Enabling Autonomous Vehicles: Gaps in Research and Education Infrastructure

Abstract

Digital Artificial Intelligence (AI), exemplified by Large Language Models (LLMs) such as ChatGPT, has achieved remarkable progress across a wide range of applications, driven not only by advances in algorithms but also by the emergence of a shared research ecosystem built upon commodity computing platforms, standardized software frameworks, open-source models, benchmark datasets, cloud infrastructure, and broadly accessible educational resources. In contrast, Autonomous Vehicles (AV), AI systems that perceive, reason, and act in the physical world, have advanced more slowly despite substantial public and private investment. Progress remains constrained by fragmented research and educational infrastructure that limits reproducibility, interoperability, scalable validation, and workforce development. This paper surveys the current state of the AV research ecosystem, including hardware platforms, autonomy software stacks, datasets, simulation environments, digital twins, testing and validation frameworks, and educational programs. Drawing lessons from the evolution of Digital AI, the paper identifies key gaps in accessibility, standardization, integration, and openness across the AV technology stack and outlines opportunities to develop shared research testbeds, modular open platforms, interoperable software and data ecosystems, common benchmarks, and interdisciplinary educational programs that can accelerate autonomous vehicle innovation. Finally, the paper provides a framework for evaluating AV research and educational infrastructure which identifies priorities for future investment.

1. Introduction

Digital AI Ecosystem:

Digital AI, especially in the form of Large Language Models (LLMs), has made significant progress in recent years [1]. The modern LLM research ecosystem is best understood as a layered, globally distributed infrastructure that combines open model hubs, datasets, training frameworks, evaluation systems, and scalable compute. As shown in Figure 1, at its center is Hugging Face [2], which functions as a coordination layer—hosting models, datasets, and tools. The result enables researchers worldwide to share, reproduce, and extend work. Around this hub are a rapidly expanding set of open-weight foundation models (e.g., from Meta [3], Alibaba [4], and Mistral AI [5]), along with large-scale datasets derived from sources like Common Crawl [6]. This ecosystem has fundamentally shifted LLM research from being concentrated in a few frontier labs to a broad, collaborative network, where universities, startups, and independent researchers can all participate in advancing state-of-the-art systems. Beneath this collaboration layer lies the technical backbone: frameworks such as PyTorch [7], DeepSpeed [8], and Megatron-LM [9] enable large-scale training and efficient fine-tuning, while LLMOps tools like LangChain [10] and LlamaIndex [11] support deployment, retrieval-augmented generation, and evaluation.

The education ecosystem around LLMs has developed almost as rapidly as the research stack itself, and it mirrors the same open, layered structure. At the foundation are openly accessible, hands-on courses led by platforms like Hugging Face and DeepLearning.AI [12]. The Hugging Face LLM course, for example, is a free, end-to-end curriculum that teaches not just theory but how to actually use models, fine-tune them, and deploy applications using real tools like Transformers and Datasets. In parallel, DeepLearning.AI and similar platforms offer a broad catalog of courses ranging from prompt engineering and RAG systems to full LLM training and evaluation, often with certificates and project-based learning. These resources have effectively created a global, self-service learning pathway, where individuals can move from basic concepts to building production-grade systems without needing access to a formal lab.

Structurally, Digital AI progress for Natural Language Processing (NLP) LLM-based solutions has been accelerated by an infrastructure with the following key characteristics (Figure 1):

• Open-source Technology: Hugging Face’s open-source approach makes NLP tools and libraries accessible to everyone, fostering a collaborative environment that accelerates advancements in AI.
• Transformers Library: The core of Hugging Face, offering thousands of pre-trained models that simplify the implementation and fine-tuning of NLP models without starting from scratch.
• Community Collaboration: A thriving ecosystem where users can share models, datasets, and code, enabling rapid innovation in AI and NLP.
• Model Hub: A central repository of pre-trained models for various NLP tasks, simplifying the process of finding and deploying suitable models.
• Training and Deployment: Tools for efficient training and deploying NLP models, with a user-friendly interface that makes model training accessible even to those with limited ML experience.
• Datasets Library: A vast collection of datasets for NLP tasks, ensuring users can easily find the data they need to train their models effectively.
• Educational Resources: Numerous tutorials, guides, and courses that help users of all levels understand and implement NLP models and techniques.

Autonomous Vehicle Research Issues:

In contrast to Digital AI, Physical AI encompasses a broad class of autonomous systems that can be characterized along three orthogonal dimensions. The first is physical embodiment, including fixed manipulators, wheeled and tracked ground vehicles, legged robots, humanoids, aerial vehicles, marine systems, spacecraft, and wearable robotic systems. The second is the operating environment, including ground, air, water, underwater, and space. The third is the operational or regulatory domain, including consumer, industrial, transportation, healthcare, agriculture, infrastructure, and defense applications, each operating under distinct safety, certification, and regulatory requirements. Autonomous Vehicles are a class of Physical AI characterized by an embodiment of wheeled ground vehicles, operating on public streets, and under the governance structure of traditional motor vehicles.

SAE research reports [13,14] outline the distinctive challenges of AVs and link them directly to these three dimensions. Because AVs are physically embodied AI systems rather than a purely digital application, safe operation depends on the reliable interaction of computing hardware, sensing technologies, communications, and vehicle actuation under real-world operating conditions. Unlike traditional software systems, AV performance is directly influenced by sensor limitations, actuator dynamics, environmental variability, and hardware degradation. Cameras, LiDAR, radar, GNSS, inertial sensors, and V2X communications must continue to operate despite adverse weather, changing lighting conditions, electromagnetic interference (EMI), hardware aging, and partial failures. Significant open research issues remain in topics such as electromagnetic interference (EMI) [15], sensor degradation in weather [16], and infrastructure dependencies [17]. In terms of operating environment (ground physics and human interaction), driving is not merely a vehicle control problem but a dynamic interaction with a highly uncertain physical and social world populated by human drivers, pedestrians, cyclists, animals, infrastructure, and continuously changing environmental conditions. Critical research challenges remain on topics such as defining clearly the Operational Design Domain (ODD) for the purpose of Validation and Verification [18,19,20]. Especially challenging are issues of External Human Machine Interface (eHMI) which require humans (ego driver, other drivers, pedestrians) and AVs to coordinate in a shared space [21,22,23,24,25]. Finally, the operational and regulatory environment envelope for AVs starts with the traditional passenger car. However, unlike conventional vehicles, AI-enabled Software Defined Vehicles (SDVs) evolve continuously through over-the-air software updates, making one-time certification increasingly inadequate and shifting emphasis toward continuous validation throughout the vehicle lifecycle. Furthermore, software-defined fleets introduce correlated systemic risk, where a single software defect can simultaneously affect thousands or even millions of vehicles, creating new challenges for large-scale fault detection, staged deployment, rollback strategies, and post-deployment safety assurance. Developing regulatory frameworks, safety standards, and operational governance capable of managing continuously evolving AV systems therefore remains one of the central research challenges for autonomous transportation.

