“世界模型，重点不是渲染的画质是否好看，而是模拟器是否能忠实还原现实物理规律和行为的因果”。11 月 3 日，在深圳举行的百度 AI Day·Apollo 专场上，百度 Apollo 首次披露面向物理 AI 的智能体驱动全模态世界模型 Apollo World，并提出世界模型“2+3”判断标准。目前，该模型已用于萝卜快跑全球化测试与验证，助力无人驾驶规模化落地。
百度智能驾驶事业群组(IDG)首席架构师、IDG 技术委员会主席王亮表示，世界模型不能只做“生成炫酷视频的画师”，更要成为验证无人驾驶安全性的“考官”。萝卜快跑是中国唯一以端到端技术规模化运行 L4 无人驾驶的出行平台，端到端全无人运营超百万公里，Apollo World 为其提供虚拟训练环境，帮助“学生” AI Driver 学习、验证和迭代。
“Apollo World 将虚拟场景内所有交通参与者建模为具备决策与交互能力的 AI 智能体，支持无人车在虚拟城市任意起终点通行，打造提前验证‘无限未来’的虚拟场景订单。”百度智能驾驶事业群组(IDG)杰出架构师、世界模型负责人杨泽进一步介绍。此外，Apollo World 的技术栈还可拓展至机器人等更广泛的物理 AI 场景。
行业人士分析，Apollo World 可“日行十万公里”，其一天完成的虚拟路测里程相当于一辆配备安全员的真实路测车连续测试约三年。依托世界模型带来的能力跃升，叠加数据飞轮效应，行业将呈现强者恒强的格局，这一趋势也正把行业推向等待已久的“AGI 时刻”。
世界模型不能只是“画得像”，Apollo 抛出世界模型“2+3”评判标尺
摩根士丹利将世界模型视为 AI 的“想象引擎”，让 AI 能在行动前先在虚拟空间推演，实现“先思考、后行动”。但把这种能力迁移到无人驾驶并不容易。李飞飞曾提出“汽车是带轮子的机器人”，但自动驾驶汽车更像是二维避障方盒子，与三维家务机器人差距明显。
王亮表示，会跳舞、叠衣服的机器人很厉害，但学会开车是另一重难题。无人驾驶从来不是具身智能的“EASY 模式”：机器人只需在受限空间执行规划动作，无人驾驶则身处开放道路，持续与海量不可预判的交通参与者博弈。对一台近两吨重、几十公里时速的车辆而言，它的每一次判断和动作都直接关系到交通安全，这对安全、实时性和冲突博弈的要求高出了一个量级——开放世界容错空间几乎为零，安全决策周期要严格限定在 1/30 秒内。这意味着，世界模型不能只满足于“像”。
站在物理 AI 视角，Apollo 提出“2+3”判断标准：“2”是决定算不算世界模型的基础门槛，“3”是衡量模型能力高下的进阶指标。“2”项基础门槛：一是因果性，遵循时间单向性，由过去决定未来，基于当前状态预测下一刻，而非拼凑画面反推过去；二是交互性，能动态响应外部干预，注入动作影响未来世界状态。“3”项增益属性：物理准确性，模拟的状态有多可靠，是否符合世界规律，学习获得的能力回到物理世界还能不能用；持久性，持续运行的时间和稳定性能不能跟上，长时间运行物体和细节是否发生漂移；实时性，低时延，生成世界复杂度的增长不影响模型推理时间，获得高模拟测试效率。
Apollo World 的三大核心能力：智能体驱动、全模态感知、行程闭环
杨泽介绍：“Apollo World 希望赋予交通参与者智能，让世界能够回应无人车的每一个动作。”
传统仿真里，其他交通参与者更像是按脚本运行的道具。在 Apollo World 构建的虚拟城市中，车辆、行人等交通参与者均被建模为具备决策和交互能力的 AI 智能体，能够根据周围交通变化调整行动。无人车减速、变道或让行，都可能引发其他参与者的不同反应，使无人驾驶系统在持续变化的交通环境中接受训练与检验。
与此同时，Apollo World 可实时生成相机、激光雷达等多模态观测数据，让无人驾驶系统直接感知虚拟环境。系统作出驾驶决策后，环境随之变化，再向系统提供新的反馈，形成端到端闭环。
杨泽表示：“行业里不少世界模型仍以独立场景模拟为主，Apollo World 则更关注完整任务的闭环模拟。”在 Apollo World 构建的虚拟城市中，无人驾驶车辆能够在不同起终点之间连续行驶，完成从接单、载客到送达的完整出行任务，使萝卜快跑能够在跨路段、跨场景的连续驾驶过程中接受训练与检验。
让 AI 教 AI 开车，世界模型叩开无人驾驶产业“AGI 时刻”
无人驾驶要走向规模化，必须同时兼顾安全、智能与规模，而这三者之间存在着一个“不可能三角”。长尾场景难靠实车里程充分覆盖，新区域部署存在测试成本与试错风险，跨路段、跨场景连续驾驶能力又需规模化验证。Apollo World 把真实运营中的复杂交互与风险场景转化为可重复、可控制、可规模化运行的虚拟测试环境，并生成现实尚未覆盖的新场景，让无人车在实车部署前先在虚拟世界完成驾驶推演，把测试成本和部署风险留在虚拟世界先行消化，让 AI 教 AI 开车。
要让虚拟验证真正贴近现实，离不开萝卜快跑持续积累的真实道路数据。萝卜快跑已落地全球 28 座城市，累计完成超 2300 万次出行服务，自动驾驶总里程突破 3.5 亿公里，其中全无人驾驶里程超过 2.4 亿公里，并已在左右舵市场开展无人驾驶实践。依托真实驾驶数据，Apollo World 提取高价值场景，虚拟测试结果再反哺车端迭代，形成真实运营与虚拟验证相互促进的闭环。
物理 AI 市场正快速增长。Grand View Research 预计，全球物理 AI 市场规模将从 2025 年的约 816 亿美元增至 2033 年的约 9604 亿美元，2026 年至 2033 年复合年增长率为 36.1%。从“生成画面”推向“验证行动”，破解“不可能三角”，Apollo World 正为无人驾驶规模化落地提供新的技术标尺，助推无人驾驶抵达 AGI 时刻。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
