: Prioritizing explainability [[8]] in AI
T HINGS (I OT) E COSYSTEM FOR AUTONOMOUS VEHICLES
models, using interpretable algorithms, and provid-
ing clear documentation on model behavior. Artificial Intelligence (AI) plays a crucial role in shaping
and enhancing the capabilities of the Internet of Things
5) User Privacy:
(IoT). Here’s an overview of how AI contributes to the IoT
• Challenge: AI systems often process vast amounts Ecosystem for Autonomous vehicles
of personal data, raising concerns about user pri- In the realm of Connected and Autonomous Vehicles
vacy. (CAVs), AI and IoT converge to create a seamless network
• Mitigation: Implementing privacy-preserving tech-
of intelligence and connectivity, transforming the driving
niques, obtaining informed consent, and adhering experience. Vehicles become intelligent agents, processing
to data protection regulations (e.g., GDPR [[9]]) to sensor data in real-time to make informed decisions: predicting
safeguard user privacy. traffic patterns, optimizing routes, detecting anomalies, and
6) Security Concerns: even adapting to changing road conditions with dynamic
• Challenge: AI models can be vulnerable to adver- adjustments. This intelligent ecosystem extends beyond indi-
sarial attacks, posing security risks. vidual vehicles, interconnecting with infrastructure and other
 5

Generative Artificial Intelligence (Gen AI): This technol-

ogy uses machine learning algorithms to produce new and
original outputs based on the patterns and information it has
learned from training data. In the context of vehicles, genera-
tive AI can be applied to various aspects, including natural lan-
guage processing for in-car voice assistants, content generation
for infotainment systems, and even simulation scenarios for
testing autonomous driving systems. Large Language Models
(LLMs) are a specific class of generative AI models that
are trained on massive amounts of text data to understand
and generate human-like language.In vehicles, LLMs can be
employed for natural language understanding and generation,
Fig. 2. AI’s role in the Internet of Things (IOT) Ecosystem for Autonomous allowing for more intuitive and context-aware interactions
vehicles
between the vehicle and its occupants. This can enhance
features like voice-activated controls, virtual assistants, and
vehicles to optimize traffic flow, anticipate potential hazards, communication systems within the vehicle.
and personalize the driving experience.
Key AI-powered IoT capabilities in CAVs include the V. R EVIEW OF E XISTING R ESEARCH AND U SE C ASES
following:
• Real-time data processing and analysis for insights into
H. J. Vishnukumar et. al. [[12]] introduced that traditional
traffic, road conditions, and vehicle health. development methods like Waterfall and Agile, fall short when
• Predictive analytics for proactive maintenance, efficient
testing intricate autonomous vehicles and proposes a novel
resource allocation, and informed decision-making. AI-powered methodology for both lab and real-world testing
• Enhanced automation for autonomous driving tasks,
and validation (T&V) of ADAS and autonomous systems.
adaptive cruise control, and dynamic route optimization. Leveraging machine learning and deep neural networks, the
• Efficient resource management for optimizing energy
AI core learns from existing test scenarios, generates new
consumption, bandwidth usage, and load balancing. efficient cases, and controls diverse simulated environments
• Security and anomaly detection for identifying potential
for exhaustive testing. Critical tests then translate to real-
threats and preventing cyberattacks. world validation with automated vehicles in controlled set-
• Personalized user experience through customized set-
tings. Constant learning from each test iteration refines fu-
tings, preferences, and tailored insights. ture testing, ultimately saving precious development time and
• Edge computing for real-time decision-making, reducing
boosting the efficiency and quality of autonomous systems.
latency and improving responsiveness. This methodology lays the groundwork for AI to eventually
handle most T&V tasks, paving the way for safer and more
Challenges to address include ensuring data privacy, security,
reliable autonomous vehicles.
interoperability, and overcoming resource constraints in con-
Bachute, Mrinal R et. al. [[13]] described the algorithms
nected vehicles. The seamless integration of AI and IoT holds
crucial for various tasks in Autonomous Driving, recognizing
the potential to revolutionize transportation, leading to safer,
the multifaceted nature of the system. It discerns specific
more efficient, and sustainable [[11]] mobility solutions.
algorithmic preferences for tasks, such as employing Rein-
forcement Learning (RL) models for effective velocity control
A. Enhancing User Experience in car-following scenarios and utilizing the ”Locally Decor-
Personalization and Recommendation Systems in-cabin: related Channel Features (LDCF)” algorithm for superior
AI-driven personalization and recommendation systems in pedestrian detection. The study emphasizes the significance
Autonomous vehicles use machine learning models to analyze of algorithmic choices in motion planning, fault diagnosis
user behavior and preferences, creating personalized recom- with data imbalance, vehicle platoon scenarios, and more.
mendations for tools, libraries, and vehicle maneuvers. They Notably, it advocates for the continuous optimization and
collect and pre-process user data, create individual profiles, expansion of algorithms to address the evolving challenges
generate tailored suggestions, and continuously adapt based in Autonomous Driving. The paper serves as an insightful
on real-time interactions, aiming to enhance user experience foundation, prompting future research endeavors to broaden
and developer productivity. the scope of tasks, explore a diverse array of algorithms, and
Natural Language Processing (NLP) in-cabin: NLP en- fine-tune their application in specific areas of interest within
ables software to comprehend and process human language. the Autonomous Driving System.
This includes chat bots [[12]], virtual assistants and voice Y. Ma et. al. [[14]] explained the pivotal role of
recognition systems that understand and respond to natural artificial intelligence (AI) in propelling the development
language queries in vehicle cabins. It allows the vehicle sub- and deployment of autonomous vehicles (AVs) within
systems to analyze and derive insights from user requirements the transportation sector. Fueled by extensive data from
and structuring requirements effectively to create responses diverse sensors and robust computing resources, AI has
and certain vehicle maneuvers. become integral for AVs to perceive their environment and
 6

