VariSAC: V2X Assured Connectivity in RIS-Aided ISAC via GNN-Augmented Reinforcement Learning

Huijun Tang, Wang Zeng, Ming Du, Pinlong Zhao, Pengfei Jiao, Huaming Wu, Hongjian SunarXiv 2025 · DOI · arXiv

The Challenge of Urban V2X Connectivity

In the evolving landscape of autonomous driving, vehicles must simultaneously communicate with infrastructure and other vehicles while sensing their immediate environment. This paradigm, known as Integrated Sensing and Communication (ISAC), is essential for safety and coordination. However, urban environments present significant hurdles: dense buildings and foliage create Non-Line-of-Sight (NLoS) conditions, and high vehicle mobility leads to rapid signal fluctuations and frequent link failures.

Traditional approaches often prioritize instantaneous signal quality, such as peak data rates or signal-to-noise ratios (SNR) at a single point in time. While useful, these metrics fail to capture the temporal continuity required for safety-critical autonomous operations. A high-speed connection that drops every few seconds is far less valuable than a stable, lower-speed link that remains uninterrupted.

System model of RIS-aided ISAC in an urban environment, highlighting the Continuous Connectivity Ratio (CCR) metric.

To address this, the authors introduce the Continuous Connectivity Ratio (CCR). This unified metric measures the proportion of time a vehicle maintains a connection above a required threshold over a sliding temporal window. By focusing on sustained connectivity, CCR bridges the gap between the high-quality environmental perception needs of Vehicle-to-Infrastructure (V2I) links and the low-latency, reliable delivery requirements of Vehicle-to-Vehicle (V2V) safety messages.

The VariSAC Framework

The proposed VariSAC framework leverages advanced machine learning to maintain these stable links in chaotic urban settings. The architecture is built upon three core components:

  1. Graph Construction and Feature Extraction: The V2X network is modeled as a dynamic graph where vehicles, base stations, and Reconfigurable Intelligent Surfaces (RIS) are nodes. This captures the complex spatial dependencies inherent in moving traffic.
  2. GNNRA (GNN with Residual Adapters): Standard Graph Neural Networks (GNNs) can suffer from "over-smoothing," where node features become indistinguishable as the network grows. VariSAC introduces Residual Adapters within the GNN architecture to preserve distinct node features and learn more intricate spatial relationships from high-dimensional data.
  3. Soft Actor-Critic (SAC) Agent: The decision-making "brain" of the system uses Deep Reinforcement Learning (DRL). The SAC agent jointly optimizes channel allocation, transmit power, and the phase shifts of RIS nodes. By dynamically "bending" signals around obstacles, the RIS creates "virtual" line-of-sight paths, significantly mitigating NLoS issues.
The VariSAC framework architecture, integrating graph construction, GNNRA for feature extraction, and SAC for decision-making.

Key Insights and Results

VariSAC was rigorously tested using simulated urban roadmaps and real-world vehicle trajectories from the Didi GAIA Open Dataset. These evaluations yielded several critical insights:

  • Superior Reliability: VariSAC consistently outperformed baseline models. It achieved an 8.38% improvement in total connectivity ratio over DDPG-based agents and a 15.54% improvement over traditional greedy algorithms.
  • Scalability in Dense Traffic: As vehicle density increased, VariSAC's performance remained robust. While other models struggled with the exponential increase in interference and search-space complexity, VariSAC’s GNN-based spatial awareness allowed it to maintain high connection reliability.
  • The Power of Spatial Awareness: The inclusion of the GNNRA module proved vital. By explicitly modeling the network topology, the agent could better understand how its actions (like adjusting an RIS) would affect connectivity across the entire local graph, rather than just for a single vehicle.

Conclusion and Future Directions

VariSAC demonstrates that "continuous connectivity" is a viable and necessary goal for future V2X systems. By shifting the focus from instantaneous performance to temporal stability and leveraging the spatial intelligence of GNNs, the framework ensures that autonomous vehicles remain connected even in the most challenging urban environments.

Future research may explore decentralized control schemes to reduce signaling overhead in massive-scale deployments and integrate more sophisticated sensing models to further enhance the "seeing" capabilities of the ISAC paradigm.

Summary updated

Abstract

The integration of Reconfigurable Intelligent Surfaces (RIS) and Integrated Sensing and Communication (ISAC) in vehicular networks enables dynamic spatial resource management and real-time adaptation to environmental changes. However, the coexistence of distinct vehicle-to-infrastructure (V2I) and vehicle-to-vehicle (V2V) connectivity requirements, together with highly dynamic and heterogeneous network topologies, presents significant challenges for unified reliability modeling and resource optimization. To address these issues, we propose VariSAC, a graph neural network (GNN)-augmented deep reinforcement learning framework for assured, time-continuous connectivity in RIS-assisted, ISAC-enabled vehicle-to-everything (V2X) systems. Specifically, we introduce the Continuous Connectivity Ratio (CCR), a unified metric that characterizes the sustained temporal reliability of V2I connections and the probabilistic delivery guarantees of V2V links, thus unifying their continuous reliability semantics. Next, we employ a GNN with residual adapters to encode complex, high-dimensional system states, capturing spatial dependencies among vehicles, base stations (BS), and RIS nodes. These representations are then processed by a Soft Actor-Critic (SAC) agent, which jointly optimizes channel allocation, power control, and RIS configurations to maximize CCR-driven long-term rewards. Extensive experiments on real-world urban datasets demonstrate that VariSAC consistently outperforms existing baselines in terms of continuous V2I ISAC connectivity and V2V delivery reliability, enabling persistent connectivity in highly dynamic vehicular environments.
BibTeX
@misc{tang2025varisac,
  author = {Huijun Tang and Wang Zeng and Ming Du and Pinlong Zhao and Pengfei Jiao and Huaming Wu and Hongjian Sun},
  title = {{VariSAC: V2X Assured Connectivity in RIS-Aided ISAC via GNN-Augmented Reinforcement Learning}},
  year = {2025},
  eprint = {2509.06763},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.06763},
}
Figure