Title: RESBev: Enhancing the Resilience of BEV Perception

Abstract: Bird’s-eye-view (BEV) perception has become a foundational element of autonomous driving architectures, offering a structured, ego-centric framework that is essential for subsequent planning and control mechanisms. Nevertheless, the practical implementation of these systems is hindered by sensor degradation and adversarial attacks, threats that can trigger significant perceptual errors and jeopardize vehicle safety. To mitigate these risks, we introduce RESBev, a robust and modular approach designed to be seamlessly integrated into current BEV perception systems to bolster their resistance to a wide array of disruptions. Our methodology redefines perception robustness as a task of predicting latent semantics. By building a latent world model, we capture spatiotemporal dependencies across sequential BEV inputs, allowing the system to learn underlying state transitions and forecast pristine BEV features for the restoration of damaged observations. This framework functions at the semantic feature stage of the Lift-Splat-Shoot pipeline, facilitating recovery that remains effective across both natural disturbances and adversarial incidents, all without altering the core backbone. Comprehensive evaluations on the nuScenes dataset reveal that, even with minimal fine-tuning, RESBev markedly increases the durability of existing BEV perception models against various external interferences and attacks.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC