NVIDIA's nanousd-labs Generates USD Runtimes from Spec with AI Agents

NVIDIA published nanousd-labs, an experimental project that generates a lightweight USD runtime directly from the USD Core Specification using AI agents. The project, part of NVIDIA Omniverse Labs, was built during an internal hackathon and demonstrates a workflow where agents read the formal standard, write code to satisfy its behavior, and validate against a test suite derived from the same spec.

Spec as contract

The USD Core Specification, developed by the Alliance for OpenUSD (AOUSD), is a versioned, machine-readable standard defining how USD data models are composed and resolved across layer stacks. Unlike adapting a large existing codebase, nanousd-labs treats the spec as the contract: agents consume it section by section, generate implementations for each behavior, and iterate until output matches requirements. The methodology aims for compliance built into the process, not tied to any single implementation. Engineers still own performance, tradeoffs, and architectural decisions; agents handle mechanical spec-to-code work like parsing, scene composition, and value resolution.

Lightweight runtime with stable C ABI

nanousd is a C++ implementation exposing a stable C ABI, callable from any language. It parses, composes, queries, and writes USD scenes but stops before rendering. The runtime carries only what a specific workload requires, with memory, threading, and ABI left as implementation choices. Client code compiles against a fixed C API and loads its implementation at runtime, so backends can be swapped without changing calling code. The project includes nanousd-python, a Python package requiring no GPU and running headlessly.

Two entry points

Developers can clone and build nanousd directly, using the Python package to open stages, traverse prims, and author assets. Agents grounded in the spec can author stages and validate compliance, returning composed scene structures and noting corrections. The second entry point is a roughly 10-minute tutorial that walks through codifying human directions into reusable skills, generating a USD ASCII parser that meets the Core Specification. The skill graph captures prompts, recipes, and tests, making the workflow reusable rather than reinvented per implementation. The project does not yet cover the entire specification, and multi-part cohesion remains a frontier, but the standard provides a durable contract for agent-driven development.