Section 2 positions AIDVs within the vehicle evolution; Section 3 surveys industry and academic work related to AIDVs; Section 4 formalizes their main principles; Section 5 introduces a reference architecture for AIDVs; Section 6 discusses representative pitfalls; Section 7 outlines a roadmap that integrates regulatory, safety, security, privacy, trust, and social awareness considerations; finally, Section 8 concludes the article and provides directions for future work.

The carmaker XPENG defines AIDVs as a new class of smart vehicles where AI is central to both the driving and the user experience (XPENG Motors, 2024).

The P7+ model has been presented as an early instantiation of this vision, integrating large AI models for advanced driver assistance and intelligent cockpit functions. From a cloud and platform perspective, Amazon Web Services (AWS) similarly positions AIDVs as an evolution of SDVs driven by the integration of generative AI, the Internet of Things (IoT), and edge computing (Raghavan et al., 2025), emphasizing AI as a primary driver of vehicle behavior and interaction. Along similar lines, a recent industry commentary emphasizes that automakers must proactively adapt their strategies to remain competitive in the era of AIDVs, highlighting the need for accelerated investment in AI capabilities, stronger data infrastructures, and partnerships across the mobility ecosystem (Just Auto, 2024).

Qualcomm, in particular, frames AIDVs as agentic systems built on on-device foundation models capable of perception, reasoning, and action under real-time constraints, highlighting tight coupling between AI inference, hardware acceleration, and edge autonomy to enable scalable, safety-aware intelligence in production vehicles (Qualcomm Technologies Inc., 2026). This scope differs from the approach followed in this work. While Qualcomm takes a product-oriented view, seeing AIDVs as an agentic, platform-driven realization enabled by tightly integrated edge AI and orchestration infrastructure, this work adopts a class-of-systems perspective, defining AIDVs as a distinct vehicle class characterized by AI-based autonomous reasoning as an architectural invariant, continuous interaction-driven learning, and explicit integration of digital twins, governance, and post-deployment evolution.

Beyond industry work, several academic contributions begin to outline research directions that resonate with the emerging notion of AIDVs, including multi-agent systems, safe reinforcement learning, and runtime verification.

Recent surveys on connected and autonomous vehicles highlight increasing reliance on learning-based and data-driven approaches across perception, decision-making, and control, motivating research on adaptive and AI-centric vehicle architectures. Murphey et al. (2023) provide an extensive overview of AI-enabled technologies for autonomous and connected vehicles, analyzing advancements in perception, decision-making, vehicle control, and connectivity. Their survey emphasizes how learning-based approaches, foundation models, and data-driven control are gradually transforming autonomous systems from static pipelines into adaptive, self-improving platforms. Their work also highlights the increasing computational demands of modern vehicles, the convergence of autonomy, electrification, and connectivity, and the role of AI as a coordinating technology across sensing, planning, and safety. From a system perspective, Teixeira et al. (2025) emphasize determinism, predictability, and reliability as foundational requirements for SDVs, explicitly constraining how adaptive or learning-based components may be integrated to preserve safety guarantees. Complementing this view, Richenhagen et al. (2025) examine how generative AI can be concretely realized within SDVs, offering practical insights into engineering integration, software-centric organizational structures, and AI-enabled development processes. Their work demonstrates that deploying large AI models in vehicles requires not only technical enablers, DevOps pipelines (Nouri et al., 2025), structured system engineering backbones, and consistent toolchains, but also new business models and cross-functional coordination.

This reflects AI’s architectural role in AIDVs, beyond SDVs add-on integration. Other recent studies further illustrate that AI is already deeply embedded in SDVs, primarily as a development-time and lifecycle enabler rather than as a runtime decision-making agent. Nguyen et al. (2025) show how Large Language Models (LLM) can support SDV software engineering through few-shot code generation, improving developer productivity and accelerating feature implementation within software-defined stacks. Similarly, Zyberaj et al. (2025) demonstrate the use of generative AI to automate test case generation and execution in SDV platforms, reinforcing AI’s role in validation, verification, and Continuous Integration and Continuous Deployment (CI/CD) pipelines (Joshi, 2025; Author Anonymous, 2025).

Collectively, these works confirm that AI is already integral to SDVs, but remains largely confined to tooling, engineering support, and controlled optimization. In contrast, AIDVs extend beyond this paradigm by embedding AI as a first-class architectural element governing runtime interactions, perception, reasoning, and adaptation, thereby introducing fundamentally different autonomy, accountability, and safety challenges, as detailed in Section 2.