make informed decisions while in motion. While existing TABLE I

research has explored various facets of AI application AUTONOMOUS D RIVING : K EY TECHNOLOGIES EVOLUTION
in AV development, this paper addresses a gap in the Technology Developed over years Future
literature by presenting a comprehensive survey of key Very challenging. Needs more
DL for object detection research to better detect objects
studies in this domain. The primary focus is on analyzing Environmental YOLOv3 K-means cluster- in blurry, extreme and rare con-
Perception ing ditions in real time.
how AI is employed in supporting crucial applications in PVANET and RCNN
AVs: 1) perception, 2) localization and mapping, and 3) Pedestrian model for object detection OrientNet, RPN, and Predictor-
detection during blurry weather Net to solve occlusion problem
decision-making. The paper scrutinizes current practices multisensor fusion system, along
to elucidate the utilization of AI, delineating associated with an INS, a GNSS, and a Li-
DL algorithm based on DAR system, would be used to
challenges and issues. Furthermore, it offers insights into Path Planning CNN implement a 3D SLAM.
potential opportunities by examining the integration of AI Vehicle Cyber- Remote control of AV deploying
security Security testing and TARA IoT sensors
with emerging technologies such as high-definition maps, Grey prediction model utilising
Motion Hidden Markov model Q- and Advanced model predictive
big data, high-performance computing, augmented reality Planning learning algorithm control for effective lane change
(AR) and virtual reality (VR) enhanced simulation platforms,
and 5G communication for connected AVs. In essence,
this paper serves as a valuable reference for researchers a) AI (Artificial Intelligence) :
seeking a deeper understanding of AI’s role in AV research, • Expert Systems: Rule-based systems that mimic human
providing a comprehensive overview of current practices expertise for decision-making [[17]].
and paving the way for future opportunities and advancements. • Decision Trees: Hierarchical structures for classification
and prediction.ex: prognostics area
VI. AI AND L EARNING A LGORITHMS STATISTICS FOR • Search Algorithms: Methods for finding optimal paths
AUTONOMOUS VEHICLES or solutions, such as A* search and path planning algo-
rithms.
This Section extends the analysis of Artificial Intelligence
• Generative AI: To create scenarios for training the system
(AI) and Learning Algorithms in autonomous vehicles, build-
and for balancing data on high severity accident/non-
ing upon previous work as described in Section V. The focus is
accident cases. (CRSS dataset). Create a non-existent
on providing additional statistical insights into the following: -
scenario dataset. Supplement the real datasets. Simulation
evolution of different types of AI and learning algorithms over
testing.
the years, - research trends in application of AI in all fields vs.
• NLP: AI Assistant (Yui, Concierge, Hey Mercedes, etc.,)
autonomous vehicles, - creation of a parameter set crucial for
- LLMs
autonomous trucks versus cars, - evolution of AI and learning
algorithms at different autonomy levels, and - changes in the b) ML (Machine Learning):
types of algorithms, software package size etc., over time. • Supervised Learning: Algorithms that learn from labeled
data to make predictions, such as:
A. Stat1: Trends of usage of AI, ML and DNN Algorithms – Linear Regression: For predicting continuous values.
over the years – Support Vector Machines (SVMs): For classification
Today, a vehicle’s main goal is not limited to transportation, and outlier detection.
but also includes comfort, safety, and convenience. This led – Decision Trees: For classification and rule genera-
to extensive research on improving vehicles and incorporating tion.
technological breakthroughs and advancements. – Random Forests: Ensembles of decision trees for
As per prior work done for the development for architecture improved accuracy.
and ADAS technology it is evident that the research till now • Unsupervised Learning: Algorithms that find patterns in
has limitations. These limitations are pertaining either to the unlabeled data, such as:
author’s elaboration of his/her knowledge or not having proper
– Clustering Algorithms (K-means, Hierarchical): For
sources. Thus its a good exercise to have a look at the trends
grouping similar data points.
over the years as our capabilities to develop these ML models
– Dimensionality Reduction (PCA, t-SNE): For reduc-
have gotten better and also the access to better computing
ing data complexity.
units[[15]][[16]] has led to the evolution of the algorithms.
In the Table I , we have summarized different modelling c) DNN (Deep Neural Networks):
algorithms for various standard components of the ADAS • Convolutional Neural Networks (CNNs): For image and
algorithm. The second column illustrates the technology that video processing, used for object detection, lane segmen-
exists in today’s date and the third column predicts potential tation, and traffic sign recognition.
future development which is efficient than the current. • Recurrent Neural Networks (RNNs): For sequential data
Below we have derived series of plots pertaining to re- processing, used for trajectory prediction and behavior
search publications in AI(Artificial Intelligence), ML(Machine modeling.
Learning) and DNN(Deep Neural Network) domains. Brief • Deep Reinforcement Learning (DRL): For learning
explanations have been provided before to understand what through trial and error, used for control optimization and
topics come under these domains. decision-making.
 7

