您正在使用IE低版浏览器，为了您的雷峰网账号安全和更好的产品体验，强烈建议使用更快更安全的浏览器
此为临时链接，仅用于文章预览，将在时失效
0
导语：从世界模型、Robotaxi、端侧智能等方向，讨论AI如何真正进入物理世界。
10月26日，第八届智能驾驶大会暨2026环秀湖物理AI产业发展推介会在苏州相城举行。
上一届，车路云一体化、智能驾驶出海和无人驾驶商业化是讨论重点，而到了今年，大会的讨论边界进一步从智能驾驶向物理AI延伸。
世界模型、强化学习、端侧智能、虚拟仿真、Robotaxi之外，机器人、低空飞行器等不同形态的物理智能开始出现在同一场大会上。这背后，是智能驾驶行业正在面对的一组新问题。
一方面，高阶辅助驾驶已经进入规模化上车阶段，Robotaxi也开始从技术验证走向运营和商业化；另一方面，当AI真正进入汽车、机器人、飞行器等物理设备，如何获得足够的数据、如何训练和验证模型、如何降低开发成本，以及如何保证系统在真实世界中的安全性，正在成为新的技术难题。
当天，包括Momenta、曹操出行、面壁智能等企业，以及高校和科研机构的多位嘉宾，从世界模型、强化学习、端侧大模型、虚拟仿真、车路云和物理AI基础设施等方向给出了自己的判断。
总结这些发言不难发现，智能驾驶正在成为物理AI最早进入现实世界的一块试验场。而从一辆车的自动驾驶，到机器人、低空飞行器以及城市交通系统，不同技术路线正在回答同一个问题：AI如何从数字世界真正进入物理世界，并形成可以规模化运行的产品。
关键词：智能驾驶安全、车路云一体化、单车智能瓶颈
李克强认为，汽车智能化已经进入大规模应用和商业闭环阶段，当前首要问题是智能驾驶系统安全。仅依靠单车自身的传感器和计算能力存在物理边界，高级别自动驾驶需要从单车智能进一步向车云协同、车路云一体化演进。
他将当前自动驾驶产业的瓶颈归纳为三类：一是部分产品定义与国家、国际标准存在偏差；二是高安全、高可靠自动驾驶的关键技术仍待突破；三是行业投入与商业产出仍不匹配。
技术层面，他认为主要存在三个问题：
感知能力仍有边界。如果车辆“看不见、看不远、看不全”，再强的AI决策也无法弥补信息缺失，因此需要通过超视距感知、数字信号上车和协同感知增加信息来源；
AI模型训练数据仍不完整。自动驾驶正在进入多模态和大模型阶段，但数据类型和场景覆盖不足仍可能产生安全问题；
开发效率较低。当前大量系统仍采用“烟囱式”架构，缺乏统一底座，存在开发门槛高、复用率低和验证周期长等问题。
李克强提出，车路云一体化可以提供两类核心能力：一是建立标准化“数据轨道”，向车辆提供广域、多维、实时交通信息；二是形成面向AI大模型的数据服务体系，为模型持续提供高可靠数据。
目前相关方案已经进入产业化验证。包括北京、重庆在内的城市正在开展车企协同开发，应用场景覆盖前方静止障碍物、“鬼探头”、绿波通行等。部分项目已经进入量产阶段，今年已有7家车企推进相关应用，中国一汽也已公布量产导入。
他认为，高级别自动驾驶未来需要解决三个问题：正视安全瓶颈、突破单车智能的物理边界，并通过车路云一体化降低开发成本、提高研发效率。
关键词：物理AI、数字孪生、飞行汽车、需求响应公交
曲小波认为，物理AI正在推动数字孪生从“映射和监视现实世界”，进一步走向“理解、推演并实时控制现实世界”。
传统数字孪生主要完成物理实体到数字实体的映射，但数字端推演往往需要数秒，很难实时控制物理系统。相关国家重点研发项目希望重点解决四个问题：实现具有泛化能力的数字推演、将推演时间压缩到100毫秒以内、建立数字端与物理端互操作机制，以及通过跨尺度、多层次架构降低末端执行误差。
其技术体系分成三个层级：
- 组件级：提高电机等执行机构的设计效率和执行精度。通过物理AI学习既有电机数据，可以在约10%误差下，将部分单点计算从小时级缩短到毫秒级；
- 载具级：解决Sim-to-Real Gap、不同机型与动力学适配以及快速验证，应用于自动驾驶、飞行器自主巡航和工程机械自主作业；
- 系统级：实现多种异构载具之间的协同，包括需求响应公交、低空交通管控、矿山工程机械等。
在飞行汽车方面，团队与奇瑞合作开发了自主巡航复合翼模块化飞行汽车，由机翼、座舱和底盘组成，可以分别组合成自主巡航飞行器和自动驾驶汽车。
工程机械方面，通过世界模型生成候选轨迹并完成轨迹选择，同时优化起铲点和卸料位置，使工程机械能够完成环境建模、铲装、运输和卸料的连续作业。
需求响应公交目前已在湖南株洲职教城常态化运行18个月。当地约有10万名师生，超过50%已经注册使用。与出租车相比，点到点行程时间只相差约2分钟，但平均单次成本约为2元，对比出租车约9.16元。
低空交通方面，则通过构建起降区域全息环境，进行冲突预测、起降排序和资源调度，目前已经在重庆将通航机场改造成飞行汽车起降测试场。
关键词：世界模型、仿真训练、合成数据、虚实闭环
任冬淳介绍了Yootta WorldLoop世界模型，这是一套面向智能驾驶和物理AI的虚实融合闭环训练与评测系统。
整个系统覆盖数据采集、数据生产、算法训练、算法评测、真机验证和数据回流，并由场景生产与编辑、算法训练、仿真评测与真机验证三个平台组成。
其核心技术包括四个模块：
- CAD和3D结构数据物理化；
- 生成式世界模型；
