Beyond the Basics: OpenUSD for Advanced Physical AI Simulation

In this talk

1. 0:06

   OpenUSD’s foundations

   Luke traces USD’s development from Pixar’s Presto animation system and describes its spread into robotics, industrial digitalization, and autonomous vehicles. He highlights three capabilities: composing data from different sources, extending the model with domain schemas, and using storage methods beyond files.

2. 5:04

   How standards move forward

   Luke outlines a path from community proposals to tested implementations, adoption, and formal AOUSD specifications. He describes a standards structure that starts with the domain-neutral core, adds specifications for areas such as geometry and physics, then provides guidance for particular industries. He says adoption should build before a specification is standardized.

3. 15:40

   SimReady and the physical AI data gap

   Luke argues that simulation needs more digitally created training scenarios because real-world data capture is limited. NVIDIA’s SimReady work aims to make simulation assets accurate, modular, and consistent across solvers. The announced SimReady SDK is intended to provide specifications, validation guidance, and reference pipelines and runtimes.

4. 20:28

   Partners, participation, and learning

   Luke cites FANUC robot assets, Lightwheel’s physics measurements, Cadence’s thermal specification work, and PTC’s factory and PLM contributions. He points listeners to guides, samples, learning resources, certification, and AOUSD participation as ways to contribute. He emphasizes defining the underlying problem before choosing a technical solution.

5. 34:23

   Interoperability and collaboration questions

   In the Q&A, Luke says Functional Mock-up Interface (FMI) could connect Functional Mock-up Units with USD, allowing assets to include behavior implementations from different solvers; he notes that its roadmap placement is undecided. On large-scene collaboration, he compares factory planning roles with departments in film production and says teams can lose time by addressing the wrong problem.