To address these research challenges, there is a need for shared, collaborative infrastructure to enable and accelerate progress. While the structure of this flow is different from the Digital AI ecosystem, the key learnings which can be leveraged are open components that allow for quick customization, significant core capabilities upon which one can build new capability, the ability to test updates, and finally, a central place for publishing results for community reuse.

Autonomous Vehicle Research Ecosystem:

The structure of an operational AV system functionality is commonly organized into a sequence of tightly coupled capabilities. The sensing layer acquires raw observations from cameras, radar, LiDAR, sonar, inertial measurement units, GNSS receivers, and environmental sensors. The perception layer converts these observations into semantic understanding of objects, terrain, obstacles, and traffic participants. Localization determines the vehicle’s position and orientation using techniques such as GNSS, SLAM, visual odometry, and inertial navigation. Planning and decision-making select safe and efficient trajectories while satisfying mission objectives and operational constraints. Finally, the control layer executes these decisions through motors, steering systems, propulsion, or other actuators, completing the closed control loop with the physical environment. Underlying this architecture are significant components for managing conventional automotive functions, such as vehicle dynamics, tire-road interaction, and suspension. In order to evaluate the value of innovation anywhere in the system, one needs to measure effectiveness at the system level.

Figure 2 illustrates this closed-loop infrastructure topology, highlighting the interactions among operational autonomy software, world-model generation, validation and assurance mechanisms, human supervision, regulatory constraints, and the physical environment throughout the system lifecycle.

AV Research Infrastructure

Given this structure, system-level research and development in autonomous vehicles depend on a layered infrastructure ecosystem composed of four tightly integrated components. At its foundation is the hardware infrastructure, comprising heterogeneous computing platforms, sensor suites, communication networks, and real-time control hardware that provide the computational and physical interface to the environment. Built upon this foundation is the autonomy software stack, which supplies reusable frameworks for perception, localization, mapping, planning, decision-making, and vehicle control. The next layer consists of the data and world-modeling infrastructure, where large-scale collections of real-world operational data are combined with synthetic data generation to capture rare and hazardous scenarios, enabling the construction of high-definition maps, simulation environments, digital twins, and continuously evolving world models that reflect changes in the operating environment. Finally, the testing, validation, and assurance infrastructure provides the mechanisms required to establish confidence in autonomous behavior before deployment. This is done by integrating simulation, digital twins, hardware-in-the-loop testing, physical proving grounds, automated scenario generation, formal verification where applicable, and safety assurance methodologies.

Section 2 of this paper provides a structured ecosystem methodology to evaluate the research and educational infrastructure supporting autonomous vehicles (AVs) along each of these dimensions. Section 3 examines the current educational infrastructure for autonomous vehicles, including university programs, open educational resources, training platforms, and workforce development initiatives, and assesses their strengths and limitations. Section 4 identifies the major structural gaps in both the AV Ecosystem and the educational infrastructure that currently limit broader participation, reproducibility, and technology transfer. Finally, Section 5 discusses future opportunities for building a more open, interoperable, and sustainable ecosystem capable of accelerating research, education, commercialization, and safe deployment of AV systems.

2. AV Research Infrastructure

To evaluate the current state, there is a need to define the scope and a rubric for evaluation. For the scope, rather than surveying every commercially available technology, the focus is on reusable research infrastructure that enables broader participation, reproducibility, technology transfer, and collaborative innovation. Specifically, the survey examines infrastructure components—including hardware platforms, autonomy software stacks, datasets, world-model generation tools, simulation environments, validation frameworks, standards, and educational resources—that are accessible to researchers outside the originating organization and can be reused, extended, or integrated into new research projects. Information was collected from peer-reviewed literature, standards organizations, vendor technical documentation, open-source repositories, university research laboratories, and publicly available educational resources. Platforms were included based on their demonstrated adoption within the research community, availability to external researchers, openness of interfaces or software, support for interoperability with broader open-source ecosystems, and potential for reuse across multiple institutions. Purely proprietary internal systems, research prototypes that are not publicly accessible, and commercial technologies lacking meaningful mechanisms for external research use were generally excluded. Comparative evaluations were then performed using qualitative rubrics assessing openness, open-source ecosystem support, and research suitability. The objective is not to rank individual technologies, but to identify the shared infrastructure that has the greatest potential to accelerate collaborative research and education in autonomous vehicles. In terms of evaluation, key ideas are openness (hardware and software), ecosystem support, and standardization.

2.1. Hardware Research Infrastructure

One of the most significant differences between Digital AI and autonomous vehicle (AV) research lies in the underlying hardware environment. Digital AI research benefits from a highly standardized computing ecosystem centered on commodity servers and workstations running widely adopted operating systems such as Linux and Windows, with AI workloads accelerated by general-purpose GPUs and mature software frameworks. Researchers can typically reproduce experiments using commercially available CPUs, GPUs, standardized drivers, cloud computing platforms, and well-supported development environments. In contrast, AV hardware devices are embedded edge devices and are typically offered as development boards with associated board support packages.

Table 1. Comparison of automotive hardware AI platforms by openness and ecosystem support.

