Generative AI-Driven Dynamic Information Prioritization for Enhanced Autonomous Driving

With the continuous upgrading of driving hardware, the amount and complexity of information received by vehicles operating in complex scenarios have significantly increased. Large-scale models have recently attracted considerable attention in advanced autonomous driving applications due to their strong capabilities in information induction the process of extracting meaningful patterns from diverse data sources, analysis, and decision-making. They have demonstrated preliminary support for autonomous driving systems. However, when a vehicle fails to promptly analyze critical information and make effective decisions within constrained time intervals, severe safety incidents may occur. To address this issue, we propose a “Cognitive Tree” framework. This framework employs a tree-like structure to assign weights to various environmental factors, thereby guiding the model to prioritize processing information with higher weights. As a result, the reasoning process becomes more focused on key factors, achieving more real-time and safer decision-making in dynamic environments. Our approach integrates multiple modules that collaboratively consider environmental information across multiple dimensions, including speed, distance, direction, and future trajectories. By assigning distinct weights to different types of information and leveraging the Cognitive Tree framework in conjunction with Retrieval-Augmented Generation, our method progressively filters information, ensuring that the model remains centered on critical elements. Experimental results on the nuScenes dataset demonstrate that, compared with traditional deep learning models and other large language model-based solutions, our framework achieves approximately a 25% reduction in collision rates.