d) Specific Examples in Autonomous Vehicles: Shifting Trends in IEEE Publications: Interestingly, post-
• Object Detection (DNN): CNNs like YOLO [[18]], SSD 2021, the upward trend in LMM and DNN publications
[[19]], and Faster R-CNN are used to detect objects (identified through filters aligned with our previous analysis)
around the vehicle. appears to plateau. This suggests a potential shift in research
• Lane Detection (DNN): CNNs are used to identify lane focus within computer vision (CV) following the emergence
markings and road boundaries . of Generative AI (GenAI) and other advanced technologies.
• Path Planning (AI): Search algorithms like A* and RRT While LMM and DNN remain foundational, their prominence
are used to plan safe and efficient routes. as primary research subjects within classic CV might be
• Motion Control (ML): Regression models declining.
[[20]][[21]][[22]] are used to predict vehicle dynamics Considering CVPR Publications: Initially, we considered
and control steering, acceleration, and braking. including CVPR publications in our analysis. However, we
• Behavior Prediction (ML): SVMs or RNNs are used to ultimately excluded them due to significant overlap with the
anticipate the behavior of other vehicles and pedestrians. IEEE dataset. As a significant portion of CVPR papers are
subsequently published in IEEE journals, including both sets
In the Figure 3 , we evaluated the papers from
would introduce redundancy and potentially skew the analysis.
[[23]][[24]]and found the trends to be as shown. One can
Figure 4 focuses on all of AI/ML publications related
observe that year 2013 the number of algorithms in DNN
to IEEEXplore [[25]], MDPI (Multidisciplinary Digital Pub-
surpasses that in generic AI and ML. This shows more research
lishing Institute) [[26]] and SAE (Society Of Automotive
with deep neural networks and the traction it received in the
Engineers) [[27]] Figure 5 focuses on the trend changes in
AI community. However the main takeaway from the graph
publications for autonomous vehicles. Figure6 focuses on
is the exponential upward trend in the number of algorithms
Science Direct [[28]] where we see the publications are in
over the years developed for AI applications.
thousands with very little presence for autonomous vehicles.
This is an indication of how AI/ML applications have sur-
passed engineering and are used everywhere from medical to
defence.
From the graphs we see comparatively less publications in
starting years 2014-2018. There is a huge surge in 2018 where
also we see Autonomous vehicles with advanced self driving
features gained traction. From the trend, we expect in future
a similar exponential rise. However we do expect additional
parameters(for ex:data got introduced) to be in the list. With
AI/ML applications coming up in every industry along with
automotive, the future for research in the area is promising.

Fig. 3. Trends of usage of AI,ML and DNN algorithms over the years

Some research was done considering platforms of IEE-

EXplore, SAE Mobilus, MDPI and Science Direct to find
out the published research in AI/ML and also particularly in
Autonomous vehicles.
When filtering the MDPI Journals and articles, one can
observe that there is an additional filter relating to Data that
pops up after 2021. This indicates that pre-2020 not much
papers related to data handling and analysis were published
as the collected data was not huge. One also observes that
the year 2020 (year of the COVID pandemic) for MDPI; saw
minimal papers for autonomous technology. While several fac- Fig. 4. No. of Publications related to AI/learning Algorithms in all Fields
tors may contribute to the rise in model deployments observed
in 2021, a possible explanation is the limited opportunity
for previous models to undergo real-world testing through B. Stat2: Parameters for AI model (Trucks vs. Cars)
vehicle deployment. Notably, the number of deployed models As per American Trucking Associations (ATA), there will be
surged to 737 in 2021, representing nearly a twofold increase a shortage of over 100,000 truck drivers in the US by 2030,
compared to earlier years. which could potentially double by 2050 if trends continue.
From the IEEE publications, one can see that although effec- Bureau of Labor Statistics (BLS stats), while not explicitly
tive research in AI/ML increases over time not much research predicting a shortage, the BLS projects a slower-than-average
has been published towards autonomous vehicle technology. job growth for truck drivers through 2030, indicating potential
 8

