DriftingPlanner: Drifting Generative Models for One-Step Trajectory Planning in End-to-End Autonomous Driving

TL;DR

DriftingPlanner introduces a one-step trajectory planner using drifting generative models that learn an attraction-repulsion field per scene to generate diverse, high-quality trajectories without iterative inference. By decoding antithetic noise pairs and averaging their outputs, it cancels out shell deviation, achieving state-of-the-art performance with single-step inference.

Why It Matters

This approach enables fast, high-accuracy trajectory planning in end-to-end autonomous driving, eliminating the need for slow iterative methods like diffusion or flow matching while maintaining superior sample quality and diversity.

Key Findings

• DriftingPlanner achieves 93.75 PDMS on NAVSIM navtest, matching top diffusion/flow matching methods with only one inference step.
• The method uses a per-scene attraction-repulsion field to avoid cross-scene contamination and maintains trajectory diversity via a shell structure.
• Antithetic sampling and averaging cancel linear displacement from the shell center, enabling accurate one-step decoding without regression loss.