Nguyen et al. (2018) survey deep reinforcement learning techniques for coordinated decision-making among multiple agents, while Garg et al. (2024) analyze safety, verification, and control challenges in multi-robot systems, providing insights into distributed autonomy and interaction. Safe reinforcement learning has been extensively studied to constrain learning-based policies under safety requirements; Kushwaha et al. (2025) offer a comprehensive survey of constrained Markov decision processes and safety-aware learning methods. Finally, runtime verification has been proposed as a complementary mechanism for monitoring learning-enabled systems post-deployment. Mannucci and de Oliveira Filho (Mannucci and de Oliveira Filho, 2023) investigate runtime verification of reinforcement learning properties, while Taleb and Khoury (2023) review uncertainty-aware runtime monitoring techniques. Although these academic works do not explicitly target AIDVs, they provide rigorous theoretical and methodological foundations for reasoning about autonomy, safety, and accountability in AI-driven vehicular systems.

More recently, in the academic literature, Zheng et al. (2025) explicitly adopt the term “AIDV,” proposing a multi-agent reinforcement learning framework in which fleets of ride-hailing vehicles jointly fulfill mobility demands while performing distributed, edge-based fine-tuning of urban foundation models. Their formulation highlights a key aspect of AIDVs: vehicles as active participants in city-scale AI ecosystems, simultaneously providing transport services and contributing to the continuous improvement of foundation models. To the best of current academic knowledge, within the vehicular domain, only Yu (2025) explicitly introduces a closely related notion, “Agentic Vehicles” (AgVs), as systems endowed with agency, contextual reasoning, tool use, and value-aligned decision-making. Although Yu focuses on the conceptual axis of agency and its socio-technical implications, the work above foregrounds the technical foundations enabling vehicles to learn, adapt, and participate in AI-driven ecosystems. In contrast to both approaches, our work defines AIDVs as a distinct vehicle class characterized by architectural requirements, integrated reasoning agents, digital-twin closed-loop validation, and post-deployment evolution. It further maps these properties onto a modular stack while systematizing AIDV-specific pitfalls and an incremental ITS roadmap.

AIDVs are built upon AI-driven solutions that govern Intelligent Driver Agents (IDAs), based on those already envisioned in the early 2010s (Group, 2008). This includes interactions and connectivity with the surrounding infrastructure. This paradigm is referred to as AI-mediated Interaction, where the system inherently supports real-time decision-making, autonomous optimization, orchestration, and continuous learning. In this design, agents interact not only with other vehicles to exchange data and coordinate actions, but also with non-vehicle entities, such as pedestrians, roadside units, and other road users, enabling the system to continually adapt and evolve.

Learning from new data over time (including traffic patterns, driving conditions, and user behaviors), vehicles can adapt to different environments, weather conditions, traffic flows, and user preferences. They can also handle unfamiliar situations, such as newly encountered road segments, by leveraging the generalization capabilities of AI and collaborating with the surrounding infrastructure. In this way, essential navigation information is obtained not only from the vehicle’s own sensors but also directly from the infrastructure (Wágner et al., 2023; De Vincenzi et al., 2025).

Vehicles may be paired with digital twins in a virtual environment featuring low-cost, high-definition mapping. These digital counterparts enable testing, training, and optimization of vehicle behavior. They can also be linked to digital twins of AI-enabled road infrastructure to support real-time mobility optimization decisions. This integration allows vehicles to anticipate and adapt to future scenarios, validate updates before deployment, and increase resilience by safely exploring rare or critical events that may be impractical to test in the physical world, such as construction zones.

AIDVs not only enable predictive maintenance, but also self-maintenance, especially for software components, and, where supported by emerging materials and hardware technologies, selected physical subsystems. Equipped with auto-regeneration materials, Electronic Control Units (ECUs), and self-healing materials (Sabet, 2024), AIDVs possess self-diagnosis and auto-repair capabilities, improving reliability and reducing the need for owner intervention or manual maintenance, especially for software components. The self-maintenance service can interact with external agents, such as those provided by the carmaker, to perform the necessary operations.

To preserve a clear separation between safety-critical driving logic and user-oriented service optimization, a dedicated recommender agent is introduced and explicitly differentiated from the IDA, which is responsible for driving-related and safety-critical functions. While the IDA governs vehicle motion and compliance with safety constraints, the recommender agent focuses on optimizing user experience and service-level interactions. It personalizes entertainment, navigation, trip planning, and service suggestions by analyzing user behavior, context, and preferences. Through continuous and adaptive interaction, the agent enhances the Human-Vehicle Interface (HVI), aligning vehicle services with driver habits and gradually shaping them over time. In doing so, it transforms the vehicle into a proactive service provider. The agent can also communicate with external agents and services, enabling coordination with other users or infrastructures, which is particularly relevant for shared mobility and the provision of personalized services.

