Accelerating the Uptake of 5G for Automotive: Real-World Trials from the TARGET-X Project

Abstract

As the automotive industry transitions toward high-level autonomy, the demand for connectivity offering deterministic low latency and high reliability becomes paramount. This paper presents the end-to-end design, implementation, and experimental validation of three advanced Vehicle-to-Everything (V2X) use cases within a real-world 5G network environment: Cooperative Perception, Automotive Digital Twin, and Predictive Quality of Service (pQoS) for Tele-operated Driving (ToD). Trials at the 370-hectare IDIADA proving ground in Spain benchmarked 5G and Multi-access Edge Computing (MEC) against 4G and cloud alternatives. Experimental results demonstrate that 5G provides substantial performance gains, achieving average latency reductions up to 90% for V2X messages compared to 4G. The integration of MEC halved the average service latency, consistently maintaining it within the 40–50 ms range required for safety-critical services, whereas cloud-based hosting exhibited uncontrollable fluctuations. In the pQoS for ToD, the implementation of a Network Digital Twin (NDT) and exposure APIs reduced the average video jitter by up to 63%, preventing service collapse in degraded coverage zones. Finally, the automotive Digital Twin (DT) achieved high-fidelity synchronization with temporal deviations consistently below 10%. These findings underscore the necessity of edge-centric architectures and proactive network telemetry for the resilient deployment of future safety-critical mobility services, effectively charting the course for the design of 6G.

1. Introduction

The evolution of Connected, Cooperative, and Automated Mobility (CCAM) has reached a critical juncture where basic connectivity is no longer sufficient. As the industry moves from driver-assist features toward full Level 4 and 5 autonomy (where it is not required that the driver take over driving) [1], the network must support heterogeneous services with conflicting requirements: low latency for safety-critical collision avoidance and cooperative perception, and high throughput and predictive reliability for remote operation [2,3].

One of the key challenges in transforming CCAM into commercial reality is the deployment of V2X services over large geographic areas. This requires extremely high reliability and availability with the presence of vast numbers of Connected and Automated Vehicles (CAVs). To address this, next-generation mobile networks must integrate environment-aware capabilities, where the infrastructure perceives road and traffic conditions, leverages Multi-access Edge Computing (MEC) for localized, high-speed computation, and exposes real-time network telemetry to Vehicle-to-Everything (V2X) applications.

The 5G ecosystem and its evolution toward 6G are engineered to ensure that these stringent requirements for latency, throughput, reliability, and availability are consistently met to enable safe and autonomous mobility [4]. In [5], the authors survey the application of 5G to advanced V2X scenarios and examine key technological enablers like MEC and network slicing. The work highlights critical implementation hurdles, ranging from radio channel phenomena and mobility management (handovers) to resource management and data fetching strategies. Furthermore, recent literature has sought to mitigate service continuity disruptions in cross-border scenarios [6,7].

The TARGET-X project (Trial Platform for 5G Evolution–Cross-Industry On Large Scale) aims to accelerate this uptake by validating 5G in real-world industrial and automotive environments. Moving beyond laboratory simulations, the TARGET-X trials utilize the IDIADA Connected Vehicle-Hub (CVH), a world-class facility that allows for high-speed, safety-critical testing. The facility was used to evaluate the performance of three use cases using 4G and 5G technologies: cooperative perception, automotive Digital Twin (DT), and predictive QoS (pQoS) for Tele-operated Driving (ToD).

Cooperative perception allows vehicles to “see” through the sensors of other road participants or infrastructure units (e.g., Road Side Units–RSUs). While early V2X standards focused on Cooperative Awareness Messages (CAMs) for basic status updates, the European Telecommunications Standards Institute (ETSI) has recently standardized the Collective Perception Service (CPS) and the corresponding Collective Perception Messages (CPM) format [8]. Research by Huang et al. [9] highlights that cooperative perception can compensate for sensor blind spots at intersections, but the sheer volume of data from LiDAR and camera sensors requires 5G’s high-bandwidth Sidelink or Uu interfaces. Recent studies suggest that a market penetration rate of at least 25% is necessary for cooperative perception to significantly reduce serious accident rates [10].

DT technology has transitioned from manufacturing to real-time vehicular operations. An Automotive DT is a virtual replica that mirrors the state, behavior, and environment of a physical vehicle [11]. In the context of 5G, DTs are used for Scenario-in-the-Loop (SciL) testing, where virtual actors (pedestrians or other vehicles) interact with a physical ego-vehicle through low-latency communication [12]. The main challenge in the state of the art is maintaining high-fidelity synchronization between the physical and virtual spaces, particularly in high-dynamic scenarios [13].

ToD serves as the ultimate safety layer for autonomous fleets, allowing a remote operator to take control during edge-case scenarios. However, ToD is extremely sensitive to packet loss and jitter. To mitigate this, the 5G Automotive Association (5GAA) has proposed the concept of pQoS [14]. Unlike reactive Quality of Service (QoS), pQoS utilizes Network Exposure Function (NEF) and Machine Learning models to provide advance notification of network degradation [15]. This allows the application layer to adjust video bitrates or initiate a safe stop maneuver before the link fails.

This paper addresses the critical requirements of next-generation Intelligent Transport Systems (ITS) by presenting the experimental validation of 5G-based IoT services for autonomous mobility in a large-scale real-world environment. Oriented toward the advancement of 5G-and-Beyond connectivity, the study offers five main contributions:

• Unified Network Exposure API and pQoS: We implement a Network Digital Twin (NDT) and a Network Exposure Application Programming Interface (API) to provide pQoS alerts regarding impending performance degradation. Although demonstrated in a ToD context (reducing video jitter by up to 63%), this proactive mechanism can serve as a unified cross-layer solution to mitigate network-induced reliability drops across various ITS use cases.
• Environment-Adaptive Cooperative Perception: We evaluate the efficacy of two different V2X messaging strategies based on environmental constraints, comparing CAM and CPM. While CAMs are used to maintain mutual awareness in scenarios where direct sensor visibility is obstructed (e.g., zero-visibility intersections), CPMs are utilized to detect and report non-connected road users, effectively extending the perception horizon of connected vehicles. The trials indicate that while CPMs are essential for comprehensive environmental models, they exhibit a 24.2% higher average trigger latency compared to CAMs under the tested conditions.
• Vehicle-in-the-Loop (ViL) via Digital Twin: We introduce an Automotive DT framework specifically designed to bridge physical and digital domains enabling high-fidelity ViL validation. This system allows physical ego-vehicles to interact with virtual actors in a synchronized digital environment, facilitating the safe, energy efficient, and repeatable testing of high-risk maneuvers with temporal synchronization deviations below 10%.
• Real-Life 5G and MEC Evaluation: We evaluate three ITS use cases at the 370-hectare IDIADA proving ground using a functional 5G Non-Standalone (NSA) network. Results demonstrate that 5G NSA and MEC reduce service latency by up to 90% for safety-critical messages compared to legacy 4G and cloud infrastructures, establishing a performance benchmark for resilient future internet mobility services.
• Identification of 6G Design Imperatives: Based on the trial results, we define critical architectural gaps and lessons learned that serve as inputs for 6G research. This includes the identification of the “TDD Frame Mismatch” for uplink-heavy services, the “Static Allocation Gap” in current slicing, and the requirement for sub-second streaming telemetry to achieve deterministic performance.

