Robust Lane Change Decision for Autonomous Vehicles: A Causal-Invariant Adversarial Reinforcement Learning Approach

Abstract

Despite the rapid development of autonomous driving technology, the lane-changing polices of autonomous vehicles still face two major challenges: perceptual uncertainty and vulnerability to adversarial attacks. Existing approaches often rely on empirical regularization or adversarial training to handle input perturbations, but they lack a systematic modeling of the causal relationship between perturbation mechanisms and policy deviations, making it difficult to fundamentally improve robustness. To address these issues, this paper proposes a novel causal-invariant adversarial reinforcement learning (CIARL) approach to enhance the robustness and safety of autonomous vehicles under optimal observational perturbations. The approach primarily comprises two core modules: the attacker employs a closed-box perturbation generator based on Bayesian optimization to construct optimal adversarial perturbations in the feature representation mapping space; the defender incorporates a causal-invariant representation learning mechanism to mitigate attacks by reducing the influence of non-causal features. This mechanism constrains the impact of adversarial perturbations on policy and cost within a controllable range while maximizing rewards, thereby enabling robust policy learning. Finally, the proposed CIARL approach is evaluated in the SUMO simulator across complex highway and on-ramp scenarios. Experimental results demonstrate that our method significantly outperforms several state-of-the-art baselines in both lane-change robustness and collision safety, verifying the core role of the causal-invariant mechanism in adversarial robust policy modeling.