Platform	Vendor	Type	Typical Form Factor	Openness—HW/SW	Open-Source Ecosystem Support
DRIVE AGX (Orin/Thor) [26,27]	NVIDIA	Full AV compute platform	In-vehicle computer + dev kits	Medium, closed HW, semi-open SW	Strong, CUDA, ROS, Omniverse, open Models
Jetson (Orin/Xavier) [28,29]	NVIDIA	Edge AI/robotics	Dev kits (small form factor)	Medium-High	Very strong, ROS/ROS2, Isaac, PyTorch
Snapdragon Ride [30]	Qualcomm	Automotive SoC platform	ECU/dev platforms	Medium	Moderate, Linux, some ROS integration
EyeQ (Mobileye) [31]	Mobileye (Intel)	Integrated ADAS SoC	Embedded automotive	Low	Limited, closed stack, APIs only
TDA4/Jacinto [32]	Texas Instruments	Embedded ADAS processor	EVM boards	Medium-High	Good, TI SDKs, Linux, ROS support
S32/BlueBox [33]	NXP	Automotive compute platform	Dev boards/ECUs	Medium-High	Good, AUTOSAR, Linux, ROS bridges
R-Car [34]	Renesas	Automotive SoC	Starter kits	Medium	Moderate, Linux, limited AI frameworks
Cvflow [35]	Ambarella	Vision AI SoC	Dev kits	Medium	Moderate, vision pipelines, limited ROS

below outlines the major platforms with an evaluation around two topics: the openness of the hardware and the open-source ecosystem support of the associated software ecosystem. The rubric for hardware openness is:

• Very High: Open hardware specifications (or open ISA), open firmware, open drivers, open SDKs, open APIs, permissive licensing, unrestricted development.
• High: Proprietary hardware but open SDKs, Linux support, documented APIs, third-party toolchains, minimal restrictions.
• Medium: Proprietary hardware with documented SDKs and APIs; some proprietary compilers, firmware, or runtime libraries required.
• Low: Mostly proprietary hardware and software; limited APIs; restricted SDKs; NDA components common.
• Very Low: Closed hardware and software stack; little documentation; vendor-controlled development environment.

The rubric for the open-source ecosystem support is:

• Very Strong: Native support for ROS2, PyTorch, TensorFlow, ONNX, Docker, Kubernetes, OpenCV, simulation frameworks, active GitHub community, extensive tutorials.
• Strong: Supports most major AI frameworks and robotics middleware with active community examples.
• Moderate: Supports common Linux AI frameworks and some ROS integration, but ecosystem is incomplete.
• Limited: Limited framework support; proprietary SDK dominates development.
• Minimal: Little meaningful integration with open-source software; primarily vendor tools.

For hardware platforms, a fundamental tension in AV Research infrastructure is that platforms optimized for commercial deployment often sacrifice transparency and extensibility, while research-oriented ecosystems derive much of their value from interoperability with open-source middleware, simulation environments, and shared tooling. Platforms such as NVIDIA Jetson and DRIVE have become dominant in academia because they combine high-performance compute with broad support for Linux, ROS/ROS2, PyTorch, and simulation frameworks, enabling rapid prototyping, reproducibility, and cross-institutional collaboration. Their ecosystems extend beyond hardware to include reusable models, datasets, and development tools, lowering the barrier to entry for complex autonomy research. However, the hardware and the lower levels of software are closed. This limits innovation in these domains.

In contrast, platforms like Qualcomm Snapdragon Ride and NXP S32 offer strong capabilities and increasing openness through Linux-based SDKs and modular architectures, making them suitable for applied and systems-level research, particularly where networking, real-time constraints, or automotive integration are important, but do not provide comparable integration with open AV software ecosystems. At the other end of the spectrum, Mobileye EyeQ represents a highly optimized, production-grade solution with limited transparency into internal algorithms and restricted software extensibility, which constrains its usefulness for exploratory or publishable research despite its commercial success. Embedded platforms such as TI Jacinto strike a middle ground, offering relatively open development environments with strong real-time and safety features, making them attractive for research into embedded autonomy and edge constraints, but limited support for open AV stacks.

Overall, while many platforms exist, each has its own ecosystem with little standardization of critical interfaces to allow for innovation in the hardware space. In Digital AI, a hardware vendor can develop a solution, run it on existing benchmarks, and demonstrate progress. For AV hardware innovators, this task is very difficult because of the lack of a set of open benchmarks or standards which connect to autonomy software stacks.

While hardware platforms represent the foundation of the broader autonomy ecosystem, the ability to build reproducible and extensible AV systems also depends heavily on the openness and modularity of the software autonomy stack—including perception, localization, planning, control, and simulation interfaces.

Table 2. Comparison of autonomous driving software stacks based on openness, ecosystem support, and research suitability.

Platform	Type	Stack Coverage	Openness	Open-Source Ecosystem	Key Strengths	Research Suitability
NVIDIA DRIVE [36]	Commercial, with open components	Full stack, perception, planning, control + simulation	Medium, closed core, open APIs/tools	Strong, CUDA, ROS, Omniverse, AI frameworks	Tight HW/SW integration, high performance, production-grade	strong but semi-closed
Autoware [37]	Fully open-source	Full stack (LiDAR-first perception, localization, planning, control)	High	Very strong, ROS/ROS2-native	Modular, academic-friendly, transparent algorithms	top academic platform
Apollo [38]	Open-source, with enterprise extensions	Full stack + HD maps + cloud integration	Medium-High	Strong, ROS-like middleware, tools	Scalable architecture, strong perception + mapping	research + industry bridge
openpilot [39]	Open-source, ADAS-focused	Partial stack, driver assistance: perception + control	High	Moderate, Python/C++, community-driven	Real-world deployment, lightweight, low-cost	good for applied/edge research

therefore compares major autonomous driving software platforms in terms of stack coverage, openness, ecosystem integration, and research suitability. The rubric for software openness is:

• Very High: Fully open-source platform with permissive licensing, public development, complete source code, and unrestricted modification and redistribution.
• High: Mostly open-source platform with minor proprietary components, well-documented APIs, and active community contributions.
• Medium: Platform provides public SDKs and APIs but retains significant proprietary components or licensing restrictions.
• Low: Platform is primarily proprietary with restricted APIs, documentation, or development tools that limit external contributions.
• Very Low: Platform is almost entirely closed, providing little or no access to source code, development tools, or internal interfaces.

