PaperlayerView exact revision on arXiv

Source-linked AI summary

VIGOR: Verifiable Industrial Geometry Optimization and Repair in Scene Programs

Fuqiang Wang, Zheng Guo, Xiaoming Wu, Hanlin Kong, Nayu Liu, Chuanting Zhang, Daojun Liang, Jingxuan Wei

arXiv:2610.13145v1cs.AI

TL;DR

Executable industrial scene programs can still violate coupled engineering constraints in their final states. VIGOR combines supervised generation with execution-grounded verification and staged geometric refinement; it achieves an Overall score of 91.20, 10.42 points above the strongest baseline.

  • Problem

    Successful execution does not ensure that an industrial scene satisfies all task-specified engineering constraints, whose shared device transforms can make corrections interact.

    Source: §1 · pp. 1–2

  • Method

    VIGOR combines supervised scene-program generation with execution-grounded verification, local refinement for independent violations, and joint refinement for residual coupled conflicts.

    Source: §1 · p. 2

  • Results

    91.20 Overall score, 10.42 points above the strongest baseline, DeepSeek-V4.1-Flash.

    Source: §5.2 · p. 7

  • Takeaways & Limitations

    Complete engineering feasibility concerns whether the final executed scene satisfies all task-specified spatial and structural constraints, not merely whether its program executes.

    Source: §1 · p. 3, §6 · p. 9

    The study uses a controlled industrial setting; extending to broader asset libraries, richer programs, dynamic process constraints, and live simulation or editor environments remains future work.

    Source: §6 · p. 10

Abstract

from arXiv · show

Large language models (LLMs) increasingly generate structured layouts and executable scene programs from natural-language instructions. However, executability does not imply engineering feasibility: an industrial scene program may execute successfully while its final state still violates coupled boundary, collision, clearance, aisle, or structural constraints. We identify this mismatch as the executability-engineering-feasibility gap, where locally plausible placements can remain globally infeasible because corrections to one device may introduce violations elsewhere. To make this gap measurable in a controlled setting, we construct an execution-grounded industrial scene-program dataset with 1,749 constructible scenarios, complete executable programs, and 94 final-state geometric checks per scenario. We further propose VIGOR, which combines supervised scene-program generation with execution-grounded verification and staged geometric refinement. Supervised fine-tuning establishes complete program generation, local projection resolves independently correctable violations, and joint conflict-graph repair coordinates residual multi-object conflicts while preserving structural relations. Experiments show that VIGOR achieves 96.86% Complete Feasibility and an Overall score of 91.20, outperforming the strongest evaluated baseline by 10.42 points in Overall. These results show that, under controlled industrial layout conditions, reliable scene-program generation requires not only executable programs, but also explicit verification and coordinated resolution of coupled engineering constraints.

1 INTRODUCTION

The introduction frames industrial scene generation around a gap between successful program execution and complete engineering feasibility, then presents a controlled benchmark and VIGOR’s staged verification and refinement approach.

1 INTRODUCTION

  • Successful execution does not guarantee engineering feasibility, and shared device transforms mean a local correction can create violations elsewhere.The setting evaluates the final scene against coupled boundary, aisle, clearance, attachment, and structural constraints.Source: §1 · pp. 1–2, Figure 1 · p. 2
  • Ablations indicate that supervised fine-tuning establishes structural and geometric priors, local refinement resolves independent violations, and joint refinement addresses residual coupled conflicts.Together, the components yield the strongest complete-feasibility performance.Source: §1 · p. 2
  • The benchmark contains 1,749 constructible scenarios with complete structured programs and final-state checks of structural consistency and geometric feasibility under varied engineering conditions.Each scenario has 11 required instances, a 25-operation program, and 94 geometric checks covering boundaries, collisions, clearances, and the main aisle.Source: §1 · pp. 2–3
  • VIGOR combines supervised scene-program generation with execution-grounded verification and complementary local–joint geometric refinement.Its stages target independently correctable violations and residual coupled conflicts.Source: §1 · pp. 2–3

2 RELATED WORK

Related work spans executable scene generation, constraint-aware layout methods, and execution-guided refinement; this paper focuses on complete feasibility in the final executed scene.