TABLE II
PARAMETER SET FOR DIFFERENCES IN USING AI IN AUTONOMOUS
TRUCKS AND CARS

Parameter Sub-class Trucks Cars

Environment Traffic operate on highways encounter diverse, often
density with predictable traffic congested, urban environ-
patterns, ments.
Road navigate primarily deal with varied road con-
infrastructure on well-maintained ditions and potentially un-
highways marked streets.
Weather con- may prioritize stability may prioritize maneuver-
ditions and visibility for cargo ability for passenger com-
safety fort.
Vehicle Size and Larger size and weight Smaller size and weight in
characteris- weight present different sensor comparison to Trucks.
tics ranges and dynamic re-
sponse complexities
Cargo require AI to manage This is not a concern for
handling cargo weight distribution cars
and safety and potential shifting
Fig. 5. No. of Publications related to AI/learning Algorithms for Autonomous cargo
Vehicles Fuel Truck AI prioritizes ef- Car AI may prioritize
efficiency ficient fuel consumption smoother acceleration and
and emissions due to long-distance travel deceleration for passenger
comfort.
Operational Route require long-distance generally focus on shorter,
considera- planning and route planning with dynamic routes with real-
tions: optimization considerations for time traffic updates.
infrastructure limitations,
rest stops, and cargo
delivery schedules.
Communication may rely on dedicated in- primarily use existing cel-
and frastructure for communi- lular networks.
connectivity cation (platooning, V2X)
Legal and reg- Regulations regarding au- regulations impacting AI
ulatory land- tomation and liability are and deployment are differ-
scape tight. ent than trucks.
AI Perception may prioritize radar and benefit from high-
algorithm and sensor LiDAR for long-range de- resolution cameras
and fusion tection for near-field obstacle
hardware avoidance.
needs
Decision- AI focuses on safe, fuel- AI prioritizes dynamic
making and efficient navigation and route adjustments,
planning traffic flow optimization pedestrian/cyclist
detection, and passenger
Fig. 6. No. of Publications related to AI/learning Algorithms for Autonomous comfort.
Vehicles vs. all Fields in Science Direct Redundancy may have stricter fail-safe have safety protocols with
and safety measures due to cargo redundant systems
protocols risks.
Additional Public Public trust in truck au- Public trust in car automa-
factors perception tomation might be slower tion is higher due to lesser
challenges in meeting future demand. Aging work force, and to build due to size and risks
acceptance potential cargo risks.
demanding job conditions and regulatory hurdles are few of Economic and automation models automation may focus on
the reasons which contribute towards the same. business mod- may involve fleet ride-sharing and individ-
els management and logistics ual ownership
The above two results give a good business case for driver- optimizations
less trucks in comparison to driverless cars. This is also contra-
dictory to the belief that truck drivers may loose jobs over the
self-driving technology. In fact as per [[29]],[[30]],driverless to turn the industry upside down. The paper clearly states
trucks can drastically reduce the driver costs, increase truck backed up by a study that level 4 automation will be reached
utilization and improve truck safety. Inspite of this, one can by 2030 followed by level 5 in 2040. Based on the interview
see not enough research has been done on the impacts of self- results conducted in [[29]] and the delphi-based scenario study
driving trucks compared to passenger transport [[31]]. There with projections for the next 10 years, it is evident one needs
is a need to ensure road freight transport has alignment with to seriously consider the impact of automation on trucks.
its current operations retaining its value chain. One cannot Lots of research revolves around passenger cars with many
think of cost reductions by taking out the driver cabin as competitors in the market. We found that not much data exists
most self-driving technology developing trucking companies for self- driving for trucks.
are focusing on hub to hub transport and unlike passenger This parameter set as shown in Table II serves as a starting
cars, not from source to destination. One would still need a point for understanding the key differences in how AI is
driver in the start and end of the journey. This refocuses on applied to autonomous trucks and cars. Each parameter can
the statement above for the need of truck drivers in future but be further explored and nuanced based on specific scenarios
eliminating the other drawbacks of long haul freight transport. and applications. Currently, the autonomous trucking has been
As mentioned in [[30]] above, according to Daimler Ex- expanding in 4 major categories such as Highway Trucking
CEO Zetsche, future vehicles need to have four characteristics; ODD, Regional Delivery ODD, Urban logistics ODD and
connected, autonomous, shared, and electric, a so-called CASE Mining and Off-Road ODD. There are 3 different categories
vehicle. Nevertheless, each of these points has the potential as well based on the different stages of logistics to handle
 9