The rubric for ecosystem is:

• Very Strong: Natively supports major open-source frameworks and middleware with a large, active developer community and extensive third-party resources.
• Strong: Supports most widely adopted open-source tools with mature documentation and an active ecosystem of users and contributors.
• Moderate: Provides compatibility with common open-source frameworks but requires additional integration effort or has limited community support.
• Low: Supports only a small subset of open-source tools and relies primarily on vendor-specific software and development environments.
• Very Low: Offers little meaningful integration with the broader open-source software ecosystem.

As with hardware platforms, AV software stacks include both commercial and open-source environments. Many closed commercial systems, including those from Waymo, Tesla, and others, are not available to the research community. Among accessible platforms, NVIDIA provides a full stack tied closely to its hardware, supported by open APIs and development tools. Researchers working within the NVIDIA ecosystem can innovate at higher levels of system functionality, but lower-level components remain largely closed. Autoware and openpilot provide more traditional open-source alternatives, while Apollo offers open-source components alongside more advanced capabilities in its commercial versions.

While Table 2 compares autonomy platforms from the perspective of openness, ecosystem integration, and research suitability, these characteristics alone do not fully capture the functional scope of each stack. In practice, AV platforms differ not only in architectural philosophy and ecosystem maturity, but also in the extent to which they implement the core autonomy pipeline—including perception, localization, planning, control, simulation integration, and cloud-connected data workflows.

Table 3. Functional comparison of major autonomous driving platforms across main functionalities.

Function	NVIDIA DRIVE	Autoware	Apollo	Openpilot
Perception	Yes, multi-sensor AI	Yes, LiDAR-heavy	Yes, multi-modal	Partly, camera-centric
Localization	Yes	Yes, GNSS + LiDAR SLAM	Yes, HD maps + GNSS	Limited
Planning	Yes	Yes	Yes	Partly, lane-following focus
Control	Yes	Yes	Yes	Yes
Simulation Integration	Yes, omniverse	Yes, CARLA, LGSVL	Yes, Cyber RT + sim tools	Limited
Cloud/Data Loop	Strong	Limited	Yes, strong	Minimal

therefore provides a functional comparison of major autonomous driving platforms coverage across the primary operational capabilities required for closed-loop autonomous systems.

Each of the software stacks offers divergent capabilities in terms of the basic AV capabilities. NVIDIA offers a complete flow within a proprietary structure. This extends to the “cloud/data loop” which connects field data with the AV stack development. Open-source platforms such as Autoware and openpilot generally do not have the completeness of the commercial platforms but are particularly well-suited for academic research because they expose the complete autonomy pipeline—perception, localization, planning, and control—allowing researchers to modify algorithms, insert experimental modules, and reproduce results across institutions. Their integration with widely adopted middleware (especially ROS/ROS2) and compatibility with open simulation environments (discussed in Section 2.3) enables rapid experimentation and “crowd-sharing” of improvements.

2.2. Data and World-Model Infrastructure

While hardware and software autonomy stacks form the operational core of AV systems, their effectiveness increasingly depends on the surrounding data infrastructure used for training, benchmarking, simulation, and validation. Unlike many Digital AI domains, AV systems require large-scale multimodal datasets that capture not only nominal operational conditions, but also rare events, environmental variability, and safety-critical edge cases. In parallel, advances in digital twins and synthetic environment generation are transforming how physical environments are represented, simulated, and replayed throughout the development lifecycle.

As a result, datasets, world models, and digital reconstruction pipelines have become foundational infrastructure layers within modern AV ecosystems.

Table 4. Comparison of autonomous driving datasets by sensor modalities, scale, openness, and primary research applications.

Dataset	Organization	Sensor Modalities	Scale/Focus	Openness	Key Use Cases
nuScenes [40]	Aptiv	Camera, LiDAR, radar	1M+ frames, urban driving	High	Perception, tracking, sensor fusion
Waymo Open Dataset [41]	Waymo	Camera, LiDAR	Very large-scale, diverse environments	High	End-to-end autonomy, behavior
KITTI [42]	Karlsruhe Institute of Technology	Camera, LiDAR	Benchmark dataset	High	SLAM, detection, stereo vision
Argoverse 2 [43]	Argo AI	Camera, LiDAR, HD maps	Urban + motion forecasting	High	Prediction, mapping
BDD100K [44]	UC Berkeley	Camera (video)	100K videos, diverse conditions	High	Vision, weather/edge cases
ApolloScape [45]	Baidu	Camera, LiDAR	Large annotated dataset	High	Segmentation, 3D understanding
Safety Pool [46]	University of Warwick	Scenario-based (derived from real + synthetic data)	Edge cases, safety-critical scenarios	Medium-High (research access)	Scenario testing, safety validation, coverage analysis
ISEAUTO [47]	TalTech	Camera, LiDAR from Shuttle	Training	High	Shuttle Use Case

compares currently available autonomous driving datasets in terms of sensor modalities, scale, openness (same rubric as above for software openness), and primary research applications.

The dataset landscape for ground autonomy spans two complementary roles: large-scale, open real-world datasets for training and benchmarking, and scenario-centric datasets for safety validation. Widely used resources such as nuScenes, Waymo Open Dataset, KITTI, Argoverse 2, BDD100K, and ApolloScape provide richly annotated multi-sensor data (camera, LiDAR, radar) that enable development and comparison of perception, localization, and prediction algorithms across diverse driving conditions. These datasets are highly open and have become the foundation for reproducible research. However, they are primarily designed around component development and capturing typical driving distributions. For safety, systematically covering rare or hazardous edge cases is very important. This gap is addressed by emerging resources, such as Safety Pool from the University of Warwick, that focus on curated, safety-critical scenarios derived from real and synthetic data. Overall, the current landscape can be characterized as a collection of niches aimed at component training in specific situations, yet missing broad system benchmarks with performance metrics.

