PaperlayerView exact revision on arXiv

Source-linked AI summary

CADFather: Autonomous CAD Reconstruction through Coordinated Tool Use

Gennadiy Savrasov, Maksim Elistratov, Nikita Gavrilov, Albert Garifullin, Oleg Pavlov, Soslan Kabisov, Vladimir Frolov, Anton Konushin, Andrey Kuznetsov, Dmitrii Zhemchuzhnikov

arXiv:2610.09127v1cs.AIcs.LG

TL;DR

Recovering editable CAD from a mesh requires reconstructing executable operations and parameters, while the useful source of proposals can vary across parts and reconstruction stages. CADFather coordinates learned and algorithmic proposals with parameter optimization through a vision-language assistant that uses execution feedback and preserves alternative candidates. Across the reported benchmarks, it improves several reconstruction metrics against published comparisons, with validity and cost trade-offs varying by benchmark.

  • Problem

    A mesh does not encode the modeling operations and parameters needed for an executable, editable CAD program, and useful reconstruction actions vary across geometries and stages.

    Source: §1 · pp. 1–2

  • Method

    CADFather uses a vision-language assistant to coordinate learned and algorithmic operation proposals and parameter optimization across candidate programs, using executed geometry as feedback.

    Source: §1 · p. 2, §6 · p. 15

  • Results

    Across the full DeepCAD, Fusion360, and MCB test sets, CADFather improves mean IoU and GMS and reduces median Chamfer distance versus published CADENA-RL sampling results, with valid outputs for every part under its stricter validity rule.

    Source: §6 · p. 15

  • Takeaways & Limitations

    The results support coordinating complementary reconstruction tools rather than relying on a single source of CAD operations.

    Source: §6 · p. 15

    The evaluation measures geometric reconstruction and program validity, not recovery of original design history, manufacturing intent, or engineering constraints, and covers individual parts rather than assemblies.

    Source: §5 · p. 14

Abstract

from arXiv · show

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.

1 Introduction

Mesh-to-CAD reconstruction must recover executable, editable programs rather than shape alone. CADFather coordinates complementary proposal and refinement tools through visual and execution feedback, maintaining alternative candidates while evaluating the system and its components.

1 Introduction

  • Recovering an executable CAD program exposes construction operations and parameters, such as bore diameter or wall thickness, for inspection and modification.This is useful when only a scan, legacy mesh, or model without construction history is available; the reconstruction must also be executable and structurally coherent.Source: §1 · p. 1
  • Learned proposals can miss features or dimensions, while algorithmic construction and parameter optimization provide complementary ways to improve candidate programs.The useful action depends on the current reconstruction: add an operation, refine parameters, or return to an earlier candidate.Source: §1 · p. 2
  • CADFather uses a vision-language assistant to select candidates, tools, and learned-proposal counts from rendered geometry and execution feedback.It can extend or refine candidates, end reconstruction, and accept executed programs into a pool that preserves earlier candidates and the best valid result.Source: §1 · p. 2
  • CADFather combines learned proposals, geometry-derived proposals, parameter optimization, and persistent retention of the best valid result in a shared framework.The operation generator and assistant require no additional training; visual-agent interaction with executable CAD environments had already been explored by CAD-Assistant and IterCAD [Mallis et al., 2025, Hu et al., 2026].Source: §1 · p. 2
  • The evaluation covers six datasets, compares published methods, and ablates algorithmic proposals and parameter optimization individually and together.It assesses geometric fidelity, validity, computational cost, and the contribution of making different tools available.Source: §1 · p. 2

2 Related Work

Prior CAD reconstruction work spans generative representations, iterative construction, parameter refinement, and tool-using agents. CADFather targets executable programs for specified meshes and distinguishes itself by coordinating specialized tools while searching across alternative candidates.

Generative CAD modeling

Geometric reconstruction and executable CAD programs

  • Reconstruction methods recover boundary representations, constructive solid geometry, or executable programs, with outputs differing in whether they provide general-purpose code and an explicit modeling sequence.CAD-Recode maps point clouds to CadQuery code [Rukhovich et al., 2025], while CADFather combines learned and geometry-based operation proposals through a common tool interface.Source: §2 · pp. 2–3

Iterative and stepwise CAD reconstruction

  • Iterative methods use intermediate geometry, but vary between revising complete programs and adding operations stepwise; CADFather makes tool selection a separate assistant decision.CADReasoner edits complete programs [Kabisov et al., 2026a], whereas CADENA generates one additional operation from a rendered partial program [Kabisov et al., 2026b].Source: §2 · p. 3