- 4D数据渲染器；
- 仿真物理引擎。
其中，3D资产除几何和颜色信息外，还会加入质量、摩擦力、弹力和材质等物理属性，使机器人能够真正与虚拟资产发生交互。
目前相关系统已经覆盖家居、工业、商业零售、智能驾驶行车和泊车五类场景。
在具身智能场景中，可以加载不同机器人的URDF模型，并设置抓取、导航等任务。任冬淳表示，如果通过真机采集1000条抓取数据，每条需要约1分钟，总耗时超过1000分钟；采用GPU并行仿真，可以将同等规模的数据生产和训练缩短到数分钟至十几分钟。
智能驾驶方面，可以利用车端或路端传感器重建真实道路环境，再改变天气、光照、车辆数量等条件，用于行车和泊车算法训练与验证。
目前Yootta已经与尚品宅配开放500套家居场景，每套包含数百至上千个可以与机器人交互的资产。下一阶段计划构建5000个虚实对应场景，发布100万个真实资产、500万小时合成数据，并适配约50款机器人硬件本体。
关键词：RoboX、Robotaxi、物理AI Agent、天地空一体化出行
龚昕认为，AI时代出行行业的竞争正在从流量入口转向一体化供给能力。
曹操出行由此提出RoboX战略，希望把网约车、出租车、巴士、低速物流车、货运卡车和eVTOL等不同移动能力，统一封装成可以被AI Agent实时调用的服务。
在这一体系中，用户未来可能不再需要单独打开叫车APP，而是由AI根据日程、位置和个人需求，自动选择并调用相应出行服务。
曹操出行将物理AI在出行领域的商业模式分为四类：
- 基础出行服务：统一调度Robotaxi、高铁、低空飞行器等交通工具；
- 超级物理AI Agent：让车辆成为旅游、餐饮、娱乐等现实服务入口；
- 实时立体数据服务：利用车内外传感器形成实时环境数据，在合规前提下服务企业和城市管理；
- 分布式算力：进一步利用车辆闲置算力。
支撑RoboX战略的是Robo OS，负责连接线上AI Agent与线下移动载具，并将用户偏好、出行场景、履约结果、位置和路线等真实数据反馈给AI。
目前曹操出行Robotaxi已经覆盖六座城市，车队规模超过200辆；Robo One在长沙、成都部署超过50台，并计划进入苏州。RAS远程中心已经在杭州和苏州落地。
按照规划，到2030年，曹操出行计划在苏州投放数千台Robotaxi和Robo One，实现7×24小时常态化运营。
其发展路径分为三个阶段：2026年跑通Robotaxi和Robo One载人、运货的准入和运营闭环；2027年随着正向研发Robotaxi车型下线开始规模复制；到2030年，将苏州模式向全国及海外市场推广。
关键词：Agnes Harness、物理AI、Test Time Training、硬件协议
杨通认为，AI正在出现三项变化：通用大模型进入更多垂直领域，模型开始通过Skill和训练反馈实现自我进化，以及大模型开始直接连接物理设备。
基于这一趋势，爱诗人工智能开发了Agnes Harness，希望降低通用模型向垂直智能体和物理AI应用迁移的成本。
Agnes Harness采用前后端解耦设计，将模型能力与MaaS、记忆、沙箱、搜索与工具、上下文管理、团队协作等能力封装起来，再通过Plugin和Skill机制适配不同应用。
在物理AI层面，其核心能力之一是MHS，即通过标准硬件协议，让大模型能够用自然语言连接显微镜、医疗设备、摄像头、无人机等可编程设备。
目前已经进行小车与机器狗协同实验，两台设备依靠摄像头感知环境，再由大语言模型直接参与决策，不使用传统遥控，也不需要完整预设动作流程。
另一个重点是ISI，主要包括两部分：
- 将模型运行过程中的能力持续沉淀为Skill；
- 通过TTT（Test Time Training）让模型在实际运行过程中继续更新。
杨通表示，传统模型数据采集、训练和推理由不同团队完成，模型上线以后往往长期不变；TTT希望把数据采集、执行反馈和模型更新连接起来，使模型能够针对具体环境持续适应。
目前相关实验使用约0.9B参数模型进行验证，部分模型约每半小时更新一次。在游戏测试中，持续积累Skill后，部分任务胜率从约3%提升到90%，三对三场景从0提升到约60%。
目前爱诗人工智能模型月调用量约65万亿Token，月图片生成量约5000万张，视频生成量约1.7亿秒。Agnes相关系统已经被用于芯片设计、算子优化、无人机集群、蛋白质合成等场景。
关键词：高阶辅助驾驶、世界模型、强化学习
顾功尧认为，中国高阶辅助驾驶已经进入快速上量阶段，城区NOA方案开始进入10万元及以下车型，全球市场则整体晚于中国数年。
Momenta的业务模式可以概括为“一飞轮、两条腿”：一条是量产辅助驾驶，另一条是L4规模化无人驾驶。量产车型持续产生真实道路数据，再反哺Robotaxi等L4产品。
目前，搭载Momenta辅助驾驶系统的乘用车已经超过130万辆，累计定点超过240个，交付120多款车型，累计真实行驶里程约200亿公里。
Momenta将物理AI理解为“世界模型+强化学习”。
其R7世界模型主要分为三个阶段：
- 预训练：利用上百亿公里真实道路数据，让世界模型学习物理规律、交通行为和道路常识；
- 仿真推演：模拟自车和其他交通参与者的行为变化；