2.3. Simulation, Testing, and Validation Infrastructure

For AV, real-world testing alone is insufficient because it is expensive, difficult to scale, and incapable of systematically covering long-tail edge cases. As a result, modern AV development increasingly depends on integrated validation infrastructures that combine simulation, scenario generation, hardware-in-the-loop testing, synthetic environments, and formal verification methods. These validation layers serve complementary roles within the broader AV ecosystem. Simulation platforms enable scalable closed-loop experimentation, scenario-generation frameworks improve coverage of rare events and adversarial situations, while formal methods and hardware-in-the-loop systems support safety assurance under tightly constrained operational conditions. Together, these infrastructures form the validation backbone of AV systems and are central to enabling reproducible, scalable, and continuously evolving autonomy workflows.

Table 5. Overview of major simulation, scenario generation, and validation platforms for autonomous systems, including organization, type, openness, core capabilities, stack integration, and research roles.

Platform	Organization	Type	Openness	Core Capability	Integration with Stacks	Research Role
CARLA [48]	Intel Labs + University of Barcelona	Open simulator	High	Photorealistic simulation, sensor modeling	Strong, Autoware, ROS, custom stacks	Benchmarking, closed-loop testing
NVIDIA DRIVE Sim/Omniverse [49]	NVIDIA	Digital twin/ simulation	Medium	High-fidelity rendering, synthetic data	Very strong, NVIDIA DRIVE, ROS bridges	Scalable Synthetic testing
SUMO (Simulation of Urban Mobility) [50]	DLR	Traffic simulation	High	Large-scale traffic flow modeling	Moderate, via APIs, ROS bridges	Scenario generation, traffic behavior
PreScan [51]	Siemens	Commercial simulation	Low	Sensor modeling, ADAS validation	Strong, MATLAB/Simulink	Industrial validation
dSPACE ASM/VEOS [52]	dSPACE	HIL/SIL simulation	Low	Real-time simulation + hardware-in-loop	Strong, AUTOSAR, ECUs	Safety validation, embedded testing
IPG CarMaker [53]	IPG Automotive	Vehicle simulation	Low	High-fidelity vehicle dynamics	Strong, industry toolchains	Vehicle-level validation
Scenic (with simulators) [54]	UC Berkeley	Scenario generation	High	Probabilistic Scenario specification	Strong, CARLA, others	Test generation, edge-case discovery
Open SCENARIO [55]	ASAM	Scenario standards	High	Standardized scenario + map definitions	Broad, industry + simulators	Interoperability, reproducibility
PolyVerif [56]	Florida Polytechnic University + TalTech + Embry-Riddle Aeronautical University	Formal verification/testing	Medium	Coverage-driven testing, formal validation	Moderate, integrates with sim + logs	Safety assurance, edge-case validation

summarizes representative platforms and standards across the simulation, scenario-generation, and validation ecosystem. The rubrics for the platform software are for software openness, and the platform integration stack for ecosystem openness.

Open platforms such as CARLA and SUMO provide accessible, extensible environments for perception, planning, and large-scale interaction studies, forming the backbone of academic experimentation. Higher-fidelity platforms such as NVIDIA Omniverse, PreScan, dSPACE, and IPG CarMaker enable sensor-accurate and hardware-in-the-loop validation, though typically within more closed industrial ecosystems. Complementing these are scenario-generation frameworks like Scenic, which allow systematic construction and sharing of test cases across tools. In addition, the work of ASAM (Association for Standardization of Automation and Measuring Systems) [57] has become foundational to ground autonomy by providing a common language and data model for describing roads, scenarios, and testing workflows. Through standards such as OpenDRIVE [58] (road network geometry), OpenSCENARIO (dynamic driving situations and behaviors), and OpenCRG [59] (road surface conditions), ASAM enables interoperability across simulation tools, autonomy stacks, and validation pipelines.

For AV research, this standardization is particularly important because it transforms testing from an ad hoc process into a structured, shareable, and increasingly automated discipline, supporting the broader goal of ‘crowd-sourced’ validation and safety assurance in complex, human-centric driving environments. However, many gaps exist in the current situation. As an example, not all simulators support all the standards for test generation, and this same situation exists between simulators, AV Stacks, and underlying hardware. Finding a coherent stable system across all these dimensions is very difficult today.

2.4. Digital Environment Reconstruction and World Models

Finally, a critical challenge in the use of simulation is the data required to build a fully rendered digital twin which leads to a need for tools for building world models. These world-building technologies bridge the gap between operational data and simulation infrastructure, enabling researchers to construct scalable, interactive, and increasingly photorealistic representations of physical environments.

Table 6. Comparison of major 3D reconstruction and digital twin generation methods for autonomous systems, including inputs, techniques, outputs, tools, openness, and research applications.

Method	Input Data	Core Technique	Output Representation	Key Tools/Platforms	Openness	Research Role
Photogrammetry (SfM/MVS) [60]	Camera (RGB images, video)	Structure-from-Motion + Multi-View Stereo	Dense 3D point clouds/meshes	COLMAP, OpenMVG, Meshroom	High	Urban reconstruction, mapping
LiDAR Mapping (SLAM/HD Maps) [61]	LiDAR + IMU (+GNSS)	LiDAR SLAM, scan matching	Point clouds, HD maps (lanes, objects)	LOAM, Cartographer, Autoware	High	Localization, map building
Sensor Fusion Mapping [62]	Camera + LiDAR + GNSS	Multi-modal fusion (deep learning + geometric methods)	Semantic HD maps, labeled environments	Apollo, Autoware, NVIDIA DRIVE	Partly	High-fidelity autonomy maps
Satellite-Based Reconstruction [63]	Satellite Imagery	Remote sensing + photogrammetry + GIS	Large-scale terrain + road networks	Google Earth Engine, QGIS	Medium-High	Regional/ global context
Neural Radiance Fields (NeRF) [64]	Camera (multi-view images)	Neural implicit scene representation	Continuous 3D scene (view synthesis)	Instant-NGP, Nerfstudio	High	High-fidelity rendering, sim
Gaussian Splatting [65]	Camera (+optional depth/LiDAR)	Real-time neural rendering	Dense, photorealistic 3D scenes	3D Gaussian Splatting frameworks	High	Real-time digital twins
Procedural World Generation [66]	Maps + rules (OpenDRIVE, GIS)	Rule-based synthesis	Structured road networks, traffic scenes	CARLA, SUMO, OpenDRIVE tools	High	Scenario generation
Hybrid Digital Twin (Data + Simulation) [67]	Real data + synthetic Augmentation	Fusion of real-world capture + simulation engines	Interactive, physics-enabled environments	NVIDIA Omniverse, CARLA, Unreal Engine	Partly	Closed-loop testing, scaling

