Pith. sign in

REVIEW 3 major objections 4 minor 40 references

Stigmergic Graph Memory: An Environment-Aware Approach for Many-to-Many Multi-Agent Pickup and Delivery

T0 review · 3 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

Pith's one-line read A decaying memory of recent warehouse traffic raises many-to-many pickup-and-delivery throughput 20.5–36.7%.

desk verdict A genuinely new decision interface with a careful but conditional empirical case; the headline gains hold against reconstructed baselines, not validated state-of-the-art M2M. read the letter →

arxiv 2607.15182 v1 pith:EOJYI3TY submitted 2026-07-16 cs.MA cs.RO

classification cs.MAcs.RO
keywords stigmergymulti-agentpickupanddeliverymany-to-manyMAPDtaskallocationgraphmemorycongestioncontrolrolling-horizonplanningwarehouserobotics
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper argues that in many-to-many MAPD, where a request names a product rather than fixed pickup and delivery points, the controller's choice of which endpoints become planner goals is a congestion-control decision. It introduces SGM, a bounded, decaying memory layer that records recent waits, blocks, and traffic on warehouse nodes and directed edges, and uses that memory to rank feasible endpoint pairs and to bias routing costs. Across five layouts, three fleet sizes, and 25 request streams per condition, SGM beats two reconstructed baselines in all 15 map–load cells, with paired throughput gains of 20.5–36.7%. The decisive ablation shows that keeping only the endpoint-steering memory and dropping route guidance retains essentially all of the throughput gain, while keeping only route guidance does not; this points to a lever that prior methods leave unexploited because they steer traffic only after goals are fixed.

What carries the argument

SGM is a typed, decaying memory layer over the warehouse graph. Separate channels record node-level waiting, endpoint pressure, and completion, and directed-edge-level traversal, delay, blocking, and directional flow, each updated as Mt+1 = ρMt + Δt. Endpoint steering adds a memory-derived path penalty to the baseline allocation cost, ranking feasible agent–source–destination candidates; route guidance converts the same memory into bounded positive edge costs for the planner. The load-bearing mechanism is the recursive update with channel-specific retention factors (0.85–0.95), letting recent traffic dominate stale events; the endpoint-only ablation isolates the endpoint-pressure/waiting/blo

What would settle it

Re-run the benchmark against the original published many-to-many MAPD implementation (or its authors' own reported numbers) under the same paired request replay. If the original baseline matches SGM's completed-task counts, the claimed gains are an artifact of the reconstruction or the changed queue-rebuild policy. A second check: if the endpoint-only ablation's throughput advantage disappears when queue preservation is also removed (i.e., both use rebuild), then the gain is due to queue continuity, not memory-guided endpoint selection.

Watch

Extended reading notes

Core claim

The central claim is that recent execution history, recorded as decaying signals on warehouse nodes and directed edges, can be used to rank feasible source–destination instantiations in many-to-many MAPD, and that this endpoint steering—not routing—accounts for the throughput improvement. The paper reports paired throughput gains of 20.5–36.7% over two reconstructed many-to-many allocation baselines across all 15 map–load conditions, with the endpoint-only ablation keeping 99.3–100.5% of full SGM throughput. Route guidance, in turn, reduces planner time, waiting, blocked moves, and blocked replans without changing throughput. The mechanism is a bounded preference layer: it changes only scala

Load-bearing premise

The two baseline allocators, reconstructed from the published many-to-many code and run under the common harness, faithfully represent the published methods; if the reconstructions are weaker than the originals, the reported 20.5–36.7% gains overstate the improvement.

