以下文章来源于深蓝AI ，作者深蓝学院 深蓝AI . 专注于人工智能、机器人与自动驾驶的学习平台。 来源 | 深蓝AI 点击下方 卡片 ，关注" 自动驾驶之心 "公众号 戳我-> 领取 自动驾驶近30个 方向 学习 路线 >>自动驾驶前沿信息获取 → 自动驾驶之心知识星球 本文只做学术分享，如有侵权，联系删文 开源是最诚实的答案。 本文 精选了 2025 年高价值的开源项目 ，不以会议论英雄，只看 复现热度 与 工程参考性 。这些仓库提供了从数据清洗、训练配方到闭环评测的全套 方案，是快速上手端到端自动驾驶的最佳捷径。 需要说明的是，2025 年开源的自动驾驶项目真的太多了，随着AI技术的发展，开源项目数量正在快速增长，优秀的工作远不止本文所覆盖的范围。本文主要以 GitHub 上已公开的项目为线索进行整理，筛 选标准以代码可获取性和项目活跃度为主（具有一定Star数量），难免存在遗漏。未被提及的相关工作同样具有参考价值，读者可结合自身研究方向进一步查阅和补充。 DiffusionDrive Github Star： 0 . 9 k 机构： 德克萨斯农工大学；密歇根大学；多伦多大学 亮点： OpenEMMA ...

作者 | TryMyBest 编辑 | 自动驾驶之心 原文链接： https://zhuanlan.zhihu.com/p/1983151280315716691 点击下方 卡片 ，关注" 自动驾驶之心 "公众号 戳我-> 领取 自动驾驶近30个 方向 学习 路线 >>自动驾驶前沿信息获取 → 自动驾驶之心知识星球 本文只做学术分享，如有侵权，联系删文 本文主要概述一下地平线一段式端到端方案（HSD）的两篇核心文章: DiffusionDrive + ResAD。 DiffusionDrive给读者们梳理了整体pipeline，ResAD则着重于性能提升的关键：轨迹残差设计。两篇文章都很精彩，也感谢地平线的分享，给从业者带来很多启发。 轨迹生成 本文核心想说的就是轨迹生成部分，所谓"Truncated Diffusion"。文章指出人类驾驶行为并不是随机分布的，具备fix patterns。从这个观察出发，文中 DiffusionDrive 图1: diffisonDrive整体架构 DiffusionDrive的整体架构如图1，可以拆成三部分：1. 感知信息 2.导航信息 3.轨迹生成 感知信息 感知信息本 ...

Core Viewpoint - The article discusses the DiffusionDrive model, which utilizes a truncated diffusion approach for end-to-end autonomous driving, emphasizing its architecture and the integration of reinforcement learning to enhance trajectory planning and safety [1]. Group 1: Model Architecture - DiffusionDriveV2 employs a reinforcement learning-constrained truncated diffusion model, focusing on the overall architecture for autonomous driving [3]. - The model incorporates environment encoding, including bird's-eye view (BEV) features and vehicle status, to enhance the understanding of the driving context [5]. - The trajectory planning module utilizes multi-scale BEV features to improve the accuracy of trajectory predictions [8]. Group 2: Trajectory Generation - The model generates trajectories by first clustering the true future trajectories of the vehicle using K-Means to create anchors, which are then perturbed with Gaussian noise [12]. - The trajectory prediction process involves cross-attention mechanisms between the trajectory features and BEV features, allowing for more accurate trajectory generation [15][17]. - The model also integrates time encoding to enhance the temporal aspect of trajectory predictions [14]. Group 3: Reinforcement Learning Integration - The Intra-Anchor GRPO method is proposed to optimize strategies within specific behavior intentions, enhancing safety and goal-oriented trajectory generation [27]. - The reinforcement learning loss function is designed to mitigate instability during early denoising steps, using a discount factor to adjust the influence of rewards over time [28]. - The model incorporates a clear learning signal by truncating negative advantages and applying strong penalties for collisions, ensuring safer trajectory outputs [30]. Group 4: Noise Management - The model introduces multiplicative noise rather than additive noise to maintain the structural integrity of trajectories, ensuring smoother exploration paths [33]. - This approach addresses the inherent scale inconsistencies in trajectory segments, allowing for more coherent and realistic trajectory generation [35]. Group 5: Evaluation Metrics - The model evaluates generated trajectories based on safety, comfort, rule compliance, progress, and feasibility, aggregating these into a comprehensive score [27]. - Specific metrics are employed to assess safety (collision detection), comfort (acceleration and curvature), and adherence to traffic rules, ensuring a holistic evaluation of trajectory performance [27].