To make the AIDVs architecture more concrete, a minimal algorithm (Algorithm 1) is presented to illustrate the AI Reasoning–Digital Twin loop for a single longitudinal driving function, namely, the adaptation of the vehicle’s time headway (i.e., following distance) on a highway.

Drawing inspiration from the literature (De Vincenzi et al., 2025), where a language for improving autonomous driving systems was proposed, a set of parameters is considered to describe the surrounding environment of the target vehicle V . These parameters include lane-level map attributes (e.g., laneID, lane connectivity, and permitted maneuvers), local-coordinate geometric representations (e.g., station value s and corresponding waypoints ( x ( s ) , y ( s ) , z ( s ) ) ), traffic control logic (e.g., active signal phase Φ and associated lane groups), and timing information (e.g., phase start and end times synchronized via a common time reference). Such parameters are continuously updated during the driving process to support coordinated, interaction-driven, and adaptive mobility (Figure 1). These parameters also define the state space of the corresponding digital twin, enabling the simulation of alternative traffic evolutions, interaction patterns, and edge cases within a virtual environment that mirrors the operational context of the physical vehicle. While illustrated here for a specific longitudinal driving function, this mechanism exemplifies a general architectural pattern in AIDVs, in which post-deployment adaptation is systematically mediated through digital twin validation rather than unconstrained real-world experimentation.

Algorithm 1 provides a minimal formalization of this architectural pattern, summarizing the AI–digital twin interaction in pseudocode. Although intentionally simple, the algorithm captures the essential pattern of continuous learning, closed-loop validation, and constrained deployment that characterizes AIDVs.

Algorithm 1

AI-digital twin adaptation loop.

1: Initialize IDA with baseline headway τ 0

The algorithm considers three main components: the physical vehicle V , its digital twin D T , and the AI Reasoning Layer hosting an IDA. The vehicle V executes a baseline longitudinal controller parametrized by a target time progress τ . The goal of IDA is to refine τ over time to improve comfort and traffic throughput, while preserving safety and regulatory constraints. The digital twin D T runs a high-fidelity simulation of V and its surrounding traffic, continuously calibrated from real telemetry.

The closed loop operates in recurring adaptation cycles:

Data collection and synchronization. During normal operation, V records trajectories and sensor traces for a fixed window D (e.g., several minutes of driving on the highway), including speed, relative distance and speed to the leading vehicle, and braking events. In addition, D is augmented with environment descriptors derived from infrastructure broadcasts, such as lane identifiers, permitted maneuvers, the currently active traffic signal phase Φ , and associated phase timing information. These data are periodically uploaded to the backend and ingested by D T , which replays the observed traffic scenes and updates its internal state to remain aligned with the physical vehicle.

Although this example adjusts a small set of parameters in a highway-following controller, the same architectural pattern applies more broadly to richer AIDV behaviors. These may include multi-objective planning policies, integrated perception–planning stacks, and multi-agent coordination strategies can be refined in the digital twin, validated against safety constraints, and then deployed back to the physical fleet.

One major concern is the potential for unintended coordination between vehicles and IDAs through Vehicle-to-Everything (V2X) communication, especially when learning-based agents continuously adapt their strategies based on shared observations and feedback. Such emergent behaviors can evolve beyond the scope of human oversight, potentially enabling vehicles to collude, optimizing traffic flow or fuel efficiency at the expense of safety, fairness, accountability, or legal constraints (Motwani et al., 2025). Currently, AVs are increasingly widespread, yet remain prone to failure and still require human supervision. In contrast, AIDVs have the potential for a greater systemic impact through persistent, coordinated, and even collusive activities among AI-enabled road agents operating at scale.

Safety certification represents a central engineering challenge in the context of AI-defined vehicles, where adaptive and learning-enabled behavior must be reconciled with certification frameworks originally designed for statically verifiable systems. Existing standards for automated driving, such as ISO 26262 for functional safety (ISO, 2021), largely assume tightly bounded and predictable behavior. In contrast, AIDVs enable post-deployment adaptation, allowing AI models to update or refine policies through continuous interaction. Such self-modifying behavior challenges certification pipelines designed for determinism and traceability. Recent surveys highlight that safety assurance practices for AI-based driving systems remain immature, with regulatory and technical frameworks lagging behind AI’s dynamic characteristics (Ullrich et al., 2025). Runtime monitoring and validation of AI-based components have been proposed to provide continuous operational evidence of safety (Aslam et al., 2024), while emerging standards such as ISO PAS 8800 extend functional safety concepts toward through-life assurance of AI components (ISO. ISO/PAS 8800, 2024). To date, no fully established methodology guarantees safety for systems that learn during operation (Seshia et al., 2022; Paterson et al., 2025). Recent work therefore advocates combining formal verification with runtime assurance: formal methods can verify safety invariants over constrained adaptation spaces (Seshia et al., 2022; Wongpiromsarn et al., 2023), while runtime assurance architectures enforce these guarantees during operation via monitored supervision and safe fallback mechanisms (Chen et al., 2022).