The paper is organized as follows. Section 2 describes the three use cases and their functional architectures. In Section 3, we describe the network architecture and the ITS platform. In Section 4, we present the network Key Performance Indicator (KPI) collection tool and we analyze the results obtained for the three use cases. Before concluding in Section 6, we discuss the challenges and lessons learned from the project in Section 5.

2. Use Cases

In the automotive vertical, the TARGET-X project focuses on three distinct use cases that are implemented, tested, and showcased on IDIADA CVH: Cooperative perception, automotive DT, and predictive QoS for ToD.

2.1. Cooperative Perception

2.1.1. Use Case Description

Cooperative perception improves autonomous navigation by enabling vehicles to construct high-fidelity environmental models through the fusion of local sensor data and V2X communications. While on-board sensors provide immediate situational awareness, their efficacy is often limited by line-of-sight obstructions and non-connected road users. To bridge these perception gaps, CAVs leverage a Cooperative Intelligent Transport Systems (C-ITS) platform, deployed through 5G networks on either Cloud (AWS) or Edge infrastructure, to exchange CAM and CPM.

This framework is validated through two distinct operational scenarios (See Figure 1). In the first, a Zero-Visibility Intersection, physical topology and environmental factors eliminate direct visibility between approaching vehicles. The second scenario, Road Damaged Vehicle, addresses a roadside damaged vehicle obstructed by adverse weather conditions, such as heavy fog or snow. This use case highlights the distinction between self-reporting and collective sensing.

2.1.2. System Architecture and Methodology

The core utility of the architecture is demonstrated through the Collision Warning Service (CWS), which monitors the incoming V2X streams to identify potential hazards. Upon detecting a risk, the CWS generates a Decentralized Environmental Notification Message (DENM), which is relayed back to the vehicles. On the client side, the Vehicle-Infrastructure Data Sharing Service (V2I-DSS) and the Local Dynamic Map Service (LDMS) integrate these network alerts with the In-Vehicle System (IVS) data. This merged environmental view is then rendered on the Human–Machine Interface (HMI), allowing the vehicle to evolve its state of knowledge and adapt driving strategies in real-time. All the V2X messages are exchanged through a Message Queuing Telemetry Transport (MQTT) broker that is implemented in the C-ITS platform.

In the Zero-Visibility Intersection scenario, local perception systems (e.g., Lidar or Radar) are unable to generate CPMs for hidden objects. Consequently, the system relies on CAMs to share telemetry data with the C-ITS platform. The CWS evaluates the trajectories of the CAVs and issues preemptive DENM alerts, mitigating the risk of collision at blind junctions.

In the Road Damaged Vehicle scenario, if the disabled vehicle is connected, the CWS identifies the hazard via the stationary CAM. In contrast, if the vehicle is not connected, passing CAVs on the opposite lane utilize their own sensors to detect the obstacle, transmitting its coordinates via CPMs. The CWS notifies the CAVs of the same lane about the stopped vehicle. By integrating both approaches, the CWS ensures that approaching vehicles receive timely warnings, enabling them to exercise extreme caution and adjust maneuvers despite the lack of visual confirmation. Ultimately, this multi-layered communication strategy transforms disparate data into actionable intelligence, significantly enhancing road safety in complex, dynamic environments.

The architectural decision to deploy the C-ITS platform at the Edge (MEC) for these scenarios is driven by the temporal safety window required for collision avoidance. In the Zero-Visibility Intersection (Figure 1), a vehicle traveling at 50 km/h covers approximately 1.4 m every 100 ms. Our results (Section 4.1) show that cloud hosting introduces latency fluctuations exceeding 50 ms, which would effectively double the ’blind’ distance before a warning is rendered. By utilizing the MEC-centric architecture shown in Figure 1, the system consistently maintains the 40–50 ms latency threshold required to provide a reliable sub-meter reaction buffer for the Autonomous Emergency Braking (AEB) systems.

2.2. Automotive Digital Twin

2.2.1. Use Case Description

The automotive DT use case implements a DT based on Scenario 1 of the cooperative perception use case. By digitizing this scenarios, we enable repeatable evaluation of V2X features via a web-based visualization and replay tool. This methodology supports the shift toward digital testing to reduce human error and allows Original Equipment Manufacturer (OEM) service providers, such as IDIADA, to meet evolving OEM connectivity requirements. The TARGET-X DT merges physical and digital domains by capturing real-world V2X data from the C-ITS platform, which is then used for complex simulations, such as autonomous vehicle integration, or ViL validation.

The use case introduces virtual objects (e.g., a digital vehicle) into the system and displays them on the HMI to trigger system reactions identical to those caused by physical vehicles. This “Replay” functionality allows for the capture, storage, and subsequent reproduction of all V2X communications within a specific scenario.

2.2.2. System Architecture and Methodology

The high-level architecture for the replay functionality is shown in Figure 2. The proposed solution utilizes an external technical stack integrated with a centralized database, a recording module, and a replay service. The operational workflow is divided into two distinct functional phases:

• Recording Phase: Scenario 1 of the cooperative perception use case is executed iteratively using two physical CAVs. During these trials, the recorder captures high-fidelity V2X data, specifically CAM and DENM, directly from the MQTT broker. This telemetry is subsequently archived in a structured recording database to serve as the ground truth for the DT.
• Replay and Augmentation Phase: In the second phase, the physical test track is restricted to a single CAV. The replay system orchestrates the scenario by injecting a DT (virtual CAV) into the network environment using the previously recorded message sets. This approach allows for the evaluation of the real vehicle’s response to the virtual agent under diverse conditions, including high-speed maneuvers that would pose significant safety risks if conducted with two physical vehicles on the track.

The rationale for this two-phase recording and replay architecture is the need for deterministic high-fidelity validation. Unlike pure software-in-the-loop simulations, ViL requires the DT to ’mimic’ the physical sensor characteristics of a real vehicle. We established a quantitative design target of <10% temporal deviation for the replay system. This threshold is critical; deviations beyond this limit would result in ’ghosting’ effects, where the virtual actor’s position fluctuates inconsistently with the ego-vehicle’s perception cycle (typically 10–20 Hz), rendering the safety validation of high-risk maneuvers unreliable.

To ensure high-fidelity validation of the DT, RTK-GPS was employed as the primary ground truth source, providing a positioning accuracy of approximately 2–3 cm. Synchronization between the physical world and the digital replica was achieved using GNSS-synchronized timestamps for every data packet. During the ’Recording Phase,’ telemetry data (position, velocity, orientation) was timestamped at the source. The ’Replay System’ then utilized these timestamps to re-inject data into the C-ITS platform. The fidelity was verified by measuring the temporal deviation, defined as the lag between the expected playback time and the actual message delivery at the MEC.

