内容来源：https://aibusiness.com/robotics/nvidia-physical-ai-push-robotics-edge-ai-updates

内容总结：

英伟达全面扩张物理AI生态，从芯片到工业合作全栈布局

在东京的一场发布会上，英伟达宣布大幅扩展其物理AI与机器人产品组合，涵盖了从AI模型、仿真软件、边缘计算硬件到工业合作伙伴关系的全链条。此举旨在加速智能机器人的部署，而不仅仅是提供AI芯片，目标是成为整个机器人行业的全栈平台标准。

核心发布：新一代边缘AI硬件与“世界模型”

英伟达推出了全新的Jetson T3000和T2000边缘计算模块，基于Blackwell架构，专为人形机器人、自主移动机器人等设计，在更小体积和更低功耗下支持多模态AI工作负载。同时发布的还有Cosmos 3 Edge，这是一个拥有40亿参数的世界基础模型，可直接在边缘设备上运行，使机器人无需依赖云端即可实时理解环境、推理并生成动作指令。

破解数据难题：边缘模型的价值

分析师指出，当前部署物理AI的最大障碍并非模型本身，而是“数据”。机器人所需的操作数据（如环境感知、力反馈、故障模式）体量巨大且难以跨机器采集。Cosmos 3 Edge的优势在于支持事后训练，能更好地适应异构的机器人及工业设备集群。

日本“宇宙联盟”扩大，工业巨头集体入局

英伟达将其“Cosmos联盟”扩展至日本，吸引了发那科（FANUC）、富士通、日立、川崎重工、久保田、NEC、软银、索尼和安川电机等巨头加入，共同开发开放的物理AI模型与机器人应用。英伟达CEO黄仁勋称，此举将日本的机电一体化技术与英伟达的物理AI结合，开创“工业自动化新时代”。分析师认为，这些合作至关重要，因为物理AI的实现要求IT（信息技术）与OT（运营技术）深度融合。

丰田合作升级：从自动驾驶到工厂与智慧城市

丰田宣布扩大与英伟达的AI合作，不再局限于自动驾驶（Drive平台），而是延伸至车辆软件工程、工厂制造及城市基础设施。丰田将利用英伟达的Omniverse和Isaac Sim创建工厂的数字孪生，并开发用于交通管理的多模态视觉语言模型。

Metropolis平台升级：赋能视觉AI智能体

英伟达还更新了Metropolis平台，新增80多个开发者库和AI技能，增强了DeepStream、TAO等工具，旨在简化能够分析实时视频、生成摘要、识别事件并支持自动决策的视觉AI智能体开发。

行业观察：雄心与挑战并存

Forrester分析师指出，英伟达的芯片再好，也需要嵌入机器人、汽车或起重机中才能发挥价值，因此其工业合作伙伴提供了将AI转化为生产系统所需的运营专业知识和信誉。不过，分析师也警告，大规模部署仍取决于能否解决数据质量、系统集成和运营准备度等长期挑战。企业需持续投资理解其希望自动化的流程，寻找物理自动化、AI与人力资源的最佳平衡点。

中文翻译：

由谷歌云赞助
选择您的首批生成式AI应用场景
要开始使用生成式AI，首先应聚焦于能够改善人类与信息交互体验的领域。

这家芯片制造商正在从基础模型、边缘硬件到软件、开发工具及产业合作，全面构建其物理AI生态系统。

英伟达正大幅扩展其物理AI与机器人技术产品组合，发布了新型边缘AI硬件、机器人基础模型、开发者软件及产业合作计划，旨在加速智能机器的部署。

这些更新在东京的一场活动中公布，涵盖英伟达机器人技术的完整堆栈，从AI模型和仿真软件到边缘计算硬件及制造合作项目。

分析师指出，此次公布的广度反映了英伟达立志成为行业物理AI全栈平台的雄心，其目标是在机器人技术开发的每个阶段巩固自身地位，而不仅仅是供应AI芯片。

高德纳高级总监分析师希曼舒·库马尔·奥贾表示：“英伟达此次发布进一步将其已领先的平台扩展为物理AI的全栈赋能者。通过提供完整的端到端参考架构……英伟达正将自己定位为机器人行业的事实平台标准。”

作为更新的一部分，英伟达推出了Cosmos 3 Edge——一款拥有40亿参数的全新世界基础模型，专为在边缘设备上直接运行而设计。该模型使机器人和视觉AI系统能够理解周围环境、实时推理并生成动作，无需依赖云计算。

