acceptodds
Under review as a conference paper at ICLR 2027

From Requirements and Drawings to Optimized Warehouse Floorplans: LLM-Driven Formulation and Solver-Based Design

Abstract

Warehouse layout design must translate natural-language requirements into feasible spatial decisions while accounting for operational performance. We present FLOWS (Formulation-Led Optimization of Warehouses with Solvers), a neuro-symbolic framework that connects language-based modeling with symbolic optimization through an explicit integer-programming formulation. A formulation agent grounds decision variables, objectives, and constraints in a graph representation of warehouse geometry and a finite rack catalog. A separate strategy agent selects admissible solution procedures without changing the underlying problem. Candidate layouts undergo static validation and simulation-based acceptance; diagnostic feedback supports formulation revision while preserving the original requirements. This separation makes the model, solution procedure, and acceptance decision independently inspectable. With replayed formulations and a fixed constructive portfolio, FLOWS produces verified layouts on all 42 warehouse plans and improves mean operating performance over two adapted planners. Small component ablations identify declaration-repair gains but do not establish an advantage for adaptive strategy selection or operational feedback.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Warehouse, Symbolic

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_ZrejOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents FLOWS, a framework that turns natural-language warehouse requirements and drawings into rack layouts. An LLM formulation agent writes an explicit integer-programming formulation over a graph of the warehouse geometry and a finite rack catalog (e.g., "at least 16 slots; every rack accessible"); a separate strategy agent selects a solution procedure; a symbolic IP solver places the racks; and candidate layouts are checked by static validation (geometry, clearance, access) and an operational simulation (throughput, cycle time, travel), with diagnostic feedback for revising the formulation. The paper reports verified layouts on all 42 warehouse plans and better mean operating performance than two adapted planners.

Critical strengths
  1. The problem framing is interesting: separating formulation, solution procedure and acceptance makes each stage inspectable, and the idea of an LLM producing an explicit, checkable optimization formulation is novel for this application.
  2. The paper is clear and easy to follow; the problem (capacity vs. accessibility, physical feasibility vs. operational performance) is well motivated.
Critical weaknesses
  1. The experiments do not match the claimed role of the LLM. The LLM is said to do formulation, strategy selection and feedback-driven revision, but the experiments only cover formulation; there are no experiments on strategy selection or on revision from feedback. In the ablations the LLM components contribute little and the gains are small, so the evidence is loose relative to the claims.
  2. No clear evidence that an LLM is needed, and limited evaluation. The paper does not justify why this problem needs to be solved with a large language model rather than a conventional pipeline. The experimental cases are fixed; it is unclear how the system performs on new floor plans it has not seen, and there is no real-world deployment, only simulation and physical checks.
Questions for authors
  1. Can you provide experiments that isolate the contribution of LLM-based strategy selection and feedback-driven revision?
  2. What does the LLM provide that a non-LLM formulation pipeline would not?
  3. How does FLOWS perform on new, expert-validated floor plans not used during development?
  4. Is there any evaluation beyond simulation, e.g., in a real warehouse deployment?
4 comments

Response to Reviewer Zrej

We submitted a revised manuscript with changes marked in blue.

We thank the reviewer for recognizing the inspectable architecture. The revision adds component studies and tests on additional drawings.

Q1: Can strategy selection and feedback revision be evaluated separately? [Original: Weakness 1; Question 1]

A1: Yes. Section 4.4 and Appendix E fix reference formulations to isolate Stages II and III from Stage I.

Condensed from manuscript Table 4:

Stage II metricFixed exactLLM-selected
Verified layouts33/4233/42
Mean runtime, all cases (s)331.25164.68
Mean runtime, common successes (s)313.27105.34
Stage III procedureKPI improvements
No feedback0/22
Equal-budget continuation1/22
Revised LLM-feedback component19/22

Stage II uses identical formulations, placements, hints, and static checks under a 600-second total budget, including preparation and model calls. Both policies verify the same 33 cases; mean runtime falls by 50.3% overall / 66.4% on common successes. This compares complete policies, including stopping rules, rather than proving superiority over every deterministic alternative.