2.3. Predictive QoS for Tele-Operated Driving

2.3.1. Use Case Description

The pQoS for ToD use case is centered on a cross-layer architecture that enables the mobile network to provide V2X applications with proactive alerts regarding impending performance degradation at specific geographic coordinates. This framework aims to mitigate the challenges posed by the stochastic nature of mobile networks, ensuring that ToD maintain high Quality of Experience (QoE) and operational safety even when the underlying 5G connectivity is unstable.

The primary scenario involves a Tele-operated Vehicle (ToV) traversing heterogeneous network zones, transitioning from an area of optimal 5G coverage to a region characterized by degraded signal quality. In a conventional setup, a lack of network observability leaves the ToV vulnerable to sudden connectivity loss, which may result in vehicle immobilization and the need for physical recovery (See Figure 3).

2.3.2. System Architecture and Methodology

To address the problem, the TARGET-X framework utilizes a network exposure API to provide the ToV with real-time predictive insights (see Figure 3). By anticipating QoS fluctuations (driven by factors such as radio channel variance, vehicle mobility, and network congestion), the system provides a sufficient temporal window for the Tele-operation Center (ToC) to execute corrective actions. This allows for the proactive adjustment of ToD functions, including longitudinal speed control, steering maneuvers, and safe-stop protocols, thereby preventing abrupt emergency braking and minimizing the risk of road obstructions.

The implementation involves a pQoS function developed within the TARGET-X project, supported by specialized CAMARA-like API [16] that facilitate seamless communication between the 5G network core and the ToC (beyond ToD, pQoS feedback loops can serve as an “early warning system” for cooperative perception, alerting drivers or autonomous systems to potential network-induced reliability drops). The pQoS is based on a NDT that reflects the historical evolution of the network in terms of different network and application KPIs such as signal level, jitter, cell load, etc. The pQoS forecasts the QoS levels within both the ToV’s serving cell and the subsequent neighboring cells along its projected trajectory [17].

The CAMARA-compliant API was developed using the FastAPI framework [18] and comprises the following core components (see Figure 4):

• MASTER API: This component acts as the central orchestrator. Based on the specific parameters within a subscription request, the MASTER API determines the execution logic and the necessary transformation functions (APIs) to be invoked. In this implementation, it coordinates two primary services: the cell ID retrieval API and the QoS monitoring API. The architecture is designed for extensibility, allowing it to interface with additional APIs for varied use cases.
• Cell ID Retrieval API: This service tracks the specific cell ID to which the ToV is connected. To account for vehicle mobility, the cell ID is updated every T seconds. To ensure real-time accuracy, the API subscribes to an MQTT broker that captures and exposes the current cell ID of the connected vehicle.
• Monitoring API: This API monitors predicted metrics (e.g., signal level, latency, throughput, and cell status) for both the current serving cell and the upcoming cell in the vehicle’s path. If any metric falls below the defined thresholds, the API triggers an alarm to the MASTER API. It receives target cell IDs from the MASTER API and queries the NDT for network KPIs and cell status. Cell status is derived from the IDIADA dashboard, which fetches Operations, Administration, and Maintenance (OAM) data at 15-min intervals, providing a snapshot of deployment-grade granularities. Network KPIs (e.g., SINR, RSRP, latency, and throughput) are collected via the MQTT broker from both the primary ToV and a leading vehicle connected to the cell ahead. Notably, in a full Network Data Analytics Function (NWDAF) deployment [19], these KPIs could be retrieved directly using the provided cell IDs.

The modularity of the MASTER API and the update interval T (Figure 4) are designed to balance prediction lead-time with network overhead. In a full NWDAF deployment, the architecture considers an update interval T of 1 s for Cell ID retrieval to ensure the system remains ‘mobility-aware’ even at high speeds. The rationale for the NDT-based Monitoring API is to provide a proactive warning window of several seconds. This lead time is quantitatively necessary to initiate the ‘Safe Stop Protocol’ (Figure 3) which, depending on the vehicle’s mass and speed, requires a controlled deceleration phase that cannot be safely executed if network degradation is only detected reactively.

The pQoS algorithm deployed during the IDIADA field tests served as a deterministic proof of concept. Unlike the ML-driven models implemented in simulation environments (e.g., [17]), the field-tested version utilized a heuristic threshold-triggering mechanism due to the absence of congested operational network data in the IDIADA environment. This algorithm monitors the NDT for specific KPI degradations (such as RSRP falling below a safety-critical floor or E2E latency exceeding a threshold). To mitigate false positives in the ToD use case, a persistence-based heuristic was applied: an alarm is only triggered if the KPI violation is sustained over a defined temporal window, filtering out transient network noise or momentary signal fading.

When the system detects that a ToV is moving from a high-performance cell toward a neighbor cell unable to meet the required throughput or latency thresholds (e.g., due to a site failure or low signal-to-interference-plus-noise ratio), it triggers an a priori alarm. This notification enables the remote driver to respond to environmental constraints in a controlled, preemptive manner, ensuring that vehicle speed is synchronized with the projected network state.

3. Network Implementation

To accurately replicate the diverse and dynamic network conditions encountered in real-world automotive environments, a comprehensive multi-technology access layer is essential for the development and validation of V2X applications. From a network design perspective, this necessitates a testbed capable of supporting both Dedicated Short-Range Communications (compliant with IEEE 1609 and ITS-G5 standards) and a multi-generational cellular infrastructure spanning 2G to 5G NSA.

In alignment with these requirements, the TARGET-X project evaluated its three primary automotive use cases at the IDIADA CVH in Santa Oliva, Spain (see Figure 5). This facility provides a sophisticated proving ground environment integrated with a private 5G network, MEC nodes, and a public hyperscaler platform (AWS) hosted in Malaga. The architectural synergy between the MEC and the hyperscaler is central to the deployment of C-ITS and use-case-specific services. This tiered computing approach allows for the distribution of services based on latency requirements, with time-critical C-ITS functions residing at the MEC, while broader analytical services are handled by the cloud platform.

3.1. Network Infrastructure

The proving ground is served by four multi-standard radio base stations providing comprehensive coverage. To validate the pQoS for the ToD use case, a specific sector (TC S2) is programmatically deactivated to trigger network transitions.

The cellular infrastructure utilizes a combination of mid-band frequencies to balance coverage and capacity. The 4G LTE layer operates on the 1800 MHz (B3) and 2100 MHz (B1) bands, with channel bandwidths of 20 MHz and 10 MHz, respectively. Complementing this, the 5G deployment utilizes the 3500 MHz (N78) band with a 60 MHz bandwidth, providing the high-throughput primary carrier for data-intensive V2X services. For the Time Division Duplex (TDD) frame structure in the N78 band, a DDDSU pattern was implemented (3 Downlink slots, 1 Special slot, 1 Uplink slot) with a 5 ms periodicity. The Sub-carrier Spacing (SCS) was set to 30 kHz, resulting in a slot duration of 0.5 ms.. The cellular infrastructure allows for granular configuration of 4G/5G Carrier Aggregation, including the fallback to single-carrier operation. Multiple Input Multiple Output (MIMO) configurations can be dynamically adjusted from standard 4 × 4 arrays down to 4 × 2 or 2 × 2, shifting performance from baseline to peak rates of approximately 300 Mbps with an average 18 ms latency. By modulating these radio parameters and applying targeted power reductions across all technologies, the testbed provides a deterministic environment for evaluating system performance under non-optimal and adverse coverage conditions.

