FunCADBench

FunCADBench is a benchmark spanning 46 mechanism families that tests whether AI-generated CAD actually works. Agents interpret underspecified requests, such as a cup-access machine, and build mechanisms that must satisfy their functional requirements in physical simulation, not just look right. Families range from clamps and transmissions to mechanical counters, alarm clocks and coin-sorting devices, covering motion primitives, topology, complexity, contact and geometric difficulty.

FunCAD-Harness is the framework behind the benchmark. It translates abstract requirements into candidate-specific physical tests, binds actuation, loads and observations to each candidate mechanism, runs rigid-body physics simulation, and returns structured diagnostic feedback that guides design revision. The project page reports 75 agent-built assemblies covering all 46 families, with 38 of them passing the full FunCAD-Harness evaluation.

Reported results on the project page: Astra solved 8 of 46 tasks with simulation tools alone and 22 with the full harness, and 0 when feedback is reduced to pass/fail only (8 for Grok). Grok with retained history and up to 8 attempts solved 36 of 46, at half of Astra's API cost. The paper is arXiv:2610.26861, Can It Actually Work? Synthesizing Functional CAD Mechanisms, by Wenqian Zhang, Yangyi Huang, Zigeng Xu, Weiyang Liu and Zhen Liu, submitted 2026-10-19.