Stage III increases throughput floors by 10% and decreases cycle-time ceilings by 10%, retaining the 33% footprint floor. The arms share 22 initial KPI-failing layouts and a 1,800-second total budget. Improvement means nondecreasing throughput and nonincreasing cycle time, with one strict improvement; it does not necessarily meet every tightened target. All 19 locked improvements persist on 20 new paired streams, averaging +9.85% throughput / -9.15% cycle time. The revised controller was developed after earlier failures and combines LLM-authored restrictions with deterministic geometric search; we attribute the result to the complete feedback component. These tests support faster verified solving and operational improvement over the tested fixed procedures.

Q2: What does the LLM provide beyond a conventional formulation pipeline? [Original: Weakness 2; Question 2]

A2: The LLM composes supported objectives and constraints from facility-specific requirements; deterministic components bind geometry, compile, solve, and verify. Section 3.2 distinguishes capacity objectives from mandatory regional capacity floors; Appendix A.2 specifies spatial filters, logical combinations, access predicates, and provenance. The agent cannot replace catalog dimensions, waive checks, or introduce unsupported expressions. A fixed template remains suitable for fixed requirements; FLOWS offers flexibility within its supported modeling language.

The C04 case study illustrates adaptability to an irregular boundary and internal exclusions: 33.00% footprint utilization versus 28.85% for CBS and 12.08% for CPU, with throughput gains of 17.62% / 9.31%, respectively (Section 4.1). These outcomes support the framework's ability to accommodate complex geometry while improving storage and operations; they do not isolate LLM superiority over a matched deterministic parser.

Q3: How does FLOWS perform on new, expert-validated floorplans? [Original: Weakness 2; Question 3]

A3: Appendix B now clarifies that only C01 was used for development testing. Prompts, formulation-generation rules, and planning configuration were frozen before the other 41 cases were evaluated; their outcomes were not used for retuning. The 42-case aggregate includes C01.

Appendix H evaluates DXF drawings supplied after framework development. On the 12-drawing / 26-region common completed subset, FLOWS satisfies each region's CBS footprint floor and declared operational requirements. Against CBS, mean footprint utilization is 29.26% versus 29.20%, throughput is 1.96% higher, cycle time 4.27% shorter, and travel 8.70% shorter. This supports applicability to additional geometries, conditional on completed tasks and including corrections/recovery; it is not a frozen single-pass benchmark or overall success-rate estimate. The quantitative dataset is separate from ongoing industrial expert review; we do not claim completed expert validation of this benchmark.

Q4: Is there evidence beyond simulation or a real warehouse deployment? [Original: Weakness 2; Question 4]

A4: FLOWS is developed with a commercial company, and some generated proposals have entered expert review in the partner's engineering workflow. It improves planning efficiency by automating candidate generation and preliminary feasibility checks before expert assessment. This is practical engineering use, rather than a completed warehouse deployment or measured post-deployment operational gains. Because the ongoing review is sensitive, partner and project details cannot be disclosed at this stage. The manuscript's quantitative claims remain based on computational checks and simulation.

Thanks for the rebuttal. It addresses some of my questions. I will keep my rate.

Dear reviewer, We hope we can address all your concerns. Please let us know if any issues remain unresolved. We welcome any further feedback and would be grateful if you would consider raising your score in light of our clarifications. Best regards, all authors

Dear Reviewer,

Thanks again for your careful review and insightful question! I am writing to make sure you have seen our rebuttal. And if our rebuttal addressed your question, we sincerely hope you can consider improving your recommendation score, which will help us a lot.

Best regards,

All authors

Reviewer_PNgiOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

To address the complexity of traditional warehouse layout design and the limitations of directly using LLMs to generate layouts, this paper proposes FLOWS, a three-stage framework consisting of LLM-based integer programming formulation, adaptive solver strategy selection, and validation and feedback. The framework aims to enable AI systems to automatically formulate optimization problems, optimize warehouse layouts, and reliably validate the results based on natural language requirements and warehouse information. Experiments on 42 warehouse cases show that FLOWS achieves a 100% layout availability rate and an average storage footprint utilization of 32.28%, while improving throughput compared with two traditional baselines. In addition, the paper conducts component ablation studies to further analyze the roles of LLM-based solver strategy selection and feedback-driven revision.