The TARGET-X MEC facility provides a low-latency environment for leveraging 5G connectivity to meet the stringent end-to-end requirements of automotive use cases. The infrastructure consists of a bare-metal server running Ubuntu 22.04, managed via OpenStack as the virtual infrastructure manager. This setup provides a high-degree of network programmability, supporting advanced configurations such as VLAN tagging, NAT, and custom tunneling protocols (VPN) to meet specific automotive service requirements.

The Teltonika RUTX50 5G router was used in all use cases as a connectivity enabler and/or KPI collection tool.

Beyond steady-state message exchange, the placement of the broker at the MEC significantly enhances service agility (the speed at which a vehicle can establish a functional session). In a V2X context, connection handshakes (TCP/TLS) are highly sensitive to the distance between the client and the server. Because the 5G-MEC architecture reduces the RTT by over 50% (as will be shown in Section 4.1) compared to cloud-based alternatives, the cumulative delay of the multi-step handshake is reduced proportionally. This ensures that vehicles entering the coverage area can synchronize with the Automotive DT or register for pQoS alerts with minimal ’connection-lag,’ a critical requirement for maintaining continuity of service during high-speed mobility.

3.2. ITS Platform

The used ITS platform utilizes a proprietary platform designed to manage standarized messages in compliance with ETSI ITS-G5 standards [20]. The framework supports three primary message types essential for V2X communication:

• CAM: Broadcast periodic dynamic status data, including vehicle position, speed, and heading, to ensure mutual awareness among nearby entities.
• DENM: Triggered by specific events to alert the network of localized hazards, such as accidents, roadworks, or adverse weather.
• CPM: Facilitate the sharing of sensor-detected data regarding non-connected road users (e.g., pedestrians or cyclists), significantly extending the perception horizon of connected vehicles beyond their immediate line-of-sight.

As illustrated in Figure 6, the platform’s operation is defined by three core functional entities: the MQTT broker, the ITS message Generator (ITG), and the ITS message Extractor (ITE).

Geospatial MQTT Broker: The system employs an MQTT broker for efficient message queuing and distribution via a publish-subscribe model. To optimize delivery, a “smart subscription” approach is implemented, integrating geographically encoded information into the topic structure. This encoding utilizes a Quadtree-based data structure to leverage the broker’s native low-latency routing capabilities. By filtering messages based on the subscriber’s current location directly at the broker level, the architecture avoids the overhead of specialized location-monitoring services, ensuring high scalability and minimal end-to-end latency. The primary rationale for the Quadtree-based ’smart subscription’ approach is to shift the computational burden of geospatial filtering from the application layer to the network broker itself. Quantitatively, standard V2X topic structures (e.g., one topic per vehicle) scale poorly, leading to an exponential increase in broker CPU load as the number of CAVs grows.

ITG: The ITG acts as a protocol translation layer, converting non-standardized application data into ETSI-compliant ITS messages. It maps raw input data to standard ETSI formats and units, generating the appropriate message headers and payloads for downstream transmission.

ITE: All messages published to the MQTT broker, either from the ITS Message Manager or from the in-field stations (vehicles, RSUs), are received by this service which converts the ETSI ITS format into a JSON-based representation. This allows external entities to interpret the messages without any knowledge of the ITS standards, easing processing and analysis tasks.

For all V2X communications, MQTT QoS Level 2 was implemented. This choice was dictated by the requirement for message delivery guarantees in safety-critical scenarios (e.g., Zero-Visibility Intersections), where the overhead of a four-way handshake (PUBLISH, 407 PUBREC, PUBREL, PUBCOMP) is preferred over the risk of duplicate message processing. In our 5G-MEC architecture, the impact of this handshake on latency is mitigated by the proximity of the broker. The end-to-end latency $L_{E2E}$ reported in Section 4 can be decomposed as:

$$L_{E2E}=L_{serialization}+L_{ITG}+L_{RAN}+L_{Broker}$$

where $L_{ITG}$ (Information Technology Gateway translation) and $L_{Broker}$ (internal queuing/matching) were estimated to contribute less than 2 ms combined. The dominance of $L_{RAN}$ in the overall budget justifies the focus on 5G vs. 4G performance gains.

4. Performance Evaluation Results

This section presents the performance evaluation results obtained from tests conducted on the IDIADA network. A network KPI collection tool has been implemented to enable real-time measurement, storage, and visualization of network metrics (See Figure 7). This tool comprises the MQTT broker, the KPI generator service, and the Grafana visualization tool. The KPI generator service subscribes to the MQTT broker to collect the measured KPI by the modems integrated in the CAVs and publish it on Grafana Dashboard.

4.1. Results of Cooperative Perception Use Case

This section presents the performance evaluation results of the cooperative perception use case in terms of latency and reliability for CAM, CPM, and DENM. All results are collected across 15 independent test rounds.

4.1.1. Latency

This section presents the service-level latency results obtained from tests conducted in the two scenarios of Use Case 1. Figure 8 depicts the violin plots of CAM and DENM messages in Scenario 1 of the cooperative perception use case. (The violin plot is a statistical visualization that combines the features of a box plot with a kernel density plot. It provides a deeper look at the data distribution by showing where values are most concentrated. The width of the shaded area represents the Kernel Density Estimate (KDE). A “fat” part of the violin indicates a high frequency of data points at that latency value, while a thin part indicates fewer occurrences. The central line in the silhouette represents the median of the data, while the upper and lower lines represent the 25^th and the 75^th percentiles, respectively). As noted in Section 2.1, CPM performance was not evaluated in this scenario. The figure illustrates latency results when the C-ITS service is hosted either at the MEC directly connected to the User Plane Function (UPF) or at an AWS cloud server. Additionally, it demonstrates the performance differences between 4G and 5G connectivity.

Several key observations emerge from the results. First, the average one-way service latency is reduced by half (average dropped from 51.95 ms to 25.76 ms) when utilizing edge deployment compared to cloud hosting. It should be noted that these edge-versus-cloud comparisons are specific to this deployment configuration, where the cloud server is located in Malaga, Spain, and results may vary with different geographic distributions. Second, DENM message latency is consistently lower than that of CAM messages, with average reductions ranging from 25% to 45%. This disparity between message types can be attributed to their transmission directions: DENM messages are transmitted via downlink, while CAM messages utilize uplink channels. Uplink performance is inherently constrained by the regulated frame structure, which typically employs a downlink-to-uplink time slot ratio of up to 3:1, thereby limiting available uplink resources. Furthermore, the random-access procedure required for uplink message transmission introduces additional latency. Third, 5G connectivity demonstrates substantial performance improvements over 4G, with average latency reductions of 75% for CAM messages and 90% for DENM messages.