summarizes major approaches for digital environment reconstruction and world-model generation used in AV research with the openness evaluation based on software rubric.

Digital twin construction for ground autonomy is fundamentally a multi-modal world-building problem, where different data sources contribute complementary strengths. Camera-based methods (photogrammetry, NeRF, Gaussian splatting) provide high visual fidelity and dense reconstruction, making them well-suited for perception and rendering tasks, while LiDAR-based methods deliver precise geometry and are critical for localization and mapping in autonomy stacks. Satellite data extends this capability to large-scale environments, enabling global context and infrastructure-level modeling, though at lower resolution. Increasingly, the state of the art lies in hybrid approaches that fuse these modalities into semantically rich, interactive digital twins that can be used for simulation and testing. From a research perspective, open-source tools in photogrammetry, LiDAR SLAM, and neural rendering have significantly democratized world-building, but fully integrated digital twin platforms (e.g., Omniverse) remain partially closed. This creates an important distinction: open pipelines enable experimentation and reproducibility, while integrated platforms enable scale and realism—highlighting a key trade-off in AV research infrastructure today.

3. Education Infrastructure for AV

The infrastructure challenges discussed above have important implications for pedagogy and workforce development. Unlike many Digital AI domains, where education can rely mainly on cloud tools, datasets, and software frameworks, AV education requires hands-on experience with sensing systems, embedded computing, simulation environments, validation workflows, and real-world operational constraints. Effective AV education therefore depends on integrated ecosystems that combine theory, software, hardware, simulation, governance, and system-level experimentation. Today, three broad educational platforms exist to address this need: MIT Beaver Works [68], GRASP Lab F1/10 [69], and Duckietown [70]. Their key characteristics are shown in

Table 7. AV Pedagogical Platforms.

Platform	Hardware (CPU & Sensor Support)	AV Software Platform	Simulation	Data/World Infrastructure	Open Educational Resources
MIT Beaver Works (RACECAR)	NVIDIA Jetson Nano (earlier) or Raspberry Pi 4 (Neo); RGB camera, 2D LiDAR, IMU, wheel encoders	Custom ROS/ROS 2 educational stack; limited support for industrial AV platforms	Unity RACECAR Simulator; limited Gazebo support	Student-generated maps and ROS bags; no direct integration with KITTI, nuScenes, or Waymo Open Dataset	Limited—RACECAR software and documentation are open-source, but the complete Beaver Works curriculum is not fully open.
F1TENTH (GRASP)	NVIDIA Jetson Nano/Xavier/Orin; 2D LiDAR, RGB camera, IMU, wheel encoders	Modular ROS/ROS 2 research stack; supports Autoware and custom research frameworks	F1TENTH Gym, Gazebo, RViz	ROS bags, SLAM maps, racetrack maps; compatible with robotics datasets but not directly coupled to major AV benchmark datasets	Strong—Hardware, software, labs, tutorials, and course materials are openly available.
Duckietown	Raspberry Pi; monocular camera, wheel encoders (optional IMU/ToF sensors)	Custom ROS/ROS 2 educational stack focused on Duckietown ecosystem	Duckietown Simulator (Gym-compatible)	Synthetic Duckietown environments, lane maps, AprilTag infrastructure, educational datasets	Strong—Software, documentation, labs, and educational resources are openly available.

below.

The MIT Beaver Works Summer Institute (BWSI) represents one of the most comprehensive system-level learning platforms for autonomy which is built around the RACECAR and related robotic systems. Its strength lies in exposing students to the full autonomy pipeline—from sensing and perception through planning and control—on real hardware, complemented by simulation environments. By combining structured curriculum, team-based development, and hands-on integration. The University of Pennsylvania’s GRASP Lab ecosystem combines advanced research with hands-on experimentation, particularly through the F1/10 autonomous racing platform. This platform allows students to explore autonomy in a high-performance, real-time setting, where perception, planning, and control must operate under tight constraints. Finally, Duckietownprovides a scalable, low-cost, and fully open-source learning platform designed to make autonomy accessible to a broad audience. For all three, the ROS (Robot Operating System) and Gazebo ecosystem [71,72] functions as the foundational software infrastructure underlying much of robotics and autonomy education. Critical challenges for all three platforms are that they are largely small closed systems. Except for GRASP support for Autoware, they are built as small teaching systems. This contrasts with Digital AI where the educational resources are easily available for the major research infrastructure platforms. In addition, while components may be open-source, the structure is a distribution-only model. With the complexity of AV, an underlying true open-source collaboration model is required for educational materials.

To address these challenges, in Europe, recent educational funding initiatives increasingly emphasize interoperability, systems engineering, and cross-institutional collaboration through distributed engineering ecosystems such as the EuroTeQ Engineering University alliance [73], which connects major European technical universities around challenge-based engineering education, CPS, robotics, and digital infrastructure collaboration, and proving-ground infrastructures such as ZalaZONE [74]. The ISEAUTO project [75] at TalTech represents one example of such an ecosystem-oriented approach, combining open-source autonomy software, commodity sensing hardware, GPU-based compute platforms, and real-world autonomous shuttle deployment into a unified educational and research environment. In Asia, AV education is often more tightly integrated with industrial autonomy ecosystems and intelligent infrastructure initiatives, including the Baidu Apollo ecosystem in China.

