Can It Actually Work? Synthesizing Functional CAD Mechanisms

TL;DR

The paper introduces FunCAD-Harness, a framework that enables LLM agents to generate functional CAD mechanisms by translating high-level goals into testable physical behaviors and using simulation feedback to iteratively improve designs. It also presents FunCADBench, a benchmark with 46 mechanism families requiring complex, multi-step operations to validate functionality. Experiments show the framework significantly boosts functional success rates compared to standard approaches.

Why It Matters

This work advances the capability of AI agents to design working mechanical systems from vague human descriptions, moving beyond visually plausible but non-functional outputs. It enables more reliable automation in engineering design, reducing the need for expert intervention in early-stage prototyping.

Key Findings

• FunCAD-Harness uses simulation-based testing tailored to individual mechanism designs to detect functional flaws invisible to visual inspection.
• The framework improves functional success rates substantially over baseline simulation-augmented LLM agents across diverse mechanism types.
• FunCADBench provides a rigorous evaluation platform for functional CAD generation, capturing interdependent and multi-step mechanical behaviors.