E2E-AD: An End-to-End Implicit Interpretability Framework for Open-Road Autonomous Driving

The rapid adoption of end-to-end autonomous driving systems (ADSs) has highlighted a critical challenge: the lack of interpretability in black-box models poses significant safety and trust concerns, particularly in open-road environments. While existing explainable AI (XAI) methods provide post-hoc interpretations, they often suffer from real-time latency and fail to establish intrinsic connections between explanations and driving decisions. To address this gap, this paper proposes End-to-End Autonomous Driving (E2E-AD), an end-to-end autonomous driving framework with implicit interpretability, which seamlessly integrates decision-making and explanation generation within a unified pipeline. The core innovation lies in embedding interpretability directly into the model’s latent representations through spatiotemporal attention mechanisms. The framework consists of three key components: a multimodal feature alignment module for robust perception, an attention trajectory backtracking algorithm for decision tracing, and an online consistency validator to ensure plausibility of explanations. By jointly optimizing driving performance and interpretability, E2E-AD eliminates the need for separate explanation models, thereby reducing computational overhead. This work not only advances the practical deployment of explainable autonomous systems but also provides a methodological foundation for safety-certifiable AI in dynamic environments.