the movement of goods for autonomous trucking like Long – Waymo is expanding its robotaxi service in Phoenix,
Haul, Middle Mile and Last Mile. Understanding these cate- Arizona, with plans to eventually launch fully driver-
gorization and how the trucking industry has been evolving to less operations.
deliver more autonomous vehicles is really important for the – The company’s Waymo Via trucking division con-
future of logistics to help optimize and streamline the entire tinues testing autonomous trucks in California and
supply chain, ensuring efficient and timely delivery of goods Texas.
to their final destination. – No publicly announced timeline for nationwide de-
ployment of driverless trucks.
Overall:
C. Stat3: Usage of AI and Learning Algorithms at various • Both Kodiak and Waymo are making progress towards
Levels of Autonomy commercializing Level 4 autonomous vehicles, but pri-
marily focused on different segments (trucks vs. passen-
Autonomous vehicles operate at various levels of autonomy,
ger cars).
from Level 0 to Level 5, each presenting unique challenges
• Driverless truck deployment timelines remain flexible and
and opportunities. This section explores the diversity
dependent on regulatory approvals and further testing as
and evolution of AI algorithms across different levels of
was discussed previously.
autonomous vehicle capabilities. Autonomous vehicles are
categorized into different levels based on their autonomy, a) Key AI/Learning Components across Levels:
• Perception:
with increasing complexity and diversity of AI algorithms
as autonomy levels progress. The six levels [[34]] of AV – L0-L2: Basic object detection and lane segmentation
autonomy define the degree of driver involvement and vehicle using CNNs.
automation. At lower levels (L0-L2), driver assistance systems – L3-L4: LiDAR-based object detection, advanced sen-
primarily utilize rule-based and probabilistic methods for sor fusion algorithms for robust object recognition.
specific tasks like adaptive cruise control or lane departure – L5: 3D object mapping, robust sensor fusion and
warning. Higher levels (L3-L4) rely heavily on machine interpretation.
learning and deep learning algorithms, particularly for • Decision-Making:
perception tasks like object detection and classification using – L0-L2: Rule-based algorithms for lane change assis-
convolutional neural networks (CNNs). Advanced sensor tance, adaptive cruise control.
fusion techniques combine data from cameras, LiDAR, radar, – L3-L4: Probabilistic roadmap planning (PRM),
and other sensors to create a comprehensive understanding decision-making models for route selection.
of the environment. Furthermore, reinforcement learning – L5: Deep reinforcement learning for adaptive behav-
and probabilistic roadmap planning algorithms contribute ior prediction, high-level route planning.
to complex decision-making and route planning in L3-L4 • Control:
AVs. L5 (full automation) requires robust sensor fusion, – L0-L2: PID controllers [[22]] for basic acceleration
3D mapping capabilities, and deep reinforcement learning and braking adjustments.
approaches for adaptive behavior prediction and high-level – L3-L4: Model Predictive Control (MPC) [[35]] for
route planning. complex maneuvers, trajectory tracking algorithms.
– L5: Multi-task DNNs for real-time coordination of
Some industry relevant examples have been illustrated be- all driving actions.
low: The following Table III provides examples of AI algorithms
Kodiak used at different autonomy levels, from L0 to L5, highlighting
• Status: Kodiak currently operates a fleet of Level 4 key techniques and applications. We have considered the
autonomous trucks for commercial freight hauling on percentage of systems using AI algorithms, algorithm types,
behalf of shippers. examples of AI algorithms at each level and the key tasks
• Recent Developments: being automated at each level of autonomy. Please note that
– Kodiak is focusing on scaling its autonomous truck- at L0, the extent to which AI or learning algorithms being used
ing service as a model, providing the driving system is very minimal and not complete algorithms in themselves,
to existing carriers. although there could be some partial techniques being used
– The company recently secured additional funding to like data processing or detecting an object on road
The level of autonomy in an AV directly correlates with the
expand its operations and partnerships.
size of its software package. Imagine a pyramid, with Level 0
– No immediate news about deployment of driverless
at the base (smallest size) and Level 5 at the peak (largest size).
trucks beyond current operations.
Each level adds functionalities and complexities, reflected in
Waymo the increasing size of the pyramid.
• Status: Waymo remains focused on Level 4 autonomous Challenges and Implications:
vehicle technology, primarily targeting robotaxi services • Limited Storage & Processing Power: Current onboard
in specific geographies. storage and processing capabilities might not yet be
• Recent Developments: sufficient for larger Level 4 and 5 software packages.
 10

TABLE III • Slower download and installation times: This can be

S TATISTICS ON AI AND L EARNING A LGORITHMS IN AUTONOMOUS frustrating for users, especially in areas with limited
VEHICLES BASED ON L EVELS OF AUTOMATION
internet connectivity.
% of • Security concerns: Larger packages offer more attack
Systems Number
Using of vectors for potential hackers.
Level of AI/Learning Key Key Tasks AI/Learning
Auton- Algo- Algorithm AI/Learning Auto- Algo- Here’s how Nvidia and Qualcomm are tackling this chal-
omy rithms Types Algorithms mated rithms lenge:
L0 (No
Au- Nvidia:
toma-
tion) 0% N/A N/A N/A 0 • Drive Orin platform: Designed for high-performance AV
Adaptive
Cruise
applications, Orin features a scalable architecture that can
Control, handle large software packages.
Lane
Departure • Software optimization techniques: Nvidia uses various
Warning techniques like code compression and hardware-specific
Rule-based (LDW), Sensing,
L1 systems, Automatic basic optimizations to reduce software size without sacrificing
(Driver Decision Emergency alerts and
Assis- trees, Naive Braking interven-
performance.
tance) 50-70% Bayes (AEB) tions 3-5 • Cloud-based solutions: Offloading some processing and
Rule-based
systems, data storage to the cloud can reduce the size of the
Decision Traffic Sign Navigation, onboard software package.[[15]]
trees, Rein- Recognition, lane
forcement Highway control, Qualcomm:
L2 learning Autopilot stop-and-
(Partial (RL), Support (ACC + lane go, limited • Snapdragon Ride platform: Similar to Orin, Snapdragon
Au- Vector centering), environ-
toma- Machines Traffic Jam mental Ride is a scalable platform built for efficient processing
tion) 80-90% (SVM) Assist adaptation 5-10
Full
of large AV software packages.
control • Heterogeneous computing: Qualcomm utilizes different
Deep under
Learning specific processing units like CPUs, GPUs, and NPUs to optimize
(DL) (e.g., condi- performance and reduce software size by distributing
Convolutional tions,
Neural dynamic tasks efficiently.
Networks, envi- • Modular software architecture: Breaking down the soft-
L3 Recurrent ronment
(Con- Neural adap- ware into smaller, modular components allows for easier
ditional Networks), Urban tation,
Au- RL, Autopilot, complex updates and reduces the overall package size. [[16]]
toma- Probabilistic Valet decision-
tion) 90-95% models Parking making 10-15 Additional approaches:
Full
• Standardization: Industry-wide standards for AV software
Advanced control in
DL (e.g., specific can help reduce duplication and fragmentation, leading to
Generative environ-
Adversarial ments, smaller package sizes.
Networks, high-level • Compression algorithms: Advanced compression algo-
L4 Transform- navi-
(High ers), Multi- City gation, rithms can significantly reduce the size of data and code
Au- agent RL, Navigation, complex
toma- Sensor fusion Highway traffic
without compromising functionality.
tion) 95-99% algorithms Chauffeur scenarios 15-20 • Machine learning: Using machine learning to optimize
Full
control in software performance and resource utilization can help
Advanced all envi- reduce the overall software footprint.
DL, Multi- ronments,
agent RL, self- The battle against AV software package size is ongoing, and
Hybrid learning
algorithms and adap- both Nvidia and Qualcomm are at the forefront of developing
(combining tation, innovative solutions. As technology advances and these ap-
L5 (Full various human-
Au- types), Universal like proaches mature, we can expect to see smaller, more efficient
toma- Explainable Autonomy decision-
tion) 100% AI (XAI) making 20+ AV software packages that pave the way for wider adoption
of self-driving vehicles.
Here’s a deeper dive as shown in the Table IV into this
• Download and Update Challenges: Updating these larger relationship[[36]].
packages may require longer download times and poten-
tially disrupt vehicle operation. TABLE IV
• Security Concerns: The more complex the software, the S OFTWARE PACKAGE S IZE BASED ON THE L EVELS OF AUTONOMY
higher the potential vulnerability to cyberattacks, neces- Level of Autonomy Package Size
sitating robust security measures. 0 Few MB
AV software package size is a major challenge for develop- 1 100s of MB
2 100s MB to Few GBs
ers like Nvidia and Qualcomm, as larger packages require: 3 Few GB to 10’s of GB
• Increased processing power and memory: This translates 4 10’s of GB to 100’s of GB
to higher hardware costs and potentially bulkier systems. 5 100’s of GB to TBs
 11