Editorial extensions

If this is right

  • Throughput in many-to-many MAPD can be improved by shaping which feasible endpoints become goals, before path planning begins.
  • The memory layer is planner-agnostic and validity-preserving, so it can be added to existing rolling-horizon control stacks without changing collision constraints.
  • Route guidance is an operational-efficiency lever (planning time, waiting, blocked motion) even when it does not change throughput.
  • Gains are consistent across structurally different aisle and corridor layouts and across fleet occupancy levels from 30% to 90% of parking capacity.
  • Queue preservation contributes a smaller but non-negligible part of the gain; removing it reduces throughput relative to full SGM but still beats the baselines.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If endpoint instantiation is the dominant lever, then methods that learn or optimize endpoint-selection policies from demand forecasts might show comparable or larger gains than pure recent-traffic memory.
  • The decay rates and weights are hand-tuned per channel; a self-tuning variant could adapt to nonstationary demand or layout changes without manual reconfiguration.
  • The sensitivity sweep shows that raising the endpoint-pair cap from 32 to 64 increases the gain from 29.2% to 40.5%, suggesting search budget is a practical lever that could be traded against compute in real deployments.
  • Because routing alone does not change throughput, the paper implies that congestion costs are best internalized at assignment time; an analytical model of how endpoint pressure propagates through the planner could predict when such memory will help.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper introduces Stigmergic Graph Memory (SGM), a bounded, decaying memory layer that records recent execution signals on warehouse nodes and directed edges and uses them to rank feasible endpoints and adjust route-cost preferences in many-to-many Multi-Agent Pickup and Delivery (MAPD). The central empirical claim is that, across five layouts, three fleet levels, and 25 paired seeds per condition, SGM improves completed-task throughput in all 15 map–load conditions over two reconstructed many-to-many allocation baselines (M2M and M2M-wSKU), with paired gains of 20.5–36.7%. The paper further argues, via the endpoint-only ablation, that the dominant mechanism is memory-guided endpoint instantiation rather than routing or queue handling. The experimental design includes paired request replay, Holm-corrected Wilcoxon tests, ablations, routing controls, sensitivity sweeps, matched-cap controls, and a medium-scale transfer study.

Significance. If the result holds, the paper makes a useful conceptual contribution: in many-to-many MAPD, congestion can be shaped before goals are fixed, at the endpoint-instantiation stage, rather than only after goals are committed. The internal comparison is unusually careful: paired request streams, 25 seeds, preplanned Holm corrections, and a battery of controls make the within-harness advantage of SGM over the reconstructed baselines credible. The main reservation is external validity: the baselines are reconstructions from the official M2M repository but are never validated against the originally reported M2M throughput, so the headline gains are only as strong as the fidelity of those reconstructions. The paper's own sensitivity and ablation evidence partly mitigates related confounds such as queue handling, but the reconstruction question remains load-bearing for the quantitative claims.

major comments (3)
  1. [Section 5, 'Baseline provenance'; Table S2] The primary quantitative claim is measured against two 'reconstructed' M2M baselines, but the paper provides no evidence that these reconstructions reproduce the published M2M/M2M-wSKU behavior. No reproduced throughput number from Schneider et al. (2026) is reported, and no check is given that the official repository's allocator, under the shared RHCR/PBS harness, yields results comparable to the original paper. Because the 20.5–36.7% gains are relative to these baselines, the headline result is conditional on reconstruction fidelity. Please add a validation subsection: either reproduce one or more originally reported conditions, or, if that is not possible, explicitly state that the baselines are unverified reconstructions and consistently use 'reconstructed' qualifiers in all text, tables, and figures (e.g., 'Best reconstructed M2M' in Table 3 and Figure 3). Without this, the numerica
  2. [Section 4.1, Eq. (1); Table S4] The memory configuration — retention factors ρ_c, path-memory weight λ_p=1.75, route-guidance weight λ_r=0.2, cap κ=0.1, and shortlist caps 64/64/32 — is described as 'empirically derived' without specifying the selection protocol. If the same 25-seed benchmark was used both to choose these values and to report final gains, the results are at risk of selection bias. The one-factor sensitivity sweeps in Table S9 are reassuring (gains remain large across varied settings), but the paper should state whether the configuration was fixed before running the test seeds, or provide a split-half/hold-out analysis. This is needed to support the strength of the numerical claims.
  3. [Section 6.3, Table S7] The no-queue-preserve ablation is a useful control and shows SGM with memory but rebuild still outperforms M2M-wSKU by 23.6–30.5%, so the queue-handling difference is not the sole driver of the headline gain. However, this ablation still uses the same reconstructed baselines as comparators. If those reconstructions are weaker than the published M2M implementation, the no-queue-preserve result inherits that weakness and does not fully rescue the external claim. This is listed as a separate comment to make the dependency explicit.
minor comments (4)
  1. [Eq. (3)] The notation π(a_i,s) ∪ π(s,d) is ambiguous: it is read as a set union of paths, but the intent is the concatenation of the edge sequences along the two paths. Consider writing it as a concatenation operator or as a multiset of edges.
  2. [Table 3 and Figure 3] The label 'Best M2M' in Table 3 and 'Best M2M' in Figure 3 should consistently read 'Best reconstructed M2M' to match the provenance statement in Section 5 and the abstract.
  3. [Section 4.5] The 'Proposition (feasibility preservation)' is a direct consequence of the design (SGM changes only positive costs and ranks feasible candidates) and does not require a proof. Consider presenting it as a remark rather than a proposition, so as not to overstate the formal contribution.
  4. [Section 6, transfer study (Table S10)] The transfer study shows that the endpoint-only ablation is not robust at medium scale (−8.6±21.0% versus M2M-wSKU), while full SGM retains a positive gain. The main text discusses this, but the Abstract's broad statement that 'endpoint instantiation is the primary throughput-control interface' could be read as universal; please add a qualifier noting that the transfer evidence shows the full controller is needed at larger scales.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical benchmark study; the main risks are baseline reconstruction and configuration tuning, not definitional or citation-based circularity.