Figure 9 presents the results obtained from Scenario 2. CPM performance was exclusively evaluated using 5G connectivity with the C-ITS service deployed in the cloud. The results corroborate the findings from Scenario 1, demonstrating superior performance with 5G compared to 4G (Average dropped from 30.46 ms to 23.83 ms or 21.7% for CAM and from 22.52 ms to 20.21 ms or 10% for DENM) and with edge deployment compared to cloud hosting (Average dropped from 54.15 ms to 23.83 ms or 55.9% for CAM and from 28.67 ms to 20.21 ms or 29.5% for DENM). Consistent with Scenario 1, DENM messages exhibit lower latency than CAM messages. Furthermore, CPM demonstrates latency characteristics closely resembling those of CAM, which is expected given that both message types are transmitted via the uplink channel.

The reduction in mean latency from 51.3 ms (Cloud) to 25.8 ms (MEC) in Scenario 1 and from 54.15 ms (Cloud) to 23.83 ms (MEC) in Scenario 2 is not merely a performance gain but a safety requirement. At an urban speed of 50 km/h, a reduction of 25.5 ms in Scenario 1 in message delivery equates to an additional 0.35-m safety buffer. While this seems small, in the context of the Zero-Visibility Intersection scenario (Figure 1), this buffer can be the deciding factor between a successful AEB intervention and a collision. Furthermore, the 5G MEC configuration maintained a 95^th percentile latency below 60 ms, ensuring that even in ’worst-case’ jitter scenarios, the vehicle receives cooperative perception data within one-and-a-half perception cycles (assuming a 20 Hz sensor rate).

Beyond average values, the statistical distributions in both scenarios provide critical insights into the system’s determinism and worst-case reliability, as summarized in

Table 1. V2X Performance Comparison (Latency in ms).

	SCENARIO 1		SCENARIO 2
	95%-Tile	Max	95%-Tile	Max
CAM CLOUD–5G	55.8	56.6	58.8	61.1
CAM EDGE–5G	27.7	28.9	25.8	26.7
CAM EDGE–4G	36.0	37.0	32.6	33.6
DENM CLOUD–5G	29.7	29.7	32.8	33.0
DENM EDGE–5G	20.9	21.2	23.1	24.6
DENM EDGE–4G	22.9	23.2	24.7	26.0
CPM CLOUD–5G	—	—	56.7	59.3

:

• Determinism and Jitter: Across both scenarios, the 5G-Edge (MEC) configuration demonstrated superior stability, with a Standard Deviation (STD) consistently between 1.0 and 1.6 ms. This is reflected in the narrow gap (typically <1.5 ms) between the 95^th percentile and the absolute maximum latency. In contrast, 5G-Cloud configurations exhibited nearly double the jitter (STD up to 2.8 ms), illustrating how public backhaul fluctuations introduce non-deterministic delays.
• Safety-Critical Reliability: The worst-case performance (maximum latency) remained well within safety-critical bounds at the Edge. For DENM messages, the maximum recorded latency was 21.2 ms (Scenario 1) and 24.6 ms (Scenario 2), ensuring emergency alerts are delivered in less than half the time of a standard 50 ms perception cycle. Even for more complex CAM and CPM messages, the 5G-Edge maximum (28.9 ms) stayed significantly below the 95^th percentile of the 4G-Edge alternative (32.6 ms), proving that 5G-MEC is essential for maintaining a sub-50 ms reaction buffer.
• Protocol Efficiency: The analysis of CPM in Scenario 2 validates the efficiency of the ITG translation and MQTT broker. Despite the higher data complexity of CPM compared to CAM, the maximum latency (59.3 ms) remained within a 3% margin of the CAM Cloud maximum (61.1 ms). This confirms that the serialization processes and Quadtree-based filtering are highly optimized, preventing complex payloads from inducing non-linear delays that would jeopardize the V2X validity window.

Beyond one-way latency, we evaluated ’trigger latency’ defined as the temporal interval between the reception of a DENM by a vehicle and the initial CAM/CPM transmission that triggered the collision warning. Crucially, trigger latency incorporates internal processing overhead, making it inherently greater than the arithmetic sum of the individual CAM/CPM and DENM transmission delays. To provide a rigorous definition of the metrics evaluated in the trials, Figure 10 presents the sequence diagram for the V2X trigger latency. The measurement interval begins at $T_{start}$, defined as the timestamp when a CAM or CPM is generated at the application layer of the source vehicle. The interval concludes at $T_{end}$, when the responding DENM is received and parsed at the application layer of the recipient vehicle.

This ‘Service-level’ approach incorporates the end-to-end communication delay and the internal processing overhead of the C-ITS platform (ITG translation and MQTT broker routing). Consistent with our objective to benchmark 5G and MEC performance, this measurement excludes exogenous factors such as raw sensor data collection (Perception Layer) and mechanical braking response (Control Layer), which are dependent on specific vehicular hardware rather than the network architecture.

Figure 11 presents the trigger latency distributions via violin plots for both cooperative perception scenarios. The results mirror the trends observed in one-way latency, demonstrating the superior performance of 5G over 4G; specifically, average latency decreased by 8.9% (from 140.33 ms to 127.85 ms) in Scenario 1 and by 5.6% (from 152.27 ms to 143.73 ms) in Scenario 2. Edge deployment further optimized performance compared to cloud hosting, reducing average latency by 28.7% (from 179.47 ms) in Scenario 1 and by 43.7% (from 255.27 ms) in Scenario 2 when utilizing CAM. Notably, in Scenario 2, the use of CPM resulted in a significantly higher average trigger latency compared to CAM, representing a 24.2% increase (from 255.27 ms to 317.15 ms), which is mainly due to the processing time difference between the two types of messages.

The design of the ITG and the geospatial broker prioritizes low-computational complexity to minimize the ‘Internal Processing Overhead’ mentioned in Section 4. While the end-to-end trigger latency is the primary metric reported, it encompasses the full computational chain: application-layer serialization, ITG protocol translation, broker routing, and final delivery. The processing impact of the middleware is observed in the variance between message types; CPM messages, which require more intensive serialization of environmental objects than CAM messages, exhibited a 24.2% higher latency. This indicates that while middleware overhead is a measurable component of the E2E budget, the Quadtree-based broker architecture ensures it does not become the primary bottleneck in 5G-MEC deployments.

4.1.2. Reliability

The reliability is defined as as the success rate of message transmission, representing the quotient of correctly received packets over the total packets sent during the evaluation period. In all cases the reliability is higher than 99.9%. The exceptionally low standard deviation and high average reliability (approaching 100%) resulted in insufficient visual variance for graphical distribution analysis, and therefore violin plots were omitted. These results demonstrate that the 5G RAN and the MEC-hosted broker effectively manage the trade-off between speed and reliability, ensuring that the low-latency occurs within a robust transmission environment that adheres to the safety-critical requirements.

4.2. Results of Automotive DT Use Case

This section evaluates the repeatability of the automotive DT through the analysis of temporal deviation, serving as the primary KPI for synchronization fidelity. The deviation is defined as the average deviation in message periodicity between the original and the replayed messages.

