以下文章来源于BigBite思维随笔 ，作者BigBite BigBite思维随笔 . Big Bite Small Talk, 杂谈随笔，聊科技，AI，成长，理财，经验杂谈。Stay Hungry 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 作者 | BigBite 编辑 | 自动驾驶之心 本文只做学术分享，如有侵权，联系删文 >>自动驾驶前沿信息获取 → 自动驾驶之心知识星球 上周柱哥转载知乎文章的那篇《 有消息称FSD不是端到端One Model 》，引起了大家很广泛的讨论。批评和支持的声音都很多，当时分享这篇文章主要是提供一个 角度看端到端： 一个VA的端到端不是万能的，One Model可以承担95%的主要功能（甚至更高），但也会增加一些小模型辅助做优化。 今天给大家分享一篇业内工程师BigBite视角下的FSD解析， 观点主要如下： 以下是原文： green发现的FSD模型参数文件信息 这里可以发现B核神经网络参数远多于A核，同时A，B共用的只有61个参数文件，也就是说早年AI Day上分享的A，B核互成冗余的设计实际上在Tesla V12 ...

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 21号，轻舟首个基于单征程6M的城市NOA方案，已正式上车理想L系列智能焕新版。23号轻舟开了一场发布 会，里面技术的部分，给大家分享一下。 单J6M实现一段式端到端+强化学习，说实话是有点东西的。 和大家一起拆解下整体的网络架构： 以上的部分是一个常见的OneModel架构，下面是不一样的地方： 后续利用Safe RL（增加规则的判断）进一步优化自车轨迹。这一套架构整体上来说，其实不复杂，难的是在 J6M 128TOPS的算力上实现。第一时间就有人问柱哥这是不是真的。 DiffusionDrive和Flow Matching已经是多家公司验证过可量产的算法了。有两个算法也推荐一下，Diffusion Planner和Flow Planner，Flow Planner是Diffusion Planner的改进版本，是清华AIR詹仙园老师团队下面的工作。 轻舟也放了几个困难场景的demo。下图是L2实车的表现，严重错位道路和复杂路口的无保护左转，效果都很 不错。严重错位的道路很考验静态的基本功，不止是道路/车道 ...

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 端到端不再是头部玩家的"特权"。 这周和几家公司交流了一下，柱哥深切的感受到了这句话的含金量，没进入端到端的车企和Tier1，都在加速 转型。有资源的公司想做一段式端到端，中小公司则选择两段式作为切入点。 这两周的交流，有几个关键的 点是大家关心的： 就目前各家公开的信息来看，200万Clips数据对于两段式来说可以做一个不错的车端模型，所需要的训练资 源也相对友好（百卡训练），而一段式则要对标几家头部公司的千万clips（千卡训练）。 当然大家还对VLA、VLM在车端/云端的作用、强化学习如何做损失权重配比比较感兴趣，哪些论文可以给落 地做参考，未来如何做方向预研。 这些问题也正是我们自动驾驶之心持续关注和分享的，柱哥已经邀请了多位业内专家在星球内部和大家交 流。 也欢迎大家加入自动驾驶之心知识星球，和大家一起交流， 我们准备了大额新人优惠券...... 扛内卷，一个足够有料的社区 对于很多想入门的同学来说，试错成本有点高。没时间和缺乏完整的体系是最大问题，这也容易导致行业壁 垒越来越高，如果想要卷赢那就更 ...

Core Viewpoint - The article discusses the growing demand for algorithm teams in the field of 3DGS (3D Gaussian Splatting) for autonomous driving, highlighting the need for skilled professionals and the development of a comprehensive training course to address this gap [2][3]. Group 1: Industry Demand and Job Roles - Companies are looking to invest in headcount (HC) for testing and closed-loop simulation in the autonomous driving sector, indicating a clear need for algorithm teams ranging from 5 to 20 members to support optimization in closed-loop simulations [2][3]. - The demand for cloud data production is also noted, particularly for static road surface reconstruction, which requires a minimum team size of around 10 people to meet basic functional needs [3]. Group 2: 3DGS Development and Learning Path - The article outlines a structured learning path for 3DGS, starting from static reconstruction to dynamic reconstruction and surface reconstruction, culminating in mixed scene reconstruction and feed-forward GS [3]. - A course titled "3DGS Theory and Algorithm Practical Tutorial" has been developed to provide a detailed roadmap for understanding 3DGS technology, covering principles and practical applications [3]. Group 3: Course Structure and Content - The course consists of six chapters, covering topics such as background knowledge, principles and algorithms of 3DGS, technical explanations for autonomous driving, important research directions, and feed-forward 3DGS [6][8][9][10][11][12]. - Each chapter is designed to build upon the previous one, ensuring a comprehensive understanding of 3DGS and its applications in the industry [8][9][10][11][12]. Group 4: Target Audience and Prerequisites - The course is aimed at individuals with a background in computer graphics, visual reconstruction, and related technologies, as well as those familiar with Python and PyTorch [17]. - Participants are expected to have a foundational understanding of probability theory and linear algebra, which are essential for mastering the 3DGS technology stack [17].

