CausalPlanner: A Causality-Enhanced Planning Framework for Generalizable Autonomous Driving

Abstract

Imitation learning (IL) has been widely adopted for autonomous driving planning because of its data efficiency and stable optimization. Yet IL-based planners often suffer from causal confusion, fitting spurious correlations instead of genuine causal mechanisms, which leads to unreliable planning behavior in complex traffic environments. Existing approaches addressing causal confusion often rely on manual labeling or simulator interventions and can only adapt to specific traffic conditions, lacking scalability. To overcome these limitations, we propose CausalPlanner, a novel planning framework that explicitly integrates causal reasoning into the imitation learning pipeline to achieve more reliable and generalizable driving behavior. Our approach integrates counterfactual reasoning with causal intervention, employing a prior guided weakly supervised strategy to autonomously identify causal elements that substantially impact the ego vehicle's planning and quantify their influence across heterogeneous traffic elements. Building on this foundation, we construct a unified framework that jointly optimizes causal elements identification and planning generation, with the quantified influence guiding the creation of contrastive samples to enforce causal constraints via contrastive learning, thereby systematically avoiding causal confusion. Comprehensive experiments on the large-scale NuPlan benchmark demonstrate that CausalPlanner achieves state-of-the-art performance across multiple scenarios, exhibiting exceptional robustness, generalization, and reliability. Our work provides a new solution for developing interpretable and trustworthy autonomous driving systems.