NVIDIA Physical AI Agent Bootcamp: 2 of 4 Courses Are Live

The first course in the NVIDIA Physical AI Agent Bootcamp opens with a story. A glowing Memory Cube has gone missing in the Old Attic of a museum, and you're the "agentic engineer on call" who has to make the scene ready for a robot. The tool you work with isn't a robot. It's a coding agent, and the goal is to learn how to point it at real NVIDIA software.

The NVIDIA Physical AI Agent Bootcamp is a set of free, self-paced courses. NVIDIA's own summary says you "build practical physical AI workflows one step at a time" with Omniverse libraries and an AI agent. "Physical AI" here means software for robots, sensors and simulated worlds. OpenUSD is the open scene format those worlds are written in.

Four lines, repeated for every step

The method NVIDIA teaches is a four-field prompt, which it calls the Patterned Prompt Method: Goal (what should be true after this step), Skills (which SKILL.md files the agent should read), Context (scene, hardware, what already exists) and Done when (an observable check, such as a saved frame or a printed label).

The agent reads those skill files, which point it to tested source code, and then writes the program. NVIDIA's setup page adds a rule that applies beyond this course: read the generated diff and run it only once it makes sense to you. A "Done when" line is what stops "the code ran" from being counted as "the work is right."

What the two live courses build

Course 1, "Ray Trace Your Way to a Better Life," produces an RTX viewport app that loads and renders OpenUSD scenes in your browser. The server renders with ovrtx, the browser only displays the stream, and ovstream carries frames and input between them. You then add robot-facing features: a replay-safe move of the cube, a focus camera, semantic segmentation and a lidar sensor, and you package the workflow as a reusable agent skill.

Course 2, "SimReady or Not, Here Comes Gravity," continues in the same viewer. You add ovphysx for physics, run a drop test, validate assets with a SimReady report, apply corrections without overwriting the source files, and prove the result with collider overlays, before-and-after drop tests and a persistent report with runtime telemetry. The course says the evidence has to be "not a rendered picture."

Two courses are still "coming soon"

The overview page lists four courses. "Blender, Oranges, and OVRTX," an RTX-rendered scene with physics inside Blender, and "Isaac Sim and the Fine Art of Picking Things Up," a robot pick-and-place task, are marked "Coming soon." Two other items on the same page carry the same label: the local developer setup path, and Brev instructions for running on a cloud GPU. Until then, NVIDIA says to use the environment linked from each course's overview.

That affects the headline promise. NVIDIA's promotional copy for the bootcamp mentions rendering scenes, validating assets and running real-time physics, which is what the two live courses cover. The robot task that gives "physical AI" its meaning isn't there yet.

What your machine needs

For the released courses, NVIDIA lists Windows 11 or Ubuntu 22.04/24.04, an Intel Core i7/i9 or AMD Ryzen processor, 16 GB of RAM, an NVIDIA GeForce RTX 3070 or better, and 250 GB of free space. You also need a coding agent that supports Agent Skills, such as Cursor, Claude Code or Codex, an account with that provider, and permission for the agent to "read and edit the course folder and run terminal commands." If you install skills separately, use skills CLI v1.5.16 or newer.

The upcoming Isaac Sim course asks for more: Linux on x86_64, 32 GB of RAM, a GeForce RTX 4080 or better with at least 16 GB of VRAM, and 50 GB of SSD space. NVIDIA says GPUs without RT cores, "including NVIDIA A100 and H100 GPUs, are not supported," which rules out data-center cards such as those.

From a live lab to a self-paced course

The material first appeared as hands-on labs. NVIDIA's developer forum lists the Physical AI Agent Bootcamp at SIGGRAPH 2026 on Wednesday, July 22, as four sessions: the browser-streamed viewer, simulation-ready assets, Blender with Omniverse libraries, and Isaac Sim pick-and-place. The documentation we opened is the self-paced version of those labs.

What the pages don't say

The pages give no cost for running an agent through a course, no completion time and no results from learners. We found no NVIDIA social post to link; the short link in the promotional copy resolves to the documentation above. You'll need an account with your agent provider, and the page doesn't say what that costs. The bootcamp itself is free.

If you try it, start with Course 1 and read every diff before the first run. The permission to run terminal commands is the same permission you'd be giving the agent in your own project.

Source: NVIDIA

nvidia omniverse openusd physical ai ai coding agents

AI

About the author

AI Briefing Daily Editorial. The editorial desk of AI Briefing Daily, an independent publication covering artificial intelligence. Every story is researched against primary sources, checked by hand and reviewed by a person before publication.