CADFather: Autonomous CAD Reconstruction through Coordinated Tool Use

AI summary

Overview

Research area: Artificial intelligence for computer-aided design (CAD) reverse engineering, combining vision-language agents with parametric CAD modeling.

Technical level: Advanced. The paper assumes familiarity with CAD boundary representations, sketch-and-extrude modeling sequences, differentiable or numerical shape optimization, and agentic tool-use frameworks.

Scope: The paper presents CADFather, an autonomous agent that coordinates learned, algorithmic, and optimization-based tools to reconstruct executable parametric CAD programs from 3D meshes, and evaluates it on six datasets.

What This Paper Is About

A 3D mesh describes a shape's surface but does not record the modeling operations and dimensions used to build it, so a mesh alone cannot expose quantities such as a bore diameter or wall thickness as editable parameters. The paper's goal is to recover an executable CAD program from a target mesh, where the reconstructed program must be runnable, structurally coherent, and usable as an editable model. The authors argue that no single source of CAD operation proposals works equally well across different part geometries and stages of reconstruction, so their system lets a vision-language assistant decide, at each step, which candidate program to develop and which specialized tool to apply.

Key Contributions

1. A mesh-to-CAD reconstruction agent that uses visual and execution feedback to select candidate programs, choose reconstruction tools, and allocate how many learned proposals to sample.
2. A shared reconstruction framework combining learned proposals, algorithmic proposals, parameter optimization, and persistent retention of the best valid result, with all tools producing programs in a common CadQuery-based representation so a candidate from one tool can be extended or refined by another.
3. Evaluation on six datasets — DeepCAD, Fusion360, MCB, CADENA-Bench, CADBench, and BenchCAD — with published-method comparisons.
4. Ablations removing algorithmic proposals, parameter optimization, or both, to measure each tool's contribution; the system uses pretrained generation and assistant models without additional training.

Main Findings

• Full test-set gains: On DeepCAD, CADFather reaches mean IoU 0.987 against 0.966 for CADENA-RL with sampling; on Fusion360, 0.976 against 0.952; on MCB, 0.913 against 0.895. The IoU gain over sampling is of similar size on all three datasets (0.018–0.024).
• Zero invalidity ratio on the three main test sets: IR is 0.0000 on DeepCAD, Fusion360, and MCB, meaning every part ends with a closed solid. Mean GMS is also higher (0.983, 0.960, and 0.766 against 0.974, 0.942, and 0.751), and median CD at 30000 points is lower (0.039, 0.034, and 0.083 against 0.042, 0.036, and 0.089).
• The gain is not from a stronger generator: The learned tool uses the same CADENA-RL checkpoint, so the improvement comes from tool coordination rather than model quality.
• Ablation: algorithmic proposals matter most. Removing algorithmic proposals lowers the mean score by 0.012–0.110 on the 1000-part subsets and raises IR from zero to 0.007–0.124. Removing the optimizer lowers the mean score by only 0.004–0.006 and leaves IR at zero.
• CADENA-Bench: Across all 3396 parts, IR is 0.0000, mean IoU 0.909, mean GMS 0.721, compared with 0.876 and 0.670 for CADENA-RL greedy, the strongest published row. Mean IoU is 0.909 over the 2240 targets with defined IoU. Shaft-and-bushing (0.943 IoU, 0.764 GMS) and lever-and-profile (0.934, 0.819) families reconstruct best; gear-and-bearing (0.880, 0.660) and spring-and-fastener (0.890, 0.672) families have the lowest IoU and GMS among families with enough parts to compare.
• Metric coverage differs by domain: Only 66% of CADENA-Bench targets have a defined IoU, and 58% on MCB, against 95–99% on DeepCAD and Fusion360, so GMS carries a larger part of the comparison on mechanical parts.
• Volume matching does not imply surface matching: Mean GMS on MCB remains 0.766, well below the IoU, which the authors say suggests that matching volume does not imply matching surface structure on mechanical components.
• CADBench and BenchCAD: Under the paper's protocol, CADBench yields IR 0.0007, IoU 0.904, GMS 0.815; BenchCAD yields IR 0.0000, IoU 0.948, GMS 0.873. Twelve of the 18000 CADBench parts end without a valid prediction because their dense target meshes make IoU computation exceed the 30 s limit. Quality varies by family: the Base family of sketch-and-extrude parts is reconstructed almost exactly (IoU 0.980, GMS 0.974), while the Organic family is hardest (IoU 0.759, GMS 0.548). On BenchCAD, mean IoU is 0.948 over 16350 targets with defined IoU, and IoU and GMS decrease monotonically from easy to hard tier (0.956 to 0.938 and 0.886 to 0.859).
• CADBench under the benchmark's own evaluation code: CADFather reaches IoU 0.930, SIoU 0.787, CD 0.029, and VSR 0.968, exceeding CADFit (0.859, 0.679, 0.038, 0.981) on the three quality metrics and on each of the 14 complexity splits, but its VSR is lower than CADFit's because 576 parts exceed the evaluator's 30 s limit.
• BenchCAD Vision2Code comparison: CADFather reaches voxel IoU 0.968 with IR 0.0 on a 1000-part random subset, against 0.910 and IR 0.9 for CADENA-RL on all 17900 parts.
• Cost analysis: The paper reports voxel IoU against estimated token cost per part on a log scale, evaluated after 1, 2, 3, 4, 6, 8, 12, and 24 assistant steps, with assistant and generator tokens priced at the OpenRouter rate for Qwen3.8-27B ($0.425 and $2.55 per million input and output tokens), and analyzes the trade-off between cost and reconstruction quality.
• Remaining weaknesses: The lowest scores fall on helical springs and twist drills, which remain poor even on targets with defined IoU.