Core Viewpoint - NVIDIA is transitioning from a hardware supplier to a comprehensive provider of autonomous driving solutions, focusing on a full-stack approach that includes cloud training, simulation, and in-vehicle inference capabilities [4][7]. Group 1: Three Pillars of Full-Stack Solutions - NVIDIA's automotive strategy is built on three main components: DGX for AI model training, OVX for simulation, and AGX for in-vehicle inference [8][20]. - DGX serves as an AI model training factory, utilizing a supercomputing cluster of thousands of GPUs to process vast amounts of driving data [11][12]. - OVX creates a virtual world that mirrors real-world conditions, allowing for extensive testing of autonomous driving algorithms without the risks and costs associated with real-world testing [13][14][16]. - AGX represents NVIDIA's well-known in-vehicle computing chips, which have evolved to provide significantly higher processing power, becoming standard in various flagship models [18][20]. Group 2: Business Model Evolution - NVIDIA's revenue model has shifted from solely selling hardware to offering engineering services, which include deep involvement in automakers' production projects [21][23]. - The company charges a one-time engineering service fee, akin to a "coaching fee," to assist automakers in optimizing their algorithms on NVIDIA's platform [24][25]. - This service model fosters a win-win situation, enhancing automakers' capabilities while providing NVIDIA with valuable feedback for continuous product improvement [25]. Group 3: Open Source Strategy - In early 2025, NVIDIA announced the open-sourcing of its Alpamayo series, which includes a large-scale reasoning model and a comprehensive simulation framework [28][29][30]. - This strategic move aims to lower industry barriers, expand the ecosystem, and establish NVIDIA as a leader in defining the next generation of autonomous driving technology [34][35]. - The open-source approach also serves to mitigate geopolitical risks by transforming core technologies into global public assets [34]. Group 4: Demand from the Chinese Market - NVIDIA's accelerated pace in the automotive sector is largely driven by demand from the Chinese market, which is ahead of overseas automakers by two to three years in smart vehicle development [38][40]. - The rapid iteration and high expectations for functionality from Chinese automakers have prompted NVIDIA to develop specialized tools like TensorRT-LLM for Auto in record time [38][40]. Group 5: Competitive Landscape - NVIDIA maintains confidence against competitors by emphasizing that the ultimate competition in smart driving lies in systemic engineering capabilities and a continuously evolving ecosystem [41][42]. - The company has built a comprehensive stack that includes chips, safety certifications, operating systems, middleware, and development tools, creating a high barrier to entry for competitors [42][44].

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 昨晚的星友面对面收获颇多，跟着嘉宾了解到很多具身行业最新的动态。几个关键点分享给大家： 对于已经有端到端和大模型经验的从业人员，比较好转行。现在具身行业非常喜欢模仿学习、强化学习背景的同学加入。 对于想入行的校招生来说，要求并不高，刷题以力扣为主，配合一些前沿算法的掌握，pi0.5/GROOT N1.5/pi*0.6等等。 扛内卷，一个足够有料的社区 但同时也要考虑风险，业内已经在考虑走向落地和应用，但形势尚不明朗。风险确实是比较高的，要有等额的可预期回报才值得入场。 本场交流我们还讨论了近三十个问题，涉及VLA、World Model、locomotion等技术交流。欢迎加入自动驾驶之心知识星球进一步交流...... 如果你也想和我们一起推动自驾领域的进步，欢迎加入我们的社区团队，和我们一起推动！ 我们准备了大额优惠券给大家，欢迎微信扫码领取，还有少量～ 社区内部还经常为大家解答各类实用问题：端到端如何入门？自动驾驶多模态大模型如何学习？自动驾驶VLA的学习路线。数据闭环4D标注的工程实践。快速解 答，方便 ...

雷峰网 . 洞见智能未来，共与产业变迁 作者丨 倪萍 链接丨 新智驾独家丨小鹏副总裁、汽车互联网中心负责人魏斌休假 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 >>国内首个自动驾驶全栈交流社区 ： 自动驾驶之心知 识星球 （戳我） ★ 魏斌曾任高德地图产品总监，2021年底加入小鹏汽车。 26年刚开始，真是热闹。据雷峰网《新智驾》消息： 小鹏副总裁、互联网中心负责人魏斌目前处于休假状态。 前几天是美团无人车换帅，上周理想组织结构大变 动。新的一年，都要好好琢磨如何应对新局势了。自动驾驶和具身智能，汽车行业都不想落下。 以下文章来源于雷峰网 ，作者倪萍 从组织层面看，魏斌所负责的互联网中心，与自动驾驶中心、整车软件平台之间存在高度协同关系。随着小鹏近年来不断强调"AI 汽车"定位，智能座舱逐渐从传统意 义上的人机交互系统，演变为车端AI能力的重要载体，其战略权重也随之提升。 在2024年以来的多次公开表态中，小鹏汽车反复强调，将进一步聚焦核心技术、提升组织效率，并推动智能驾驶、智能座舱与整车平台的深度融合。 自动 驾 驶 之心 世世世世世世世世界界界界界 ...

