ExplainDrive: A Multimodal Chain-of-Thought Reasoning Approach for Explainable Automated Driving Systems

End-to-end decision models based on deep learning have become increasingly prominent in automated driving systems. However, their black-box nature poses significant challenges to interpreting decision processes, especially in dynamic and complex scenarios. Existing approaches largely focus on post-hoc analyses or isolated single-step explanations, lacking comprehensive explanations from scenario understanding to decision-making and failing to address the complexity of real-world scenarios with coherent reasoning. To address these limitations, we propose ExplainDrive, a multimodal Chain-of-Thought reasoning framework that integrates causally optimized temporal representations with explainable decision-making. ExplainDrive follows a three-stage pipeline: (i) extracting spatio-temporal features via a Causal Temporal Former, (ii) constructing hierarchical scenario understanding, and (iii) progressively deriving driving decisions with interpretable rationales. This design enhances transparency at each intermediate step and mitigates spurious correlations through causal feature selection. Extensive experiments on the BDD-X and nuScenes datasets demonstrate that ExplainDrive consistently improves the quality of decision explanations and outperforms compared models across multiple key evaluation metrics.