Parameter optimization in CAD programs

  • Parameter optimization changes program geometry without changing its operation structure, unlike selecting or reconstructing modeling operations.CADFather exposes optimization as a separate tool and does not introduce a new numerical optimizer.Source: §2 · p. 3

Tool-using CAD agents

  • Tool-using CAD agents already plan and execute actions with visual feedback, so CADFather's distinction is target-mesh reconstruction over alternative candidate programs.Its assistant uses renders and geometric evaluation to choose both a candidate to continue and a specialized tool, rather than relying on a fixed operation source or repeatedly editing one program.Source: §2 · p. 3

3 Method

CADFather coordinates learned proposals, algorithmic fitting, and parameter optimization to reconstruct executable CAD programs from target meshes. A vision-language assistant directs the search across alternative candidates, while execution, evaluation, and protected-best selection govern the output.

3.1 Problem Formulation and System Overview

  • CADFather represents each candidate as an executable program in a shared DSL, letting proposals from one tool be extended or refined by another.The system keeps alternative candidates and admits programs for further reconstruction only after validation and visual acceptance.Source: §3.1 · p. 3

3.2 Candidate Pool

  • The pool preserves earlier candidates and presents up to twelve eligible programs by score, alongside the protected best and empty start when needed.The assistant can revisit earlier programs or request a different first operation; textual deduplication blocks identical programs, not geometrically equivalent ones.Source: §3.2 · p. 4

3.3 Assistant Decision Step

  • For each part, the assistant uses a bounded candidate table, reconstruction transcript, and selected renders to choose actions for multiple candidates or finish.The transcript records executed actions, and the renders show the target, best candidate, and recent search state rather than every candidate.Source: §3.3 · p. 5

3.4 Reconstruction Tools

  • Learned calls sample alternative next operations, while algorithmic fitting proposes first or warm operations for missing or extra regions and returns four ranked alternatives per call.Learned sample counts are chosen by the assistant within remaining limits; algorithmic extrusions that add little target overlap are dropped.Source: §3.4 · pp. 5–6
  • Parameter optimization adjusts sketch coordinates, radii, and extrusion depths while fixing planes, axes, and angles, fitting signed-distance mismatch for supported programs.It handles extrusions, holes, and revolves; the resulting child is judged by the reconstruction score, not the distance-fit objective.Source: §3.4 · p. 6

3.5 Execution, Measurements, and Protected Best

  • Candidates must build as watertight positive-volume solids; the score averages available IoU and normalized GMS, or uses GMS alone when IoU is unavailable.Targets and candidates are placed in a common frame, with candidate scale fixed rather than normalized to its own bounding box.Source: §3.5 · pp. 6–7
  • The final output is the protected best valid candidate, retained independently of the branch the assistant continues.A valid candidate takes precedence over an invalid one, and the best-of-N choice uses the reported metrics.Source: §3.5 · p. 7

3.6 Budgets, Recovery, and Termination

  • Per-tool budgets constrain requests, while malformed replies and failed executions trigger recovery actions; the part ends at assistant finish or fixed step and time limits.The assistant may finish as “shape” or “stalled,” and a finish call is ignored until at least one valid candidate exists.Source: §3.6 · p. 7

4 Experiments

CADFather is evaluated across established CAD reconstruction benchmarks, with comparisons, ablations, and cost analyses examining both reconstruction quality and execution validity. It performs strongly across datasets, while results also reveal harder mechanical parts, resource limits, and the value of algorithmic proposals.

4.1 Datasets and Setup

  • Evaluation covers the full DeepCAD, Fusion360, and MCB test sets, plus CADENA-Bench, CADBench, and BenchCAD under specified per-part execution and tool budgets.The three full test sets contain 8,046, 1,725, and 5,000 parts; limits include 24 assistant steps and 1,000 seconds per part.Source: §4.1 · pp. 7–8

4.3 Main Results

  • CADFather achieves the best reported IR, mean IoU, mean GMS, and median CD at both sampling densities across DeepCAD, Fusion360, and MCB.Mean IoU is 0.987, 0.976, and 0.913, respectively, versus 0.966, 0.952, and 0.895 for CADENA-RL sampling; IR is zero on all three.Source: Table 1 · p. 8, §4.3 · p. 9