Methodology in Plain English

The system takes a target mesh and searches for a program whose executed geometry matches it.

Candidate pool: Every candidate program records its identifier, parent, source tool, program, generated mesh, number of operations, validity, and measurements. Extending a candidate creates a child rather than modifying stored history; textual deduplication prevents an identical program from creating a new entry. At each step the assistant sees a bounded table of up to twelve candidates in descending score order, restricted to candidates that built, are watertight, and still admit a legal action; the protected best and the empty start are also shown.

Decision step: One decision is made for the whole part, and the assistant may act on several candidates at once. Its input has four parts: a fixed rule description (task, objective, tools with approximate cost, answer format); a transcript of the reconstruction so far recording the assistant's one-sentence rationale and the actions actually executed; a set of renders showing the target, the best candidate so far, and the best candidate produced by the previous step; and the candidate table. The assistant issues up to four act calls, each naming a tool and a candidate, plus a sample count for learned proposals, and may issue a finish call with reason shape or stalled.

Three tools: Learned proposal generation samples next-operation proposals from a pretrained CADENA-RL generator, with the assistant choosing how many samples to draw. Algorithmic proposal generation fits operations to the target geometry without a learned model: on the empty start it proposes a revolve for bodies of revolution or extrusions found by cutting the target with planes, and on an existing candidate it searches for missing material to add and extra material to cut. Each call returns the next four proposals from a ranked list, and each proposal becomes a separate alternative program. Parameter optimization refines the sketch coordinates, radii, and extrusion depths of an existing program while keeping sketch planes, revolve axes, and angles fixed, using numerically estimated gradients of a mean-squared signed-distance mismatch, fitting coarse points first and then points near the surface.

Validation and measurement: A candidate is invalid if its program does not build, its mesh is not as a whole a watertight solid with positive volume, or execution or measurement exceeds the time limit; this is stricter than CADENA's rule, which accepts a prediction if any connected component is a closed solid. The target is centered and scaled by its largest dimension, and candidates are brought to the same scale by a single fixed transform rather than by fitting their own bounding boxes, so a wrongly sized candidate scores lower instead of having its scale error normalized away. Score is the mean of available metrics — IoU and GMS, or GMS alone — and invalid candidates score zero. The final output is the protected best, retained independently of the assistant's later choices.

Budgets and recovery: Per part, the limits are 24 assistant steps, at most 12 operations in a program, and 1000 seconds of wall time, with budgets of 80 generator samples, 6 algorithmic-proposal calls, 10 optimizer calls, and 120 program executions; a single learned call requests at most 32 samples, and building and measuring one candidate is limited to 30 seconds. Malformed replies trigger a request for a revised action, then a fallback to the cheapest available action on the best candidate; failed executions are reported back and the protected best is never lost.