A potential pitfall of AIDVs lies in their sustainability implications. Because AIDVs rely on pervasive interactions among multiple AI agents, their operation presupposes the widespread availability of computational resources, continuous connectivity, and large-scale data exchange across vehicles and infrastructure. This AI-by-design paradigm may significantly increase energy consumption, hardware complexity, and networking overhead, both on-board and at the edge (Schwartz et al., 2020). Without careful architectural choices, such as workload distribution, selective activation of agents, and energy-aware reasoning, these requirements could translate into higher environmental costs, increased operational expenses, and reduced scalability, challenging the long-term sustainability of large-scale AIDV deployment.

AI agents can prioritize driver preferences, convenience, or commercial objectives at the expense of collective safety, social equity, or broader societal interests, particularly when optimization goals are poorly specified or misaligned. Highly persuasive recommender agents may influence user decisions, such as destinations or purchase choices, in subtle ways that are difficult to detect or contest. Without robust mechanisms for transparency, explainability, and accountability, these systems can produce opaque outcomes that reinforce existing biases and complicate the assignment of responsibility in the event of failures (Mehrabi et al., 2021). Over time, such influence may erode individual autonomy and meaningful consent by systematically shaping ethically questionable behavior (Harari, 2020).

AIDVs expand the attack surface beyond traditional vehicle subsystems, introducing vulnerabilities not only at the sensor and perception level, but also across learned models, interaction protocols, and adaptive decision-making pipelines. Adversaries, for example, can subtly manipulate sensor inputs, e.g., by altering road signs, to induce IDAs to misinterpret speed limits or navigation cues (Eykholt et al., 2018; Sato et al., 2023; MohajerAnsari et al., 2025). Because these systems depend heavily on image recognition and multi-sensor fusion, even small perturbations can propagate through the perception pipeline, posing significant risks to public safety. Moreover, excessive reliance on AI-based security mechanisms may foster unwarranted confidence in their robustness. Factors such as sampling bias, spurious correlations, and evaluations limited to controlled settings are often underestimated, yet they can critically compromise real-world effectiveness and system trustworthiness (Arp et al., 2021).

AI systems depend on a broad, layered infrastructure that spans beyond software and vehicles to include energy supply, specialized hardware, data centers, model training platforms, and application environments. As noted from an industry perspective by NVIDIA founder and CEO Jensen Huang, AI “is becoming the foundation of the largest infrastructure buildout in human history” with a “five-layer cake” comprising energy, chips and computing infrastructure, cloud data centers, AI models, and applications, each requiring dedicated investment, construction, and governance to support safe, scalable AI development (Huang, 2026). In the context of ITS, this layered infrastructure imperative spans short to long-term phases and underpins regulatory, safety, and deployment frameworks across the AIDV ecosystem.

Taken together, this roadmap emphasizes that the transition to AIDVs is not a single technological leap, but a staged co-evolution of architecture, infrastructure, governance, and societal trust.

Although AIDVs hold transformative potential compared to SDVs, their deployment must start with robust regulation, rigorous testing, and ethical design to mitigate socio-technical risks.

As the deployment expands, a key challenge is ensuring that different vehicles can interact safely and effectively within dynamic environments.

The technical and sustainability limits of large-scale learning, coordination, and communication must be addressed to create an AI-driven mobility ecosystem.

At the ecosystem level, harmonized global standards for AIDV interoperability, liability attribution, and policy alignment will be required to ensure long-term trustworthiness and large-scale deployment. Unlike earlier phases focused on vehicle-level safety or fleet-level interaction, this stage assumes the coexistence of heterogeneous AIDVs, legacy vehicles, and AI-enabled infrastructure across jurisdictions with different regulatory, cultural, and economic constraints. Consequently, international coordination bodies and cross-border regulatory frameworks will play a central role in defining shared interfaces, certification envelopes, and accountability mechanisms for adaptive, learning-enabled vehicles.

This long-term vision also addresses critical systemic mobility challenges, such as alleviating urban congestion, improving energy efficiency, promoting sustainability, and mitigating workforce shortages in domains such as long-haul trucking and logistics (O’Connor, 2025). At this stage, unified Mobility-as-a-Service (MaaS) platforms will act as integrative layers, coordinating demand, routing, and pricing across public transport, shared mobility, and autonomous fleets. In parallel, cloud–edge orchestration layers will be essential to manage distributed learning, digital-twin synchronization, and policy enforcement across geographically dispersed fleets, enabling AIDVs to operate as coordinated agents within a global ITS.