Critical strengths
  1. The paper integrates LLM-based formulation, integer programming solvers, and simulation-based validation into a unified framework. Through structured constraint representations and independent verification mechanisms, it improves the reliability of warehouse layout generation and demonstrates strong practical applicability.
  2. The paper achieves a 100% physically valid layout generation rate across 42 warehouse cases, with improvements in storage utilization and operational efficiency. It also provides baseline comparisons, component ablation studies, and case-level analyses, strengthening the credibility of the experimental results.
Critical weaknesses
  1. The overall effectiveness of the framework is reasonably well demonstrated, but the independent contributions of its key components remain unclear. By integrating LLM-based formulation, symbolic optimization, and simulation-based validation, the paper achieves promising overall performance on warehouse layout design tasks. However, the existing experiments do not sufficiently disentangle the contributions of individual components to these improvements. In particular, adaptive solver strategy selection and feedback-driven revision do not demonstrate clear advantages over simpler deterministic approaches. Consequently, the necessity and independent technical value of these key components within the complete framework remain insufficiently supported by empirical evidence.
Questions for authors
  1. The current results do not clearly demonstrate the independent advantages of these components over deterministic approaches. Under what types of warehouse layout design tasks are these adaptive mechanisms expected to provide practical benefits? Do the existing experiments offer concrete evidence to support this expectation?
  2. For the types of requirements covered in the current experiments, what specific advantages does LLM-based formulation offer over these simpler approaches? Are there representative cases demonstrating its ability to handle changing requirements or complex constraints?
3 comments

Response to Reviewer PNgi

We submitted a revised manuscript with changes marked in blue.

We thank the reviewer for recognizing FLOWS's reliability and practical relevance. We provide component evidence and clarify its advantages over fixed procedures.

Q1: When do adaptive selection and feedback help, and what evidence supports them? [Original: Weakness 1; Question 1]

A1: Stage II targets solving cost; Stage III targets operationally inadequate but physically feasible layouts. In Section 3.3, the strategy agent selects supported encodings and solver settings according to model structure and budget, without changing the objective or constraints. For example, complete connectivity encoding trades upfront cost against repeated solves under incremental enforcement. Section 3.4 uses simulated KPI failures to revise reversible search restrictions, preserving the original acceptance conditions.

The new experiments fix reference formulations to separate these effects from Stage I (Section 4.4, Appendix E):

Condensed from manuscript Table 4:

Stage II metricFixed exactLLM-selected
Verified layouts33/4233/42
Mean runtime, all cases (s)331.25164.68
Mean runtime, common successes (s)313.27105.34
Stage III procedureKPI improvements
No feedback0/22
Equal-budget continuation1/22
Revised LLM-feedback component19/22

Stage II gives both policies identical formulations, placements, hints, and static checks under a 600-second total budget, including preparation and model calls. The same 33 cases succeed, with 50.3% lower mean runtime overall and 66.4% lower runtime on common successes. This is a complete-policy comparison, including stopping rules, not proof of superiority over every deterministic policy.

Stage III tightens throughput floors by 10% and cycle-time ceilings by 10%, retaining the 33% footprint floor. All arms share the 22 initial KPI failures and a 1,800-second total budget. Improvement requires nondecreasing throughput and nonincreasing cycle time, with one strict improvement; it does not imply meeting every tightened target. All 19 locked improvements persist on 20 new paired streams, averaging +9.85% throughput / -9.15% cycle time. Feedback combines LLM-authored restrictions with deterministic geometric search; its controller was revised after earlier failures, so the result concerns the complete component.

Compared with the tested fixed solving and continuation policies, FLOWS obtains verified layouts sooner and improves poor operational outcomes without relaxing the original requirements. These are the practical benefits supported by the component tests.

