Robust Decision-Making Method for Autonomous Driving Based on Multi-Modal Fusion and Uncertainty Modeling

With the continuous advancement of intelligent technology, autonomous driving shows significant potential for enhancing road safety and traffic efficiency. As a vital component of autonomous driving, the performance of the decision-making system directly affects the autonomous driving level of the vehicle. However, existing methods fail to fully exploit the complementarity of multi-modal features and overlook the impact of model parameter uncertainty, resulting in insufficient model robustness in extreme weather scenarios and making autonomous vehicles prone to making erroneous decisions, thereby increasing safety risks. To this end, we propose a robust decision-making method based on multi-modal fusion and uncertainty modeling, which extracts multimodal features through various encoders and employs an attention mechanism to fuse image and LiDAR features. Meanwhile, a fuzzy encoder is designed to extract robust fuzzy state representations, which are then defuzzified into crisp state representations that could be directly used for downstream decision-making tasks. These representations are further fused with other modal features to obtain a comprehensive scene representation. Additionally, a real-time coordination optimization mechanism between perception and decision-making is designed, where the perception process learns the scene representation, the decision-making process generates actions based on this representation, and feedback is employed to update the environment state, achieving dynamic coordination. Experimental results demonstrate that the proposed model achieves superior performance compared to baseline models in extreme weather scenarios, and exhibits strong robustness. In addition, the ablation experiments validate the effectiveness of each module.