然而，Omdia（Informa TechTarget旗下机构）机器人技术分析师亚历克斯·韦斯特表示，部署物理AI的最大障碍在于数据本身，而非模型。

他说：“企业面临的最大挑战是试图在数据仍然混乱的环境中应用AI。”

他指出，尽管数据准备已成为部署大语言模型的先决条件，但机器人技术需要更大量的操作数据，包括环境感知、力反馈及故障模式等难以收集且难以在机器之间传输的信息。

韦斯特认为，这正是Cosmos 3 Edge可能展现价值之处，其可进行后训练的能力使其更适合异构机器人车队及工业设备。

英伟达还将Cosmos联盟扩展到日本，FANUC、富士通、日立、川崎重工、久保田、NEC、软银、索尼和安川电机加入该倡议，共同开发开放的物理AI模型及机器人应用。

英伟达CEO黄仁勋在媒体预沟通会上发布的声明中称，此举将日本的机电一体化技术与英伟达的物理AI相结合，旨在开创“工业自动化的新时代”。

这些合作凸显了物理AI向基于生态系统的发展模式的更广泛转变：AI开发者、机器人制造商和工业企业各自贡献技术栈的不同部分。

韦斯特说：“这些合作不仅对英伟达至关重要，从更广范围看对物理AI的演进也意义重大。它们突显了物理AI需要信息技术与操作技术的融合。”

他表示，机器人制造商通常缺乏构建基础模型所需的AI专业知识，而AI公司则普遍缺乏大规模部署机器人所需的行业知识、客户关系及市场策略。

Forrester副总裁兼首席分析师保罗·米勒表示，这些合作最终将决定物理AI进入企业环境的速度。

他说：“世界上最优秀的芯片，如果没有机器人、汽车或起重机来嵌入它，用处也不大。”他认为英伟达的产业合作伙伴提供了将AI技术转化为生产系统所需的操作专业知识和公信力。

此外，丰田宣布扩大与英伟达的AI合作，从自动驾驶领域延伸到工厂与智慧城市。

在去年达成的使用英伟达Drive开发下一代驾驶辅助系统的协议基础上，丰田现将在车辆软件工程、制造和城市基础设施中全面使用英伟达技术。

扩展后的合作包括使用Omniverse和Isaac Sim创建工厂数字孪生、利用AI模型辅助软件开发，以及丰田子公司Woven开发的多模态视觉语言模型来支持交通管理和城市智能。

英伟达还发布了Jetson T3000和Jetson T2000——两款基于Blackwell架构的新型边缘计算模块，旨在为类人机器人、自主移动机器人及其他智能机器提供算力。这些模块在保持对多模态AI工作负载（包括大语言模型、视觉语言模型和机器人基础模型）支持的同时，相比旗舰产品Jetson AGX Thor体积更小、能效更高。

英伟达还更新了Metropolis平台，新增80多个开发者库和AI技能，包括对DeepStream、TAO和视觉AI软件的增强。

这些新增功能旨在简化能够分析实时视频、生成摘要、识别事件并支持自动化决策的智能视觉AI系统的开发。

综合来看，这些发布展现了英伟达的雄心：提供构建和部署物理AI所需的大部分底层基础设施，涵盖基础模型、仿真软件、边缘计算及产业合作。

米勒表示，这一战略可能加速企业采用，但他提醒，广泛部署仍将取决于能否克服数据质量、系统集成和运营准备度方面的长期挑战。

米勒说：“我们还需要持续投入，以理解我们可能希望自动化的流程。在Forrester，我们讨论‘自动化三角’（平衡物理自动化、AI和人类劳动力）。找到这些能力的最佳组合，仍是目前很少有企业能始终做对的事情。”