Core Insights - The article discusses the upgrade of DiffusionDrive to version 2, highlighting its advancements in end-to-end autonomous driving trajectory planning through the integration of reinforcement learning to address the challenges of diversity and sustained high quality in trajectory generation [1][3][10]. Background Review - The shift towards end-to-end autonomous driving (E2E-AD) has emerged as traditional tasks like 3D object detection and motion prediction have matured. Early methods faced limitations in modeling, often generating single trajectories without alternatives in complex driving scenarios [5][10]. - Previous diffusion models applied to trajectory generation struggled with mode collapse, leading to a lack of diversity in generated behaviors. DiffusionDrive introduced a Gaussian Mixture Model (GMM) to define prior distributions for initial noise, promoting diverse behavior generation [5][13]. Methodology - DiffusionDriveV2 introduces a novel framework that utilizes reinforcement learning to overcome the limitations of imitation learning, which previously led to a trade-off between diversity and sustained high quality in trajectory generation [10][12]. - The framework incorporates intra-anchor GRPO and inter-anchor truncated GRPO to manage advantage estimation within specific driving intentions, preventing mode collapse by avoiding inappropriate comparisons between different intentions [9][12][28]. - The method employs scale-adaptive multiplicative noise to enhance exploration while maintaining trajectory smoothness, addressing the inherent scale inconsistency between proximal and distal segments of trajectories [24][39]. Experimental Results - Evaluations on the NAVSIM v1 and NAVSIM v2 datasets demonstrated that DiffusionDriveV2 achieved state-of-the-art performance, with a PDMS score of 91.2 on NAVSIM v1 and 85.5 on NAVSIM v2, significantly outperforming previous models [10][33]. - The results indicate that DiffusionDriveV2 effectively balances trajectory diversity and sustained quality, achieving optimal performance in closed-loop evaluations [38][39]. Conclusion - The article concludes that DiffusionDriveV2 successfully addresses the inherent challenges of imitation learning in trajectory generation, achieving an optimal trade-off between planning quality and diversity through innovative reinforcement learning techniques [47].

Core Insights - The article discusses the advancements in end-to-end autonomous driving, emphasizing the importance of multi-modal fusion architectures and the introduction of GMF-Drive as a new framework that improves upon existing methods [3][4][44]. Group 1: End-to-End Autonomous Driving - End-to-end autonomous driving has gained widespread acceptance as it directly maps raw sensor inputs to driving actions, reducing reliance on intermediate representations and information loss [3]. - Recent models like DiffusionDrive and GoalFlow demonstrate strong capabilities in generating diverse and high-quality driving trajectories [3]. Group 2: Multi-Modal Fusion Challenges - A key bottleneck in current systems is the integration of heterogeneous inputs from different sensors, with existing methods often relying on simple feature concatenation rather than structured information integration [4][6]. - The article highlights that current multi-modal fusion architectures, such as TransFuser, show limited performance improvements compared to single-modal architectures, indicating a need for more sophisticated integration methods [6]. Group 3: GMF-Drive Overview - GMF-Drive, developed by teams from University of Science and Technology of China and China University of Mining and Technology, includes three modules aimed at enhancing multi-modal fusion for autonomous driving [7]. - The framework combines a gated Mamba fusion approach with spatial-aware BEV representation, addressing the limitations of traditional transformer-based methods [7][44]. Group 4: Innovations in Data Representation - The article introduces a 14-dimensional pillar representation that retains critical 3D geometric features, enhancing the model's perception capabilities [16][19]. - This representation captures local surface geometry and height variations, allowing the model to differentiate between objects with similar point densities but different structures [19]. Group 5: GM-Fusion Module - The GM-Fusion module integrates multi-modal features through gated channel attention, BEV-SSM, and hierarchical deformable cross-attention, achieving linear complexity while maintaining long-range dependency modeling [19][20]. - The module's design allows for effective spatial dependency modeling and improved feature alignment between camera and LiDAR data [19][40]. Group 6: Experimental Results - GMF-Drive achieved a PDMS score of 88.9 on the NAVSIM benchmark, outperforming the previous best model, DiffusionDrive, by 0.8 points, demonstrating the effectiveness of the GM-Fusion architecture [29][30]. - The framework also showed significant improvements in key sub-metrics, such as driving area compliance and vehicle progression rate, indicating enhanced safety and efficiency [30][31]. Group 7: Conclusion - The article concludes that GMF-Drive represents a significant advancement in autonomous driving frameworks by effectively combining geometric representations with spatially aware fusion techniques, achieving new performance benchmarks [44].

