This paper aims to overcome the limitations of end-to-end autonomous driving, which struggles to generate adaptive, robust, and interpretable decisions across diverse scenarios. We propose a knowledge-driven diffusion policy, called KDP (Knowledge-Driven Diffusion Policy), which integrates generative diffusion modeling and a sparse expert-mixed routing mechanism. The diffusion component generates temporally consistent action sequences, while the expert routing mechanism enables modular knowledge organization by activating context-specific and reusable experts. Extensive experiments across diverse autonomous driving scenarios demonstrate that KDP consistently achieves higher success rates, reduced collision risk, and smoother control compared to existing approaches.