- 强化学习：在世界模型构建的仿真环境中进一步训练车端驾驶模型。
Momenta目前使用统一模型架构支撑L2++到L4产品，量产辅助驾驶、Robotaxi等产品共用同一条模型和数据主线，只根据法规和安全要求设置不同冗余。
R7已经应用于多款量产车。例如复杂山路中，系统需要同时处理骑行者、隔离带和对向车辆，寻找短暂的安全超车窗口；在夜间城区绕过违停车辆时，还需要提前识别远处隔离墩，并规划重新回到原车道的轨迹。
Momenta正在预研下一代R9世界基础模型。R7参数规模约10亿级，R9计划提升到约千亿级，同时车端芯片算力也从百TOPS级向千TOPS级演进。
R9希望进一步理解更加细微的物理变化，例如不仅识别山体落石，还能够推演落石后续运动是否符合自由落体规律。
关键词：密度法则、端侧智能、全模态感知、物理AI
刘知远认为，大模型过去八年的核心变化并不只是参数规模持续扩大，而是形成了一套可以不断吸收不同类型数据的通用技术路线。
随着模型继续扩大，数据和算力正在成为新的约束。
在数据侧，从2024年前后开始，大规模强化学习推动模型由模仿学习向探索式学习发展，使模型可以主动探索并生成新的训练数据。
而在算力侧，如果仍然只依靠扩大参数规模提升能力，计算需求最终难以持续。因此，大模型还需要提高“智能密度”，即使用更少参数和算力达到相同智能水平。
面壁智能由此提出“密度法则”。
根据其观察，全球开源大模型的智能密度大约每100天提升一倍，即大约每100天就可以用一半参数和一半计算量，实现此前相近的智能能力。
刘知远认为，这类似于半导体行业的摩尔定律。未来模型一方面会不断变小，手机、汽车、机器人等终端算力另一方面会继续提高，两条趋势最终交汇，使过去只能运行在服务器上的大模型逐步进入终端。
他将未来AI分为三种形态：超级智能、云侧智能和端侧智能，其中端侧智能将成为物理AI的重要载体。
人工智能进入物理世界还存在三个趋势：
- 高智能密度：降低模型运行所需的参数和算力；
- 全模态感知：通过视觉、声音和其他传感器理解现实世界；
- 自主智能体：在开放环境中持续决策、学习和行动。
面壁智能已经将端侧模型应用到手机、汽车座舱、卫星和机器人等场景，并与华为、三星等企业开展合作。
按照密度法则推算，如果智能密度继续维持约100天翻倍的速度，不到两年时间，同等能力所需的参数量和算力可能下降两个数量级，使原本万亿级模型的部分能力压缩到百亿级，并部署到终端设备。
刘知远认为，未来大模型将持续进入手机、PC、汽车、机器人、可穿戴设备和IoT等终端，端侧智能将成为AI进入物理世界的重要路径。雷峰网雷峰网(公众号：雷峰网)
雷峰网原创文章，未经授权禁止转载。详情见转载须知。
为了您的账户安全，请验证邮箱
请验证您的邮箱
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
You are using a low-end IE browser and strongly recommend a faster and safer browser for your Thunder Peak account security and better product experience.
This is a temporary link, only for article preview, which will expire at the time
0
Introduction: From the world model, Robotaxi, end-end intelligence, to discuss how AI really enters the physical world.
On 26 October, the 8th Smart Driving Congress and the Association for the Development of the Artistics of Lake 2026 was held in Sang-City, Suzhou.
In the previous session, the integration of road clouds, smart driving to sea and the commercialization of unmanned aerial vehicles had been the focus of discussions, and by this year, the General Assembly had reached a further extension of the boundaries of discussions from smart driving to physical AI.
World models, enhanced learning, end-end intelligence, virtual simulations, and different forms of physical intelligence, such as robotics, low-altitude aircraft and so forth, began to appear at the same conference. Behind this is a new set of problems that the intelligent driving industry is facing.
