Particle.news
Download on the App Store
4 ARTICLES
·
2d ago

Carnegie Mellon Unveils LegoGPT: AI for Physically Stable Lego Designs

The open-source system combines AI and physics to generate buildable Lego structures, achieving near-perfect stability in tests.

Lego Bricks in child's hands with Lego Duplo blocks and toys background
Image
In case a Lego structure is not stable, the model can roll back to the unstable brick

Overview

  • LegoGPT, developed at Carnegie Mellon University, uses AI to create stable, buildable Lego designs from text prompts.
  • The system integrates a fine-tuned LLaMA-3.2-Instruct model with Gurobi optimization for structural stability analysis.
  • A publicly available dataset, StableText2Lego, includes over 47,000 Lego designs and was used to train the model.
  • Testing confirmed 99.8% of generated structures were stable, with rollback features improving stability rates from 24% to 98.8%.
  • Currently supporting an 8-brick library and a 20×20×20 grid, future updates aim to expand design capabilities and brick types.