2.1 LANGUAGE-DRIVEN EXECUTABLE SCENE GENERATION

  • Language-driven scene generation has progressed from object layouts and semantic relations toward programmatic, executable, industrial, and simulation-ready scenes.The paper studies complete executable scene generation from natural-language industrial requirements, emphasizing final-state spatial and structural feasibility.Source: §2.1 · p. 3

2.2 SCENE CONSTRAINT SATISFACTION AND COMPLETE FEASIBILITY

  • Constraint-aware scene methods use spatial, physical, and relational conditions, structured representations, optimization, and global validation to reduce layout inconsistencies.This work instead asks whether the final executed scene satisfies all task-specified spatial and structural constraints simultaneously.Source: §2.2 · p. 3

2.3 EXECUTION-GUIDED SCENE GENERATION AND SCENE LEGALIZATION

  • Program synthesis and scene or CAD generation use execution feedback, geometric verification, search, and legalization to refine generated artifacts.This work integrates legalization into language-to-scene generation while preserving program structure, instance identities, hierarchy, and attachment consistency.Source: §2.3 · p. 3

3 METHOD

VIGOR treats final-state engineering feasibility as distinct from successful scene-program execution and addresses it with execution-grounded verification and staged repair. Its pipeline combines complete program generation with local corrections and joint resolution of residual conflicts.

3.1 PROBLEM FORMULATION

  • VIGOR evaluates complete feasibility on the final reloaded scene, because improving one constraint can degrade another through shared device transforms.The constraint set includes workspace, boundary, collision, clearance, aisle, and structural constraints.Source: §3.1 · p. 4

3.2 VIGOR: GENERATION WITH LOCAL–JOINT REFINEMENT

  • For executable initial programs, VIGOR verifies violations, applies local refinement, then coordinates residual multi-object conflicts through joint refinement while preserving structure and attachments.Nonexecutable outputs bypass refinement and remain failed.Source: §3.2 · p. 4, Figure 2 · p. 4

3.3 SUPERVISED SCENE-PROGRAM GENERATION

  • Supervised fine-tuning learns complete programs from reference programs, covering device creation, transform assignment, hierarchy construction, and scene saving.At inference, a single greedy pass produces the initial program, which is executed and verified before residual violations are refined.Source: §3.3 · pp. 4–5

3.4 LOCAL–JOINT GEOMETRIC REFINEMENT

  • Local refinement re-executes boundary, aisle, and clearance corrections, admitting candidates that reduce targeted violations or deficits while preserving execution and structural correctness.Among admissible candidates, it favors fewer modified devices and smaller displacement.Source: §3.4 · p. 5
  • Joint refinement groups residual violations through conflict-graph components and uses beam search over valid successors that preserve satisfied parent links and attachments.Search stops at feasibility or budget exhaustion, retaining a state only if it improves the input objective and preserves attachment consistency.Source: §3.4 · p. 5

4 DATASET

The dataset pipeline standardizes public assets and engineering conditions, screens for constructible layouts, converts reference layouts into executable programs, and validates their executed states. It contains 1,749 constructible scenarios with varied constraint profiles and feasibility margins.