4. Gaps and Challenges

The previous sections showed that progress has already been made in building the technical and educational foundations of AV. Open-source autonomy stacks, datasets, simulation environments, validation frameworks, and hands-on educational ecosystems are increasingly available to researchers and students worldwide. Despite significant progress in autonomous vehicle research, the survey reveals several fundamental structural gaps that continue to limit the pace of innovation.

First, openness remains uneven across the ecosystem. Although many software components have become available through open-source initiatives, critical portions of the technology stack—particularly at the hardware, firmware, and low-level systems layers—remain proprietary. Even within open software platforms, robust supporting ecosystems are often fragmented or immature, limiting interoperability and broader community participation. A second challenge is functional fragmentation. As demonstrated throughout the survey, numerous high-quality solutions exist for individual components such as hardware platforms, autonomy software stacks, datasets, simulation environments, and validation tools. However, these components have largely evolved independently and are rarely designed to operate together as an integrated ecosystem. Unlike the Digital AI community, which benefits from coordinating platforms such as Hugging Face that seamlessly connect models, datasets, tools, benchmarks, and educational resources, the AV community lacks a comparable integration layer that simplifies experimentation, promotes reuse, and accelerates collaborative innovation.

A third gap is the absence of robust benchmarks and evaluation methodologies. One of the defining characteristics of Digital AI has been the availability of widely accepted benchmark datasets, standardized evaluation metrics, and openly published performance results, allowing researchers to compare competing approaches on a common basis. In autonomous vehicles, however, comparable system-level benchmarks remain largely unavailable. Even the definition of success varies considerably across organizations, making objective comparison and reproducible evaluation difficult. Closely related is the lack of standardized data abstractions and simulation fidelity metrics. Effective evaluation of autonomous systems requires a clear understanding of the abstraction level represented by simulation models and digital twins. At present, simulation platforms employ different assumptions regarding sensor fidelity, environmental realism, traffic behavior, and physical modeling, yet these abstraction levels are rarely formalized or standardized. As a result, it is difficult to compare results across simulation environments or to determine whether conclusions obtained from one platform generalize to another [76].

Finally, access to physical testing infrastructure remains a significant barrier. Real-world testing requires substantial financial investment, specialized facilities, regulatory approvals, and comprehensive safety procedures. These costs are particularly burdensome for research involving sensing technologies, communications, and telematics, where meaningful evaluation often requires large-scale operational deployments. Consequently, many researchers are unable to validate promising innovations under realistic operating conditions.

As a consequence of this fragmentation, it remains difficult for a hardware innovator to easily test effectiveness against a standard set of AV system benchmarks, an expert in perception to show effectiveness in interesting edge conditions, an HMI expert to test effectiveness of protocols against AV algorithms, or even a physics-informed neural network for path planning to test against conventional safety standards. Overall, no existing platform simultaneously provides open hardware, an open autonomy stack, standardized datasets, interoperable validation pipelines, and educational resources. This fragmentation is the central structural limitation of today’s AV ecosystem.

The survey also identifies several important shortcomings within the current educational ecosystem for autonomous vehicles. First, there is a need for a more collaborative open educational community. While many universities and organizations publish course materials, most educational resources follow a traditional “develop and distribute” model rather than the collaborative development model that has proven successful for open-source software. Given the complexity and rapid evolution of AV, educational content will likely need to evolve through larger community-driven collaborations that remain tightly coupled to the underlying hardware and software ecosystems. Second, existing curricula place relatively little emphasis on the broader discipline of systems engineering for safety-critical autonomous systems. Current programs generally excel at teaching perception, planning, control, and machine learning, but provide comparatively limited coverage of governance, regulatory frameworks, industry standards, safety assurance, system validation, certification, and lifecycle engineering. These topics are increasingly essential as autonomous systems transition from research prototypes to deployable commercial products. Finally, Human Machine Interaction (HMI) remains significantly underrepresented within current educational programs despite its central importance to the safe deployment of autonomous systems. Subjects such as human factors, trust, explainability, shared autonomy, operator interaction, and the validation of human-AI interfaces receive relatively little attention. As a result, graduates may possess strong technical skills while lacking sufficient preparation to design autonomous systems that interact safely, effectively, and transparently with human users. Overall, to successfully research and provide solutions in the AV space, these three aspects need to be integrated into the pedagogical frameworks. Early attempts have been funded by the European Union with projects such as SafeAV [77].

5. Opportunities and Recommendations

The gaps identified in the previous section suggest that advancing AV requires more than incremental improvements in algorithms or isolated research platforms. Progress increasingly depends on coordinated investment in shared infrastructure, interoperable ecosystems, validation methodologies, and interdisciplinary workforce development. The following recommendations outline several opportunities through which academia, standards organizations, industry, and public–private partnerships can help accelerate scalable and reproducible AV research and education. Traditionally, the drivers of these efforts come from three sources—government, industrial consortiums, and private industry.

