This paper proposes a knowledge-based diffusion policy (KDP) to address the challenges of multimodal action generation, temporal stability, and generalization across diverse scenarios in end-to-end autonomous driving. KDP integrates generative diffusion modeling and a sparse expert mixed routing mechanism to generate temporally consistent, multimodal action sequences and activate context-specific, specialized, and reusable experts, enabling modular knowledge construction. Experimental results across various driving scenarios demonstrate that KDP achieves higher success rates, lower collision risks, and smoother control than existing methods. Further analysis confirms the effectiveness of sparse expert activation and the Transformer backbone, as well as the structural specialization and cross-scenario reuse of experts. These results demonstrate that the diffusion model with expert routing is a scalable and interpretable paradigm for end-to-end autonomous driving.