On the one hand, high-level assistive driving has moved into a scalding phase, and on the other hand, Robotaxi has moved from technical validation to operational and commercialization; on the other hand, when AI actually enters physical equipment such as cars, robots and aircraft, how to obtain adequate data, how to train and validate models, how to reduce development costs and how to ensure the safety of the system in the real world, it is becoming a new technical challenge.
On that day, a number of guests, including Momenta, Cao Cao's travels, face-to-face intelligence, and institutions of higher learning and scientific research, gave their own judgement in the direction of world models, intensive learning, end-end LLM, virtual simulations, road clouds and physical AI infrastructure.
Summarizing these presentations, it is easy to see that smart driving is becoming the first test site for physical AI to enter the real world. From the auto-driving of a car to robots, low-altitude aircraft and urban transportation systems, different technical routes are answering the same question: how does AI actually enter the physical world from the digital world and develop products that can be scaled up.
Keywords: Smart driving safety, road cloud integration, bicycle smart bottlenecks
In Li Keqiang ' s view, the intellectualization of a car has entered the stage of large-scale applications and commercial closed loops, and the first priority at this time is the security of a smart driving system. The physical boundary exists only on the bicycle ' s own sensors and computational capacity, and high-level auto-driving requires further evolution from a bicycle ' s intelligence towards a combination of a car cloud and a road cloud.
He grouped the current bottlenecks in the auto-driving industry into three categories: first, the partial definition of the product differs from national and international standards; secondly, the key technologies for high-security, high-reliability auto-driving still need to be broken; and thirdly, the industry ' s input still does not match commercial output.