The level of autonomy directly influences the size of an for drawing meaningful conclusions. The paper concludes by
AV’s software package. While higher levels offer greater presenting a clear image of the evolving AI landscape in
convenience and potential safety benefits, they come with the autonomous vehicles, stressing its critical role in efficient and
challenge of managing increasingly complex and computation- safe transportation solutions. It identifies key challenges and
ally intensive software packages that would require large stor- suggests areas for future research, contributing to a road map
age spaces that the current processors cannot accommodate. for researchers, practitioners, and enthusiasts interested in the
Hence, the transformation towards zonal-based architectures is dynamic relationship between AI, learning algorithms, and the
desirable with multiple but small number of processors that are forefront of contemporary transportation.
tasked to accomplish a particular function or task providing
the ample of storage space needed for moving towards Level R EFERENCES
5 along with supporting connected and automated vehicle [1] U. Bordoloi, S. Chakraborty, M. Jochim, P. Joshi, A.
concepts. Raghuraman, and S. Ramesh, ”Autonomy-driven Emerg-
ing Directions in Software-defined Vehicles,” in 2023
Design, Automation & Test in Europe Conference &
VII. C ONCLUSION
Exhibition (DATE), Apr. 2023, pp. 1-6.[CrossRef]
This paper presents a comprehensive analysis of the role [2] Z. Liu, W. Zhang, and F. Zhao, ”Impact, challenges
of AI and learning algorithms in autonomous vehicles, diving and prospect of software-defined vehicles,” Automotive
into various aspects such as the shifts from rule-based systems Innovation, vol. 5, no. 2, pp. 180-194, 2022.[CrossRef]
to deep neural networks due to improved model capabilities [3] A. Pugliese, S. Regondi, and R. Marini, ”Machine
and computing power. The specific needs for trucks vs. cars learning-based approach: global trends, research direc-
have been detailed like the trucks prioritizing hub-to-hub tions, and regulatory standpoints,” Data Science and Man-
transport and efficient long-haul journeys, while passenger agement, vol. 4, pp. 19-29, 2021. [CrossRef]
cars aiming for source-to-destination autonomy. Increasing [4] S. Khastgir et al., ”Distributed ODD Awareness for Con-
complexity from basic object detection (L0-L2) to 3D mapping nected and Automated Driving,” Transportation Research
and adaptive behavior prediction (L3-L5) has been explained. Procedia, vol. 72, pp. 3118-3125, 2023.[CrossRef]
Challenges and implications such as limited storage and [5] S. Grigorescu et al., ”A Survey of Deep Learning
processing power, software update concerns, and increased Techniques for Autonomous Driving,” arXiv preprint
security vulnerabilities at higher autonomy levels have been arXiv:1910.07738v2, Mar. 2020. [CrossRef]
discussed. Some of the key conclusions include that AI is [6] H. Le et al., ”Coderl: Mastering code generation through
crucial for achieving different levels of autonomous vehicle pretrained models and deep reinforcement learning,” Ad-
functionality. Advanced techniques like deep learning and rein- vances in Neural Information Processing Systems, vol. 35,
forcement learning are essential for higher levels with complex pp. 21314-21328, 2022. [CrossRef]
decision-making and adaptable behavior. Truck and car AI [7] M. Allamanis et al., ”Learning natural coding conven-
applications have distinct requirements. Trucks focus on route tions,” in Proceedings of the 22nd ACM SIGSOFT In-
optimization and fuel efficiency, while passenger cars prioritize ternational Symposium on Foundations of Software Engi-
passenger comfort and dynamic adaptation to urban environ- neering, Nov. 2014, pp. 281-293. [CrossRef]
ments. Software package size grows with autonomy level. This [8] P. Hase and M. Bansal, ”Evaluating explainable AI: Which
poses challenges for storage, processing, and software updates, algorithmic explanations help users predict model behav-
highlighting the need for efficient architectures and robust ior?,” arXiv preprint arXiv:2005.01831, 2020.[CrossRef]
security measures. Further research is needed on self-driving [9] M. Vogel, T. Bruckmeier, and F. Di Cerbo, ”General data
trucks despite their promising business potential. This area protection regulation (GDPR) infrastructure for microser-
lags behind passenger car research, but its development could vices and programming model,” U.S. Patent 10,839,099,
significantly optimize logistics and address driver shortages. 2020.[CrossRef]
Certain aspects fall outside the scope of this paper, including [10] J. Mökander and L. Floridi, ”Operationalising AI gov-
emerging technologies and AI algorithms like quantum AI, ernance through ethics-based auditing: an industry case
transfer learning, and Meta-Learning. Currently, these appli- study,” AI and Ethics, vol. 3, no. 2, pp. 451-468,
cations in robotics and physical systems, such as vehicles, are 2023.[CrossRef]
limited. However, the paper does not rule out their potential [11] R. Nishant, M. Kennedy, and J. Corbett, ”Artificial intel-
use within vehicles for tasks like edge computing or learning ligence for sustainability: Challenges, opportunities, and
human behaviors through transfer learning. In addition, the a research agenda,” International Journal of Information
study on trends in published research in artificial intelligence Management, vol. 53, p. 102104, 2020.[CrossRef]
(Stat 1) only considers the four most popular platforms within [12] H. Vishnukumar et al., ”Machine learning and deep
a broader automotive industry. The paper acknowledges the neural network — Artificial intelligence core for lab and
existence of other journals and conferences, such as NeurIPS real-world test and validation for ADAS and autonomous
[[37]], which may have more publications. A quick study on vehicles: AI for efficient and quality test and validation,”
these new platforms revealed similar trends to those presented pp. 714-721, 2017. [CrossRef]
in the paper. The exclusion of discussions on emerging trends [13] M. R. Bachute and J. M. Subhedar, ”Autonomous driving
is intentional to narrow the scope and present relevant studies architectures: insights of machine learning and deep learn-
 12