4.1 DATASET CONSTRUCTION

  • The dataset starts with public assets and candidate conditions over aisle width, device clearance, boundary clearance, and aisle orientation, after standardizing geometry and scene interfaces.Assets and environments come from NIST ARIAC, Poly Haven, and AWS RoboMaker Small Warehouse.Source: §4.1 · p. 6
  • Constructibility screening removes conditions without a complete valid layout, and a deterministic search selects a validated reference layout for each retained condition.The selected layout serves as the intermediate representation for scene-program conversion.Source: §4.1 · p. 6
  • Each reference layout becomes a complete scene program.v1 program encoding object identities, transforms, and parent–child relations without changing its spatial configuration.Programs are executed, save/reloaded, and validated for completeness, structural relations, boundaries, collisions, clearances, and the main aisle.Source: §4.1 · p. 6

4.2 DATASET ANALYSIS

  • From 2,754 candidate conditions, the dataset retains 1,749 constructible scenarios: 1,399 training, 175 interpolation-test, and 175 compositional-test samples.Each scenario has 11 instances and a fixed-length 25-operation program; all 1,749 reference programs pass 94 geometric checks.Source: §4.2 · pp. 6–7, Table 1 · p. 7
  • Three similarly sized constraint profiles vary by aisle width, equipment clearance, and boundary clearance, with each spanning Loose, Moderate, and Tight feasibility margins.Constraint composition and feasibility tightness are complementary dimensions of dataset variation.Source: §4.2 · p. 6, Figure 4 · p. 6

5 EXPERIMENTS

Experiments compare VIGOR with generative and code/layout methods, finding stronger overall performance and complete feasibility. Ablations and case analyses indicate that staged refinement addresses violations and coupled conflicts that program generation alone leaves unresolved.

5.1 EXPERIMENTAL SETUP

  • The evaluation uses single-pass greedy decoding and compares open-weight, API, and published code/layout methods against VIGOR.The base model is Qwen3-VL-8B-Instruct, fine-tuned for three epochs on 1,399 samples.Source: §5.1 · p. 7

5.2 MAIN RESULTS

  • 96.86 Complete Feasibility for VIGOR contrasts with 11.71 for GPT-5.5 and 0.00 for Claude Sonnet 5, despite their higher GVB and Bottleneck Geometry scores.The results expose a gap between geometric quality and complete satisfaction of coupled engineering constraints.Source: §5.2 · p. 7
  • OpenCodeReasoning-Nemotron-7B and CodeRL+ score 0.25 and 1.46 Overall, while ReSpace scores 14.26, indicating that code or layout capability alone is insufficient.VIGOR’s Overall score is 91.20.Source: §5.2 · pp. 7–8
  • 91.20 Overall, 10.42 points (12.9%) above strongest baseline DeepSeek-V4.1-Flash; Table 2 also reports VIGOR’s best Bottleneck Geometry, Functional Layout, and Tight-FQ scores.VIGOR scores 72.39 in Functional Layout versus 42.24 for DeepSeek-V4.1-Flash, despite comparable GVB and Bottleneck Geometry scores.Source: §5.2 · p. 7, Table 2 · p. 8

5.3 ABLATION STUDY

  • VIGOR-SFT reaches 14.57 Feasibility, local refinement raises it to 56.86, joint refinement to 81.43, and full VIGOR to 96.86.The ablation reports complementary roles: SFT supplies structured initialization, while local and joint refinement resolve remaining violations and coupled conflicts.Source: Table 3 · p. 8, §5.3 · p. 8

5.4 CASE STUDY AND ERROR ANALYSIS

  • In the representative task, VIGOR preserves feasibility, while GPT-5.5 violates aisle and clearance requirements and Claude Sonnet 5 produces overlapping equipment.The case illustrates the challenge of coordinating coupled multi-object constraints.Source: §5.4 · pp. 8–9, Figure 5 · p. 9

5.5 CONSISTENCY WITH HUMAN JUDGMENTS

  • LLMJudge and human ratings align closely across evaluated methods, with near-perfect correlation (r = 0.9773).VIGOR occupies the top-right region under both evaluation protocols.Source: §5.5 · p. 9

6 CONCLUSION

The paper studies complete engineering feasibility in executable industrial scene programs and introduces an execution-grounded benchmark and VIGOR’s deterministic local–joint refinement. Its scope is controlled layouts with fixed assets and a structured program schema.

