An AI agent called ArtisanCAD can learn from expert CAD recordings and apply that knowledge to generate production-ready 3D models from underspecified prompts.
Researchers introduced ArtisanCAD, a CAD agent built around a new intermediate representation called CAD-IR. The system ingests expert procedure recordings — CATIA operation logs, macro files, drawing notes — and distills them into reusable parameterized skills. When a user prompt is vague or describes only high-level design intent, CAD-IR scaffolds the gap into a complete, executable sequence of CAD operations. The agent then runs those operations through a CATIA backend, checks results against multi-view visual feedback, and iterates until the output meets production standards.
Most text-to-CAD research targets clean, fully specified prompts — a setting that rarely matches real engineering work, where intent is often partial or implicit. ArtisanCAD targets that harder, more practical case: on the Text2CAD benchmark, it reduced mean Chamfer Distance on intermediate prompts from 14.83 to 9.88, a meaningful accuracy gain on a metric that measures how closely generated geometry matches a reference shape. The authors also tested it on four complex automotive components, where the system successfully converted expert CATIA recordings into skills and used them to generate editable variant models.
The broader pattern here is knowledge distillation applied to professional workflows rather than model weights — the expertise lives in the procedure library, not a fine-tuned model. Whether that generalizes beyond CATIA-heavy automotive shops, or requires a fresh library for every new domain, is the question the paper leaves open.