full rationale

The paper's central claim is an empirical benchmark result rather than a derivation. SGM is defined by Eqs. (1)-(4) as a memory layer that augments an M2M-style allocation cost C0 with bounded memory penalties; throughput is then measured in simulation. Nothing in these equations assumes the reported 20.5-36.7% gains. The memory configuration is described as 'empirically derived', and the main results use that configuration, but the paper reports one-factor sensitivity sweeps (Table S9) showing the advantage persists across path-memory weights, retention factors, and route-memory weights, and the endpoint-only ablation retains SGM throughput while routing-only does not. Tuning a controller on the evaluation benchmark is a potential overfitting/validity concern, not definitional circularity, because the measured endpoint variable (completed tasks under replayed request streams) is not contained in the fitted constants by construction. The reference list contains no papers by the present authors, so self-citation patterns are absent; no uniqueness theorem is imported from the authors' prior work, and no ansatz is smuggled in via citation. The baseline reconstructions are explicitly described as 'reconstructed from the official repository of Schneider et al. (2026)' and the paper limits its claim to the shared RHCR/PBS harness in the 'Scope of the claim' passage. The unvalidated reconstructions and the Rebuild-vs-Preserve queue-handling difference are genuine external-validity risks for the unqualified 'outperforms M2M' phrasing, but they concern whether the baselines are representative, not whether the derivation reduces to its own inputs. No circular step can be identified under the stated criteria.

Assumptions & free parameters 8 free parameters · 5 assumptions · 1 invented entities

The central claim depends on a substantial set of tuned parameters and on the assumption that reconstructed baselines represent the original M2M methods. No new physical entities are introduced; the SGM memory is an algorithmic construct. The sensitivity sweeps mitigate, but do not eliminate, the concern that headline results use a configuration selected for those benchmarks.