6 CONCLUSION

  • Executable programs are insufficient unless the executed scene satisfies spatial, structural, and functional constraints coupled through shared device transforms.The benchmark contains 1,749 scenarios and verifies executed, saved, and reloaded programs against final scene states.Source: §6 · p. 9
  • The study is limited to a controlled industrial layout setting with a fixed asset inventory, structured program schema, and deterministic state-based verification.Broader asset libraries, richer programs, dynamic process constraints, and live simulation or editor environments remain future directions.Source: §6 · pp. 9–10

AI USE STATEMENT

  • Generative AI tools assisted only with generating and organizing structured task instances; authors designed the benchmark and independently verified generated data and reported results..Source: AI USE STATEMENT · p. 10

REPRODUCIBILITY STATEMENT

  • The paper reports benchmark construction, program specification, execution protocol, evaluation criteria, training, inference, and baseline settings to support reproduction..Source: REPRODUCIBILITY STATEMENT · p. 10

A PERFORMANCE RESULTS WITH RADAR CHART ANALYSIS

The appendix compares process and executed-scene capability profiles across model families. Existing methods can score highly on program construction while lagging in final-scene quality, whereas VIGOR is comparatively balanced.

A PERFORMANCE RESULTS WITH RADAR CHART ANALYSIS

  • Program Process Score measures structured program construction, Final Scene Score evaluates executed-scene quality, and Unified Capability Score summarizes both.These three dimensions form the process–result capability profile.Source: A PERFORMANCE RESULTS WITH RADAR CHART… · p. 13
  • CodeRL+ scores 100.0 in Program Process but 2.0 in Final Scene quality, illustrating that program-level correctness alone is insufficient under coupled geometric constraints.Several general models also approach saturated Program Process scores while showing weaker Final Scene performance.Source: A PERFORMANCE RESULTS WITH RADAR CHART… · p. 13
  • DeepSeek-V4.1-Flash reaches 79.7 in Final Scene score, above GPT-5.5 at 69.5 and Claude Sonnet 5 at 70.1, but their profiles remain imbalanced..Source: A PERFORMANCE RESULTS WITH RADAR CHART… · p. 13
  • VIGOR’s profile is more balanced, with 99.7 Program Process, 89.3 Final Scene, and 94.5 Unified Capability.The paper frames successful program construction as a prerequisite, with execution-grounded verification and coordinated conflict resolution also needed for reliable scene generation.Source: A PERFORMANCE RESULTS WITH RADAR CHART… · p. 13

B FINE-GRAINED TASK CAPABILITY ANALYSIS

The task-level analysis evaluates capability across generalization, constraint density, aisle orientation, and constraint stress. VIGOR’s scores remain comparatively high and stable across the reported categories.

B FINE-GRAINED TASK CAPABILITY ANALYSIS

  • VIGOR scores 90.9 on interpolation and 90.3 on the held-out compositional split, versus DeepSeek-V4.1-Flash’s 80.3 and 80.6.The small split difference indicates performance on held-out combinations of engineering conditions.Source: B FINE-GRAINED TASK CAPABILITY ANALYSIS · p. 13
  • Across loose-, medium-, and tight-space planning, VIGOR scores 91.3, 89.6, and 88.8, showing limited degradation as geometric margins narrow.The passage notes that tightly coupled boundary, collision, clearance, and aisle constraints remain difficult for general-purpose models.Source: B FINE-GRAINED TASK CAPABILITY ANALYSIS · p. 13
  • VIGOR scores 90.3 on X-aisle planning and 90.7 on Z-aisle planning, with little sensitivity to orientation changes.Several weaker models show greater orientation-dependent variation.Source: B FINE-GRAINED TASK CAPABILITY ANALYSIS · p. 13
  • Under wide-aisle, high-clearance, and large-boundary-margin stress, VIGOR scores 91.2, 90.2, and 89.6, while DeepSeek-V4.1-Flash remains around 80.The categories test constrained placement space where satisfying one requirement can reduce neighboring devices’ feasible regions.Source: B FINE-GRAINED TASK CAPABILITY ANALYSIS · p. 14
  • Across all ten capability categories, VIGOR ranges from 88.8 to 91.3, contrasting with lower and more variable profiles from existing methods.The analysis attributes reliable scene generation to final-state verification and coordinated resolution of coupled constraints as well as program construction.Source: B FINE-GRAINED TASK CAPABILITY ANALYSIS · p. 14

