HAPS-Enabled V2X Networks Overview

• HAPS-enabled V2X networks are vehicular communication systems that integrate high-altitude platform stations as aerial relays to extend coverage and ensure continuous connectivity.
• They employ a hybrid V2I/V2V/V2H radio access model combined with deep reinforcement learning and Age of Information control to optimize resource allocation and maintain data freshness.
• These networks support corridor-scale intelligent transportation applications such as autonomous driving, platooning, and emergency response by providing reliable, low-latency communication even in infrastructure-limited areas.

HAPS-enabled Vehicle-to-Everything (V2X) networks are vehicular communication systems in which high-altitude platform stations (HAPS) operate as a non-terrestrial network (NTN) layer that complements terrestrial infrastructure for safety-critical and service-rich transportation applications. In the cited literature, HAPS is positioned as an aerial communication platform, aerial relay, “super-macro base station,” and “aerial roadside unit,” with roles spanning V2I/V2N connectivity, platooning support, sensing and monitoring, caching, and computing. The common motivation is that autonomous driving, platooning, accident reporting, and corridor-scale intelligent transportation systems (ITS) require communication continuity, wide-area coverage, and fresh state information even in remote, rural, disaster-stricken, or otherwise infrastructure-constrained areas (Ince et al., 21 Jul 2025, Jaafar et al., 2021).

1. Architectural definition and operational scope

HAPS-enabled V2X extends conventional vehicular networking by inserting a stratospheric infrastructure layer, typically around 20 km altitude, into the communication path. In the platoon-oriented formulation, the network consists of one HAPS, an RSU, and multiple autonomous vehicle platoons, each with one platoon leader (PL) and multiple followers. The PL is the communication agent responsible for intra-platoon and inter-platoon exchanges and can dynamically select among three modes: $\theta_j^t = 0$ for V2I, $\theta_j^t = 1$ for V2V, and $\theta_j^t = 2$ for V2H. In the numerical setting of the AoI-oriented study, the simulated urban network contains five PLs, each with six followers, and the HAPS provides an aerial communication option when terrestrial links are insufficient or when wide-area connectivity is needed (Ince et al., 21 Jul 2025).

A broader systems view appears in the HAPS-ITS architecture for trans-continental highways. There, HAPS is not limited to radio relaying. It is modeled as an integrated multi-function node with communication, storage, computing, and imaging payloads; H2V/V2H, H2G, and H2H links; gateways in cities; roadside IoT sensors; and traffic management centers. The architecture supports vehicles, roadside sensing systems, backhaul systems, and control centers in a corridor-wide service mesh. The paper explicitly maps HAPS to V2I, V2N, infrastructure-assisted V2V backup, and broader V2X orchestration (Jaafar et al., 2021).

Two architectural patterns therefore recur. The first is a hybrid V2I/V2V/V2H radio access model centered on PL decision-making and resource allocation. The second is a corridor-scale ITS model in which HAPS acts as an active infrastructure layer for transportation services. A plausible implication is that “HAPS-enabled V2X” is best understood not as a single protocol stack, but as a family of NTN-assisted vehicular architectures spanning communication-only and communication-computing-sensing deployments.

2. Communication model, links, and channel assumptions

In the platoon-based HAPS-V2X formulation, radio resources are organized over a set of $K$ orthogonal sub-channels using OFDM. For platoon leader $j$, the relevant channel gains on sub-channel $k$ at time $t$ are $h_{j,I}^t[k]$ for infrastructure links, $h_{j,v}^t[k]$ for follower or vehicular links, and $h_{j,H}^t[k]$ for HAPS links. Each PL chooses a communication mode $\theta_j^t = 1$0, a sub-channel assignment indicator $\theta_j^t = 1$1, and transmit power $\theta_j^t = 1$2, with the explicit restriction that each agent can use only one sub-channel at a given time (Ince et al., 21 Jul 2025).

The HAPS link is modeled as Rician fading, justified by likely line-of-sight conditions. The printed expression is

$\theta_j^t = 1$3

where $\theta_j^t = 1$4 includes scintillation loss, atmospheric attenuation, clutter loss, shadow fading, and free-space path loss, while $\theta_j^t = 1$5 and $\theta_j^t = 1$6 denote LoS and non-LoS probabilities. The source synthesis explicitly notes typographical inconsistencies or missing parentheses in the paper, but preserves the intended Rician-like composite HAPS channel. By contrast, V2I and V2V channels are represented as

$\theta_j^t = 1$7

with large-scale fading $\theta_j^t = 1$8 and small-scale fading $\theta_j^t = 1$9 (Ince et al., 21 Jul 2025).

