CRAFT BOT: Physical Prototyping Through Large Language Models

From brief via plans to construction, with AI

This project investigates whether large language models can
participate meaningfully in architectural design when constrained to operate through executable CAD code. We introduce CraftBot, an AI system that generates architectural geometry in Blender via Python scripts, forcing all design reasoning to manifest as procedural construction. Thirteen experiments explore how CraftBot interprets heterogeneous reference material, including photographs, drawings, and construction manuals, and how it incrementally refines geometry through an iterative loop of code generation, execution, visual feedback, and revision. The experiments show that LLMs can assemble coherent architectural structures, absorb domain-specific construction logic, and maintain parametric relationships over multiple iterations when code becomes the primary representational medium. At the same time, CraftBot remains largely incapable of autonomously detecting its own geometric errors, requiring human guidance to identify inconsistencies and resolve ambiguities. 

The project presents how architectural relevance for AI can emerge from integration of the real-world knowledge in expert domains and its participation in the design process through a constrained, executable construction logic.

The authors documented the process of CraftBot development and its results in a research paper and are currently looking for publication venue. One of the procedural models generated by CraftBot will be built as a scaled building model and exhibited together with visual outputs of other designs.

....
 

The project author is Luka Piškorec, previously a lecturer at Aalto University and researcher at ETH Zürich. He is a co-founder of TEN Studio (Zürich and Belgrade) and {protocell:labs} (Helsinki), practices that work at an intersection of architecture, design, digital art and research.

The project is part of art-ai-fact initiative.