MTG-RPD: Multimodal Trajectories Generation with Rule-Based Prior Diffusion for End-to-End Autonomous Driving

Xinyu Wang, Heng Wang, Huanyu Shen

Keywords: End-to-end Autonomous Driving, Trajectory Planning

Abstract: Replicating human driving behaviors in complex and authentic real-world environments remains a key challenge in autonomous driving. While end-to-end autonomous driving technologies have advanced substantially, generating safe and diverse multimodal trajectories poses a persistent hurdle. In recent years, diffusion-based methods have demonstrated remarkable potential across image generation, robotics, and autonomous driving—with trajectory generation approaches based on diffusion models also emerging. However, balancing real-time performance and reconstruction accuracy remains an unresolved issue. To address these limitations, we propose MTG-RPD, an innovative trajectory generation method that integrates rule-based prior knowledge. The approach first generates trajectory anchor points via rule-based prior clustering, then leverages a conditional diffusion model to transform an anchored Gaussian distribution into a multimodal trajectory distribution under scene information guidance. Notably, the diffusion model is specifically designed to facilitate agent-agent and agent-environment interactions. On the planning-based NAVSIM dataset, MTG-RPD achieved a PDMS of 88.5 when evaluated using the ResNet-34 backbone network.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 9429