Models and data: The CADENA-RL checkpoint is served in bfloat16, and the assistant is Qwen3.8-27B with FP8 weights and explicit thinking disabled; both models are frozen. The full test sets contain 8046 (DeepCAD), 1725 (Fusion360), and 5000 (MCB) parts; ablations use fixed 1000-part subsets.

Why This Matters

Impact on research. The paper argues that tool choice and candidate selection should be explicit decisions made against intermediate geometry, rather than relying on a fixed operation source or repeated editing of a single program. Because both the generator and the assistant are frozen, the reported gains over CADENA-RL isolate the effect of coordinating complementary tools in a shared representation, and the ablations attribute most of the benefit to algorithmic proposals rather than to the optimizer. The zero invalidity ratio on DeepCAD, Fusion360, and MCB, and the strong CADENA-Bench results, provide a reference point for future work on mesh-to-CAD reconstruction.

Real-world applications:

• Recovering editable parameters such as bore diameters or wall thickness from scanned, legacy, or history-free components.
• Reverse engineering where a component is available only as a scan or a mesh without its construction history.
• Reusing an existing mesh as the starting point for modification, since an executable program can be inspected and edited rather than only viewed.
• Feeding recovered programs into downstream mechanical workflows that require a buildable solid, given the system's emphasis on validity, watertightness, and positive volume.

Industry relevance. The evaluation spans mechanical-part families from multiple sources and industrial part families, and the paper reports computational cost and a cost-versus-quality trade-off — practical considerations for deployment. The reported failure cases (dense meshes exceeding a 30 s IoU limit, the 576 CADBench parts exceeding the evaluator's time limit, and weak results on helical springs and twist drills and the Organic family) mark where engineering effort would be needed before broad industrial use.

Future Directions

• Broaden parameter optimization coverage: The optimizer handles extrusions, holes, and revolves, and programs with other operations cannot be refined; extending it to a wider operation set is an open problem.
• Close the surface-matching gap: Mean GMS on MCB (0.766) and on CADENA-Bench (0.721) trails IoU substantially, and the authors note the two means cover different subsets; improving surface and normal agreement on mechanical parts remains unresolved.
• Fix the remaining failure modes: Helical springs, twist drills, and the Organic family of object-like meshes score lowest, and twelve CADBench parts fail because dense target meshes make IoU computation exceed the execution limit.
• Address cost and time budgets: The paper presents a token-cost analysis and notes that 576 CADBench parts exceed the evaluator's 30 s limit, so reducing cost while retaining quality is a stated direction through its cost–quality analysis.

Target Audience

Researchers and practitioners working on CAD reverse engineering, parametric modeling, 3D shape reconstruction, and tool-using vision-language agents. It is also relevant to engineers who need to convert legacy or scanned meshes into editable CAD programs, and to readers who want a worked example of combining learned models, geometric algorithms, and numerical optimization inside a single agentic search loop. The paper's density of metrics, budgets, and protocol details makes it most useful to readers already comfortable with CAD representations and benchmark methodology.

Note: the provided paper content is truncated mid-sentence within the BenchCAD Vision2Code comparison, so any results appearing after that point in the original paper are not reported here.

Authors’ abstract

Reconstructing an editable CAD model from a 3D shape remains a challenging engineering task. Existing methods can propose CAD operations, but no single source of proposals works equally well across different part geometries and stages of reconstruction. We introduce CADFather, an autonomous agentic system that coordinates complementary tools to recover parametric CAD programs from 3D meshes. A vision-language assistant inspects renders of the target and intermediate reconstructions, then decides which candidate CAD programs to extend, which tools to invoke, how many proposals to generate, and when to finish. Learned and algorithmic tools propose CAD operations, while numerical optimization refines the parameters of existing programs. Proposed or refined programs are executed and evaluated to provide feedback for subsequent decisions. The agent maintains alternative candidate programs for each target part and preserves the best valid result throughout reconstruction. CADFather uses pretrained generation and assistant models without additional training. We evaluate reconstruction quality and execution validity on the full DeepCAD, Fusion360, and MCB test sets, as well as on CADENA-Bench, CADBench, and BenchCAD. We additionally analyze computational cost and the trade-off between cost and reconstruction quality.