anteloc/ldraw-nova

What it solves

ldraw-nova allows AI agents to design buildable, physical LEGO models. It overcomes the difficulty LLMs have with complex 3D geometry math by providing a Python-based tooling system that allows agents to write code that generates the model, rather than calculating coordinates manually.

How it works

The system uses an agentic workflow where the AI reads instructions and documentation on the LDraw language and LEGO building. The agent then plans the model, creating a plan.json file. Instead of writing the LDraw source code directly, the agent creates a Python generator script (generate.py) that produces the final LDraw source files.

To refine the model, the agent iteratively renders images of the model, inspects them, and adjusts the positioning and aesthetics until the model is complete. The system includes tools for finding suitable parts, detecting collisions and gaps, and headless rendering for visual inspection.

Who it’s for

Users who want to use AI to generate 3D LEGO models in the LDraw format, as well as developers interested in agentic workflows for physical design.

Highlights

• Code-based generation: Agents generate Python scripts to produce 3D models, sidestepping direct geometry math.
• Visual feedback loop: The agent can render and inspect images of its own creations to iteratively improve the design.
• Multi-format output: Produces LDraw source code, glTF files for Blender, and supports VR interactive views via Meta Quest 3.
• Semantic search: Integrates with jev-rerank for semantic search and re-ranking to help agents find the correct LEGO parts.