4.4 Evaluation on CADENA-Bench

  • On CADENA-Bench, CADFather has zero invalid predictions and mean IoU 0.909 and GMS 0.721, versus 0.876 and 0.670 for CADENA-RL greedy.Only 66% of targets have defined IoU; gear-and-bearing and spring-and-fastener families have the lowest IoU and GMS among sufficiently sized families.Source: Table 2 · p. 9, §4.4 · p. 9

4.2 Metrics and Comparison Protocol

  • IoU is computed only for watertight, nonzero-volume targets, whereas GMS and CD cover all targets; reported quality metrics otherwise use valid predictions.The validity rule is stricter than CADENA’s, so its published invalidity ratios are not directly comparable.Source: §4.2 · pp. 8–9

4.5 CADBench and BenchCAD

  • On CADBench, CADFather exceeds published mesh-to-CAD methods on IoU and SIoU and has lower CD, but its valid shape rate is below CADFit’s.It is better than CADFit on these three metrics across all 14 splits; 576 parts exceed the evaluator’s 30-second limit.Source: Table 4 · p. 10, §4.5 · p. 10
  • On BenchCAD, CADFather reaches voxel IoU 0.968 with no invalid predictions, versus 0.910 for CADENA-RL; four assistant steps reach 0.943 at $0.016 per part.The full-budget result costs $0.046 per part; comparisons with frontier agents are contextual because their inputs and evaluated subsets differ.Source: Table 5 · p. 10, §4.5 · p. 11, Figure 4 · p. 11

4.6 Ablations

  • Removing algorithmic proposals lowers mean score more than removing optimization, with the largest effect on MCB, where invalidity rises from zero to 0.12.Algorithmic proposals can supply a valid starting base when sampled first operations produce no closed solid.Source: Table 6 · p. 12, §4.6 · p. 12

4.7 Efficiency and Qualitative Analysis

  • Mechanical parts require more reconstruction effort: MCB averages 12.3 assistant steps versus 6.4 on DeepCAD, and CADENA-Bench averages 13.0 steps and 307 seconds per part.Typical failures include incorrect initial topology, missing fine surface structure despite high IoU, and termination at resource limits.Source: Table 7 · p. 12, §4.7 · pp. 12–14

5 Limitations

CADFather’s reconstruction is constrained by the operations its representation supports and the proposals its tools produce. Its evaluation also leaves important engineering contexts outside scope and incurs computational costs.

5 Limitations

  • CADFather cannot correct an unsuitable operation sequence through parameter optimization, and retaining candidates cannot help if useful alternatives were never generated.Search limits may also end trajectories that could have succeeded with more computation.Source: §5 · p. 14
  • Assistant decisions rely on incomplete visual and numerical evidence and do not guarantee that the best reconstruction will be found.Decisions can depend on the prompt, model behavior, and which candidates are available for inspection.Source: §5 · p. 14
  • The evaluation measures geometry and program validity, not original design history, manufacturing intent, engineering constraints, assemblies, tolerances, or simulation-validated designs.It covers individual parts, and CADFather incurs planning, inspection, and alternative-execution costs.Source: §5 · p. 14
  • Ablations use fixed budgets on 1,000-part subsets, and direct comparisons with adapted CAD-Assistant and IterCAD remain future work.Source: §5 · p. 14

6 Conclusion

CADFather coordinates learned and algorithmic proposals with parameter optimization, using a vision-language assistant and execution feedback to reconstruct executable CAD programs. Across the reported benchmarks, it improves on specified published baselines, while direct comparisons with other CAD agents remain future work.

6 Conclusion

  • A vision-language assistant uses execution feedback to select candidate programs and tools, while learned and algorithmic proposals supply operations and optimization refines existing programs.The system preserves the best valid candidate without additional training of either model.Source: §6 · p. 15
  • Across the full DeepCAD, Fusion360, and MCB test sets, CADFather improves mean IoU and GMS and reduces median Chamfer distance relative to published CADENA-RL sampling results.It produces valid outputs for every part in these tests under its stricter validity rule; it also improves on published CADENA-RL greedy results on CADENA-Bench, and on CADBench mesh-to-CAD baselines in IoU, surface IoU, and Chamfer distance, though its valid shape rate is below CADFit’s.Source: §6 · p. 15
  • On BenchCAD, CADFather achieves higher voxel IoU than the compared frontier models at up to 100× lower estimated inference cost.Source: §6 · p. 15
  • Direct comparisons with other CAD agents and controlled evaluation of the assistant’s tool-selection policy remain future work.Source: §6 · p. 15
Loading 2610.09127v1…