Pith. sign in

REVIEW 2 major objections 5 minor 59 references

Enhancing Lifelong Multi-Agent Path-finding by Using Artificial Potential Fields

T0 review · 2 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Adding artificial potential fields to existing planners yields up to a 7-fold throughput increase in lifelong multi-agent path finding, while leaving one-shot MAPF performance unchanged.

desk verdict A solid empirical paper on APF-based cost shaping for lifelong MAPF, with an honest negative result for one-shot MAPF; the large throughput claim needs variance reporting and parameter robustness checks before I'd trust the 7x number. read the letter →

arxiv 2505.22753 v1 pith:DRYQJL5D submitted 2025-05-28 cs.AI cs.MAcs.RO

Arseniy Pertzovsky , Roni Stern , Ariel Felner , Roie Zivan This is my paper
classification cs.AIcs.MAcs.RO
keywords multi-agentpathfindinglifelongMAPFartificialpotentialfieldsthroughputMAPF-LNS2PIBTrolling-horizoncollisionresolutioncongestionavoidance
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

Artificial potential fields (APFs) are a classic robot-motion trick: obstacles and other agents push, goals pull. The paper asks whether adding such repulsion costs to modern multi-agent pathfinding (MAPF) planners helps them avoid congested areas. It argues and experimentally shows that in one-shot MAPF the answer is no, but in lifelong MAPF, where agents receive new goals continuously and the same corridors are reused by many plans, the same APF costs raise system throughput substantially, up to roughly seven times for one planner configuration. If true, this gives a cheap, planner-agnostic way to cut congestion in warehouse-style task fulfillment without changing the underlying search.

What carries the argument

The central object is a time-indexed repulsion potential field around each already-planned agent path: $APF_i(v,t) = w\cdot\gamma^{-d(v,\pi_i[t])}$ when the distance $d(v,\pi_i[t])$ from vertex $v$ to the path at time $t$ is below a cutoff $d_{\max}$, and $0$ otherwise; total APF cost is the sum over all other agents' fields. This is a named class of objects, Artificial Potential Fields (APFs), with parameters $w$ (strength), $\gamma$ (decay), and $d_{\max}$ (radius of influence). The field is added not to the heuristic but to the accumulated edge cost ($g$-value) in TA*, to both the soft-collision count and $g$-value in SIPPS (taking the max over the safe interval), and to the neighbor-ordering heuristic in PIBT. Because the search space and conflict rules are unchanged, completeness is preserved; only the expansion order is biased. These fields are what carry the argument from 'avoid other agents' paths' to 'reduce congestion in lifelong replanning.'

What would settle it

Run the same lifelong experiments on the same four maps with the APF radius set to zero (which the paper identifies with plain TA*) and with window and horizon varied beyond 5; if the throughput gap between APF and vanilla LNS2 collapses when $d_{\max}=0$ but persists when $d_{\max}>0$, the field itself is responsible. If instead the 7-fold gap depends on the particular window, horizon, and tuned $w$ and $\gamma$ values and vanishes under small perturbations, the robustness claim would be undercut.

Watch

Extended reading notes

Core claim

On the paper's own terms: injecting APF-based costs into the low-level single-agent searches of existing MAPF algorithms changes which paths agents choose, biasing them away from the planned paths of other agents. The authors integrate this into Temporal A* (adding the APF cost to the g-value), into SIPPS (adding it to the collision count and g-value), and into PIBT (adding it to the vertex-ordering heuristic, with LaCAM inheriting the change). In the lifelong setting inside the Rolling-Horizon Collision Resolution framework, this bias repeatedly steers agents around congested regions; on the empty-32-32 grid, LNS2 using TA*+APF reaches a throughput of about 1400 with 450 agents, roughly seven times vanilla LNS2, and PIBT+APF outperforms both vanilla PIBT and the guide-path baseline. The same APF augmentation is reported not to help, and sometimes to hurt, when solving a single offline MAPF instance.

Load-bearing premise

The load-bearing premise is that the throughput gains come from the APF mechanism itself, rather than from the specific RHCR settings (window=horizon=5, AllAgents+iStay+Persist, 100-step limit) and the per-algorithm APF parameters that were tuned on the same four benchmark maps used for evaluation.