Figure 12 illustrates the distributions of these deviations (specifically for the average, median, and 95^th percentile) across 17 independent replays within the DT. The results are presented as (a) absolute temporal offsets in milliseconds and (b) percentage deviations relative to real-time measurements. The results indicate high fidelity in the DT’s performance; the central tendencies (average and median) demonstrate high accuracy, with percentage deviations consistently remaining below 10% (under 7 ms) and averages frequently dropping below 5% (approximately 4 ms). While the percentage deviation for the 95^th percentile exhibits greater variance, the absolute value plot shows this is only 14 ms. In the context of C-ITS, where message intervals are often 100 ms, a 14 ms gap indicates that the DT remains “safe” for testing most collision-avoidance logic.

The experimental results for the DT synchronization demonstrate exceptional fidelity, with a median temporal deviation below 7 ms. To put this into perspective, a standard automotive perception loop operates at 10–20 Hz (50–100 ms intervals). A 7 ms deviation implies that the DT is synchronized within less than 15% of a single sensor frame cycle. Quantitatively, this means that even at high speeds (e.g., 100 km/h), the positional error attributed to network and processing lag is less than 20 cm. This level of precision validates our architectural choice of placing the DT synchronization logic at the MEC, as it successfully eliminates the jitter and ’ghosting’ effects that would otherwise compromise the safety-critical validation of AEB and other V2V maneuvers.

4.3. Results of the pQoS for ToD Use Case

This section presents the performance evaluation results of the pQoS for ToD. The performance study focused on the throughput and the jitter of video transmission on uplink and the latency of the command on downlink. The scientific validity of the reported latency values is supported by the high reliability of the communication links, measured via reliability. The reliability for critical control commands was in all cases 100%, meaning all steering and braking instructions were successfully delivered without erasure. For the high-bandwidth video feedback stream required for ToD, the average reliability was measured at 99.9%.

In addition to the evaluation in IDIADA CVH environment, the pQoS was also evaluated through simulation using network KPIs provided by a European operator for a dense urban area in [17]. In this paper, we compared the following three scenarios:

• Scenario 1 (No NDT): In this scenario, Sector TC S2 in Figure 5 is turned off and the network exposure API is deactivated.
• Scenario 2 (UE-based NDT): In this scenario, Sector TC S2 is turned off and the network exposure API is activated using the information collected from the CAVs located in the big roundabout with no coverage.
• Scenario 3 (Network-based NDT): In this scenario, Sector TC S2 is turned off, and the network exposure API is activated using the information collected from the network through the network OAM.

4.3.1. Video KPIs

This section analyzes the uplink throughput and jitter results for the three cameras utilized across all experimental scenarios.

Figure 13a presents the violin plots for uplink video throughput. In all scenarios, the average throughput remained within the 10–50 Mbps range; specifically, 16.4 Mbps for the No NDT scenario, 19.2 Mbps for the UE-based NDT scenario, and 19 Mbps for the Network-based NDT scenario. A critical distinction arises in the 5^th percentile observed values: while scenarios 2 and 3 maintained values above 16 Mbps (16.6 Mbps and 18.1 Mbps respectively), the throughput in scenario 1 plummeted to 6.7 Mbps.

A key finding is the significantly higher stability of video throughput when predictive QoS is implemented. Compared to the baseline (Scenario 1), the standard deviation of throughput decreased by 67% in Scenario 2 and 76% in Scenario 3. This reduction in performance fluctuation highlights the efficacy of proactive notifications in maintaining link quality.

To further investigate this behavior, Figure 14 illustrates the throughput evolution during a controlled cell deactivation event. As the vehicle approached the low-coverage zone in Scenario 1, the throughput dropped below 5 Mbps (well under the 10 Mbps threshold required for effective remote visualization). Lacking predictive alerts, the remote driver continued into the degraded coverage area, leading to an abrupt collapse of throughput, video artifacts, and eventual loss of the feed, which necessitated an emergency intervention by the safety driver. Conversely, in Scenarios 2 and 3, the network exposure API provided timely notifications. This allowed the remote driver to halt the vehicle before entering the dead zone, ensuring that throughput remained stable and the user experience was uninterrupted. Consequently, the driver successfully executed a controlled parking maneuver without service degradation.

The impact on video stability is further evidenced in Figure 13b, which depicts the violin plots for uplink video jitter. The “No NDT” scenario exhibited an average jitter of 20.8 ms, with extreme peaks reaching 167 ms. In contrast, jitter remained consistently below 23 ms, with averages of 7.7 ms for network-based NDT and 8.8 ms for UE-based NDT (A reduction of 63% and 58%, respectively). Analysis of the 95^th percentile further validates this: the network-based and UE-based NDT recorded values of 10.4 ms and 17.4 ms, respectively, while the “No NDT” scenario reached 92.6 ms. Given that low jitter is paramount for ToD, these results underscore the critical role of exposure APIs in guaranteeing the smooth, low-latency video streaming necessary for safe remote vehicle operation.

It should be noted that the video uplink average latency was much lower than the threshold of 100 ms in all scenarios: the highest was in no NDT scenario with 58.3 ms. This is because the safety driver stopped the ToV before the latency becomes very high.

4.3.2. Command KPIs

This section analyzes the downlink latency results across all experimental scenarios.

The experimental results for command latency across the three evaluated scenarios are illustrated in Figure 15. In the two scenarios utilizing network exposure APIs, namely UE-based NDT and network-based NDT, the average latency values remained consistently within the 20–50 ms range, recording averages of 38 ms and 36 ms, respectively. Notably, the 95^th percentile values for these scenarios also stayed within this critical range, at 49 ms and 50 ms. These results align with the stringent low-latency requirements necessary for maintaining responsive control in ToD applications.

In contrast, the baseline scenario (No NDT) demonstrated significantly degraded performance during cell deactivation. In this case, the average command latency rose to 65 ms, while the 95^th percentile surged to 149 ms. Such high latency values and extreme outliers are generally considered unacceptable for ToD, as they can lead to delayed vehicle responses and compromised safety.

4.3.3. Scalability Analysis

To evaluate the scalability of the proposed architecture, we analyzed the resource footprint of the 20 Mbps ToD video stream relative to the 5G NR physical layer capacity. The deployment utilized a 60 MHz channel with a 30 kHz SCS, yielding $N_{PRB}=162$. Under the DDDSU TDD frame configuration, which allocates approximately 20% of the temporal resources to the uplink, the theoretical peak UL throughput for the sector is approximately 100 Mbps. Given the 95^th percentile recorded bitrate of 20 Mbps for the ToD video feedback, a single vehicle occupies roughly 20% of the available uplink PRB pool. This utilization rate indicates that the 5G-MEC architecture can concurrently support five tele-operated vehicles per cell sector. This level of scalability is sufficient for initial commercial fleet deployments. For higher-density urban scenarios, capacity could be further expanded by transitioning to a more uplink-centric TDD pattern (e.g., DDSUU) or by utilizing the wider bandwidths available in the mmWave bands.

5. Challenges and Lessons Learned

While the TARGET-X trials demonstrate that 5G NSA and MEC provide the foundational gains necessary for initial CCAM services, achieving average latency reductions of up to 90% compared to legacy 4G, the results also expose fundamental architectural bottlenecks that define the 6G research frontier. Transcending the incremental ’5G-over-4G’ paradigm, our empirical findings identify critical architectural gaps and lessons learned that constitute the primary design imperatives for beyond-5G and 6G networks.

