CADFather: Autonomous CAD Reconstruction through Coordinated Tool Use

CADFather. Autonomous CAD Reconstruction through Coordinated Tool Use. By Gennadiy Savrasov and nine others, from Lomonosov Moscow State University, Innopolis University, and one other institution. Published on the sixth of October.

Turning a 3D shape back into an editable CAD model is still a hard engineering task. Existing methods can propose CAD operations, but the trouble is consistency. No single source of proposals works equally well across different part geometries, or across the different stages of a reconstruction. A method that does well on one kind of part, or early in the process, may do poorly on another kind, or later on.

CADFather does not try to build the one proposer that wins everywhere. It is an autonomous agentic system that coordinates complementary tools, to recover parametric CAD programs from 3D meshes. The coordination is the method.

At the centre is a vision-language assistant. It inspects renders of the target and of intermediate reconstructions, and then decides which candidate CAD programs to extend, which tools to invoke, how many proposals to generate, and when to stop. Two kinds of tool do the work. Learned and algorithmic tools propose CAD operations, while numerical optimisation refines the parameters of programs that already exist. Every proposed or refined program is executed and evaluated, and that result feeds the next decision.

The agent keeps alternative candidate programs for each target part, and preserves the best valid result throughout the reconstruction, so a poor branch never overwrites a good one. It uses pretrained generation and assistant models, with no additional training.

Reconstruction quality and execution validity are evaluated on the full DeepCAD, Fusion360 and MCB test sets, as well as on CADENA-Bench, CADBench and BenchCAD. The authors also analyse computational cost, and the trade-off between cost and reconstruction quality. The abstract does not report the scores themselves.

The bet is that the right combination of existing tools, chosen step by step by an assistant looking at the result, can do what no one tool does alone. It also means the system's strength comes from orchestration rather than from new training: every model inside it is pretrained, and what CADFather adds is the judgment about which one to call, how often, and when the reconstruction is good enough to end. Six benchmarks are where that bet is tested, with its cost measured alongside its quality.