USDAgent

Orchestrate USD scenes via natural language.

usdagent run --stage warehouse.usda \
  "Create a warehouse with 50 randomized shelf units in a 20x10m floor area, \
   add concrete material to the floor, metal material to the shelves, vary heights 2-3m"

[planner] Decomposed into 6 steps
[tool] create_stage warehouse.usda
[tool] create_prim /World/Floor type=Mesh
[tool] create_material /World/Looks/Concrete
[tool] bind_material /World/Floor → Concrete
[tool] scatter_on_surface target=/World/Floor count=50 ...
[tool] validate_stage → OK
[done] 47s · 6 tool calls · scene saved.

---

Install

pip install usdagent
export ANTHROPIC_API_KEY=sk-ant-...

Requires Python 3.11+ and usd-core.

Quick Start

from usdagent import run

result = run(
    prompt="Add a 10x10m concrete floor with 20 scattered crates",
    stage_path="scene.usda",
)

Features

• Natural language → USD stage operations
• Procedural scatter on surfaces and in volumes
• Material creation and binding (UsdPreviewSurface)
• Scene validation — catches broken refs and layer issues
• Works headless — no Omniverse app required
• Optional Omniverse Kit extension for live preview
• Supports Anthropic Claude, OpenAI, and local Ollama models
• Full Python API + CLI

CLI

usdagent run "Create a scene"            # NL → USD
usdagent validate scene.usda            # integrity check
usdagent info scene.usda                # metadata

Architecture

User Prompt → Planner LLM → Tool Selector → Tool Execution → Validate → Reflect → done/loop

Built on LangGraph. Stage handle lives in agent state — never serialized into LLM context. Every mutation batch ends with USD validation.

License

Apache 2.0