英文来源：

Sponsored by Google Cloud
Choosing Your First Generative AI Use Cases
To get started with generative AI, first focus on areas that can improve human experiences with information.
The chipmaker is fleshing out its physical AI ecosystem, from foundation models and edge hardware to software, developer tools and industrial partnerships.
Nvidia is broadly expanding its physical AI and robotics portfolio, unveiling new edge AI hardware, robot foundation models, developer software and industrial partnerships aimed at accelerating deployment of intelligent machines.
Unveiled during an event in Tokyo, the updates span Nvidia's entire robotics stack, from AI models and simulation software to edge computing hardware and manufacturing collaborations.
Analysts said the breadth of the announcements reflects Nvidia's ambition to establish itself as the industry's full-stack platform for physical AI, strengthening its position across every stage of the robotics development pipeline rather than simply supplying AI chips.
"Nvidia's announcements further expand their already leading platform toward a full-stack enabler of physical AI," Himanshu Kumar Ojha, a senior director analyst at Gartner, said. "By providing complete, end-to-end reference architectures ... Nvidia is positioning itself as the de facto platform standard for the robotics industry."
As part of the updates, Nvidia launched Cosmos 3 Edge, a new 4 billion parameter world foundation model designed to run directly on edge devices. The model enables robots and vision AI systems to interpret their surroundings, reason in real time and generate actions without relying on cloud computing.
However, Alex West, a robotics analyst at Omdia, a division of Informa TechTarget, said the biggest obstacle to deploying physical AI is data rather than the models themselves.
"The biggest challenge for enterprises is trying to apply AI in environments where data is still messy," he said,
While data readiness has become a prerequisite for deploying large language models, he said robotics requires much larger volumes of operational data, including environmental awareness, force feedback and failure modes that are difficult to collect and transfer between machines.
That is where Cosmos 3 Edge could prove valuable, West said, arguing its ability to be post-trained makes it better suited to heterogeneous fleets of robots and industrial equipment.
Nvidia is also expanding its Cosmos Coalition into Japan, with FANUC, Fujitsu, Hitachi, Kawasaki Heavy Industries, Kubota, NEC, SoftBank, Sony and Yaskawa Electric joining the initiative to develop open physical AI models and robotics applications.
In a statement released at a media pre-briefing, Nvidia CEO Jensen Huang described the move as bringing together Japan's mechatronics and Nvidia's physical AI to create "a new era of industrial automation."
The partnerships underscore a broader shift in physical AI toward ecosystem-based development, where AI developers, robotics manufacturers and industrial firms each contribute different parts of the technology stack.
"These partnerships are crucial, not just for Nvidia, but more broadly for the evolution of physical AI," West said. "They highlight the fact that physical AI requires convergence of IT and OT."
Robot manufacturers often lack the AI expertise needed to build foundation models, he said, while AI companies generally lack the industrial knowledge, customer relationships and go-to-market strategies needed to deploy robotics at scale.
Paul Miller, vice president and principal analyst at Forrester, said those partnerships will ultimately determine how quickly physical AI reaches enterprise environments.
"The best chip in the world is not a lot of use without a robot or a car or a crane in which to embed it," he said, arguing Nvidia's industrial partners provide the operational expertise and credibility needed to turn AI technology into production systems.
Separately, Toyota announced an expanded AI partnership with Nvidia, moving beyond autonomous driving into factories and smart cities.
Building on last year's agreement to develop next-generation driver assistance systems using Nvidia Drive, Toyota will now use Nvidia technologies across vehicle software engineering, manufacturing and urban infrastructure.
The expanded partnership includes the use of Omniverse and Isaac Sim to create digital twins of factories, AI models to assist software development, and a multimodal vision-language model developed by Toyota subsidiary Woven to support traffic management and urban intelligence.
Nvidia also unveiled the Jetson T3000 and Jetson T2000, new Blackwell-based edge computing modules aimed at powering humanoid robots, autonomous mobile robots and other intelligent machines. The modules offer a smaller, more power-efficient alternative to the flagship Jetson AGX Thor while maintaining support for multimodal AI workloads, including large language models, vision-language models and robot foundation models.
Nvidia also updated its Metropolis platform, adding more than 80 developer libraries and AI skills, including enhancements to DeepStream, TAO and Vision AI software.
The additions are designed to simplify the development of agentic vision AI systems capable of analyzing live video, generating summaries, identifying incidents and supporting automated decision-making.
Taken together, the announcements demonstrate Nvidia's ambition to provide much of the underlying infrastructure needed to build and deploy physical AI, from foundation models and simulation software to edge computing and industrial partnerships.
Miller said the strategy could accelerate enterprise adoption, but cautioned that widespread deployment will still depend on overcoming longstanding challenges around data quality, system integration and operational readiness.
"We also need continued investment in understanding the processes we might want to automate," Miller said. "At Forrester, we talk about the Automation Triangle [balancing physical automation, AI and the human workforce]. Finding the best mix of those capabilities is something that few firms consistently get right yet."