For V2H, the achievable capacity on sub-channel $\theta_j^t = 2$0 is written as

$\theta_j^t = 2$1

with co-channel interference

$\theta_j^t = 2$2

Similar capacity formulas are stated to apply for V2I and V2V, although the paper does not print them explicitly. This yields a coupled wireless environment in which local decisions interact through shared sub-channels and interference (Ince et al., 21 Jul 2025).

The corridor-scale HAPS-ITS paper expands the link taxonomy. It identifies H2V/V2H as the main aerial vehicular access link; H2G as HAPS-to-gateway backhaul, proposed with FSO or mmWave; H2H as HAPS-to-HAPS interconnection for resilience and traffic sharing; vehicle-to-roadside-sensor links using Wi-Fi, Bluetooth, or LoRaWAN; and V2V as the conventional direct vehicle link that HAPS may back up or augment. It also mentions HAPS-to-satellite as a future integration path rather than a central component (Jaafar et al., 2021).

A recurring misconception is that HAPS-enabled V2X is merely satellite-style connectivity applied to cars. The surveyed papers do not support that simplification. They repeatedly distinguish HAPS from more distant non-terrestrial platforms through relatively low latency, likely LoS access, and corridor-scale integration with terrestrial RSUs, gateways, and sensors (Ince et al., 21 Jul 2025, Jaafar et al., 2021).

3. Age of Information as the central control objective

The most explicit formalization of HAPS-enabled V2X in the provided material is AoI-aware. The cited study argues that throughput or latency alone is insufficient for autonomous vehicular systems because stale status information may remain unsafe even when average throughput is acceptable. Accordingly, it adopts Age of Information (AoI) as the principal optimization lens for platoon-based autonomous vehicle systems (Ince et al., 21 Jul 2025).

The paper does not use a continuous-time sawtooth definition. Instead, AoI is defined through a slot-based recursion: $\theta_j^t = 2$3 The interpretation given in the synthesis is direct: successful V2I or V2H transmission resets AoI to one slot, otherwise AoI increases by $\theta_j^t = 2$4. The paper explicitly supports V2V mode at the system level, yet the printed AoI reset equation includes only V2I and V2H; the synthesis notes that the paper does not explain this omission (Ince et al., 21 Jul 2025).

The average AoI objective is

$\theta_j^t = 2$5

and the overall optimization is written as

$\theta_j^t = 2$6

As summarized in the source block, this combines three aims: minimize average AoI, maximize the probability that cumulative communication capacity satisfies the minimum data transmission requirement $\theta_j^t = 2$7, and minimize average transmit power. The reliability-related term is therefore tied to cumulative capacity sufficiency rather than a separate packet error or outage formalism (Ince et al., 21 Jul 2025).

This formulation makes freshness a first-class control variable in HAPS-supported V2X. It also narrows the scope. The paper does not provide peak AoI, weighted AoI, queue-dependent AoI, retransmission modeling, HARQ, or a full queueing discipline. The freshness process is slot-level and update-level rather than packet-generation-level. This suggests that the work is best read as an AoI-driven radio resource allocation model, not as a complete transport-layer theory of information timeliness.

4. Deep reinforcement learning for decentralized resource allocation

The optimization variables in the HAPS-V2X platoon model are sub-channel assignment $\theta_j^t = 2$8, communication mode selection $\theta_j^t = 2$9, and power allocation $K$0. The corresponding constraints enforce minimum HAPS and infrastructure capacities, valid discrete choices, one-sub-channel use, and maximum transmit power. In the cleaned rendering provided in the synthesis, these appear as $K$1 through $K$2, with $K$3, $K$4, $K$5, $K$6, $K$7, and $K$8. The formulation contains mixed discrete-continuous decisions and interference coupling, and the paper motivates DRL as a response to dynamic vehicular environments and autonomous online adaptation requirements (Ince et al., 21 Jul 2025).

Each platoon leader is an RL agent. The local state is

$K$9

combining current V2V, V2I, and V2H channel states, previous-step interference, current AoI, remaining message load, and remaining time budget. The local reward is

$j$0

where $j$1 penalizes transmission power and $j$2 is a stepwise function enforcing minimum capacity satisfaction. The explicit algebraic forms of $j$3 and $j$4 are not given in the paper (Ince et al., 21 Jul 2025).

Training follows a standard actor-critic loop with local replay buffers $j$5, exploration noise, target networks, critic loss

$j$6

Bellman target

$j$7

and deterministic policy gradient

$j$8

The paper reports batch size 64, discount factor $j$9, actor layers $k$0, critic layers $k$1, actor learning rate $k$2, and critic learning rate $k$3 (Ince et al., 21 Jul 2025).