Q2: What does LLM formulation offer for changing requirements or complex constraints? [Original: Question 2]

A2: Its advantage is a flexible interface from facility-specific requirements to composable, geometry-grounded constraints. Section 3.2 distinguishes maximizing capacity from enforcing a mandatory regional capacity floor: the former is an objective, while the latter must remain a constraint. The supported language also composes spatial filters, logical conditions, and access predicates (Appendix A.2). Deterministic compilation binds declarations to fixed geometry and catalog facts; independent checks enforce validity. This avoids restricting every request to a prescribed layout family, although a fixed template can suffice for a fixed requirement structure. New constraint types outside the language still require compiler and validator support.

C04 illustrates adaptability to complex geometry, with an irregular boundary and internal exclusions. FLOWS achieves 33.00% footprint utilization versus 28.85% for CBS and 12.08% for CPU. Against these baselines, throughput increases by 17.62% / 9.31%, and cycle time decreases by 16.82% / 10.24% (Section 4.1). Thus, the framework accommodates geometric restrictions while improving both storage and operations. C04 is an illustrative case, not a complexity-stratified benchmark; capacities and destination distributions differ.

Appendix H further combines regional CBS footprint floors with throughput of at least 40 orders/h and cycle time of at most 600 minutes on additional drawings. Its 12-drawing / 26-region common completed subset meets these requirements, allowing corrections and recovery. This supports practical adaptability, not an overall success-rate estimate.

Together, these examples show the framework's flexibility across geometry and requirement configurations. They do not isolate the formulation LLM's contribution from optimization and validation or establish superiority over a matched deterministic parser. We will state that distinction explicitly.

Thank authors for their detailed response and detailed information from the rebuttal, it addressed many of my concerns, I keep my origin score, thanks.

Dear Reviewer,

Thanks again for your careful review and insightful question! I am writing to make sure you have seen our rebuttal. And if our rebuttal addressed your question, we sincerely hope you can consider improving your recommendation score, which will help us a lot.

Best regards,

All authors

Reviewer_hTfaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces FLOWS, a system for warehouse layout planning that considers storage capacity, accessibility, and efficient workflows. A language model-based formulation agent translates requirements into an integer optimization model. Potential rack placements are constrained by the building geometry and a fixed catalog of rack types. A strategy agent selects a feasible solution method, and a symbolic solver determines the rack arrangement based on the formulated objectives and constraints. The resulting layouts are checked for spatial feasibility and evaluated in a simulation with respect to operational performance. Feedback is used to revise the formulations. The main contributions are an executable, geometry-constrained formulation interface; a clear separation between problem formulation, solution methods, and result validation; and an evaluation that relates spatial and operational layout quality to computational effort. The main evaluation covers 42 warehouse floor plans and uses previously generated large language model formulations and a fixed selection of constructive methods. In this configuration, FLOWS generates physically validated layouts for all floor plans and achieves, on average, higher throughput and shorter processing times than two adapted baseline methods. Additional experiments involving 107 comparison tasks show positive average differences in throughput. Component-level tests, however, indicate that LLMs provide no advantage in error correction compared to fixed repair rules.

Critical strengths
  1. FLOWS does not treat the LLM output as a finalized warehouse layout. Instead, a formulation agent translates the requirements into a mathematical model, which is then used to compute the rack arrangement and validate it separately. Fixed geometric data and a predefined rack catalog constrain the search space and, consequently, the placement decisions.

  2. The paper examines not only the utilization of rack space but also the impact of the layout on simulated warehouse operations, including throughput and transport distances. An additional comparison requires FLOWS to achieve at least the same rack footprint as the respective baseline method. This investigates whether improvements in operational performance persist even without reducing the rack footprint.

Critical weaknesses
  1. The main comparison uses previously generated LLM formulations and a fixed selection of planning methods. The complete adaptive architecture is not evaluated. In the supplementary component-level tests, the fixed repair rule achieves the same results as the language model when correcting known formulation errors. The benefit of feedback from operational performance also remains unproven. The results demonstrate that the tested overall configuration works, but there is no concrete evidence that LLM-driven decisions provide added value.

