[2603.18561] CausalVAD: De-confounding End-to-End Autonomous Driving via Causal Intervention

Computer Science > Computer Vision and Pattern Recognition arXiv:2603.18561 (cs) [Submitted on 19 Mar 2026 (v1), last revised 10 Apr 2026 (this version, v2)] Title:CausalVAD: De-confounding End-to-End Autonomous Driving via Causal Intervention Authors:Jiacheng Tang, Zhiyuan Zhou, Zhuolin He, Jia Zhang, Kai Zhang, Jian Pu View a PDF of the paper titled CausalVAD: De-confounding End-to-End Autonomous Driving via Causal Intervention, by Jiacheng Tang and 5 other authors View PDF HTML (experimental) Abstract:Planning-oriented end-to-end driving models show great promise, yet they fundamentally learn statistical correlations instead of true causal relationships. This vulnerability leads to causal confusion, where models exploit dataset biases as shortcuts, critically harming their reliability and safety in complex scenarios. To address this, we introduce CausalVAD, a de-confounding training framework that leverages causal intervention. At its core, we design the sparse causal intervention scheme (SCIS), a lightweight, plug-and-play module to instantiate the backdoor adjustment theory in neural networks. SCIS constructs a dictionary of prototypes representing latent driving contexts. It then uses this dictionary to intervene on the model's sparse vectorized queries. This step actively eliminates spurious associations induced by confounders, thereby eliminating spurious factors from the representations for downstream tasks. Extensive experiments on benchmarks like nuScenes show C...