AI Radar

Noteworthy AI conversations and writing, summarized from publisher notes.

How OpenUSD and AI Are Building Smarter Virtual Worlds

· NVIDIA AI Podcast

In episode 268 of the NVIDIA AI Podcast, host Noah Kravitz speaks with Aaron Luk, NVIDIA's Director of Product Management for Simulation Technology and lead of the OpenUSD ecosystem effort. Aaron explains that OpenUSD (Universal Scene Description) was open-sourced by Pixar in 2016 and has since become a foundational framework for 3D worldbuilding across filmmaking, manufacturing, robotics, and digital twins. He details how USD's layer-based composition and schema system enable non-destructive collaboration, allowing multiple disciplines and teams to build on each other's work without overwriting it, while abstracting away the underlying data sources. The conversation centers on why USD is a natural fit for physical AI. Aaron describes how USD provides a unified way to describe robots, their environments, and the scenarios they encounter, integrating cleanly with NVIDIA Cosmos for synthetic data generation and Omniverse (with sensor RTX) for physically accurate sensor simulation. He addresses the sim-to-real gap, noting that while USD does not solve physics simulation itself, it standardizes the inputs to multiple solvers so different physics engines can co-simulate different parts of a scene. He also discusses digital twins for industrial customers like Lowe's and Siemens, and the emerging standards work through AOUSD, where he chairs the core specification working group. Aaron closes with practical onboarding advice: NVIDIA's Deep Learning Institute offers a Learn OpenUSD curriculum with a certification path, complemented by blueprints on build.nvidia.com, docs.omniverse.nvidia.com, and the open-source USD GitHub repository. He notes that even non-coders can begin using USD via copilots and Omniverse-integrated tools, and reflects on USD's evolution from a 2012 Pixar pair-programming project into a broad cross-industry standard.

AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.