A Unified Candidate Set with Scene-Adaptive Refinement via Diffusion for End-to-End Autonomous Driving

TL;DR

This paper tackles multimodal trajectory planning for end-to-end autonomous driving by addressing the limitations of fixed candidate vocabularies and aggressive scene-adaptive refinements. It introduces CdDrive, which unifies a static vocabulary with diffusion-based scene-adaptive refinements and evaluates all candidates via a shared decision module, enhanced by the Horizon-Aware Trajectory Noise Adapter (HATNA) to ensure smooth diffusion. The key contributions are diffusion-based refinement over anchors, a unified candidate set with a latent world-model rollout for scoring, and horizon-aware noise modulation to improve continuity; these are validated on NAVSIM v1 and v2, showing state-of-the-art performance and robust ablations. The approach offers practical impact by enabling robust planning across routine and highly interactive driving scenarios while maintaining computational efficiency and stability in real time.

Abstract

End-to-end autonomous driving is increasingly adopting a multimodal planning paradigm that generates multiple trajectory candidates and selects the final plan, making candidate-set design critical. A fixed trajectory vocabulary provides stable coverage in routine driving but often misses optimal solutions in complex interactions, while scene-adaptive refinement can cause over-correction in simple scenarios by unnecessarily perturbing already strong vocabulary trajectories.We propose CdDrive, which preserves the original vocabulary candidates and augments them with scene-adaptive candidates generated by vocabulary-conditioned diffusion denoising. Both candidate types are jointly scored by a shared selection module, enabling reliable performance across routine and highly interactive scenarios. We further introduce HATNA (Horizon-Aware Trajectory Noise Adapter) to improve the smoothness and geometric continuity of diffusion candidates via temporal smoothing and horizon-aware noise modulation. Experiments on NAVSIM v1 and NAVSIM v2 demonstrate leading performance, and ablations verify the contribution of each component.

Figures (5)

• Figure 1: Candidate construction paradigms. (a) Static trajectory vocabulary followed by direct online decision. (b) Scene-adaptive refinement on the vocabulary, followed by online decision. (c) CdDrive: unify static vocabulary candidates and diffusion-refined candidates into one candidate pool for a shared decision module.
• Figure 2: Overview of CdDrive. CdDrive forms a unified candidate set by combining vocabulary-anchor candidates with diffusion-refined, scene-adaptive candidates. HATNA performs horizon-aware noise adaptation prior to denoising refinement to improve geometric continuity and smoothness. A shared trajectory decision module evaluates all candidates and selects the executed trajectory.
• Figure 3: Qualitative examples where scene-adaptive refinement does not necessarily improve over vocabulary trajectories. Green denotes the ground-truth trajectory, red denotes the selected vocabulary trajectory, and blue denotes the paired diffusion candidate. In routine (low-interaction) scenes, refinement can introduce over-corrections (highlighted in red circles), where the diffusion candidate deviates from the selected vocabulary trajectory and drifts away from the ground-truth trend.
• Figure 4: Qualitative comparison of diffusion candidates with and without HATNA. Uniform noising across timesteps often yields kinked trajectories, while HATNA produces smoother diffusion candidates.
• Figure 5: Qualitative comparison in highly interactive scenarios. Diffusion-refined candidates better align with interaction-induced constraints than static vocabulary trajectories.