DiDrive Enhances Safe Autonomous Driving with Risk-Aware AI

Autonomous driving systems face significant challenges, including managing distribution shifts in real-world data, handling heavy-tailed risk signals, preventing out-of-distribution actions, and dealing with high-dimensional state redundancy. This research introduces DiDrive, a novel offline diffusion framework specifically designed to enhance safety in autonomous driving through a distribution-guided approach. DiDrive integrates two key components: the Risk-Aware Hierarchical Diffusion (RHDif) architecture and the 3DICE policy optimization paradigm. RHDif employs a low-level risk-gated encoder and a high-level contextual modulator to filter out irrelevant environmental data and focus on critical safety threats. Concurrently, 3DICE mitigates issues like overestimation and gradient oscillation by using in-sample calibrated guidance, spatiotemporal optimization, and ensemble-based candidate ranking. Evaluations on the CARLA benchmark, particularly in dense traffic with 60 vehicles, demonstrated DiDrive's superior performance, achieving an 85% success rate and significantly higher average rewards compared to existing baselines.