Editorial extensions

If this is right

  • On the empty-32-32 grid, LNS2 with TA*+APF reaches a throughput of about 1400 at 450 agents, roughly 7 times vanilla LNS2.
  • PIBT+APF outperforms vanilla PIBT and the L-PIBT+GP baseline on all four test maps, with about 20% higher average throughput on room-32-32-4.
  • APF integration preserves completeness for TA* and SIPPS because the underlying search space is unchanged; only node-expansion order and path preference change.
  • In one-shot MAPF, APF variants are not beneficial (identical or inferior results) across PrP, LNS2, PIBT, and LaCAM, so the gain is specific to the lifelong, repeated-replanning regime.
  • Throughput improves for algorithms with TA*, SIPPS, and PIBT on all four benchmark maps, suggesting the effect is not an artifact of one planner.

Reading between the lines

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

  • A natural next test: vary the RHCR window and horizon while holding APF parameters fixed; if the relative gain shrinks when the horizon is long enough to see congestion in advance, the mechanism is confirmed as a short-horizon congestion-forecast substitute.
  • The APF cost could be combined with learned congestion predictors or with the guide-path baseline; because the field is a simple parameterized function, its weights could be tuned by search or learning rather than hand-set per algorithm.
  • The paper tunes APF parameters per algorithm on the same four maps used for evaluation; transferring these parameters to new, denser, or larger maps may require re-tuning, and the published sensitivity analysis suggests w and gamma in particular are load-bearing.
  • A direct ablation would quantify how much of the 7-fold gain comes from reducing deadlocks versus from shorter effective path lengths; the paper does not separate these, but the iStay+Persist failure handling makes deadlock avoidance a plausible major component.
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

2 major / 5 minor

Summary. The paper investigates whether Artificial Potential Fields (APFs) can improve Multi-Agent Path Finding (MAPF) and Lifelong MAPF (LMAPF). It proposes a myopic direct APF method, then integrates APF-based repulsion penalties into the g-values and priority functions of TA*, SIPPS, PIBT, and LaCAM/LaCAM*. Experiments on four grid maps from the MAPF benchmark show that APFs do not help one-shot MAPF but can substantially increase LMAPF throughput, with the headline claim being an approximately 7-fold throughput improvement for LNS2 with TA*+APF on empty-32-32 with 450 agents. A parameter sensitivity analysis is included.

Significance. If the central empirical claim holds, the contribution is valuable: a simple, low-overhead cost-shaping mechanism that can be dropped into several existing LMAPF planners to improve throughput, while honestly reporting that the same mechanism does not help one-shot MAPF. The paper provides complexity overhead estimates for each integration and includes a sensitivity analysis, which is a strength. The main limitation is that the headline gains are demonstrated only under parameters tuned on the same benchmark maps, with no error bars or significance testing, so the generalizability of the 7-fold claim is not yet established. The paper does not ship code, but the experimental setup is described in sufficient detail to be reproduced.