C TASK CONSTRUCTION AND CONSTRUCTIBILITY SCREENING

The benchmark fixes the industrial scene inventory and varies explicit spatial conditions, screening candidate conditions for constructible layouts before partitioning the retained scenarios.

C TASK CONSTRUCTION AND CONSTRUCTIBILITY SCREENING

  • Each task uses a fixed 7.5 × 7.0 m workspace and eleven required instances, with condition variation controlled by aisle width, equipment clearance, boundary clearance, and aisle orientation.The parameter tuple is (w, c, b, o), with o ∈ {X, Z}.Source: C TASK CONSTRUCTION AND… · p. 14
  • The parameter grid contains 2,754 candidate conditions, varying distances in 0.05 m increments while keeping inventory, asset geometry, workspace, and attachment relations fixed.The grid is summarized in Table 4.Source: C TASK CONSTRUCTION AND… · p. 15, Table 4 · p. 15
  • Deterministic layout generation and validation retain 1,749 conditions and record 1,005 unsuccessful constructions, which do not prove mathematical infeasibility.Each condition is tried with two predefined construction variants in a fixed order.Source: C TASK CONSTRUCTION AND… · p. 15
  • For each retained condition, the reference pipeline selects the highest-ranked legal candidate and converts it into a complete program without changing transforms or attachment relations.Reference programs, layouts, execution traces, and verification results are kept separate from model-visible inputs.Source: C TASK CONSTRUCTION AND… · p. 15
  • The 1,749 conditions are partitioned into 1,399 training, 175 interpolation-test, and 175 compositional-test samples, with the latter holding out joint width–clearance combinations.Interpolation samples are selected by stratified systematic sampling, and each has a same-orientation training neighbor one grid step away.Source: C TASK CONSTRUCTION AND… · pp. 15–16

D SCENE-PROGRAM SPECIFICATION AND GENERATION INPUT

The model receives an answer-free, parameterized task description and must produce a complete scene program; dataset quality assurance then checks execution, structure, geometry, and saved-state consistency.

D SCENE-PROGRAM SPECIFICATION AND GENERATION INPUT

  • The model input supplies public task constraints, asset geometry, and the program interface without reference layouts, traces, or evaluation scores.Condition-dependent fields specify clearance thresholds, workspace bounds, and the reserved aisle.Source: D SCENE-PROGRAM SPECIFICATION AND… · p. 16
  • The required output is a complete ordered scene_program.v1 JSON program with explicit transforms, required parent links, and one final save operation.The reference program contains 25 operations for eleven instances and two attachments.Source: D SCENE-PROGRAM SPECIFICATION AND… · pp. 16–18

E EXECUTION-BASED DATASET QUALITY ASSURANCE

  • Programs are schema-checked and executed from an empty state; successful execution must reach the final save operation and produce a new state file.Unknown operations or assets, duplicate identifiers, invalid numbers, non-positive scales, cyclic parent relations, and operation-order violations are rejected.Source: E EXECUTION-BASED DATASET QUALITY… · p. 18, D SCENE-PROGRAM SPECIFICATION AND… · p. 18
  • Post-execution checks verify all eleven identities, asset assignments, transforms, parent links, and placement anchors, including the two required attachments.Reference information is used for verification, not to complete or repair the submitted program.Source: E EXECUTION-BASED DATASET QUALITY… · p. 18
  • Independent geometry verification applies 94 checks per scene for boundary clearance, aisle preservation, collisions, and equipment clearance, alongside structural checks.World collision AABBs are recomputed by transforming the eight corners of each public local AABB.Source: E EXECUTION-BASED DATASET QUALITY… · p. 18
  • All 1,749 retained conditions pass program conversion, execution, save–reload, identity and attachment, geometry, and hard-constraint checks.This yields 164,406 passing geometric checks under the public asset geometry and Python execution semantics, not a claim of Unity execution for every scene.Source: E EXECUTION-BASED DATASET QUALITY… · p. 19

F MACHINE EVALUATION PROMPTS

The appendix specifies machine-judge prompts for two separate visual dimensions and describes how methods are evaluated across rule-based, visual, and aggregate measures.