At the technical level, three main problems were identified:
There are still boundaries to sensory capacity. If vehicles are “observed, invisible and insufficiently visible”, stronger AI decision-making will not make up for the lack of information and will therefore require increased information sources through over-the-horizon perceptions, digital signals and synergetic perceptions;
The AI model training data is still incomplete. Automatic driving is entering the polymorphic and LLM phases, but the type of data and inadequate coverage of the scene may still create security problems;
Development is less efficient. A large number of systems still use “coastal” structures, lack a unified base, high development thresholds, low reuse rates and long certification cycles.
Li Keqiang suggested that the integration of road clouds could provide two core capabilities: a standardized “data track” to provide wide-area, multidimensional and real-time traffic information to vehicles; and a data service system for AZLLM to provide continued, reliable data for models.
The programmes are now undergoing industrialization certification. Cities, including Beijing and Chongqing, are working on the collaborative development of cars, using scenes to cover the front of static barriers, “ghost scouts”, green waves, etc. Some of the projects are already in the production phase, with seven cars promoting their applications this year, and China's one-size-fits-all imports.
In his view, the future of high-level auto-driving needed to address three issues: facing security bottlenecks, breaking the physical boundaries of a bicycle ' s intelligence, and reducing development costs and improving research and development efficiency through the integration of road clouds.
Keywords: Physical AI, digital twins, flying cars, demand response bus
According to Xinbo, physics AI is pushing digital twinning from “mapping and monitoring the real world” to “understanding, evolving and controlling it in real time”.
Traditional digital twinning mainly completes the mapping of physical entities to digital entities, but the digital end-to-end often takes seconds to control physical systems in real time. The country-focused R & D project aims to focus on four issues: achieving a broad-based digital rollover, reducing the rollover time to within 100 milliseconds, establishing a digital-physical interoperability mechanism and reducing implementation errors at the end through a multiscale, multilayered structure.
The technology system is divided into three levels:
- Component level: Increased design efficiency and implementation accuracy of implementing agencies such as electric power. (a) Learning from existing electrical data through Physical AI, which reduces the number of single-point calculations from hourly to milliseconds at about 10 per cent error;
- Loading stages: the resolution of Sim-to-Real Gap, the adaptation of different types of aircraft to kinetics and rapid validation for automatic driving, autonomous vehicle cruises and mechanical engineering;
- System level: to achieve synergy among various isomers, including demand-response bus, low-altitude traffic control, mine engineering machinery, etc.
In the area of flying vehicles, the team, in cooperation with Chire, developed an autonomous cruise composite-wing modular flight vehicle, comprising wings, cabins and chassis, which can be combined into an autonomous cruise vehicle and an automatic driving vehicle.
For engineering machines, candidate tracks are generated through world models and trajectory selection is completed, while at the same time optimizing removal points and discharges to enable engineering machines to complete the continuous operation of environmental modelling, shoveling, transport and discharges.
Demand-responsible bus traffic has now been in regular operation in Hunanjia for 18 months. There are approximately 100,000 teachers and students, over 50 per cent of whom are registered for use. Compared to taxis, the point-to-point travel time is only about two minutes, but the average single-time cost is about $2, compared to taxis about $9.16.
For low-altitude transportation, conflict forecasting, landing sequencing and resource mobilization are being carried out through the construction of a holographic environment for the off- and off-off areas, which is now being converted into a flight vehicle landing test site in Chongqing.
Keywords: world model, simulation training, synthetic data, false closed loop
Ren Dong-soon presented the Yootta World Loop World Model, a virtual integrated closed-ring training and evaluation system for smart driving and physics AI.
The system covers data collection, data production, algorithm training, algorithm assessment, real-time validation and data back-flow, and consists of three platforms for scenario production and editing, algorithm training, simulation and validation.
Its core technology consists of four modules:
- Physicalization of CAD and 3D structural data;
- Generating world models;
- 4D data renderer;
- Simulation physics engine.
Of these, 3D assets, in addition to geometric and colour information, add physical properties such as mass, friction, ballistics and materials, which enable robots to interact with virtual assets.
The five types of home, industry, commercial retail, smart driving and parking are now covered by the relevant systems.
In a smart scene, the URF model of different robots can be loaded and tasks such as capture, navigation, etc. set up. Ren Dong-soon indicated that it would take about one minute each to collect data from a genuine machine, for a total of more than 1,000 minutes, and that using the GPU parallel simulation would reduce data production and training on the same scale to a few minutes to ten minutes.
