Large language models (LLMs) can generate structured layouts and executable scene programs from natural‑language instructions, yet executability does not guarantee engineering feasibility. An industrial scene program may run successfully while its final state still violates boundary, collision, clearance, aisle, or structural constraints. We call this the executability‑engineering‑feasibility gap, where locally plausible placements remain globally infeasible because correcting one device can introduce violations elsewhere.
To measure this gap, we built an execution‑grounded industrial scene‑program dataset containing 1,749 constructible scenarios, complete executable programs, and 94 geometric checks per scenario.
Based on the dataset we propose VIGOR, which combines supervised scene‑program generation, execution‑grounded verification, and staged geometric refinement. The pipeline consists of:
- Supervised fine‑tuning for full program generation;
- Local projection to resolve independently correctable violations;
- Joint conflict‑graph repair to coordinate remaining multi‑object conflicts while preserving structural relations.
Experiments show VIGOR achieves 96.86% complete feasibility and an overall score of 91.20, outperforming the strongest baseline by 10.42 points, demonstrating that reliable scene‑program generation under controlled industrial layout conditions requires not only executable code but also explicit verification and coordinated resolution of coupled engineering constraints.
Review