Summary

DiffusionDriveV2 addresses the mode collapse issue in generative diffusion models for end-to-end autonomous driving, where models often produce conservative or homogeneous behaviors. It introduces a novel approach combining reinforcement learning (RL) with truncated diffusion modeling to constrain low-quality outputs and actively explore superior driving trajectories. This method targets researchers and engineers in autonomous driving, offering a significant improvement in balancing trajectory diversity with consistent high-quality performance, setting new benchmarks on the NAVSIM datasets.

How It Works

The core innovation lies in using RL to refine diffusion model outputs. DiffusionDriveV2 employs scale-adaptive multiplicative noise to encourage broad exploration of driving behaviors. Crucially, it utilizes intra-anchor GRPO (Proximal Policy Optimization) to manage advantage estimation within specific driving intentions (anchors) and inter-anchor truncated GRPO to incorporate a global perspective across different intentions. This dual GRPO strategy prevents improper advantage comparisons between distinct intentions, mitigating further mode collapse and enhancing overall output quality while preserving multimodality.

Quick Start & Requirements

The project released its initial code and weights on December 9th, 2025, alongside documentation and training/evaluation scripts. Setup involves preparing the NAVSIM environment and then the DiffusionDriveV2 environment. Specific hardware requirements (e.g., GPU, CUDA versions) and detailed installation commands are not explicitly detailed in the provided README snippet. Links to official quick-start guides or demos are also not present.

Highlighted Details

• Achieves state-of-the-art performance with 91.2 PDMS on the NAVSIM v1 dataset and 85.5 EPDMS on the NAVSIM v2 dataset in closed-loop evaluation.
• Successfully resolves the trade-off dilemma between diversity and consistent high quality for truncated diffusion models.
• Sets new benchmark records on the NAVSIM autonomous driving datasets.

Maintenance & Community

The project's initial release, including code and weights, occurred on December 9th, 2025, accompanied by paper publication on Arxiv. Notable contributors are listed, with Xinggang Wang as a corresponding author. No community channels (like Discord or Slack) or roadmap details are mentioned in the provided text.

Licensing & Compatibility

DiffusionDriveV2 is released under the Apache 2.0 license. This license is generally permissive and allows for commercial use and integration into closed-source projects.

Limitations & Caveats

The README indicates this is an "initial version of code and weights," suggesting potential for ongoing development and refinement. No specific limitations, known bugs, or unsupported platforms are detailed in the provided text.