Reliable V2X-Enabled Autonomous Driving with Explainable Federated Neuro-Symbolic Intelligence
TL;DR
Simulation-based V2X results show that the proposed method improves decision quality, reduces missed safety-critical states, and converges faster than standard Federated Averaging (FedAvg) under vehicle-side data that are not independent and identically distributed (non-IID).
Abstract
Connected and autonomous vehicles rely on vehicle-to-everything (V2X) cooperation when local sensing is incomplete or delayed, as often occurs in dense traffic, occluded intersections, and roads with constrained communication. Yet many data-driven driving models remain difficult to interpret, and centralized learning requires large raw sensory data transfers that raise communication and privacy concerns. This paper proposes an explainable federated neuro-symbolic intelligence framework for reliable V2X-enabled autonomous driving. The framework combines neural perception, symbolic rule-based reasoning, fuzzy uncertainty modeling, and reliability-aware federated learning. Each vehicle trains a local decision agent from local driving and communication-context observations, while an edge or roadside unit aggregates model updates without collecting raw sensory data. The local agent also produces an interpretable rationale, including activated rules, uncertainty indicators, and risk factors associated with the selected decision. Simulation-based V2X results show that the proposed method improves decision quality, reduces missed safety-critical states, and converges faster than standard Federated Averaging (FedAvg) under vehicle-side data that are not independent and identically distributed (non-IID).
