arXiv:2512.07745v1 Announce Type: new 
Abstract: Generative diffusion models for end-to-end autonomous driving often suffer from mode collapse, tending to generate conservative and homogeneous behaviors. While DiffusionDrive employs predefined anchors representing different driving intentions to partition the action space and generate diverse trajectories, its reliance on imitation learning lacks sufficient constraints, resulting in a dilemma between diversity and consistent high quality. In this work, we propose DiffusionDriveV2, which leverages reinforcement learning to both constrain low-quality modes and explore for superior trajectories. This significantly enhances the overall output quality while preserving the inherent multimodality of its core Gaussian Mixture Model. First, we use scale-adaptive multiplicative noise, ideal for trajectory planning, to promote broad exploration. Second, we employ intra-anchor GRPO to manage advantage estimation among samples generated from a single anchor, and inter-anchor truncated GRPO to incorporate a global perspective across different anchors, preventing improper advantage comparisons between distinct intentions (e.g., turning vs. going straight), which can lead to further mode collapse. DiffusionDriveV2 achieves 91.2 PDMS on the NAVSIM v1 dataset and 85.5 EPDMS on the NAVSIM v2 dataset in closed-loop evaluation with an aligned ResNet-34 backbone, setting a new record. Further experiments validate that our approach resolves the dilemma between diversity and consistent high quality for truncated diffusion models, achieving the best trade-off. Code and model will be available at https://github.com/hustvl/DiffusionDriveV2

يمثل تقديم DiffusionDriveV2 تقدمًا كبيرًا في نمذجة الانتشار المقيدة بالتعلم المعزز من أجل القيادة الذاتية من البداية إلى النهاية، حيث يعالج القيود التي يعاني منها سلفه من خلال تحسين جودة المخرجات مع الحفاظ على التعددية. يستخدم هذا النموذج ضوضاء مضاعفة قابلة للتكيف مع المقياس وGRPO داخل الأنكر لتحسين توليد المسارات.

La introducción de DiffusionDriveV2 marca un avance significativo en la modelización de difusión truncada restringida por el aprendizaje por refuerzo para la conducción autónoma de extremo a extremo, abordando las limitaciones de su predecesor al mejorar la calidad de salida mientras mantiene la multimodalidad. Este modelo utiliza ruido multiplicativo adaptativo a escala y GRPO intra-ancla para una mejor generación de trayectorias.

L'introduction de DiffusionDriveV2 représente une avancée significative dans la modélisation de diffusion tronquée contrainte par l'apprentissage par renforcement pour la conduite autonome de bout en bout, en abordant les limitations de son prédécesseur en améliorant la qualité de sortie tout en maintenant la multimodalité. Ce modèle utilise du bruit multiplicatif adaptatif à l'échelle et du GRPO intra-ancre pour une meilleure génération de trajectoires.

The introduction of DiffusionDriveV2 marks a significant advancement in reinforcement learning-constrained truncated diffusion modeling for end-to-end autonomous driving, addressing the limitations of its predecessor by enhancing output quality while maintaining multimodality. This model utilizes scale-adaptive multiplicative noise and intra-anchor GRPO for improved trajectory generation.

DiffusionDriveV2: Reinforcement Learning-Constrained Truncated Diffusion Modeling in End-to-End Autonomous Driving

arXiv:2507.04049v3 Announce Type: replace 
Abstract: Most end-to-end autonomous driving methods rely on imitation learning from single expert demonstrations, often leading to conservative and homogeneous behaviors that limit generalization in complex real-world scenarios. In this work, we propose DIVER, an end-to-end driving framework that integrates reinforcement learning with diffusion-based generation to produce diverse and feasible trajectories. At the core of DIVER lies a reinforced diffusion-based generation mechanism. First, the model conditions on map elements and surrounding agents to generate multiple reference trajectories from a single ground-truth trajectory, alleviating the limitations of imitation learning that arise from relying solely on single expert demonstrations. Second, reinforcement learning is employed to guide the diffusion process, where reward-based supervision enforces safety and diversity constraints on the generated trajectories, thereby enhancing their practicality and generalization capability. Furthermore, to address the limitations of L2-based open-loop metrics in capturing trajectory diversity, we propose a novel Diversity metric to evaluate the diversity of multi-mode predictions.Extensive experiments on the closed-loop NAVSIM and Bench2Drive benchmarks, as well as the open-loop nuScenes dataset, demonstrate that DIVER significantly improves trajectory diversity, effectively addressing the mode collapse problem inherent in imitation learning.

DIVER هو إطار جديد للقيادة الذاتية من النهاية إلى النهاية يجمع بين التعلم المعزز وتوليد قائم على الانتشار للتغلب على قيود طرق التعلم بالتقليد التقليدية، التي غالبًا ما تؤدي إلى سلوكيات قيادة محافظة. تتيح هذه الطريقة المبتكرة توليد مسارات قيادة متنوعة وقابلة للتطبيق من عرض خبير واحد.

DIVER es un nuevo marco de conducción autónoma de extremo a extremo que combina el aprendizaje por refuerzo con la generación basada en difusión para superar las limitaciones de los métodos de aprendizaje por imitación tradicionales, que a menudo conducen a comportamientos de conducción conservadores. Este enfoque innovador permite generar trayectorias de conducción diversas y factibles a partir de una sola demostración de un experto.

DIVER est un nouveau cadre de conduite autonome de bout en bout qui combine l'apprentissage par renforcement avec la génération basée sur la diffusion pour surmonter les limitations des méthodes d'apprentissage par imitation traditionnelles, qui conduisent souvent à des comportements de conduite conservateurs. Cette approche innovante permet de générer des trajectoires de conduite diverses et réalisables à partir d'une seule démonstration d'expert.

DIVER is a newly proposed end-to-end autonomous driving framework that combines reinforcement learning with diffusion-based generation to overcome the limitations of traditional imitation learning methods, which often lead to conservative driving behaviors. This innovative approach allows for the generation of diverse and feasible driving trajectories from a single expert demonstration.

DiffusionDriveV2: Reinforcement Learning-Constrained Truncated Diffusion Modeling in End-to-End Autonomous Driving

Was this article worth reading? Share it

LucidQuery AI

Dyad

Dynamiq