“The world model does not focus on whether the retrofitted paint is good or not, but on whether the simulator is able to faithfully reverse the causes and consequences of actual physical patterns and behaviour”. On 3 November, at Baidu AI Day Apollo, held in Shenzhen, Baidu Apollolo disclosed for the first time the full-model world model of AI agent for physics AI, Apollo World, and proposed a world model “2+3” judgement. Currently, the model has been used for the testing and validation of radish run-globalization to assist in the scalability of unmanned aerial vehicles.
Wang Liang, Chief Architect of the Baidu Smart Driving Cluster (IDG) and Chairman of the IDG Technical Committee, stated that the world model could not simply be a “shower” for the production of cool videos, but rather a “testant” for the safety of unmanned driving. Rope-runing is the only platform in China to run L4 drones on an end-to-end scale, with over a million kilometres running from end to end, Apollo World providing them with a virtual training environment to help “students” AI Driver study, validate and overlap.
“Apollo World Models all transport participants in the virtual scene as AI AI agent, with decision-making and interactive capabilities, supports the free-riding of unmanned vehicles at the end of a virtual city and creates a virtual scenario order to pre-validate `the infinite future'.” Further presentations were made by Yang Zai, Distinguished Architect, World Model Leader, Baidu Smart Driving Group (IDG). In addition, the Apollo World technology warehouse can extend to a wider physical AI scenario, such as robotics.
Industry analyses show that Apollo World can “go 100,000 kilometres a day” and that the virtual road miles completed in one day are equivalent to about three years of continuous testing of a real road survey vehicle equipped with a security guard. The industry is also moving towards the long-awaited “AGI moment”, depending on the ability of the world's models to leapfrog with data and flyer effects.
The world model can't be just a painting. Apollo throws the world model 2+3 judge.
Morgan Stanley sees the world model as the "imaginative engine" of AI, allowing AI to advance in virtual space before action, to achieve "think, act." It is not easy, however, to move this capability to unmanned. Li Fei Fei suggested that “cars are robots with wheels”, but auto-piloting is more like a 2D shelter box, with a clear gap with 3D domestic robots.