In terms of smart driving, end-end or road-end sensors can be used to re-establish the real road environment and to change conditions such as weather, light, number of vehicles, etc., for use in driving and parking algorithm training and validation.
At present, Yootta has opened up 500 home scenes with a decent home, each containing hundreds to thousands of assets that can interact with robots. The next phase is planned to build 5,000 false equivalents, publish1 million real assets, 5 million hours of composite data, and fit around 50 robotic hardware units.
Keywords: RoboX, Robotaxi, Physical AI Agent, Sky Empty Okay.
It is argued that the competition in the AI-era travel industry is shifting from a flow point to an integrated supply capacity.
As a result of his trip, Cao Cao has proposed a RoboX strategy, which seeks to integrate the different mobile capacities of Internet-based vehicles, taxis, buses, low-speed logistics vehicles, freight trucks and eVTOL into a service that can be accessed in real time by AI Agent.
In this system, users may in the future no longer need to open a car called APP on its own, but it is AI that automatically selects and calls the corresponding travel service according to schedule, location and personal needs.
Cao Cao travels in four categories:
- Basic travel services: unified movement of vehicles such as Robotaxi, heavy iron, low-altitude aircraft, etc.;
Superphysics AI Agent: making vehicles a real service entry point for tourism, catering, entertainment, etc.;
- Real-time stereodata services: the use of internal and external sensors to generate real-time environmental data to service business and city management subject to compliance;
- Distributive computing: greater use of vehicle idle computing.
The RoboX strategy is supported by Robo OS, which connects AI Agent on the line with an offline moving vehicle and feeds back to AI the true data of user preferences, travel scenes, performance results, location and route.
At present, Cao Cao's travel to Robotaxi has covered six cities, with a fleet of over 200; Robo One has deployed more than 50 in Changsha and Cheng, and plans to enter Suzhou. RAS telecentres have landed in Hangzhou and Suzhou.
By 2030, Cao Cao ' s trip was planned to launch thousands of Robotaxi and Robo One in Suzhou, with 7x24 hours of normal operation.
Its development path is divided into three phases: running through Robotaxi and Robo One in 2026, access to and operation of cargo, and closed loops; replicating in 2027 on a scale as it moves towards the development of the lower line of the Robotaxi model; and extending the Suzhou model to national and overseas markets by 2030.
Keywords: Agnes Harness, Physical AI, Test Time Trading, Hardware Protocol
According to Yang Tong, three changes are taking place in AI: General LLM is entering more vertical areas, the model is beginning to evolve itself through Skill and training feedback, and LLM is beginning to connect physical equipment directly.
This trend led to the development of Agnès Harness, a poetic artificial intelligence that seeks to reduce the cost of moving the universal model to a vertical AI agent and a physical AI application.
Agnes Harness uses a back-to-back decoupling design, encapsulating model capabilities with those of Maas, memory, sandboxes, search and tools, context management, teamwork, etc., and adapting them to different applications through the Plugin and Skill mechanisms.
At the physical AI level, one of its core competencies is MHS, which allows LLM to connect microscopes, medical equipment, cameras, drones and other programmable equipment in natural languages through standard hardware protocols.
There has been an experiment with small cars working with machine dogs, two devices relying on a camera sensory environment and direct decision-making by large-language models, without the use of traditional remote control and without the need for a full pre-set workflow.
Another focus is ISI, which consists mainly of two parts:
- Maintaining the capacity of the model during its operation as Skill;
- To keep the model updated in practice through TTT.
Yang Tung indicated that traditional model data collection, training and justification were done by different teams and that the model tended to stay on line for a long time; TTT hoped to link data collection, implementation feedback and model updating to enable the model to adapt continuously to specific circumstances.
Currently, the relevant experiments are validated using a model of approximately 0.9B parameters, some of which are updated approximately every half hour. After the continuous accumulation of Skill in the game tests, the success of some tasks rose from about 3 per cent to 90 per cent and three-to-three scenarios from 0 to about 60 per cent.
There are currently about 65 trillion Token monthly calls for poetry artificial intelligence models, about 50 million monthly pictures and about 170 million seconds for videos. Agnes-related systems have been used in such settings as chip design, algorithm optimization, drone clusters and protein synthesis.
