Reproduce “AnchDrive: Bootstrapping Diffusion Policies with Hybrid Trajectory Anchors for End-to-End Driving” on OpenResearch

About this paper

Jinhao Chai, Anqing Jiang, Hao Jiang, Shiyi Mu, Zichong Gu, Hao Sun, Shugong Xu

Researchers at Shanghai University and Bosch Corporate Research developed AnchDrive, an end-to-end autonomous driving framework that achieves 85.5 EPDMS on the NAVSIM v2 benchmark. It uses a hybrid trajectory anchor mechanism to efficiently bootstrap a diffusion policy, reducing denoising steps to two while outperforming previous methods that relied on fixed or numerous anchors.

Key insights

• Efficiently bootstrapping diffusion policies with hybrid trajectory anchors effectively addresses the computational bottleneck, making generative models viable for real-time, safety-critical autonomous driving applications.
• The combination of dynamic, scene-specific anchors and static, broad-coverage priors enhances both adaptability to diverse driving contexts and generalization to unseen environments, improving multi-modality and robustness.
• High-quality trajectory generation can be achieved with significantly fewer computational resources and minimal denoising steps (e.g., just two), establishing a new paradigm for efficient motion planning through a refinement-based generative approach.

Summary by alphaXiv.