A methodological ambiguity is explicitly documented in the synthesis. The title emphasizes DDPG, but the body compares DDPG and FD-MADDPG; early in the DRL section DDPG is described as a single-agent independent-learning paradigm, while later it is described as using a centralized critic considering overall interference. FD-MADDPG is described as fully decentralized and without centralized training. The synthesis therefore identifies DDPG as the simpler baseline and FD-MADDPG as the decentralized multi-agent method that better handles interference and scales better. This inconsistency is itself part of the technical record and should be retained when characterizing the literature (Ince et al., 21 Jul 2025).

5. HAPS-specific services and corridor-scale ITS functions

Beyond radio resource allocation, HAPS-enabled V2X has been presented as an ITS infrastructure paradigm for sparse and trans-continental highway environments. The HAPS-ITS paper treats HAPS as a core enabler of future ITS services over long corridors where terrestrial V2X infrastructure is incomplete or absent, and where availability of power and connectivity is itself a design constraint (Jaafar et al., 2021).

The service portfolio is broad. HAPS is proposed to support traffic monitoring, accident reporting, surveillance, platooning, autonomous driving support, infotainment, fleet management, and sparse-area connectivity. In accident scenarios, vehicles send emergency data to HAPS while the platform can also capture images or video for incident assessment. The paper states that LiDAR cameras can form object pictures from 45 km away. It also proposes HAPS as an “aerial roadside unit” capable of collecting roadside sensor data, distributing road, weather, and traffic information, bridging failed V2V exchanges, and supporting connected and autonomous vehicles when local infrastructure is absent (Jaafar et al., 2021).

A particularly concrete claim concerns delay. The paper states that forwarding V2V messages through HAPS incurs an additional round-trip delay of $k$4, and argues that this remains small enough for timely reaction support. It also states that HAPS should support NR-V2X communication protocols (3GPP Release-17) for this function. For highway safety, the paper further states that vehicle response time should be below 200 ms, described as the response time of a professional driver, while HAPS communication delay is said to be below 0.33 ms (Jaafar et al., 2021).

The same paper provides capacity- and deployment-oriented figures for corridor-scale design. A HAPS node as a super-macro BS may provide coverage up to 500 km radius in general discussion, but the Trans-Sahara case study assumes a more practical 40 km radius footprint. For the Algiers-to-Lagos Trans-Sahara highway of 4,504 km, approximately 59 HAPS-ITS nodes are required under the 40 km radius assumption, with 26 in Algeria; if existing 3G/LTE coverage is leveraged only to fill gaps, the counts become 48 for the full route and 19 for the Algerian portion. For the Algeria segment, only 8 gateways are assumed deployable in cities, so several HAPS nodes along the El Menia–In Guezzam corridor rely on H2H backhaul instead of direct H2G (Jaafar et al., 2021).

The paper also reports that current HAPS communication reliability is up to 99.9%, whereas level 4–5 CAVs need reliability above 99.999%. This is a useful corrective to overly expansive claims about HAPS suitability. In the paper’s own framing, HAPS is already suitable for many assistance-oriented and corridor-support functions, but not yet sufficient by itself for the most demanding ultra-reliable high-automation safety services (Jaafar et al., 2021).

6. Sensing-assisted and intelligent-surface extensions

A related research direction connects HAPS-enabled V2X with integrated sensing and communications (ISAC) and intelligent surfaces. The cited IOS-based study does not model HAPS directly, but the provided synthesis states that its framework is directly useful for HAPS-enabled V2X because it addresses how sensing improves directional communication under mobility and channel uncertainty, while reconfigurable surfaces split and steer energy between sensing and communication functions (Meng et al., 2022).

The modeled system consists of one mmWave RSU, one mobile vehicle, an intelligent omni-surface (IOS) mounted on the vehicle, and an in-vehicle user. Each slot is divided into an S$k$5C phase and a communication-only phase. During S$k$6C, part of the incident signal is reflected toward the RSU to strengthen the sensing echo and part is refracted into the vehicle to serve the in-vehicle user; during the communication-only phase, all energy is directed into the vehicle using the improved vehicle state estimate obtained in the first phase. The average slot throughput is

$k$7

and the paper derives a sufficient and necessary condition for the existence of the sensing-and-communication phase: $k$8 If this condition is not met, the optimal solution sets $k$9, meaning that no sensing-and-communication phase should be used (Meng et al., 2022).

