BrickWright: Agentic Creation of Physically Feasible Brick Builds
BrickWright-guided agents reached 100% physical validity across 512 prompts, compared with 9%–69% for baseline agents.
The framework turns text or image prompts into LEGO builds using pretrained agents and a Python toolkit called BrickLayer. Its validator checks for real parts, collisions, floating bricks and connected structures where possible, giving agents feedback to correct their assemblies. The assembly API halves inference cost compared with working in raw LDraw. The authors report improved prompt adherence for lightweight agents and maintained adherence for stronger agents.
HF Daily Papers' note
The framework turns text or image prompts into LEGO builds using pretrained agents and a Python toolkit called BrickLayer. Its validator checks for real parts, collisions, floating bricks and connected structures where possible, giving agents feedback to correct their assemblies. The assembly API halves inference cost compared with working in raw LDraw. The authors report improved prompt adherence for lightweight agents and maintained adherence for stronger agents.
HF Daily Papers' note
score 5