free parameters (8)
  • Memory retention factors rho_c per channel = waiting/delay 0.85; blocking/endpoint-pressure 0.90; congestion/traversal/flow 0.92; completion 0.95
    Empirically derived configuration controlling how fast each execution signal decays; directly shapes endpoint and route scores. Section 4.1 and Table S3.
  • Endpoint path-memory weight lambda_p = 1.75
    Scales the memory penalty added to the M2M-wSKU-style baseline cost in Eq. S2; tuned and reported with sensitivity sweep.
  • Route guidance weight lambda_r and cap kappa = lambda_r=0.2, kappa=0.1
    Controls bounded routing-cost inflation in Eq. S3; every legal edge has cost between 1 and 1.1.
  • Corridor retention modifier = +0.05 retention, capped at 0.99
    Optional adjustment increasing retention on degree-at-most-two corridors; part of the empirically derived configuration.
  • SGM shortlist caps = source 64, destination 64, pairs 32
    Candidate budget for endpoint steering; sensitivity shows throughput varies strongly with pair cap (16: +16.0%, 32: +29.2%, 64: +40.5%).
  • M2M-wSKU baseline weights = w_b=1.0, w_s=0.25
    Baseline inventory-distribution term weights from Schneider et al. 2026; used in the reconstructed baseline and in SGM's C0 term.
  • Guidance horizon = 32 timesteps
    Horizon for discounted planned-occupancy cost in Rt; a controller setting.
  • RHCR/PBS planning window and allocator depth = window 512, allocator sequence depth 3
    Common controller settings fixed across methods; likely influence absolute throughput but not the method comparison.
assumptions (5)
  • domain assumption MAPF vertex and edge conflict constraints are the correct model of collision avoidance.
    Section 3 defines the problem with these constraints; SGM's feasibility-preservation claim relies on RHCR/PBS enforcing them.
  • domain assumption Well-formed infrastructure with designated parking cells sufficient for each agent (Ma et al. 2017).
    Section 5 states fleets do not exceed the 93 parking cells, maintaining one non-task endpoint per agent; this supports lifelong feasibility.
  • domain assumption Reconstructed M2M and M2M-wSKU accurately represent the published methods of Schneider et al. (2026).
    Section 5 'Baseline provenance' says the baselines are reconstructed from the official repository but does not validate them against the original published throughput numbers.
  • ad hoc to paper Sampled unit-cost shortest paths are adequate scoring proxies for endpoint ranking.
    Eq. 3 sums memory penalties along unit-cost shortest paths; the final RHCR/PBS paths may differ materially from these proxies.
  • domain assumption Paired request replay across the chosen seeds and layouts represents warehouse workloads of interest.
    The evaluation fixes request streams per seed and layout; nonstationary demand, different SKU distributions, and physical robot constraints are left to future work.
invented entities (1)
  • SGM typed decaying memory channels (node and directed-edge channels)
    purpose: Store recent execution signals such as waiting, endpoint pressure, blocking, delay, and flow, and use them to rank endpoints and guide routing.
    Defined in Section 4.1 as an algorithmic construct; no evidence outside the paper's own simulations establishes its validity as an independent physical or system phenomenon.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Stigmergic Graph Memory: An Environment-Aware Approach for Many-to-Many Multi-Agent Pickup and Delivery." pith.science (2026). https://pith.science/paper/EOJYI3TY

@misc{pith2026260715182,
  author       = {Pith},
  title        = {Pith review of: Stigmergic Graph Memory: An Environment-Aware Approach for Many-to-Many Multi-Agent Pickup and Delivery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EOJYI3TY}},
  note         = {Machine review of arXiv:2607.15182}
}
read the original abstract

Automated fulfillment warehouses must continuously assign and execute pickup-and-delivery work while avoiding congestion. In many-to-many Multi-Agent Pickup and Delivery (MAPD), a request specifies a stock-keeping unit rather than fixed endpoints, requiring the controller to select an agent, source, and destination before path planning. Existing graph-guidance methods primarily influence routing after goals are fixed, leaving endpoint instantiation uninformed by recent traffic. We introduce Stigmergic Graph Memory (SGM), a bounded, decaying memory layer that records recent execution signals on warehouse nodes and directed edges to rank feasible endpoints and route preferences without altering collision constraints or planner validity. Across paired request streams on five layouts, three load levels, and 25 seeds per condition, SGM outperforms two reconstructed many-to-many allocation baselines in all 15 map-load conditions, with paired throughput gains of 20.5-36.7%. These results show that recent execution memory can improve warehouse throughput by shaping which feasible goals enter the planner, not only how agents travel to already fixed goals.

Figures

Figures reproduced from arXiv: 2607.15182 by the authors.

Figure 1
Figure 1. SGM execution cycle. The current warehouse state provides released requests, agent locations, and active goals. SGM [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The five warehouse layouts, each 50 × 27 cells with 93 designated parking cells. Restricted, open-top, and open are benchmark warehouse layouts from the many-to-many MAPD setting; maze A and maze B provide structurally distinct corridor networks at the same scale. Supplementary Table S5 gives the per-layout cell counts. comparisons of SGM with M2M and M2M-wSKU using the Holm step-down procedure (Holm 1979). Paired m… view at source ↗
Figure 3
Figure 3. Cumulative completed tasks on maze A and [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 1 canonical work pages

  1. [1]

    Sturtevant and Ariel Felner and Sven Koenig and Hang Ma and Thayne T

    Roni Stern and Nathan R. Sturtevant and Ariel Felner and Sven Koenig and Hang Ma and Thayne T. Walker and Jiaoyang Li and Dor Atzmon and Liron Cohen and T. K. Satish Kumar and Roman Bart. Multi-Agent Pathfinding: Definitions, Variants, and Benchmarks , booktitle =. 2019 , doi =