The paper further reports an interior optimum with $t$0 and $t$1, and emphasizes that, especially below $t$2 W transmit power, the proposed sensing-assisted design significantly outperforms the baselines. The synthesis interprets this as highly relevant to power-constrained aerial systems because improved beam alignment can reduce required transmit power for a target throughput (Meng et al., 2022).

This line of work should not be conflated with a direct HAPS result. The same synthesis explicitly notes that channel model, distance scale, Doppler structure, coverage, multi-user effects, latency assumptions, source beamforming, and surface placement would all need modification for true HAPS scenarios. Even so, it provides a transferable design template for hybrid HAPS + RSU + intelligent-surface architectures in which macro-scale tracking and local directional access are jointly optimized (Meng et al., 2022).

7. Empirical results, limitations, and open research problems

The clearest empirical results in the supplied material come from the AoI-aware HAPS-V2X platooning study. In its simulation setup with 1 HAPS, 1 RSU, 5 platoon leaders, 6 followers per platoon, and an urban-area scenario, the paper reports that FD-MADDPG converges faster and reaches a higher and more precise reward value than DDPG, while DDPG exhibits a longer and more difficult learning process. For average AoI in the HAPS-V2X scenario, when the gap between platoons is 5 m, DDPG average AoI is approximately 13 ms and FD-MADDPG average AoI is approximately 6 ms; when platoon spacing increases to 35 m, FD-MADDPG AoI increases by only about 3 ms, while DDPG AoI increases by almost 20 ms. The paper interprets this as evidence that decentralized multi-agent learning better handles interference and channel variation and maintains fresher information in HAPS-supported V2X (Ince et al., 21 Jul 2025).

The same study is careful, however, about what it does not show. It does not specify the exact number of sub-channels $t$3, exact slot duration $t$4, exact HAPS transmit parameters, exact vehicle speeds, path-loss exponents, replay buffer size, target update coefficient, number of training episodes, or confidence intervals. It also does not provide a comprehensive ablation comparing HAPS versus no-HAPS under identical RL policies. Thus, the HAPS benefit is supported more by system design rationale and qualitative comparison than by a fully isolated quantitative study (Ince et al., 21 Jul 2025).

Across the cited papers, several limitations recur. Simplified channel abstractions are prominent: the HAPS-V2X paper uses a simplified Rician/LoS probability expression and does not model detailed NTN propagation, Doppler, blockage dynamics, or weather effects; the IOS paper uses simplified LoS and free-space assumptions rather than aerial air-to-ground fading. Mixed-action treatment is incomplete in the DDPG formulation because the action must include discrete mode and sub-channel choices as well as continuous power control, yet the exact mechanism for handling this is not clearly specified. Queueing and update processes are also simplified: the AoI model uses a slot-reset rule with no explicit packet generation process, queue evolution, service-time model, or HARQ (Ince et al., 21 Jul 2025, Meng et al., 2022).

At the system level, the HAPS-ITS paper identifies a reliability gap for highly automated vehicles, weather sensitivity of FSO and mmWave, throughput bottlenecks on H2V links in dense traffic, energy constraints from communication-plus-computing-plus-caching-plus-imaging payloads, failure and continuity issues, interference in dense HAPS constellations, and the complexity of HAPS-to-LEO integration. It also notes security risks from malicious data injection in V2V and V2I and proposes HAPS imaging as a mechanism for validating possibly falsified reports (Jaafar et al., 2021).

The future-work directions named in the source block include energy-efficient learning strategies, adaptive reward mechanisms, real-world deployments in large-scale vehicular networks, investigation of HAPS mobility, hybrid AI-driven optimization, and comparison against attention-based DRL, federated learning, and game-theoretic approaches. The highway-oriented paper additionally points toward real-time redundant HAPS constellations, hybrid FSO/RF design, interference management through power control and advanced antenna techniques, failure recovery, and multi-layer orchestration with LEO satellites (Ince et al., 21 Jul 2025, Jaafar et al., 2021).

Taken together, the literature portrays HAPS-enabled V2X as a technically credible NTN extension for vehicular networking, especially in sparse-area and infrastructure-constrained settings, but not yet as a settled architecture. The most defensible consensus is narrower: HAPS can extend coverage, provide continuity, and support fresher vehicular information exchange; AoI-aware control and decentralized learning are promising for platoon-based resource allocation; and sensing-assisted or intelligent-surface methods may become important in future hybrid aerial-terrestrial V2X designs (Ince et al., 21 Jul 2025, Jaafar et al., 2021, Meng et al., 2022).