F MACHINE EVALUATION PROMPTS

  • The machine judge scores Operational Usability and Spatial Organization in separate calls, returning an integer from 0 to 5 with an evidence-based reason.The prompts use separate cards and judgments for each dimension.Source: F MACHINE EVALUATION PROMPTS · pp. 19–20
  • The machine judge evaluates each valid anonymous scene card in separate OU and SO calls, producing a 0–5 score and concise evidence-based reason for one dimension.The two visual dimensions are scored independently.Source: F MACHINE EVALUATION PROMPTS · pp. 19–20
  • Operational Usability assesses operation and maintenance convenience, material handling and workflow, device coordination, and visible obstacles to real use.It is scored on a 0–5 integer scale.Source: F MACHINE EVALUATION PROMPTS · pp. 19–20
  • Visual judges must not re-derive exact geometry, compare layouts with a reference arrangement, or guess which method produced a scene.Exact geometric checks belong to a separate automatic scorer.Source: F MACHINE EVALUATION PROMPTS · pp. 19–21
  • Spatial Organization judges functional-zone clarity, device grouping and orientation, compactness without excessive crowding, and overall orderliness.It is scored on a 0–5 integer scale.Source: F MACHINE EVALUATION PROMPTS · pp. 20–21

G EVALUATION METRICS

  • Evaluation combines rule-based engineering metrics from executed states, visual scene-quality judgments, and an aggregate engineering score.Invalid or unscorable outputs remain in the denominator and receive the prescribed zero score.Source: G EVALUATION METRICS · p. 21

G.1 STRUCTURAL AND GEOMETRIC METRICS

The metrics separately quantify structural correctness, geometric violations and bottlenecks, functional quality, an aggregate engineering score, and visual judgments.

G.1 STRUCTURAL AND GEOMETRIC METRICS

  • Structural correctness requires both specified parent and attachment relations to be fully satisfied; Complete Feasibility requires valid execution and passing final-state hard constraints.A task without a valid scene state receives zero for the stated ratio metric.Source: §G.1 · pp. 21–22
  • Geometric Violation Burden counts unsatisfied atoms across 94 checks, while Bottleneck Geometry reports the weakest satisfaction rate among boundary, collision, clearance, and aisle families.Bottleneck Geometry takes the per-task minimum before averaging and is not a linear conversion of mean GVB.Source: §G.1 · pp. 21–22

G.2 FUNCTIONAL QUALITY AND TEST-SUBSET METRICS

  • Functional Layout averages six equally weighted evaluator scores covering process-chain, tool/kitting, rack, container, desk, and hazard-sign placement.Additional summaries report lower-tail, interpolation, compositional, and tight-condition performance.Source: §G.2 · p. 22

G.3 OVERALL SCORE

  • Overall averages structural correctness, geometric quality, functional quality, and complete feasibility after scaling the four components to 0–100.Visual judgments are reported separately and excluded from Overall.Source: §G.3 · pp. 22–23

G.4 VISUAL JUDGMENT SCORES

  • Visual evaluation uses 20 fixed dual-view cards per method; LLMJudge and Human Rating assess Operational Usability (OU) and Spatial Organization (SO) separately on 0–5 scales, with human ratings aggregated across three independent raters.LLMJudge and Human Rating are presented separately and are not combined with one another or with Overall.Source: §G.4 · p. 23
Loading 2610.13145v1…