  2. [2]

    Proceedings of the First Artificial Intelligence and Interactive Digital Entertainment Conference (AIIDE) , volume =

    David Silver , title =. Proceedings of the First Artificial Intelligence and Interactive Digital Entertainment Conference (AIIDE) , volume =. 2005 , doi =

  3. [3]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Trevor Standley , title =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2010 , doi =

  4. [4]

    Sturtevant , title =

    Guni Sharon and Roni Stern and Ariel Felner and Nathan R. Sturtevant , title =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2012 , doi =

  5. [5]

    Sturtevant , title =

    Guni Sharon and Roni Stern and Ariel Felner and Nathan R. Sturtevant , title =. Artificial Intelligence , volume =. 2015 , doi =

  6. [6]

    Proceedings of the IEEE International Conference on Robotics and Automation (ICRA) , year =

    Mike Phillips and Maxim Likhachev , title =. Proceedings of the IEEE International Conference on Robotics and Automation (ICRA) , year =

  7. [7]

    Priority Inheritance with Backtracking for Iterative Multi-Agent Path Finding , booktitle =

    Keisuke Okumura and Manao Machida and Xavier D\'. Priority Inheritance with Backtracking for Iterative Multi-Agent Path Finding , booktitle =. 2019 , pages =. doi:10.24963/ijcai.2019/76 , url =

  8. [8]

    Priority Inheritance with Backtracking for Iterative Multi-Agent Path Finding , journal =

    Keisuke Okumura and Manao Machida and Xavier D\'. Priority Inheritance with Backtracking for Iterative Multi-Agent Path Finding , journal =. 2022 , doi =

Show all 40 references
  1. [9]

    winPIBT: Extended Prioritized Algorithm for Iterative Multi-Agent Path Finding , year =

    Keisuke Okumura and Yasumasa Tamura and Xavier D\'. winPIBT: Extended Prioritized Algorithm for Iterative Multi-Agent Path Finding , year =. 1905.10149 , archivePrefix =

  2. [10]

    Hang Ma and Jiaoyang Li and T. K. Satish Kumar and Sven Koenig , title =. Proceedings of the International Conference on Autonomous Agents and Multiagent Systems (AAMAS) , year =

  3. [11]

    Multi-Agent Path Finding with Kinematic Constraints , booktitle =

    Wolfgang H. Multi-Agent Path Finding with Kinematic Constraints , booktitle =. 2016 , doi =

  4. [12]

    Durham and T

    Jiaoyang Li and Andrew Tinka and Scott Kiesel and Joseph W. Durham and T. K. Satish Kumar and Sven Koenig , title =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2021 , doi =

  5. [13]

    Harabor and Peter J

    Zhe Chen and Javier Alonso-Mora and Xiaoshan Bai and Daniel D. Harabor and Peter J. Stuckey , title =. IEEE Robotics and Automation Letters , volume =. 2021 , doi =

  6. [14]

    Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages =

    Qinghong Xu and Jiaoyang Li and Sven Koenig and Hang Ma , title =. Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages =. 2022 , doi =

  7. [15]

    2022 , eprint =

    Tim Tsz-Kit Lau and Biswa Sengupta , title =. 2022 , eprint =

  8. [16]

    Proceedings of the International Symposium on Combinatorial Search (SoCS) , volume =

    Xiaohu Wu and Yihao Liu and Xueyan Tang and Wentong Cai and Funing Bai and Gilbert Khonstantine and Guopeng Zhao , title =. Proceedings of the International Symposium on Combinatorial Search (SoCS) , volume =. 2021 , doi =

  9. [17]

    Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages =

    Hiroya Makino and Seigo Ito , title =. Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages =. 2024 , doi =

  10. [18]

    2026 , eprint =

    Ethan Schneider and Jingkai Chen and Tianyi Gu and Kunlei Lian and Seth Hutchinson and Sonia Chernova , title =. 2026 , eprint =. doi:10.48550/arXiv.2605.07835 , url =

  11. [19]

    2023 , eprint =

    Ming-Feng Li and Min Sun , title =. 2023 , eprint =

  12. [20]

    Stuckey , title =

    Zhe Chen and Daniel Harabor and Jiaoyang Li and Peter J. Stuckey , title =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2024 , doi =

  13. [21]

    Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI) , pages =

    Yulun Zhang and He Jiang and Varun Bhatt and Stefanos Nikolaidis and Jiaoyang Li , title =. Proceedings of the International Joint Conference on Artificial Intelligence (IJCAI) , pages =. 2024 , doi =

  14. [22]

    Stuckey and Jiaoyang Li , title =

    Hongzhi Zang and Yulun Zhang and He Jiang and Zhe Chen and Daniel Harabor and Peter J. Stuckey and Jiaoyang Li , title =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2025 , doi =

  15. [23]

    Fontaine and Stefanos Nikolaidis and Jiaoyang Li , title =

    Yulun Zhang and Varun Bhatt and Matthew C. Fontaine and Stefanos Nikolaidis and Jiaoyang Li , title =. 2026 , eprint =

  16. [24]

    La Reconstruction du Nid et les Coordinations Interindividuelles chez Bellicositermes Natalensis et Cubitermes sp

    Pierre-Paul Grass. La Reconstruction du Nid et les Coordinations Interindividuelles chez Bellicositermes Natalensis et Cubitermes sp. La Th. Insectes Sociaux , volume =

  17. [25]

    Eric Bonabeau and Marco Dorigo and Guy Theraulaz , title =

  18. [26]

    IEEE Transactions on Evolutionary Computation , volume =

    Marco Dorigo and Luca Maria Gambardella , title =. IEEE Transactions on Evolutionary Computation , volume =. 1997 , doi =

  19. [27]

    Artificial Life , volume =

    Guy Theraulaz and Eric Bonabeau , title =. Artificial Life , volume =. 1999 , doi =

  20. [28]

    Stuckey and Jiaoyang Li and Sven Koenig , title =

    Hang Ma and Daniel Harabor and Peter J. Stuckey and Jiaoyang Li and Sven Koenig , title =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2019 , doi =

  21. [29]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume =

    Pavel Surynek , title =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2010 , doi =

  22. [30]

    Proceedings of the 18th International Conference on Autonomous Agents and MultiAgent Systems (AAMAS) , pages =

    Minghua Liu and Hang Ma and Jiaoyang Li and Sven Koenig , title =. Proceedings of the 18th International Conference on Autonomous Agents and MultiAgent Systems (AAMAS) , pages =

  23. [31]

    2024 , eprint =

    Arsalan Sharifnassab and Saber Salehkaleybar and Sina Ghiassian and Surya Kanoria and Dale Schuurmans , title =. 2024 , eprint =

  24. [32]

    Ahuja and

    Ravindra K. Ahuja and. A Survey of Very Large-Scale Neighborhood Search Techniques , journal =. 2002 , doi =

  25. [33]

    Large Neighborhood Search , booktitle =

    David Pisinger and Stefan R. Large Neighborhood Search , booktitle =. 2010 , doi =

  26. [34]

    Biometrics Bulletin , volume=

    Individual Comparisons by Ranking Methods , author=. Biometrics Bulletin , volume=. 1945 , doi=

  27. [35]

    Scandinavian Journal of Statistics , volume=

    A Simple Sequentially Rejective Multiple Test Procedure , author=. Scandinavian Journal of Statistics , volume=

  28. [36]

    Wurman and Raffaello D'Andrea and Mick Mountz , title =

    Peter R. Wurman and Raffaello D'Andrea and Mick Mountz , title =. AI Magazine , volume =. 2008 , doi =

  29. [37]

    Warehousing in the E-Commerce Era: A Survey , journal =

    Nils Boysen and Ren. Warehousing in the E-Commerce Era: A Survey , journal =. 2019 , doi =

  30. [38]

    LaValle , title =

    Jingjin Yu and Steven M. LaValle , title =. Proceedings of the AAAI Conference on Artificial Intelligence , volume =. 2013 , doi =

  31. [39]

    Ayorkor Korsah and Anthony Stentz and M

    G. Ayorkor Korsah and Anthony Stentz and M. Bernardine Dias , title =. The International Journal of Robotics Research , volume =. 2013 , doi =

  32. [40]

    Lifelong Path Planning with Kinematic Constraints for Multi-Agent Pickup and Delivery , booktitle =

    Hang Ma and Wolfgang H. Lifelong Path Planning with Kinematic Constraints for Multi-Agent Pickup and Delivery , booktitle =. 2019 , doi =

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.