  2. All warehouse floor plans were known during development, and no held-out data were used. Although separate simulation streams were employed for final evaluation, generalization to previously unseen warehouse layouts remains untested.

Questions for authors

Appendix B states that all 42 warehouse floor plans were known during development. Were the prompts, formulations or planning methods adjusted based on the results for these floor plans? If so were these changes mainly made to fix feasibility issues or to improve operational performance? It is needed to clarify this in the paper, as it would make the evaluation results easier to interpret.

3 comments

Response to Reviewer hTfa

We submitted a revised manuscript with changes marked in blue.

We thank the reviewer for recognizing the geometry-grounded formulation and the footprint-matched evaluation. We clarify the development protocol and provide new component evidence.

Q1: Were prompts, formulations, or methods adjusted using the 42 floorplans, and why? [Original: Weakness 2; sole Question]

A1: The earlier phrase "all cases were development-exposed" was imprecise. Only C01 was used for development testing. Prompts, formulation rules, and planning configuration were frozen before the other 41 cases were first evaluated; their outcomes were not used for retuning. The documented construction-target adjustment addressed storage-density shortfalls, not confirmation throughput or cycle time. Within-instance candidate selection uses development simulation, followed by 20 separate confirmation streams; this differs from cross-instance configuration tuning. Appendix B now clarifies this protocol. The 42-case aggregate includes C01.

Beyond this cohort, Appendix H reports 12 additional drawings / 26 regional tasks supplied after framework development. At each region's CBS footprint floor, FLOWS improves mean throughput by 1.96% and cycle time by 4.27% over CBS. This supports transfer to additional geometries, conditional on the common completed subset and allowing corrections/recovery, rather than establishing frozen single-pass generalization.

Q2: What supports adaptive control beyond fixed planning and deterministic repair? [Original: Weakness 1]

A2: The mechanisms address different limitations. In Section 3.3, the strategy agent conditions on model structure and budget while preserving the formulation. For connectivity, complete encoding incurs upfront cost; incremental enforcement starts smaller but may require repeated solves. Choosing between them can adapt computational effort to the instance; this is a rationale, not a theoretical speedup guarantee. In Section 3.4, static validity alone cannot establish operational adequacy. Feedback uses KPI failures to revise reversible search restrictions, while the original requirements, workload, and independent acceptance checks remain fixed. Repairing a known declaration error therefore tests a different capability from improving a valid but operationally inadequate layout.

Section 4.4 and Appendix E isolate Stages II and III with fixed reference formulations:

Condensed from manuscript Table 4:

Stage II metricFixed exactLLM-selected
Verified layouts33/4233/42
Mean runtime, all cases (s)331.25164.68
Mean runtime, common successes (s)313.27105.34
Stage III procedureKPI improvements
No feedback0/22
Equal-budget continuation1/22
Revised LLM-feedback component19/22

Stage II uses identical formulations, placements, hints, and static checks under a 600-second total budget, including preparation and model calls. Runtime falls by 50.3% overall / 66.4% on common successes, with the same 33 verified cases. This compares complete policies, including stopping rules, and does not isolate the LLM from every deterministic alternative.

Stage III raises throughput floors by 10% and lowers cycle-time ceilings by 10%, retaining the 33% footprint floor. The three arms share 22 initial KPI-failing layouts and a 1,800-second total budget. Improvement means throughput does not decrease and cycle time does not increase, with one strict improvement; it need not satisfy every tightened target. All 19 locked improvements persist on 20 new paired streams, averaging +9.85% throughput / -9.15% cycle time. The controller combines LLM-authored restrictions with deterministic geometric search and was revised after earlier failures; the measured benefit belongs to the complete feedback component.

Thus, relative to the tested fixed policies, adaptive control delivers checked layouts faster and feedback improves operationally inadequate layouts without weakening requirements. Matching deterministic repair on known errors does not negate these distinct benefits.