Keywords: high-level assistive driving, world model, enhanced learning
According to Zhong Guo Yiu, China ' s high-level auxiliary driving has entered a fast upscale phase, the city NOAA programme has begun to enter the 100,000 or less model and the global market as a whole has been behind China for several years.
Momenta's business model can be summarized as “one wheel, two legs”: one is a mass-assisted driving and the other is a L4-sized unmanned driving. The volume model continues to produce real road data, feeding L4 products such as Robotaxi.
At present, more than 1.3 million passenger vehicles are carrying the Momenta auxiliary driving system, with a cumulative number of more than 240 fixed points and a delivery of more than 120 models, with a cumulative mileage of approximately 20 billion kilometres.
Momenta understood physics AI as “World Model + Enhanced Learning”.
Its R7 world model is divided into three main phases:
- Pre-training: use of 100 billion kilometres of real road data to enable world models to learn about physical patterns, traffic behaviour and road commons;
- Hypothetical simulation: simulation of behaviour changes in the self-car and other transport participants;
- Enhanced learning: further training on end-of-car driving models in a world model-building simulation environment.
Momenta currently supports L2++ to L4 products using a uniform model structure, with the same model and data main line shared by the Auxiliary Drive, Robotaxi and others, with different redundancys based on regulations and safety requirements.
R7 has been applied to multiple-volume vehicles. For example, in complex mountain roads, the system needs to deal with both cyclists, isolation belts and the search for short, secure superwindows for vehicles, and, in the case of night city bypasses, to identify in advance the remote silos and plan for re-orbiting back to the original track.
Momenta is preparing for the next generation of R9 World Foundation models. The R7 parameter size is about 1 billion degrees, the R9 plan is to be raised to about 100 billion stages, while the end chip calculus has evolved from 100 TOPS to 1,000 TOPS.
R9 hopes to further understand more subtle physical changes, such as not only identifying mountain rocks, but also being able to influence whether the fall-off follow-up movement is consistent with the law of free fall.
Keywords: Density rule, end-side intelligence, all-modular perception, physics AI
According to Liu, the core change of LLM over the past eight years is not just the continued expansion of the size of the parameters, but rather the development of a common technical route that can be used to absorb different types of data over time.
As models continue to expand, data and algorithms are becoming new constraints.
On the data side, starting around 2024, a large-scale enhanced learning drive model has evolved from imitation to exploratory learning, enabling the model to proactively explore and generate new training data.
On the other hand, when only the ability to scale up the parameters continues to be relied upon, computing demand is ultimately not sustainable. Therefore, LLM also needs to increase its “intelligence density”, i.e. to achieve the same level of intelligence using fewer parameters and algorithms.
The face wall intelligence thus proposes the "density rule".
According to its observations, the global open source LLM is about to double its intellectual density every 100 days, i.e. half of the parameters and half of the calculated amount per 100 days, to achieve the previously close intelligence capability.
Liu Zhiway believes that this is similar to the Mor law of the semiconductor industry. The future model will become smaller on the one hand and the end-calculation of mobile phones, cars and robots on the other will continue to improve, and the two trends will eventually converge so that LLM, which used to operate only on the server, will gradually enter the terminal.
He divided the future AI into three forms: super-intelligence, cloud-side intelligence and end-side intelligence, in which end-side intelligence will become an important vehicle for physical AI.
There are three other trends in artificial intelligence entering the physical world:
- High-intelligence density: to reduce the parameters and algorithms required for model operation;
- Full-modular perception: understanding the real world through vision, sound and other sensors;
- Autonomy AI agent: Continuous decision-making, learning and action in an open environment.
Face-wall intelligence has applied end-side models to such scenarios as mobile phones, car compartments, satellites and robots, and has worked with Chinese and tristar companies.
According to the density method, if the intellectual density continues at a rate of about 100 days to double in less than two years, the parameter count and the calculator required for the same capacity may fall by two orders of magnitude, compressing part of the original trillion-class model to 10 billion and deploying it to the terminal equipment.
Liu Zhiyuan believes that in the future LLM will continue to enter terminals such as mobile phones, PCs, cars, robots, wearable equipment and Iot, and that end-side intelligence will be an important route for AI to the physical world. Thunder Peak Network (WeChat Official Account:Ray Peak Network)
Original Ray Peak article, unauthorized reproduction. For further information, see the information below.
For your account security, please verify the mailbox.
Please verify your mailbox.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.