Wang Liang said that robots who danced and folded clothes were good, but learning to drive was another challenge. Unmanned drivers have never been an intelligent “EASY model”: robots only have to carry out planning exercises in restricted spaces, while unmanned drivers are on open roads and continue to compete with large, unpredictable traffic participants. For a vehicle that weighs almost two tons and is dozens of kilometres an hour, every judgment and move is directly related to traffic safety, which is one step above the requirements of safety, real time and conflict games -- - Opening the world to the wrong space is almost zero, and the security decision-making cycle is strictly limited to 1/30 seconds. This means that the world model cannot simply be satisfied with "likes".
From the physical AI perspective, Apollo proposes a “2+3” test: “2” is the basis threshold for determining whether the world model should be counted, and “3” is an indicator of progress below model capacity. The base threshold of “2” is causality, one that follows one-way timing, one that determines the future from the past and is based on current state predictions for the next moment, rather than on a patchwork of images, and one that is interactive and dynamic in responding to external intervention, and that injects action to influence future state of the world. The “3” gain properties: physical accuracy, how reliable the state of the simulation is, whether the ability to learn to return to the physical world is in keeping with world law, whether or not the ability to learn to return to the physical world can be used; persistence, the time and stability of continuous operation can keep up, and whether objects and details are drifting over a long period of time; real-time, low latency, the growth that generates the world's complexity does not affect model reasoning time and achieve high simulation testing efficiency.
Apollo World's three core competencies: AI agent-driven, all-modular perception, travel closed
Yang Zai, “Apollo World wants to give transport participants the intelligence to respond to every move of a drone.”
In traditional simulations, other transport participants are more of a script-based prop. In the virtual city built by Apollo World, transport participants, such as vehicles, pedestrians and others, are modelled as AI AI agent, capable of decision-making and interaction, capable of adjusting their actions to changes in the surrounding traffic. Drifting, diversion or diversion of unmanned vehicles can trigger different reactions from other participants, allowing unmanned systems to undergo training and testing in a continuously changing transport environment.
At the same time, Apollo World produces in real time multi-modular observational data such as cameras, laser radars and direct perception of the virtual environment by unmanned systems. When the system makes driving decisions, the environment changes and provides new feedback to the system, forming end-to-end closed loops.
According to Yang Ze, “Many world models in the industry still focus on independent scenario simulations, while Apollo World focuses more on closed loop simulations of complete missions.” In the virtual city built by Apollo World, unmanned vehicles are able to travel continuously between different endpoints, completing complete travel missions from receipt, passenger to delivery, enabling radish runaways to undergo training and testing during continuous driving across sections and scenes.
Let AI teach AI to drive, world model to open the unmanned industry "AGI Time"
To move to scale, drones must balance safety, intelligence and size, and there is a “no triangle” between them. Long tailings are difficult to adequately cover by physical drive, new area deployments are at risk of testing costs and error, and continuous driving capability across circuits and scenes needs to be scaled and validated. Apollo World transforms complex interactions and risk scenarios in real operations into a repertoire, manageable, scalable virtual testing environment, and creates new scenarios that are not yet covered by reality, allowing unmanned vehicles to finish driving in the virtual world prior to their deployment, leaving test costs and deployment risks in the virtual world to digest and allow AI to teach AI to drive.
Real road data that continue to accumulate in the absence of radish is needed to make virtual validation truly relevant to reality. Rhodships have landed in 28 cities across the globe, cumulatively completing more than 23 million trips, driving auto-piloting a total distance of 350 million kilometres, with a total unmanned mileage of more than 240 million kilometres, and has been operating unmannedly in the market for the right and right rudders. Based on real driving data, Apollo World extracts high-value scenes, and the results of virtual testing reverse the end of the car, creating a closed loop between real operations and virtual validation.
Physical AI markets are growing rapidly. Grand View Research projects that the global physical AI market size will increase from approximately $81.6 billion in 2025 to about $960.4 billion in 2033, with a compound annual growth rate of 36.1 per cent in 2026 to 2033. From "generated images" to "certification operations", deciphering the "unlike triangle" and Apollo World is providing new technological markers for the scalding of drones to help drive them to the AGI moment.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.