Introducing CoreWeave Physical AI Launchables: Run Workloads With One Command

介绍 CoreWeave 物理 AI 可启动项：一条命令运行工作负载

无需先构建环境，即可运行物理 AI 循环依赖的工作负载。

by

Justin Hodges

复制

机器人和自动驾驶（AV）工程师需要承担太多角色。从感知、与物理原型匹配的仿真、运动规划、控制系统，到所有这些背后的代码，所有工作最终都为了将真正可用的机器人交付出去。在正式运行前的那周是最昂贵的，这时需要有人搭建环境、连接存储设备并等待调度器。CoreWeave 物理 AI 可启动项通过一系列经过测试、即开即用的模板优化了这一流程，这些模板基于这些工程团队已经在使用的工具。选择一个模板，调整关键设置，然后通过一条命令即可运行。

首发时，CoreWeave 物理 AI 可启动项目录已涵盖 NVIDIA Isaac Sim、CARLA、MuJoCo、NVIDIA AlpaSim 和 NVIDIA Cosmos 3。同时包含新的 NVIDIA Cosmos Data Factory 代理技能，用于在视频、图像和文本上进行生成式 AI 数据增强，以及自动标注、数据整理和检索。更多 Cosmos Data Factory 工作流即将作为可启动项模板上线。

CoreWeave 工程师构建并测试了每个模板。无论您的团队是否曾经运行过该工作负载，环境都已经就绪。

通过一条命令运行模板

从专为物理 AI 优化的堆栈开始

您无需构建或管理工作负载底层的基础设施。CoreWeave 负责运行这些底层组件，每个组件都经过 AI 工作负载优化并相互连接：针对工作负载调优的基础设施、在运行时展示运行表现的可观测性系统，以及用于开发和改进模型与代理的 CoreWeave Forge。所有功能都与您团队现有工具保持开放兼容。

CoreWeave 物理 AI 可启动项在底层增加了一个工程层：您团队用于开发机器人、自动驾驶车辆和数字孪生的软件工具。这些软件并非我们开发。我们集成的是您团队已经选择的工具：NVIDIA Isaac Sim、CARLA 等，未来还会持续增加更多工具。

集成方式围绕工作流程展开：模型、数据集、环境、实验和评估。您选择影响结果的变量，我们处理底层基础设施。每个模板都位于物理 AI 循环的某个环节：观察、整理、改进、评估、运行。

‍

运行机器人和自动驾驶工作流

您现在可以在 CoreWeave 物理 AI 可启动项上运行的内容，以及即将上线的功能。目录按您团队已使用的软件进行组织，每个软件下都有针对具体任务的模板：生成场景、评分场景集、训练策略。每个模板都会声明其需求和输出内容，因此在运行前您可以判断是否匹配。

只需简单调整设置即可更改模板。在命令行中传递 name=value 参数对，调整场景、持续时间、并行度或待测试模型，然后指向您自己的数据。如果现有模板都无法匹配您的任务，每个工具都提供一个空白模板：已测试的环境，工作负载由您自行处理。

您无需在列表中看到自己的具体项目才能运行它。

工具

可用模板

NVIDIA Isaac Sim

• 空白：您自己项目的起点

• Franka-pick-place：Franka 机械臂抓取放置，附带视频片段

/think

• Scene-sdg: 叉车与托盘场景，带标签的合成数据

• Vehicle-dynamics: 车辆变体，驾驶和制动轨迹

• Warehouse-actors: 仓库中的工作人员和事件，带标签的多摄像头数据

CARLA

• Scenario-matrix: 跨矩阵的驾驶场景，配有视频

MuJoCo

• Playground-rl-sweep: 基于MuJoCo Warp的GPU并行强化学习训练

• Kitchen-pick-place: 使用RoboCasa的厨房操作场景

NVIDIA AlpaSim

• Closed-loop-driving: 在重建的真实场景中对NVIDIA Alpamayo驾驶模型进行闭环评估

NVIDIA Cosmos 3

• Scenario-generation: 使用Cosmos 3 Nano生成稀有或高成本场景的文本到视频

Voxel51 FiftyOne

即将推出：

• 构建高质量数据集和计算机视觉模型的开源工具

对于更深入的修改，直接在纯文本中编辑模板的应用代码。以下是NVIDIA Isaac Sim的示例：基于一个基准的变体，每个变体都是人工编辑的，并在约30秒内启动。没有任何脚本化操作。当下一个实验只需这么短的时间来设置时，你可以测试更多替代方案，并将更多时间用于分析结果。