ing algorithms,” Machine Learning with Applications, vol. [28] Science Direct, ”AI/ML publications (2014-2023),” Ac-
6, p. 100164, 2021.[CrossRef] cessed on 28th January, 2024. [CrossRef]
[14] Y. Ma et al., ”Artificial intelligence applications in the de- [29] C. Fritschy, S. Spinler, ”The impact of autonomous
velopment of autonomous vehicles: A survey,” IEEE/CAA trucks on business models in the automotive and logistics
Journal of Automatica Sinica, vol. 7, no. 2, pp. 315-329, industry–a Delphi-based scenario study,” Technological
2020.[CrossRef] Forecasting and Social Change, vol. 148, pp. 119736,
[15] L. Marie, ”NVIDIA Enters Production With DRIVE 2019.[CrossRef]
Orin, Announces BYD and Lucid Group as New EV [30] A. Engholma, A. Björkmanb, Y. Joelssonb, I. Kristof-
Customers, Unveils Next-Gen DRIVE Hyperion AV Plat- ferssond, A. Perneståla, ”The emerging technological in-
form,” Published by Nvidia Newsroom, Press Release, novation system of driverless trucks,” Transportation Re-
2022, Accessed on 26 Jan 2024.[CrossRef] search Procedia, vol. 49, pp. 145-159, 2020. [CrossRef]
[16] K. Tharakram, ”Snapdragon Ride SDK: a premium plat- [31] D. Parekh, N. Poddar, A. Rajpurkar, M. Chahal, N. Ku-
form for developing customizable ADAS applications,” mar, G. P. Joshi, and W. Cho, ”A Review on Autonomous
Published by Qualcomm OnQ Blog, Jan. 2022, Accessed Vehicles: Progress, Methods and Challenges,” Electronics,
on 26 Jan 2024. [CrossRef] vol. 11, no. 14, p. 2162, 2022. [CrossRef]
[17] D. Garikapati and Y. Liu, ”Dynamic Control Limits Ap- [32] D. Chen, Y. Lin, W. Li, P. Li, J. Zhou, and X. Sun,
plication Strategy For Safety-Critical Autonomy Features,” ”Measuring and relieving the over-smoothing problem for
in 2022 IEEE 25th International Conference on Intelligent graph neural networks from the topological view,” in Pro-
Transportation Systems (ITSC), Oct. 2022, pp. 695-702. ceedings of the AAAI conference on artificial intelligence,
[CrossRef] vol. 34, no. 04, 2020, pp. 3438-3445. [CrossRef]
[18] J. Redmon and A. Farhadi, ”YOLOv3: An Incremental [33] M. Bojarski, D. D. Testa, D. Dworakowski, B. Firner, B.
Improvement,” arXiv preprint arXiv:1804.02767, 2018. Flepp, P. Goyal, D. L. Jackel, M. Monfort, U. Muller,
[CrossRef] J. Zhang, X. Zhang, J. Zhao, K. Zieba, ”End-to-End
[19] W. Liu et al., ”SSD: Single-shot multibox detector,” in Deep Learning for Self-Driving Cars,” arXiv preprint
European conference on computer vision, 2016, pp. 21- arXiv:1604.07316. [CrossRef]
37.[CrossRef] [34] SAE International, ”J3016 202104: Taxonomy and Defi-
[20] S. Karaman and E. Frazzoli, ”Sampling-based algorithms nitions for Terms Related to Driving Automation Systems
for optimal motion planning with applications to multi- for On-Road Motor Vehicles,” Revised on April, 2021.
robot and humanoid control,” The International Journal [CrossRef]
of Robotics Research, vol. 30, no. 7, pp. 846-884, 2011. [35] E. F. Camacho and C. Bordons, Model predictive control.
[CrossRef] Springer Science & Business Media Edition 2. ISBN
[21] R. E. Kavraki, J. J. Kuffner Jr, J.-C. Latombe, and M. N. 0857293982, 9780857293985, 2013.[CrossRef]
Raghavan, ”Deformable collision detection for robot nav- [36] P. Koopman and M. Wagner, ”Challenges in Autonomous
igation,” IEEE Transactions on Robotics and Automation, Vehicle Testing and Validation,” SAE International Journal
vol. 15, no. 3, pp. 436-443, 1999.[CroosRef] of Transportation Safety, vol. 4, no. 1, pp. 15-24, 2016.
[22] K. J. Åström and T. Hägglund, PID Controllers: Theory, [CrossRef]
Design, and Tuning. Research Triangle Park, North Car- [37] Neural Information Processing Systems (NeurIPS),
olina: ISA - The Instrumentation, Systems and Automation ”AI/ML publications (2014-2023),” Accessed on: 28th
Society, 1995. [CrossRef] January, 2024. [CrossRef]
[23] S. Tuli, F. Mirhakimi, S. Pallewatta, S. Zawad, G. Casale, [38] M. Brockschmidt et al., ”Generative Adversarial
B. Javadi, F. Yan, R. Buyya, and N. R. Jennings, ”AI Networks for Software Engineering: The Next
augmented Edge and Fog computing: Trends and chal- Step in Neural Program Synthesis,” arXiv preprint
lenges,” Journal of Network and Computer Applications, arXiv:1712.05159, 2017.
p. 103648, 2023.[CrossRef] [39] K. Bimbraw, ”Autonomous cars: Past, present and future
[24] S. Elkhediri, A. Benfradj, A. Thaljaoui, T. Moulahi, a review of the developments in the last century, the
and S. Zeadally, ”Integration of Artificial Intelligence present scenario and the expected future of autonomous
(AI) with sensor networks: Trends and future research vehicle technology,” in 2015 12th international confer-
opportunities,” Journal of King Saud University-Computer ence on informatics in control, automation and robotics
and Information Sciences, p. 101892, 2023. [CrossRef] (ICINCO), Jul. 2015, pp. 191-198.[CrossRef]
[25] Institute of Electrical and Electronics Engineers (IEEE)
Explore, ”AI/ML publications (2014-2023),” Accessed on
28th January, 2024. [CrossRef]
[26] Multidisciplinary Digital Publishing Institute (MDPI),
”AI/ML publications (2014-2023),” Accessed on 28th
January, 2024.[CrossRef]
[27] Society of Automotive Engineers (SAE) International,
”AI/ML publications (2014-2023),” Accessed on 28th
January, 2024. [CrossRef]
 13