Q3: What are the limits of the additional-drawing evidence? [Original: Weakness 2]

A3: Appendix H's common completed subset attains 29.26% footprint utilization versus CBS's 29.20%, while reducing travel by 8.70% alongside the gains in A1. Formulation, strategy selection, symbolic solving, and acceptance checking are exercised on these additional geometries. Separate simulation streams test stochastic robustness; the additional drawings test applicability beyond the original geometry cohort. Because the comparison includes corrections and recovery and conditions on completed tasks, we do not interpret it as an overall success-rate estimate. We will keep this scope explicit in the final manuscript.

Thank you for the clarification and the additional experiments. These address my concerns about the floor plans in the main evaluation. The runtime gains and improvements further strengthen the paper.

However, my main concern remains the specific contribution of the LLM. While the new experiments compare complete strategies and feedback procedures, it remains unclear how reliably the LLM translates changing requirements into constraints. I therefore maintain my initial rating.

Thank you for clarifying this concern. Strategy and feedback comparisons alone do not establish faithful requirement translation. We therefore tested two scenarios on C04, S0 and S1, each in three English styles. Geometry, dock, catalogue, candidate domain and workload were fixed. Each input received one formulation call; expected constraints were specified separately and withheld from the model.

For S0, the following actual excerpts express equivalent requirements:

StyleInput excerptsGenerated constraints
Everyday“Racks must cover at least 30% of the usable floor area”; “at least 80 cm of clear width”; “at least 60 orders an hour”; “within 6 hours”Utilization >=30%; clearance >=0.8 m; throughput >=60 orders/h; mean cycle <=21,600 s
Practitioner“Minimum rack-footprint utilization is 30%”; “clear width >= 0.8 m”; “throughput must be >= 60 orders/h and mean cycle time <= 360 min”Same constraints
Mathematical“U >= 0.30”; “0.8 m minimum clearance”; “T >= 60 orders/h”; “C <= 21600 s”Same constraints

All three preserve capacity maximization and require every selected rack to have a picking-access node reachable from the dock through the 0.8 m clearance graph. Containment, non-overlap and terminal preservation are enforced by the geometry kernel. The 30% footprint bound translates to at least 10,500 slots (350 racks) in the fixed candidate catalogue.

S1 uses a 25% footprint minimum and stricter operational requirements. Its everyday input says “at least 70 orders an hour” and “within 5 hours”; the practitioner input uses “>= 70 orders/h” and “<= 300 min”; the mathematical input uses “T >= 70 orders/h” and “C <= 18000 s”. All three yield throughput >=70 orders/h, mean cycle <=18,000 s and a footprint bound equivalent to 8,760 slots (292 racks).

All six formulations passed semantic checks of capacity bounds, units, inequality direction, universal access, objective direction and KPI thresholds. We check mathematical equivalence rather than identical variable names or JSON.

We then planned under each formulation and independently verified physical feasibility and simulated performance. All three styles within each scenario selected the same layout and passed confirmation:

ScenarioRequired utilizationRequired throughput (/h)Required cycle (min)Achieved utilizationAchieved throughput (/h)Achieved cycle (min)Passed
S0>=30%>=60<=36030.00%70.30302.853/3
S1>=25%>=70<=30025.03%72.32290.613/3

“Cycle” denotes mean cycle time. Both capacity and operational requirements differ between scenarios. S1's development means met its thresholds, but its cycle-time interval was inconclusive. We locked the candidate and independently confirmed it without reselection or relaxed acceptance criteria. Its 95% intervals were [72.20, 72.43] orders/h and [286.11, 295.52] minutes, satisfying both requirements.

This study follows a general prompt clarification after initial C04 testing; no C04-specific answers were inserted. The inputs share one geometry and simulation streams, so they demonstrate consistency across authored formulations, not six independent warehouse cases. The evidence supports meaning preservation and adaptation within the supported vocabulary.

Thank you again for your valuable and constructive feedback. Your comments have substantially improved the clarity of our paper. We would be grateful if you could reevaluate our work and reconsider the score in light of these clarifications and additional results.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.