Core Algorithm - The industry has shifted towards end-to-end solutions, moving away from modular approaches, at least in public discourse [1] - The introduction of world models is prevalent, with some companies using them to generate training data, while others incorporate them into end-to-end models to enhance performance [1][8] - There is a divergence in opinions regarding the necessity of language models (VLA) in autonomous driving, with some companies arguing that language is not essential for driving tasks [1][11] Simulation and Infrastructure - The closed-loop systems have evolved from data-driven to simulation testing and training loops [2] - 3DGS is highlighted as a crucial technology for building simulation environments, as emphasized by Tesla at CVPR 2025 [5] - Infrastructure is critical, with companies like Xiaomi and Li Auto noting its benefits for development efficiency [3][14] Organizational Capability - Organizational ability is vital, as large autonomous driving teams face significant management challenges [4] - Team culture and collaboration are emphasized as essential for overcoming complex technical and management issues [5] Technical Choices Comparison - A comparison of various companies' technical choices reveals differing approaches to core technologies and the role of world models and simulation tools [9] - Companies like Li Auto advocate for a training loop that evolves from imitation to self-learning, while NVIDIA emphasizes interpretability and reasoning in AI [9] Key Non-Core Factors - R&D infrastructure and engineering efficiency are crucial for the success of autonomous driving technologies [14] - Simulation and synthetic data are becoming essential for addressing corner cases that real-world data cannot cover [14] - The scale of computing power and chip adaptation is critical, as autonomous driving is not just a software issue but also a hardware challenge [15] User Experience and Safety - User experience and safety are paramount, with companies like Xiaomi stressing the importance of balancing advanced technology with user concerns [17] - The need for a dual-stack safety mechanism is highlighted, ensuring that even aggressive end-to-end models have a fallback to traditional rule-based systems for safety [19]

Core Viewpoint - The article emphasizes the growing demand for positions in the field of autonomous driving, particularly in the areas of world models, end-to-end systems, and VLA, highlighting the importance of practical experience and advanced knowledge in these domains [2][4]. Course Overview - The course on world models in autonomous driving is being launched in collaboration with industry experts, focusing on various algorithms and applications, including Tesla's world model and the Marble project by Fei-Fei Li's team [2][4]. - The course aims to provide a comprehensive understanding of world models, covering their development history, current applications, and different approaches such as pure simulation, simulation + planning, and generative sensor input [7]. Course Structure - **Chapter 1: Introduction to World Models** This chapter reviews the relationship between world models and end-to-end autonomous driving, discussing the evolution and current applications of world models, as well as various streams within the field [7]. - **Chapter 2: Background Knowledge of World Models** This chapter covers foundational knowledge related to world models, including scene representation, Transformer technology, and BEV perception, which are crucial for understanding subsequent chapters [8][12]. - **Chapter 3: General World Model Exploration** Focuses on popular models such as Marble, Genie 3, and the latest discussions around VLA + world model algorithms, providing insights into their core technologies and design philosophies [9]. - **Chapter 4: Video Generation-Based World Models** This chapter delves into video generation algorithms, starting with notable works like GAIA-1 & GAIA-2 and extending to recent advancements, ensuring a balance between classic and cutting-edge research [10]. - **Chapter 5: OCC-Based World Models** Concentrates on OCC generation methods, discussing three major papers and a practical project, highlighting their applicability in trajectory planning and end-to-end systems [11]. - **Chapter 6: World Model Job Specialization** This chapter shares practical insights from the instructor's experience, addressing industry applications, pain points, and interview preparation for related positions [12]. Learning Outcomes - The course is designed to elevate participants to a level equivalent to one year of experience as a world model algorithm engineer, covering key technologies and enabling practical application in projects [15].

Core Insights - The article discusses the evolving landscape of the AI and autonomous driving industry, highlighting the increasing complexity and competition in algorithm engineering roles, particularly as single-module algorithm positions are expected to decline [4][5][6]. Group 1: Industry Trends - The year 2025 is anticipated to be a pivotal year for the AI industry, marked by rapid advancements in large models and a shift in autonomous driving technology from end-to-end solutions to more complex architectures [4]. - There is a growing concern among industry professionals about being left behind, as many engineers are still working on outdated solutions while only a few have the opportunity to engage in cutting-edge projects [5]. Group 2: Job Market Insights - For algorithm engineers, potential job directions include L2 assisted driving, with opportunities primarily in traditional and new automotive manufacturers, as well as suppliers. However, newer roles in VLA (Vehicle Level Architecture) may offer more core positions [9]. - The L4 commercialization is expected to gain momentum in 2026, with companies expanding teams for applications such as robotaxis and autonomous logistics. These teams are generally smaller and have clearer business models [11]. - The concept of embodied intelligence is gaining traction, with numerous startups emerging. Current recruitment focuses on VLA, reinforcement learning, and motion control, while traditional perception algorithms are less in demand [14]. Group 3: Personal Career Decisions - The author ultimately chose to pursue opportunities in the L4 direction, which, while not the most popular, aligns better with personal judgment and market trends [15].