制造业传统的设计-构建-测试流程基于一个单一假设：真实世界测试是唯一可靠的测试环境。

Manufacturing’s traditional design-build-test cycle rested on a single assumption: Real-world testing was the only reliable test environment.

制造业传统的“设计—建造—测试”循环基于一个基本假设：真实世界测试是唯一可靠的测试环境。如今，高保真仿真技术已能生成精度足以支撑量产级 AI 的合成训练数据。这使得感知系统、推理模型及具身智能工作流得以在实际工厂环境中高效运行。OpenUSD 已成为实现这一目标的连接性标准，而采用该标准的制造商已取得可量化的成果。随着物理 AI 日益融入工业运营，制造商正面临一项基础性挑战：资产无法在 3D 流程之间可靠迁移。每当资产从计算机辅助设计（CAD）工具转入仿真平台，其物理属性、几何信息与元数据均会丢失——迫使团队从头重建。 SimReady 是一项基于 OpenUSD 构建的内容标准，明确定义了具备物理准确性的 3D 资产必须包含哪些要素，方可在渲染、仿真与 AI 训练等不同流程中稳定运行。此外，NVIDIA Omniverse 库提供了具备物理精确性与照片级真实感的仿真层，AI 模型在此完成训练与验证后方可部署。该平台将机器人工作站表征为 USD 文件，并运行与其物理实体完全一致的固件，从而实现在产线建成前即开展机器人训练、零件公差测试与 AI 模型验证。合成训练变量（如光照条件与几何差异）可大规模生成，覆盖大量人工难以复现的场景。 ABB Robotics 工业业务线总经理 Craig McDonnell 表示：“我们已成功实现全技术栈的垂直整合，并将其优化至模拟版本准确率达 99% 的水平。”下游成效包括：产品导入周期最多缩短 50%，调试时间最多减少 80%，设备全生命周期总成本降低 30–40%。捷豹路虎（JLR）亦将“仿真优先”原则应用于车辆空气动力学领域。工程师基于整车产品线逾 20,000 次风洞校准的计算流体力学（CFD）仿真结果，训练神经代理模型；目前 95% 的气热负载任务已在 NVIDIA GPU 上运行。构建于 Omniverse 并部署于 JLR 的 Neural Concept Design Lab，可在设计师调整车辆几何形态时实时可视化气动变化，将原本串行的“先设计、后仿真”流程压缩为连续闭环。过去耗时四小时的结果，如今仅需一分钟即可获得。 一旦工厂投入生产，另一类智能挑战随即展开——而这单靠仿真无法解决。此外，Factory Playback 利用 NVIDIA Cosmos Reason 视觉语言模型，在本地 NVIDIA GPU 上实时解析摄像头视频流与操作员行为。该系统已部署于全球工业设备制造商 Terex（拥有逾 40 座工厂），预计可提升良率 3%，降低返工率 10%。 Tulip Interfaces 联合创始人兼首席信息官 Rony Kubat 表示：“我非常期待制造商如何借助 AI 的强大能力，增强其日常作业能力。” SimReady 资产、Omniverse 库及 NVIDIA 物理 AI 技术栈共同构成了一套基础框架，开发者可据此在任意工业应用场景中采纳、扩展并组合使用。以下为入门指南：

Manufacturing’s traditional design-build-test cycle rested on a single assumption: Real-world testing was the only reliable test environment. Today, high-fidelity simulation produces synthetic training data accurate enough for production-grade AI. This is enabling perception systems, reasoning models and agentic workflows to excel in live factory environments. OpenUSD has emerged as the connective standard that makes this practical, and the manufacturers building on it are already experiencing measurable results. As physical AI becomes integral to industrial operations, manufacturers face a foundational challenge: Assets don’t travel reliably between 3D pipelines. Every time an asset moves from a computer-aided design tool to a simulation platform, physics properties, geometry and metadata are lost — forcing teams to rebuild from scratch. SimReady is the content standard, built on OpenUSD, defining what physically accurate 3D assets must contain to work reliably across rendering, simulation and AI training pipelines. In addition, NVIDIA Omniverse libraries provide the physics-accurate, photorealistic simulation layer where AI models are trained and validated before deployment. The platform represents robot stations as USD files running the same firmware as their physical counterparts, making it possible to train robots, test part tolerances and validate AI models before a production line exists. Synthetic training variations — such as lighting conditions and geometry differences — can be generated at scale, covering scenarios that would be impractical to replicate manually. “We’ve managed to vertically integrate the complete technology stack and optimize it to a point where we’re now achieving 99% accuracy on the simulated version,” said Craig McDonnell, managing director of business line industries at ABB Robotics. The downstream outcomes: up to 50% reduction in product introduction cycles, up to 80% reduction in commissioning time and a 30-40% reduction in total equipment lifecycle cost. JLR applied the same simulation-first principle to vehicle aerodynamics. Engineers trained neural surrogate models on more than 20,000 wind-tunnel-correlated computational fluid dynamics simulations across the vehicle portfolio — with 95% of aero-thermal workloads now running on NVIDIA GPUs. The Neural Concept Design Lab — built on Omniverse and deployed at JLR — visualizes aerodynamic changes in real time as designers adjust vehicle geometry, collapsing what was a sequential design-then-simulate cycle into a continuous loop. A result that once took four hours now takes one minute. Once a factory goes into production, a different intelligence challenge begins — one that simulation alone can’t address. In addition, Factory Playback uses the NVIDIA Cosmos Reason vision language model to interpret camera streams and operator behaviors in real time, running on premises on NVIDIA GPUs. Deployed at Terex, a global industrial equipment manufacturer with over 40 plants, the system is expected to deliver a 3% increase in yield and 10% reduction in rework. “I am excited to see what manufacturers will do with the power of AI to augment their daily capabilities,” said Rony Kubat, cofounder and chief information officer of Tulip Interfaces. SimReady assets, Omniverse libraries and NVIDIA’s physical AI stack provide a foundation developers can adopt, extend and combine across any industrial application. Here’s how to get started: