acceptodds
Under review as a conference paper at ICLR 2027

Multi-Objective Layout Optimization for Multi-Robot Coordination in Automated Warehouse

Abstract

We study the problem of multi-objective layout optimization for automated warehouses, where hundreds to thousands of robots are coordinated to transport packages from one location to another. Previous works have attempted to optimize the system throughput of warehouses by optimizing their layout (e.g., the physical locations of the storage shelves) while assuming a fixed storage capacity (i.e., the number of storage shelves). However, in practice, storage capacity also significantly affects the operating cost of the warehouse, because a smaller storage capacity requires the warehouse to occupy more land. Importantly, there are trade-offs between the storage capacity and throughput, because a larger storage capacity creates more obstacles for robots, complicating the multi-robot coordination and potentially reducing throughput. In this paper, we generalize the state-of-the-art single-objective layout optimization methods based on Quality Diversity (QD) algorithms to multiple objectives, where we optimize for a Pareto front (PF) of throughput and storage capacity. We show that our optimized PF significantly outperforms vanilla multi-objective optimization algorithms across various robot task distributions, robot coordination algorithms, and layout sizes. We additionally present a Mixed Integer Linear Programming Solver (MILP) that enforces the optimized layout to be biconnected, which helps the downstream coordination algorithm move the robots more effectively. Our code will be publicly available upon acceptance.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Multi-Agent Path Finding, Multi-Robot Coordination, Multi-objective Optimization, Automated Warehouse, Layout Optimization

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_aY8xOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies multi-objective layout optimization for automated warehouses, aiming to simultaneously maximize robot throughput and storage capacity. The authors propose MO-DSAGE, a surrogate-assisted multi-objective quality-diversity optimization framework that searches over neural cellular automata (NCA) generators to produce a Pareto front of warehouse layouts. The framework integrates MO-CMA-MAE with surrogate models to reduce expensive layout evaluations. In addition, the authors introduce a MILP formulation to enforce layout validity and biconnectivity. Experiments with different multi-agent pathfinding algorithms and warehouse configurations demonstrate improved Pareto-front hypervolume in several settings.

Critical strengths

The proposed method achieves substantially better hypervolume than several optimization baselines in the main experimental setting.

Jointly optimizing throughput and storage capacity captures an important trade-off that is overlooked by previous single-objective warehouse layout optimization methods.

Critical weaknesses
  1. The core framework largely combines previous works

  2. The main extension from single-objective to multi-objective optimization appears to consist primarily of adopting an existing multi-objective optimizer and extending the surrogate predictions and archives accordingly.

  3. The paper does not clearly establish a sufficiently novel algorithmic principle beyond integration.

  4. The experiments do not adequately disentangle the contributions of surrogate modeling and multi-objective quality-diversity optimization. In Table 1, MO-CMA-MAE and CMA-MOA achieve relatively similar performance

Questions for authors

What is the essential methodological novelty of MO-DSAGE

4 comments

We thank the reviewer for the feedback and the opportunity to clarify our contributions. We have incorporated your suggestions into the paper, with revisions highlighted in red.

1-3. Essential methodological novelty. As reviewer g31h notes, “limited novelty” is not a deal-breaker for practical work: although individual components may be based on known work, putting them together is not always obvious. Our problem is motivated by a conversation with a company building multi-robot automated warehouse systems. Unlike human-operated facilities, robot-operated warehouse layouts may have fewer requirements for regular patterns, creating a larger search space and a harder optimization problem (Zhang et al.). We jointly optimize throughput and storage capacity to provide layouts with different trade-offs. The framework combines existing MO-CMA-MAE, DSAGE, and NCA components, and we introduce a novel biconnectivity MILP for layout repair. We also use NCA to optimize a layout generator at a smaller size and generate larger layouts, with a concurrently trained surrogate to speed up search.

  1. We agree that diversity brings limited improvement at the optimization size of 33 × 36; its benefit mostly appears when using NCA to generate larger layouts, as shown in Table 2. At 69 × 69, MO-CMA-MAE achieves a hypervolume of 9,868, compared with 9,154 for CMA-MOA; at 101 × 102, the values are 23,288 and 17,421, respectively. These Table 2 results show that diversity contributes to better solutions when scaling to larger layouts. Separately, surrogate modeling improves sample efficiency by reducing the number of expensive real evaluations needed during optimization. We have clarified this point in Section 5.2.

References

Zhang, Yulun, et al. “Multi-Robot Coordination and Layout Design for Automated Warehousing.” IJCAI 2023.

Dear reviewer aY8x,

Thank you for your thoughtful feedback. We have posted our responses and uploaded a revised manuscript. As the discussion period draws to a close, we would welcome any remaining questions or concerns. Thank you for your time and consideration.

Thank you for the clarification and response. I acknowledge the practical value of the integrated framework and the contribution of the biconnectivity MILP.

Thank you for your thoughtful follow-up. We appreciate your recognition of the integrated framework and biconnectivity MILP. If any substantive concerns remain that our response or revision has not addressed, we would welcome the opportunity to clarify them.

Reviewer_8hrtOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies warehouse layout optimization for automated warehouses. It optimizes both system throughput and storage capacity. The method uses surrogate-assisted multi-objective search and a MILP that enforces biconnectivity. Experiments compare the method across different planners, warehouse settings, and layout sizes.

Critical strengths
  1. The paper addresses a meaningful gap in prior MAPF-aware warehouse layout optimization by jointly optimizing throughput and storage capacity, instead of treating storage capacity as fixed.

  2. The method is practically well motivated. The surrogate-assisted search reduces expensive MILP and simulation evaluations, while the biconnectivity constraint is clearly connected to downstream multi-robot planning performance.

Critical weaknesses
  1. The benefit of QD over ordinary multi-objective optimization is not fully clear. Since the final goal is a global Pareto front, stronger evidence is needed to show that diversity preservation itself improves the final solutions.

  2. The surrogate model is central to the method, but its prediction and ranking reliability are not directly evaluated in enough detail. This is especially important for layout problems, where small structural changes can lead to large performance differences.

Questions for authors
  1. How accurate is the surrogate model, especially for ranking candidates near the Pareto front? It would be helpful to report metrics such as ranking correlation, top-k recall, or prediction error on high-quality layouts.

  2. How does surrogate accuracy change for novel or out-of-distribution layouts? Since small structural changes can cause large changes in throughput, this seems important for a diversity-driven search method.

  3. Can the authors clarify how much of the final improvement comes from QD itself? The comparison between MO-CMA-MAE and CMA-MOA does not show a very clear advantage of diversity preservation for the final global Pareto front.

  4. How sensitive is the advantage of MO-DSAGE to the real-evaluation budget? In simpler settings or with much larger computation budgets, does direct optimization catch up with the surrogate-assisted method?

  5. Could you clarify the intended scope of the novelty claim? Multi-objective formulations are common in broader facility-layout, scheduling, and operations-research problems, so it would be useful to distinguish what is new specifically for MAPF-aware automated warehouse layout optimization.

2 comments

We thank the reviewer for the thoughtful questions. We have incorporated your suggestions into the paper, with revisions highlighted in red.

1-2. We evaluate the accuracy of a surrogate model from a MO-DSAGE run with PIBT on (1) layouts on the final Pareto Front (PF), (2) in-distribution training set, and (3) out-of-distribution validation set. The training and validation sets each have 200 layouts. The training set is sampled from the 10k layouts evaluated in the MO-DSAGE run, the validation set is sampled from a different MO-DSAGE run.

We compare the predicted and ground-truth throughput by their numerical differences and their rankings. For numerical accuracy, we get an MAE of 0.08 (PF), 0.18 (training), and 0.28 (validation), as well as an RMSE of 0.11 (PF), 0.25 (training) and 0.44 (validation). For ranking, we get Spearman's ρ\rho of 0.99 (PF), 0.98 (training) and 0.93 (validation), as well as a correct pairwise ordering of 98.20% (PF), 95.49% (training) and 90.24% (validation). Overall, our surrogate model is well-trained in terms of predicting the throughput. These results are added to Section 5.2 and we have added Figure 5 to show the scatter plots the comparison between predicted and ground-truth throughput.

  1. We agree that diversity brings limited improvement at the optimization size of 33 × 36. We emphasize that its benefit is mostly seen when using the NCA to generate larger layouts, as shown in Table 2. At 69 × 69, MO-CMA-MAE achieves a hypervolume of 9,868, compared with 9,154 for CMA-MOA; at 101 × 102, the values are 23,288 and 17,421, respectively. We have clarified this point in Section 5.2.

  2. In principle, direct optimization can catch up with MO-DSAGE if given a sufficient number of evaluations. In practice, however, we cannot run enough evaluations to observe this convergence because of compute limits.

  3. Our work is motivated by the needs of automated warehouses operated by robots. Compared with human-operated facilities, robot-operated warehouse layouts may have fewer requirements for regular patterns, creating a larger search space and a harder optimization problem (Zhang et al.). Our work jointly optimizes throughput and storage capacity, uses NCA to generate larger layouts, and introduces a biconnectivity MILP for layout repair.

References

Zhang, Yulun, et al. “Multi-Robot Coordination and Layout Design for Automated Warehousing.” IJCAI 2023.

Thank you for the experiments and response. Most of my concerns are addressed, and I will maintain my original rating of weak acceptance.

Reviewer_g31hOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies the problem of generating layouts for a multi-robot warehouse (e.g., Amazon’s robot warehouse) considering multiple objectives that need to be optimized. Examples of the objectives are amount of storage shelves in the layout and throughput. These objectives may conflict with each other as increasing storage may decrease throughput via fewer locations that the same number of robots can move on. The key idea is to use a Neural Cellular Automata (NCA) for layout optimization thet generates the map, along with a MILP based repair of the map to ensure correct type of connectivity, with a multi-objective with quality-diversity search (using CMA) for the NCA’s parameters. Finally, they also use a surrogate model (DSAGE) to stand in for the MILP repair during the fast search process. The key results show that the presented method achieves better diversity in the generated layouts along the objectives of interest.

Critical strengths

I like the problem setting; although its specific to multi-robot warehouse layout optimization, the general recipe could be useful in other layout (e.g., chip design, city planning) problems and also other multi-objective problems. The paper also presents a sound approach that actually seems to work compared to the baselines. I also like the MILP repair part of the paper that could be of independent interest. While one may argue these are not traditional ICLR contributions, I would disagree. This paper presents concrete algorithmic framework for a practical application.

Critical weaknesses

I personally don’t believe that “limited novelty” is a deal-breaker for papers that address a practical problem. The paper does use existing components (I believe all the individual components in the pipeline are based on known work) but putting them together is not always obvious when first developing the solution. My main limitation is that while diversity of objectives is the motivation, the paper doesn’t seem to use many objectives in their evaluation. It’s mostly throughput vs storage count… but I would imagine in practice, there are many other objectives that practitioners would care about. For example, how many robots are used at any given point, travel times for the robots, latency in specific requests being served etc. I would imagine that many of them would be affected by the layout even if the throughput is similar. In fact, even with throughput, you could consider p50, p90, etc. type metrics. Overall, I think the multi-objective machinery is under-utilized in the settings used in the experiments.

Questions for authors

Please see my comments in the weakness above. I would like to hear about additional objectives and how effective does the method remain when we have several competing objectives.

Since the number of robots is tied to free space, part of the trade-off may come from fleet size rather than congestion. It would be nice to results with a fixed fleet to clarify this.

What I would have liked to see is how practical this method is on actual production level problem setups. Perhaps, that’s hard to do in a double-blind setting but I would love for the authors to clarify how realistic a scale of the problems that they are solving are to the real world.

4 comments

We thank the reviewer for appreciating the practical motivation of our work. We have incorporated your suggestions into the paper, with revisions highlighted in red.

  1. Our problem setting is motivated by a conversation with a real-world warehouse company. We agree that automated warehouses may consider other metrics, such as workstation idle time, robot travel time, and request latency, but these metrics can mostly be related to throughput or shelf count. For example, in a lifelong setting, robots are mostly traveling, so reducing travel time between sequential tasks directly increases throughput. Request latency could be considered by formulating lifelong MAPF with deadlines and measuring throughput within the deadline (Huang et al.). In addition, we emphasize that considering the current two objectives (throughput and number of shelves) already form a very challenging problem and serve as a starting point for future research that may consider additional objectives. Nevertheless, our work does not consider all possible objectives relevant to automated warehouses. For example, p50 and p90 of throughput could be optimized by binning throughput over time and used as additional objectives. We have added discussion of these future directions in Section 6.

  2. We compute the number of robots as a fixed ratio of traversable space because of the trade-off between fleet size and throughput. Typically, the number of robots that achieves the highest throughput in a layout is in a “sweet spot”: adding more can create congestion, while reducing the number can lower throughput. However, optimizing the sweet spot for every single layout is computationally expensive, so we make the simplifying assumption that this point can be approximated by a fixed ratio of traversable space. We have added discussion of this assumption in Section 5.1 and 6.

  3. Compared to real-world problems, our proposed approach can be applied, but it does have two main limitations: simulator realism and layout size. First, our simulator does not model robot dynamics, which would make simulation higher fidelity but slower. To mitigate this issue, the surrogate modeling approach used in this paper can be used to speed up the search. Second, real warehouse layouts can also be much larger than the 33 × 36 default optimization size. We use NCA so that we can optimize a layout generator at a smaller size and generate larger layouts, including 100 × 100. This can match some small or mid-sized warehouses, though some extra large facilities can be up to five times larger. We believe this gap can be addressed through reasonable future improvements to our method.

References

Huang, Taoan, et al. “Deadline-Aware Multi-Agent Tour Planning.” ICAPS 2023.

Dear reviewer g31h,

Thank you for your thoughtful feedback. We have posted our responses and uploaded a revised manuscript. As the discussion period draws to a close, we would welcome any remaining questions or concerns. Thank you for your time and consideration.

Dear reviewer g31h,

Thank you for your thoughtful feedback. We have posted our responses and uploaded a revised manuscript. As the discussion period draws to a close, we would welcome any remaining questions or concerns. Thank you for your time and consideration.

The authors have clarified several of the issues I'd pointed out. I will raise my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.