• Open, Modular “Academic Autonomy Stack”: A key opportunity in AV research infrastructure is the development of a standardized, open, end-to-end autonomy stack that integrates hardware, software, and validation tools into a cohesive framework. Today’s research efforts are often fragmented, requiring significant integration work across sensors, compute platforms, middleware, and autonomy algorithms. An open, modular stack—built on widely adopted components such as ROS2, Autoware, and standardized sensor/compute interfaces—would enable interoperability and reuse across institutions. Much like open frameworks transformed Digital AI, such a stack would provide a common baseline for experimentation, benchmarking, and collaboration, allowing researchers to focus on innovation rather than system assembly. A key aspect would be to build an easily executable set of AV benchmarks. Similar to Digital AI, private industry, typically through the formation of a non-profit, would be the natural place to build such a stack.
• Shared AV Testbeds (Real + Hybrid): Another major opportunity lies in the creation of shared, remotely accessible AV testbeds that combine real-world systems with digital twins. These testbeds could include instrumented campuses, controlled urban environments, or robotic labs where researchers can deploy and evaluate autonomy algorithms under consistent conditions. By integrating real vehicles or robots with simulation overlays, hybrid environments enable both realistic validation and scalable experimentation. Importantly, shared access would democratize the use of expensive infrastructure, allowing smaller institutions and global collaborators to participate in cutting-edge research while ensuring repeatability and comparability of results. Example investments include ZalaZone where large capital investments can be made by government. This contrasts with each university getting a small car or using local resources.
• Open Scenario & Edge-Case Databases: Advancing safety in AV requires moving beyond traditional datasets toward structured scenario databases that capture rare and safety-critical events. While existing datasets support perception and prediction, they often lack systematic coverage of edge cases such as unusual human behavior, emergency interventions, or complex multi-agent interactions. Open, standardized scenario repositories—building on efforts like OpenSCENARIO and Safety Pool—would allow researchers to define, share, and reuse challenging test cases. This would enable coverage-driven validation, improve benchmarking of safety performance, and support a more rigorous, community-driven approach to understanding long-tail risks. The nature of edge-case data is that this ability needs to be available to the broader research community and even the general public.
• Integrated Simulation + Validation Pipelines: A critical infrastructure opportunity is the development of integrated pipelines that unify data, simulation, scenario generation, and formal validation into a closed-loop system. Today, these components are often disconnected, leading to inefficiencies and gaps in safety assurance. By linking real-world datasets with simulation platforms (e.g., CARLA, Omniverse), scenario generation tools (e.g., Scenic), and formal verification methods, researchers can create workflows that systematically explore and validate system behavior. Such pipelines enable continuous validation, where models are iteratively tested against increasingly diverse and challenging conditions, bridging the gap between development and deployment. Traditionally, these capabilities grow from universities and then build a following with an industry sponsored consortium.
• Low-Cost, Scalable Research Platforms for Education: Finally, there is a strong opportunity to develop affordable, scalable AV platforms that support both education and early-stage research. These platforms could include small autonomous vehicles, modular sensor kits, and pre-integrated software stacks that allow students and researchers to experiment with real-world autonomy at low cost. By lowering barriers to entry, these platforms would expand participation in AV, create a broader talent pipeline, and enable universities worldwide to contribute to research and innovation in a practical, system-level manner. Traditionally, consortia have not been employed in pedagogy, but complex problems such as AV require ‘crowd-sourcing’ of ideas in concepts such as FRODO [78]. University consortiums are the natural place for this work.
• Consortiums and Standards: Consortiums and standards bodies play a critical role in enabling productivity within complex ecosystems like AV by providing a shared foundation for interoperability, collaboration, and reuse. In environments where multiple organizations develop hardware, software, data, and validation tools independently, the absence of common standards leads to fragmentation, duplication of effort, and high integration costs. Consortium-driven standards—such as those developed by groups like ASAM—establish common data formats, interfaces, and scenario definitions that allow components from different vendors and research groups to work together seamlessly. This not only accelerates development by reducing the need for custom integration, but also enables reproducibility and comparability of results across institutions, which is essential for scientific progress. Overall, the work of development of interesting interface standards is quite challenging and develops over time. Organizations such as ASAM, SAE, or IEEE are naturally built for this work based on industry input.

6. Conclusions

The rapid progress of Digital AI has demonstrated the transformative power of open, layered, and widely accessible research ecosystems. Shared datasets, open-source frameworks, reusable models, scalable compute infrastructure, and collaborative education platforms have enabled universities, startups, and independent researchers worldwide to participate in advancing the state of the art. In contrast, progress in AV has been constrained not primarily by a lack of algorithms, but by the absence of equally mature research and education infrastructure.

AV systems operate under fundamentally different constraints than purely digital systems. They must perceive and act in uncertain physical environments, integrate tightly coupled hardware and software stacks, interact safely with humans, and operate under real-time, safety-critical conditions. These requirements introduce challenges in reproducibility, validation, interoperability, lifecycle management, and regulatory compliance that cannot be solved through software advances alone. Current ecosystems remain fragmented across hardware platforms, software stacks, datasets, simulation tools, and validation methodologies, limiting collaboration and slowing innovation.

This paper has argued that enabling the next generation of AV requires coordinated investment in shared infrastructure across both research and pedagogy. Key opportunities include the creation of modular open autonomy stacks, interoperable standards, shared real-world and hybrid testbeds, scalable digital twin environments, open scenario and edge-case databases, and integrated validation pipelines that combine simulation, formal verification, and real-world testing. Equally important is the development of educational ecosystems that combine theory with hands-on experimentation, expose students to safety assurance and lifecycle engineering, and bridge the gap between academic prototypes and deployable industrial systems.

The examples discussed throughout this paper—including Autoware, ASAM standards, Duckietown, MIT Beaver Works, ISEAUTO, PolyVerif, and the SafeAV consortium—demonstrate that important building blocks already exist. What remains missing is a sufficiently integrated and open ecosystem that enables large-scale ‘crowd-sharing’ of ideas, reproducible experimentation, and collaborative validation across institutions and domains. The success of Digital AI suggests that such openness and modularity can dramatically accelerate innovation when supported by shared infrastructure and community standards.

Ultimately, AV should be viewed not merely as a software discipline, but as a systems-engineering challenge requiring deep integration across computation, sensing, communication, validation, human interaction, and real-world operations. Addressing these challenges will require sustained collaboration among universities, standards organizations, funding agencies, industry partners, and government regulators. Investments in shared infrastructure, open platforms, and interdisciplinary workforce development will be essential if AV is to achieve the same pace of scalable innovation that has characterized modern Digital AI. The central bottleneck facing AV today is therefore not simply intelligence, but infrastructure. The institutions and ecosystems that successfully lower the barriers to experimentation, validation, and education will play a defining role in shaping the future of autonomous systems, robotics, and cyber-physical intelligence.

The defining innovation of Digital AI was not merely the transformer—it was the creation of a reusable, collaborative infrastructure that allowed millions of researchers to innovate simultaneously. Autonomous Vehicles now face a similar inflection point. The next breakthrough may come not from a new planning algorithm, but from building the shared infrastructure that allows thousands of researchers to improve every layer of the autonomy stack.