一凡 发自 凹非寺 量子位 | 公众号 QbitAI 智能汽车、自动驾驶、物理AI的竞速引擎，正在悄然收敛—— 至少核心头部玩家，已经在最近的ICCV 2025，展现出了共识。 在端到端一统江湖但数据瓶颈开始成为新挑战后，必须转向强化学习，必须把云端生成式世界模型作为新基座。 春江水暖，异口同声。特斯拉和理想汽车，都在AI顶会现场分享着最新实践真知。 特斯拉自动驾驶副总裁 Ashok Elluswamy 在演讲中透露，当前特斯拉正在用世界模拟器来评估车端模型。几乎同时，理想VLA模型负责人 詹锟 也围绕世界模型，在 具身智能研讨会 做了题为 《World Model：Evolving from Data Closed-loop to Training Closed-loop》 （世界模型让我们从数据闭环走向训练闭环）的分享。 理想的观点是，当前数据闭环已经不够用了，VLA通往L4，需要训练闭环。 而这种闭环的构建方法和背后原因，詹锟也在会后更进一步的对话中分享了理想的思考和实践。 自动驾驶下半场？从数据闭环迈向训练闭环 在ICCV现场，理想在开篇就提出， 自动驾驶技术和大模型一样，都进入了下半场 。为什么这么 ...

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 理想汽在ICCV'25期间也分享了些新东西！目前还没有视频对外。 VLA团队负责人詹锟老师做了一场世界模型的presentation，名为World Model: Evolving from Data Closed-loop to Training Closed-loop。自动驾驶之心第一时间做了解 读分享给大家~ 首先是介绍下理想VLA司机大模型： 回顾了理想汽车智能驾驶的发展路线，从规则时代的轻图和无图，再到基于AI的E2E+VLM快慢双系统和VLA， 这四个方案中Nav（导航）是重点突出的模块。 下面介绍的是数据闭环的价值。左上角这张图是一个完整的数据闭环流程： 影子模式验证→经由数据触发回传到云端进行数据挖掘→有效样本进行自动标注→生 成训练集训练模型→模型下发验证性能。 这个过程已经可以做到一分钟的数据回传。 目前已经有15亿公里的驾驶数据，200+的Trigger来生产15-45s的Clip数据。 目前理想的端到端量产版本MPI已经到了220+， ...

Core Insights - The article discusses the advancements in autonomous driving technology, particularly focusing on the transition from data closed-loop systems to training closed-loop systems, marking a new phase in autonomous driving development [17][20]. Group 1: Development of Li Auto's VLA Model - Li Auto's VLA driver model has evolved through various stages, from rule-based systems to AI-driven E2E+VLM systems, with a strong emphasis on navigation as a key module [6]. - The end-to-end mass production version of MPI has reached over 220 units, representing a 19-fold increase compared to the version from July 2024 [12]. Group 2: Data Closed-Loop Value - The data closed-loop process includes shadow mode validation, data mining in the cloud, automatic labeling of effective samples, and model training, with a data return time of one minute [9][10]. - Li Auto has accumulated 1.5 billion kilometers of driving data, utilizing over 200 triggers to produce 15-45 second clip data [10]. Group 3: Transition to Training Closed-Loop - The core of the L4 training loop involves VLA, reinforcement learning (RL), and world models (WM), optimizing trajectories through diffusion and reinforcement learning [22]. - Key technologies for closed-loop autonomous driving training include regional simulation, synthetic data, and reinforcement learning [24]. Group 4: Simulation and Generation Techniques - Simulation relies on scene reconstruction, including visual and Lidar reconstruction, while synthetic data generation utilizes multimodal techniques [25]. - Li Auto's recent advancements in reconstruction and generation have led to significant improvements, with multiple top conference papers published in the last two years [26][29][31]. Group 5: Interactive Agents and System Capabilities - The development of interactive agents is highlighted as a critical challenge in the training closed-loop [37]. - System capabilities are enhanced through world models providing simulation environments, diverse scene construction, and accurate feedback from reward models [38]. Group 6: Community and Collaboration - The article mentions the establishment of nearly a hundred technical discussion groups related to various autonomous driving technologies, with a community of around 4,000 members and over 300 companies and research institutions involved [44][45].