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

2026-02-03

AI Summary

This work addresses a key dilemma in end-to-end autonomous driving: fixed trajectory vocabularies struggle in highly interactive scenarios, while scene-adaptive methods often overcorrect high-quality trajectories in simpler settings. To resolve this, we propose CdDrive, a novel framework that seamlessly integrates fixed-vocabulary trajectories with adaptive candidates generated by a diffusion model, unified under a shared scoring module for consistent selection. A core component, the Horizon-Aware Trajectory Noise Adapter (HATNA), enhances trajectory continuity and geometric consistency through horizon-aware noise modulation and temporal smoothing. Evaluated on the NAVSIM v1/v2 benchmarks, CdDrive achieves state-of-the-art performance, with ablation studies confirming the effectiveness of each module and demonstrating significantly improved robustness and accuracy across diverse driving scenarios.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Planning, Routing, and Scheduling: Planning with Language Models
• Search and Optimization: Sampling/Simulation-based Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

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. Code: https://github.com/WWW-TJ/CdDrive.

Problem

• end-to-end autonomous driving
• trajectory candidate set
• scene-adaptive refinement
• multimodal planning
• diffusion model

Innovation

• diffusion-based planning
• scene-adaptive refinement
• trajectory candidate set
• end-to-end autonomous driving
• horizon-aware noise modulation