Toward Safety-Aware Cross-Layer Multi-Agent Learning for V2X Communications

Abstract

Reliable vehicle-to-everything (V2X) communication is critical for connected and automated driving, where dynamic network topologies, varying traffic densities, and wireless interference can adversely affect latency and stability. While existing learning-based solutions have improved functions such as task offloading and resource allocation, they typically optimize protocol layers in isolation and fail to incorporate safety-oriented control constraints. This paper proposes a cross-layer reinforcement learning framework that jointly performs physical-layer adaptation, medium access control (MAC) congestion control, and edge task offloading. The framework combines constrained multi-agent reinforcement learning with a Lyapunov-based stability mechanism to regulate latency and queue dynamics under time-varying conditions, while a multi-head attention module captures interactions among communication, computation, and vehicular control states. Evaluated in ns-3 under urban and highway scenarios, the proposed approach reduces average end-to-end latency from 13.2 ms to 9.4 ms in the urban scenario and improves packet delivery ratio from 92.1% to 98.3% under the evaluated settings. Furthermore, the safety-aware policy reduces simulated control-instability events and maintains robust communication performance under dynamic traffic conditions, demonstrating the effectiveness of integrating cross-layer learning with control-aware safety constraints for reliable V2X communication and edge intelligence.