5.1. Beyond Best-Effort to 6G Determinism

Our trials confirm that the integration of 5G with MEC is essential for maintaining latencies in the 40–50 ms range for C-ITS services. However, for high-level (Level 4/5) autonomy and high-speed cooperative maneuvers, even a 50 ms “average” is insufficient. 6G orchestration must move from current “best-effort” low latency toward deterministic sub-10 ms end-to-end cycles.

The “uncontrollable fluctuations” observed in our cloud-based tests highlight a critical architectural gap to maintain performance stability under varying loads: the need for AI-powered, context-aware orchestration across the IoT-MEC-cloud continuum. The efficacy of such an orchestrator is underpinned by its ability to ingest Environmental and Operational contexts as high-dimensional input features. This allows the network to move beyond reactive load-balancing toward a proactive, deterministic framework that guarantees required QoS regardless of external volatility.

Environmental context defines the physical obstacles and signal integrity constraints facing the system. It answers: Can the sensor “see” and can the signal “reach”?. It focuses on external physical conditions (e.g., weather, terrain, or electromagnetic interference) that impact the reliability of hardware sensors and the integrity of wireless propagation. It acts as a reliability trigger, signaling the need for redundant data streams or high-bandwidth remote assistance. For example, heavy fog or rain may degrade local LiDAR/camera reliability, prompting the orchestrator to preemptively allocate high-bandwidth streams for remote sensor fusion or tele-operated assistance.

Conversely, operational context defines the stakes of the current mission. It answers: How fast must we decide and who is involved?. It focuses on dynamic system states or task-specific demands (e.g., traffic density, maneuver complexity, or safety-criticality) that dictate the necessary performance bounds and resource priority. It acts as latency-sensitivity weights that drive real-time workload migration across the IoT-MEC-cloud continuum. A vehicle entering a high-density intersection, for instance, triggers an immediate workload migration from the Cloud to the MEC to minimize processing jitter during a critical maneuver.

Furthermore, the validation of ToD use case confirms that deterministic performance is fundamentally contingent upon a tightly coupled, bidirectional data exchange between the mobile network and applications. This identifies a vital design gap for 6G: the transition from passive exposure to active synchronization.

• Architectural Symbiosis (Beyond Exposure): It is no longer sufficient to merely “expose” network statistics. Real-time telemetry regarding localized congestion and resource availability must be ingested by the application’s logic, while the application must simultaneously inject its performance requirements directly into the network’s orchestration plane. This enables a proactive “early warning system” where the network and application negotiate a graceful degradation or a “safe state” transition (e.g., for an autonomous vehicle) before a physical link failure occurs.
• Beyond CAMARA (The Negotiation Gap): While current frameworks like the CAMARA project [16] provide standardized “Northbound APIs” for status exposure, they remain largely transactional. 6G requires rich APIs that support iterative negotiation, allowing the application to request specific performance requirements and the network to provide “counter-offers” based on predictive telemetry.
• Evolution of pQoS into “Self-Healing” Orchestration: The use of pQoS as an early warning system remains a reactive control loop where the application (ToD) must adjust to a network alert. The 6G architectural gap lies in Autonomous Orchestration, where the network does not only “alert” the application to a 15-min “coarse” degradation, but uses sub-second streaming telemetry (via NWDAF-like functions) to proactively migrate MEC workloads or adjust beamforming before the user experience is impacted.

5.2. Beyond Static Pipes: Digital Twin for Context-Aware 6G V2X Slicing

The transition from 5G to 6G marks a move from rigid, “blind” resource allocation to a fluid, context-aware ecosystem. Current 5G deployments suffer from a “Static Allocation Gap,” where network slices act as inflexible pipes, oblivious to the physical reality of the vehicle. To resolve this, 6G must integrate a DT plane that allows for real-time adaptation based on environmental shifts.

• From Rigid Pipes to Context-Aware Slicing: In 5G, a slice is often a static reservation that does not differentiate between a vehicle on an empty highway and one caught in a sudden traffic jam or heavy rain. These environmental factors drastically alter the QoS. 6G architecture must bridge this gap by making slice configuration dynamic and context aware in a similar way to the orchestrator in the previous section. For instance, as vehicle density increases, the network can proactively expand the V2X slice capacity to handle the surge in basic safety messages. Another example, at the start of rain, the system can trigger a higher-reliability profile (lower latency, higher redundancy) to compensate for the increased risk of accidents and potential signal degradation.
• The “What-If” Engine: Simulating Impact via NDT: A core innovation in this 6G vision is the use of a DT as a “What-If” engine. Before any slice reconfiguration is pushed to the live network, the DT simulates the change. It should be able to provide an impact analysis by predicting how expanding a V2X slice will affect existing users (e.g., public voice users or industrial IoT sensors). It should also be able to perform conflict resolution: If the simulation predicts a drop in reliability for safety-critical services, the engine must suggest an optimized configuration that balances the load without compromising safety.

5.3. Transitioning from Coarse to Fine-Grained Analytics

A major technical barrier remains the temporal granularity of current network analytics.

• The “Minute vs. Second” Gap: Conventional network dashboards often provide updates at 15-min intervals, a “coarse” granularity sufficient for general monitoring but inadequate for safety-critical V2X applications.
• Real-time Telemetry: Future 5G/6G functions must support sub-second, fine-grained analytics (e.g., instantaneous SINR and scheduling delays) to enable responsive edge intelligence. This requires a shift from batch-based polling to streaming telemetry and standardized APIs. Whether utilizing the Network Data Analytics Function (NWDAF) or NDT, data acquisition should be on-demand and AI-triggered to minimize computational and memory overhead.

5.4. TDD Frame Mismatch: Addressing Uplink-Heavy V2X Demand

The trial results highlighted a fundamental architectural conflict between existing public network configurations and the requirements of advanced CCAM applications.

• The Asymmetry Paradox: Most commercial 5G deployments utilize a Time Division Duplex (TDD) frame configuration that is heavily biased toward the Downlink (DL) to accommodate consumer data consumption (e.g., video streaming, web browsing). However, critical V2X use cases are inherently Uplink (UL) heavy. For example, ToD massive UL throughput (ranging from 15 to 50 Mbps per vehicle) to stream high-definition camera feeds and LiDAR sensor data to the edge/cloud, while requiring minimal DL bandwidth for control commands.
• Infrastructure Recommendations: This mismatch leads to rapid UL resource exhaustion in public networks. Recent research has identified this as a critical bottleneck, proposing machine-learning-based approaches to dynamically optimize uplink scheduling and resource allocation to handle such vertical-specific demands [21]. To support large-scale automotive deployments, there is a critical need for:

  –

  Flexible TDD Configurations: Implementing more symmetric or dynamically adjustable TDD patterns within existing bands to better serve vertical industries.

  –

  Dedicated Spectrum: The allocation of specific frequency bands dedicated to industrial and automotive use cases, ensuring that safety-critical UL traffic is not throttled by consumer DL demand.