Divya Garikapati is a Senior IEEE Member

and currently serving as the standards committee
member within the IEEE Intelligent Transportation
Systems Society (ITSS) and a peer reviewer for
IEEE Intelligent Transportation Systems Confer-
ences (ITSC). She actively participates in several
industry level standards discussions within IEEE and
SAE organizations. She is also the working group
chair for the IEEE Vehicular Technology Society
standards discussions. She received her Masters in
Electrical Engineering Systems from the University
of Michigan, Ann Arbor in 2014. Prior to that, she received her Bachelors in
Electronics and Communications Engineering from Andhra University College
of Engineering, Andhra Pradesh, India. Her current work focuses on Systems
and Safety research and development for Level 2,3 and 4 Autonomous
vehicles. She has over 10 years of experience in the automotive industry.

Sneha Sudhir Shetiya is a Senior IEEE Member

and received her Maters degree in Electrical Engi-
neering with a major in computer vision and Signal
Processing from North Carolina State University, in
2021. She received her Bachelor’s degree in Elec-
tronics and Communication Engineering from the
Visvesvaraya Technological University, Karnataka,
India, in 2014. Her work involves middleware topics
for embedded development of autonomous driving
stack, automotive diagnostics, systems engineering
and functional safety. She is an active volunteer
with IEEE region 4 and takes part in activities of Women In Engineering
(WIE) groups in the region. She is part of the committee for senior member
evaluation at IEEE for 2024 and has been a proctor for IEEExtreme 24 hour
coding competition.