<div style="display: flex; flex-direction: column; gap: 16px;"> <div style="display: flex; flex-wrap: wrap; gap: 16px;"> <figure style="flex-grow: 1; flex-basis: calc((600px - 100%) * 999); min-width: 0; margin: 0;"> <div style="position: relative; padding-top: 75%;"> <iframe src="https://customer-vonlksqj0yyd6d75.cloudflarestream.com/d5ca7981086c8bf9f887ce4ac99ce9ec/iframe?muted=true&preload=true&loop=true&autoplay=true&poster=https%3A%2F%2Fcustomer-vonlksqj0yyd6d75.cloudflarestream.com%2Fd5ca7981086c8bf9f887ce4ac99ce9ec%2Fthumbnails%2Fthumbnail.jpg%3Ftime%3D%26height%3D600&controls=false&letterboxColor=transparent" title="Camera viewpoint" loading="lazy" style="border: none; position: absolute; top: 0; left: 0; height: 100%; width: 100%;" allow="accelerometer; gyroscope; autoplay; encrypted-media; picture-in-picture;" allowfullscreen="true"></iframe> </div> <figcaption>摄像头视角</figcaption> </figure> <figure style="flex-grow: 1; flex-basis: calc((600px - 100%) * 999); min-width: 0; margin: 0;"> <div style="position: relative; padding-top: 75%;"> <iframe src="https://customer-vonlksqj0yyd6d75.cloudflarestream.com/dad33198bc97f7f946e5f53cab07c1b8/iframe?muted=true&preload=true&loop=true&autoplay=true&poster=https%3A%2F%2Fcustomer-vonlksqj0yyd6d75.cloudflarestream.com%2Fdad33198bc97f7f946e5f53cab07c1b8%2Fthumbnails%2Fthumbnail.jpg%3Ftime%3D%26height%3D600&controls=false&letterboxColor=transparent" title="Warm lighting" loading="lazy" style="border: none; position: absolute; top: 0; left: 0; height: 100%; width: 100%;" allow="accelerometer; gyroscope; autoplay; encrypted-media; picture-in-picture;" allowfullscreen="true"></iframe> </div> <figcaption>暖光照明</figcaption> </figure> <figure style="flex-grow: 1; flex-basis: calc((600px - 100%) * 999); min-width: 0; margin: 0;"> <div style="position: relative; padding-top: 75%;"> <iframe src="https://customer-vonlksqj0yyd6d75.cloudflarestream.com/b7ad1b086c94f12c107bc563eec9f1ff/iframe?muted=true&preload=true&loop=true&autoplay=true&poster=https%3A%2F%2Fcustomer-vonlksqj0yyd6d75.cloudflarestream.com%2Fb7ad1b086c94f12c107bc563eec9f1ff%2Fthumbnails%2Fthumbnail.jpg%3Ftime%3D%26height%3D600&controls=false&letterboxColor=transparent" title="UR10e robot embodiment" loading="lazy" style="border: none; position: absolute; top: 0; left: 0; height: 100%; width: 100%;" allow="accelerometer; gyroscope; autoplay; encrypted-media; picture-in-picture;" allowfullscreen="true"></iframe> </div> <figcaption>UR10e机器人实体</figcaption> </figure> </div> <div style="display: flex; flex-wrap: wrap; gap: 16px;"> <figure style="flex-grow: 1; flex-basis: calc((600px - 100%) * 999); min-width: 0; margin: 0;"> <div style="position: relative; padding-top: 75%;"> <iframe src="https://customer-vonlksqj0yyd6d75.cloudflarestream.com/6dc8a027a207fec8062442f2860d0143/iframe?muted=true&preload=true&loop=true&autoplay=true&poster=https%3A%2F%2Fcustomer-vonlksqj0yyd6d75.cloudflarestream.com%2F6dc8a027a207fec8062442f2860d0143%2Fthumbnails%2Fthumbnail.jpg%3Ftime%3D%26height%3D600&controls=false&letterboxColor=transparent" title="Custom carton object" loading="lazy" style="border: none; position: absolute; top: 0; left: 0; height: 100%; width: 100%;" allow="accelerometer; gyroscope; autoplay; encrypted-media; picture-in-picture;" allowfullscreen="true"></iframe> </div> <figcaption>自定义纸箱物体</figcaption> </figure> <figure style="flex-grow: 1; flex-basis: calc((600px - 100%) * 999); min-width: 0; margin: 0;"> <div style="position: relative; padding-top: 75%;"> <iframe src="https://customer-vonlksqj0yyd6d75.cloudflarestream.com/d82fa16fa86c3ea9112aa1de36b757d1/iframe?muted=true&preload=true&loop=true&autoplay=true&poster=https%3A%2F%2Fcustomer-vonlksqj0yyd6d75.cloudflarestream.com%2Fd82fa16fa86c3ea9112aa1de36b757d1%2Fthumbnails%2Fthumbnail.jpg%3Ftime%3D%26height%3D600&controls=false&letterboxColor=transparent" title="Placement into bin" loading="lazy" style="border: none; position: absolute; top: 0; left: 0; height: 100%; width: 100%;" allow="accelerometer; gyroscope; autoplay; encrypted-media; picture-in-picture;" allowfullscreen="true"></iframe> </div> <figcaption>放入料箱</figcaption> </figure> <figure style="flex-grow: 1; flex-basis: calc((600px - 100%) * 999); min-width: 0; margin: 0;"> <div style="position: relative; padding-top: 75%;"> <iframe src="https://customer-vonlksqj0yyd6d75.cloudflarestream.com/b7ad1b086c94f12c107bc563eec9f1ff/iframe?muted=true&preload=true&loop=true&autoplay=true&poster=https%3A%2F%2Fcustomer-vonlksqj0yyd6d75.cloudflarestream.com%2Fb7ad1b086c94f12c107bc563eec9f1ff%2Fthumbnails%2Fthumbnail.jpg%3Ftime%3D%26height%3D600&controls=false&letterboxColor=transparent" title="UR10e robot embodiment" loading="lazy" style="border: none; position: absolute; top: 0; left: 0; height: 100%; width: 100%;" allow="accelerometer; gyroscope; autoplay; encrypted-media; picture-in-picture;" allowfullscreen="true"></iframe> </div> <figcaption>UR10e机器人实体</figcaption> </figure> </div> <p style="text-align: center;">Franka机械臂抓取与放置模板（使用NVIDIA Isaac Sim）</p> </div>

