Skip to content

Latest commit

 

History

History
61 lines (44 loc) · 1.99 KB

File metadata and controls

61 lines (44 loc) · 1.99 KB

geniesim_generator — LLM-driven scene generation 🏗️

A scene-language DSL + LLM tooling that lets a model write a small Python program that builds a Shape tree from a shared asset library, which the package compiles into a scene.usda + layout graph the Genie Sim stack can load.

License: Mozilla Public License Version 2.0 Agent doc: AGENTS.md Skills: skills/


📦 Install

geniesim_generator is gated behind extras because its deps are heavy (LLM / ML stacks). Install via:

pip install -e "source/geniesim/[generator]"
# or directly:
pip install -e source/geniesim_generator/

geniesim bootstrap asks before installing — default is skip.


🛠️ What it does

  • Scene Language DSL — a small Python surface (Shape, primitives, @register helpers) that an LLM can program against.
  • LLM in the loop — Open WebUI hosts the LLM, prompts ship an LLM_RESULT.py script back, the package executes it and emits a USD + layout graph.
  • Asset search — VL-based queries into the shared asset library so the DSL can name real objects, not abstract placeholders.

⚠️ Not a CLI verb. Unlike geniesim benchmark, there's no geniesim generator … subcommand. geniesim_cli treats this package as a pip-installable peer (bootstrap / deploy geniesim_generator / status); end-to-end use happens via the skills below.


🤖 Skills

Skill Purpose
generate-scene Drive the LLM to emit a new scene from a natural-language description
search-assets Query the asset library for objects matching a description
deploy-generator Build / ship the generator stack (Open WebUI + MCP servers)

🔗 Pointers