5.5. Evolution from Basic Awareness to Collective Perception

The trials highlighted the inherent limitations of the legacy CAM and DENM framework. In high-stakes scenarios, such as vehicles on a collision course at an intersection, CAM/DENM-based systems typically rely on constant-velocity assumptions and limit alerts to the immediate actors involved. To ensure robust safety in complex urban environments, the following architectural shifts are recommended:

• Rich Messaging Frameworks: There is a fundamental requirement to transition toward richer message types, such as the CPM and the Sensor Data Sharing Message (SDSM). These allow vehicles to share processed sensor data, effectively extending the “line of sight” for all connected actors.
• Centralized Intelligence via Traffic Management Center (TMC): Achieving true cooperative perception requires a cloud-based TMC. By acting as a C-ITS orchestrator, a TMC can aggregate multi-source data to infer global traffic patterns and distribute dynamic safety policies, bridging the gap between individual vehicle awareness and system-wide safety.

Results from Scenario 2 indicate that perception occlusion caused by fog renders the system incapable of identifying and reporting non-connected obstacles. To address this vulnerability, future research must prioritize the development of advanced multi-modal or indirect detection techniques to ensure safety in low-visibility environments.

5.6. Bridging the Semantic Gap: Mapping Network to Service KPIs

A significant challenge identified during the trials was the “semantic gap” between low-level network metrics (latency, jitter, packet error rate) and high-level service performance (e.g., ToD responsiveness or perception accuracy).

• Correlation Modeling: Future deployments must establish systematic, cross-layer KPI correlation models. These models should link stochastic network behavior with deterministic service outcomes under varying mobility patterns and environmental conditions.
• Automated Service-Level Agreement (SLA) Assurance: Developing these analytical models is a prerequisite for automated SLA assurance, allowing the network to proactively adjust resources before service degradation occurs.

5.7. The Scalability Gap: Performance Degradation in Dense V2X Environments

While our controlled trials at IDIADA provided a baseline for 5G performance, transitioning to large-scale urban deployments introduces non-linear performance degradation. The presence of thousands of concurrent V2X nodes affects the system through four primary vectors:

• Radio Resource Exhaustion: In our trials, the ego-vehicles had nearly exclusive access to the Physical Resource Blocks (PRBs). However, in a dense urban environment thousands of vehicles requesting high-bandwidth Uplink (e.g., for ToD) and frequent messaging (e.g., CAM/CPM) lead to PRB exhaustion. This results in increased scheduling delays and packet loss as the 5G gNodeB struggles to allocate resources within the TDD frame. The 90% latency reduction observed in our trials could quickly evaporate if the MAC-layer scheduler becomes a bottleneck.
• Signal Integrity and Interference Floor: High vehicle density significantly raises the interference floor. Even with advanced beamforming, the cumulative “noise” from thousands of sidelink (V2V) and cellular (V2I) transmissions degrades the Signal-to-Interference-plus-Noise Ratio (SINR). Lower SINR forces the system to use more robust but lower-order Modulation and Coding Schemes (MCS). This reduces the effective throughput, making high-definition sensor streaming for ToD unreliable or impossible without massive infrastructure densification.
• The “Message Storm” and MEC Processing Jitter: A single intersection with 200 connected vehicles can generate thousands of (CAMs/CPMs) per second. This “message storm” puts immense pressure on the MEC application server. We observed 40–50 ms latencies with few actors, but under mass-scale load, the MEC compute jitter increases. The time required for the Traffic Management Center (TMC) to fuse data from thousands of sources can exceed the “freshness” requirements of safety-critical DTs.
• Impact on pQoS Reliability: Our pQoS mechanism relies on the network’s ability to predict its own state. In a highly dynamic, large-scale environment, the temporal validity of a pQoS alert shrinks. If the network state changes faster than the API can “expose” it, the $63\%$ reduction in video jitter we achieved may drop. This necessitates a move toward the sub-second streaming telemetry we identified as a 6G requirement, as traditional polling-based APIs cannot keep up with urban-scale volatility.

By identifying these specific degradation vectors, it becomes clear why physical trials must be coupled with DTs. While physical testing serves as a high-fidelity ’unit test’ for individual node performance, the DT leverages these empirical parameters to simulate emergent phenomena, such as ’Message Storms’ or ’Resource Exhaustion’, across thousand-node architectures. To bridge this gap, we propose a hybrid scalability roadmap where the DT translates controlled local data into actionable insights for urban-scale complexity.

• The automotive DT as a Scalability Bridge: The primary value of the proposed ViL framework is its ability to bridge the gap between controlled trials and urban complexity. By achieving high-fidelity synchronization (below 10% deviation) in a controlled environment, we establish the DT as a “ground-truth” emulated environment. Once calibrated, these twins allow for the high-fidelity scaling of dense vehicular traffic and dynamic radio environments, to test resilience and “edge-case” scenarios without the prohibitive costs or safety risks of full-scale physical infrastructure. The proposed DT framework allows us to simulate the “Message Storms” by flooding the MEC application with thousands of virtual V2X messages, testing the limits of the Traffic Management Center’s (TMC) processing capacity.
• Hybrid Emulation Strategy: Future 6G validation must adopt a hybrid approach where the NDT simulates thousands of background V2X users (generating congestion and interference) while the physical “Ego-vehicle” interacts with this virtual load in real-time. This allows researchers to test if the pQoS and MEC orchestration logic remains robust under “urban-scale” stress that is physically impossible to replicate on a proving ground.

6. Conclusions

This paper presented the end-to-end design and validation of advanced CCAM use cases (Cooperative perception, Automotive Digital Twin, and predictive QoS) within the 5G-enabled TARGET-X trial platform. Our findings demonstrate that 5G technology is not merely a connectivity upgrade but a foundational enabler for high-fidelity vehicular synchronization and safety-critical orchestration.

The experimental evaluation at the IDIADA facility confirms that achieving seamless CCAM requires a tiered computing architecture. We demonstrated that the integration of 5G networks with MEC resources is essential for maintaining latencies within the required 40–50 ms thresholds for C-ITS services, representing a 75–90% improvement over 4G and cloud-based deployments. However, our study also reveals significant challenges in spectrum utilization; the current DL-heavy TDD configurations of public networks are ill-suited for the UL-heavy demands of ToD and sensor sharing.

Our results also highlight a critical trade-off: while cloud hosting offers broader reach, its inherent latency fluctuations and lack of sub-second controllability pose risks to cooperative perception accuracy. In summary, the TARGET-X trials provide a scalable blueprint for the transition from localized V2X experiments to resilient, large-scale automotive digital ecosystems.

While this study established the foundational proof of concept, several research avenues remain to be explored. Our current pQoS implementation served as a proof of concept without Machine Learning (ML). Future research will focus on integrating ML algorithms trained on operational network data to improve the accuracy of “early warning” alerts for reliability drops. Furthermore, we plan to refine the KPI correlation models mentioned in the recommendations to enable Automated SLA Assurance, where the network autonomously adjusts beamforming or edge resources before a service-level agreement is violated. Finally, we aim to prototype and evaluate dynamically adjustable TDD patterns in a controlled private network environment to quantify the performance gains for UL-heavy ToD services without compromising DL stability for other users.