Camera viewpoint

Warm lighting

Placement target

Pickup position

Added scene fixture

Heavier payload (250g)

Custom carton object

Placement into bin

UR10e robot embodiment

Franka arm pick-and-place Template with NVIDIA Isaac Sim

More Templates are on the way. If something essential for you is missing, reach out and tell us .

Leverage agents to run your workflows

Engineers already let agents write simulation code and drive the command line. An agent asked how to close a sim-to-real gap shouldn't spend its context working out how to stand up a cluster. It should have what it needs to do the work the team cares about.

People and agents work through the same visible workflow: its inputs, settings, runs, and results. You can inspect what an agent changed and trace the result back to the work that produced it.

NVIDIA publishes reference AI agent skills across its physical AI software. They're built to run on any cloud, so they don't prescribe how that cloud is operated. CoreWeave Physical AI Launchables ships its own specific skills on top, tuned for high performance on CoreWeave's cloud. These skills arrive with every Template, and work with Claude Code, Codex, and other coding AI agents .

What every Template gives you

Choose a Template, change the settings that matter, run it, and get the results back. The infrastructure work comes with it.

Work at the level of your domain. Start from a robotics, AV, or engineering Template and set the parameters specific to your work. The packaging, deployment, and execution happen underneath.

Reuse and adapt working Templates. Change settings or edit the application code without rebuilding what runs it. Share the result with colleagues, and the next experiment starts from something that already ran.

Keep GPUs busy on GPU work. Where the workload permits, Templates finish CPU-only preparation and post-processing before GPU execution begins, and release capacity as soon as the GPU stage ends.

Run in parallel without collisions. Each Template declares its own resource and concurrency limits, so parallel work doesn't compete with itself.

Automate through scripts and agents. Every operation is documented and every response is machine-readable. A script or a coding agent can launch a run, watch it, and retrieve the results.

Get every result back in one place. Reports, metrics, media, and diagnostics all come back through the same interface. Nothing to go digging for.

Reach CoreWeave from the tools you already use

Not everything runs through a Template. Some teams already live in a tool they like, so we’re working with software providers in physical AI to run their products on CoreWeave. The interface stays theirs and the compute runs on CoreWeave. Partners include Cadence, Convergent Science, Encord, nTop, PhysicsX, and Voxel51 , spanning vision and data tooling through engineering simulation.

"Physical AI demands a new level of performance and scale across the entire AI development lifecycle, from data enrichment through to model inference. By bringing together CoreWeave's accelerated compute infrastructure and Encord's AI data platform, we're making it easier for teams to work with multimodal data and advance their AI model development." Eric Landau, Co-founder and CEO, Encord.

"Physical AI 对基础设施提出了非凡的要求，从处理密集的多模态记录到训练和评估计算密集型模型。通过将 Voxel51 的数据工作流程与 CoreWeave 的高性能计算相结合，我们正在帮助开发者加速数据飞轮，使每次迭代都能将原始传感器数据转化为更优质的模型。" Voxel51 联合创始人兼首席执行官 Brian Moore 表示。

===== 视频 1：NVIDIA Cosmos-AV 在 CoreWeave 云上运行的 Voxel51 平台 ====

在 CoreWeave 云上运行的 Voxel51 平台上的 NVIDIA Cosmos-AV

===== 视频 2：CoreWeave 无服务器推理上的 Encord 代理分类 ====

在 CoreWeave 无服务器推理上的 Encord 代理分类

将其指向您自己的场景和数据

从最接近您工作的 CoreWeave Physical AI Launchables 模板开始：带有剧集视频的抓取和放置、生成带标签合成数据的叉车和托盘场景，或在矩阵中扫过的驾驶场景。

然后告诉我们您实际上想运行的内容。我们将引导您找到匹配的模板，并安排已成功运行该模板的工程师团队，无论是持续的预留容量项目还是单次突发任务来测试一个想法。

请求简报

想先了解硬件细节？我们的解决方案简介涵盖了如何将 GPU 与工作负载匹配，以及我们在 NVIDIA RTX PRO 6000 Blackwell Server Edition 和 NVIDIA H100 GPU 上运行 NVIDIA AlpaSim 闭环驾驶模拟时的实测数据。

阅读解决方案简介

在 CoreWeave 上，通过一个命令即可运行针对 NVIDIA Isaac Sim、CARLA 等测试配置。从最接近您工作的模板开始，并将其指向您自己的场景和数据。

分享本文： /