Core Viewpoint - The article discusses the GMF-Drive framework developed by the University of Science and Technology of China, which addresses the limitations of existing multi-modal fusion architectures in end-to-end autonomous driving by integrating gated Mamba fusion with spatial-aware BEV representation [2][7]. Summary by Sections End-to-End Autonomous Driving - End-to-end autonomous driving has gained recognition as a viable solution, directly mapping raw sensor inputs to driving actions, thus minimizing reliance on intermediate representations and information loss [2]. - Recent models like DiffusionDrive and GoalFlow have demonstrated strong capabilities in generating diverse and high-quality driving trajectories [2][8]. Multi-Modal Fusion Challenges - A key bottleneck in current systems is the multi-modal fusion architecture, which struggles to effectively integrate heterogeneous inputs from different sensors [3]. - Existing methods, primarily based on the TransFuser style, often result in limited performance improvements, indicating a simplistic feature concatenation rather than structured information integration [5]. GMF-Drive Framework - GMF-Drive consists of three modules: a data preprocessing module that enhances geometric information, a perception module utilizing a spatial-aware state space model (SSM), and a trajectory planning module employing a truncated diffusion strategy [7][13]. - The framework aims to retain critical 3D geometric features while improving computational efficiency compared to traditional transformer-based methods [11][16]. Experimental Results - GMF-Drive achieved a PDMS score of 88.9 on the NAVSIM dataset, outperforming the previous best model, DiffusionDrive, by 0.8 points [32]. - The framework demonstrated significant improvements in key metrics, including a 1.1 point increase in the driving area compliance score (DAC) and a maximum score of 83.3 in the ego vehicle progression (EP) [32][34]. Component Analysis - The study conducted ablation experiments to assess the contributions of various components, confirming that the integration of geometric representations and the GM-Fusion architecture is crucial for optimal performance [39][40]. - The GM-Fusion module, which includes gated channel attention, BEV-SSM, and hierarchical deformable cross-attention, significantly enhances the model's ability to process multi-modal data effectively [22][44]. Conclusion - GMF-Drive represents a novel end-to-end autonomous driving framework that effectively combines geometric-enhanced pillar representation with a spatial-aware fusion model, achieving superior performance compared to existing transformer-based architectures [51].

Core Viewpoints - The current development of cutting-edge technologies in autonomous driving is not yet fully mature for mass production, with significant challenges remaining to be addressed [1][27][31] - Emerging technologies such as VLA/VLM, diffusion models, closed-loop simulation, and reinforcement learning are seen as potential key directions for future exploration in autonomous driving [6][7][28] - The choice between deepening expertise in autonomous driving or transitioning to embodied intelligence depends on individual circumstances and market dynamics [19][34] Group 1: Current Technology Maturity - The BEV (Bird's Eye View) perception model has reached a level of maturity suitable for mass production, while other models like E2E (End-to-End) are still in the experimental phase [16][31] - There is a consensus that the existing models struggle with corner cases, particularly in complex driving scenarios, indicating that while basic functionalities are in place, advanced capabilities are still lacking [16][24][31] - The industry is witnessing a shift towards utilizing larger models and advanced techniques to enhance scene understanding and decision-making processes in autonomous vehicles [26][28] Group 2: Emerging Technologies - VLA/VLM is viewed as a promising direction for the next generation of autonomous driving, with the potential to improve reasoning capabilities and safety [2][28] - The application of reinforcement learning is recognized as having significant potential, particularly when combined with effective simulation environments [6][32] - Diffusion models are being explored for their ability to generate multi-modal trajectories, which could be beneficial in uncertain driving conditions [7][26] Group 3: Future Directions - Future advancements in autonomous driving technology are expected to focus on enhancing safety, improving passenger experience, and achieving comprehensive scene coverage [20][28] - The integration of closed-loop simulations and data-driven approaches is essential for refining autonomous driving systems and ensuring their reliability [20][30] - The industry is moving towards a data-driven model where the efficiency of data collection, cleaning, labeling, training, and validation will determine competitive advantage [20][22] Group 4: Career Choices - The decision to specialize in autonomous driving or shift to embodied intelligence should consider personal interests, market trends, and the maturity of each field [19][34] - The autonomous driving sector is perceived as having more immediate opportunities for impactful work compared to the still-developing field of embodied intelligence [19][34]