major comments (2)
  1. [Section 6 (LMAPF experiments and parameter sensitivity)] The APF parameters for each algorithm are selected on the same four benchmark maps used for evaluation ('observed to work best in general across all grids'), and the sensitivity analysis (Figures 4 and 7) shows that throughput is highly sensitive to these parameters: for PIBT+APF, gamma=1 performs like vanilla PIBT, t_max=2 is sharply better than t_max=1 or 3, and large w can be worse than no APFs. Because the headline '7-fold' figure is the best-case ratio on empty-32-32 with 450 agents under parameters selected on that map set, the paper does not yet establish that the improvement is due to the APF mechanism rather than to an overfit configuration. I request a holdout evaluation on maps not used for tuning, or a report of results under a fixed default parameter set alongside the tuned best case.
  2. [Section 6 (LMAPF experiments)] All throughput numbers are averages over 15 random instances, but no standard deviations, confidence intervals, or significance tests are reported anywhere in the paper. Without a measure of variance it is impossible to assess whether the reported ratios (e.g., the approximately 7-fold LNS2+TA*+APF improvement on empty-32-32) are stable or driven by a few outliers. Please include per-instance distributions, error bars/confidence intervals, or a statistical test (e.g., paired comparison against the vanilla baseline) for the main throughput comparisons.
minor comments (5)
  1. [Section 3] The text 'using DAFP is very fast' appears to contain a typo; it should read 'DAPF'.
  2. [Section 6] The phrase 'within the RCHR framework' should be 'within the RHCR framework'.
  3. [Section 6 and Appendix B] The cross-references to figures are inconsistent: the main-text sensitivity analysis is referred to as 'Figure 7' but the displayed figure is labeled Figure 4, and the appendix similarly cites 'Figure 7' for TA*+APF sensitivity while Figures 6 and 7 appear in the appendix. Please unify the figure numbering and references.
  4. [Equation (4)] The notation d(v, t, pi_i[t]) in Equation (4) is confusing because the time argument is already encoded in the node; it should be d(v, pi_i[t]) with an explicit note that v and pi_i[t] are vertices in the same grid.
  5. [Section 2] The heading 'Temporal A∗, SIPP, and SIPPSTemporal A ∗' contains a duplicated phrase and should be cleaned up.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper reports measured benchmark comparisons; parameter tuning is an overfitting risk, not a circular derivation.

full rationale

The paper introduces APF-based cost terms into existing MAPF planners (TA*, SIPPS, PIBT) and evaluates them on standard MAPF benchmark maps. The claimed central result, up to a 7-fold throughput increase in LMAPF, is an empirical measurement: throughput is computed from RHCR simulations after applying Eqs. (4)-(12), and the APF cost is not defined in terms of the measured throughput. The APF parameters (w, d_max, gamma, t_max) were chosen as 'observed to work best in general across all grids' on the same four benchmark maps, and the sensitivity analysis shows that performance is strongly parameter-dependent; this is a legitimate generalization and overfitting concern, but it is not circular because the reported throughput values are direct experimental outcomes rather than predictions forced by a fitted model. The authors' use of their own prior work [30] to justify the AllAgents+iStay+Persist failure-handling configuration is a protocol choice and is not load-bearing for the APF improvement claim. The completeness argument for TA*+APF and SIPPS+APF (Section 4.3) is an independent search-space-preservation argument. No equation is equivalent by construction to the headline claim, and no fitted parameter is renamed as a prediction.

Assumptions & free parameters 11 free parameters · 4 assumptions · 0 invented entities

The central claim rests on eleven hand-fixed numerical settings (APF parameters plus RHCR window/horizon) and on representativeness of the benchmark protocol. No new physical or algorithmic entities are postulated. The APF functions are classical, and the modifications are limited to cost functions of existing planners.

free parameters (11)
  • TA*+APF repulsion strength w = 1
    Chosen by the authors as working best across grids; controls APF penalty weight relative to movement cost.
  • TA*+APF influence radius d_max = 4
    Cutoff radius for repulsion around other agents' paths; tuned per variant on the benchmark maps.
  • TA*+APF decay rate gamma = 2
    Exponential decay rate of repulsion with distance; selected after sensitivity analysis.
  • SIPPS+APF repulsion strength w = 0.1
    Selected as best in the supplementary sensitivity analysis; low weight avoids distorting soft-constraint counts.
  • SIPPS+APF influence radius d_max = 3
    Best value in the reported SIPPS+APF sensitivity sweep.
  • SIPPS+APF decay rate gamma = 3
    Best value in the reported SIPPS+APF sensitivity sweep.
  • PIBT+APF repulsion strength w = 0.1
    Chosen as best on the room-32-32-4 map and other grids.
  • PIBT+APF influence radius d_max = 2
    Chosen as best across grids; d_max=1 or 4 gave similar results in sensitivity analysis.
  • PIBT+APF decay rate gamma = 3
    Chosen as best; gamma=1 was equivalent to vanilla PIBT.
  • PIBT+APF lookahead steps t_max = 2
    t_max=2 yielded significantly better results than all other tested values.
  • RHCR window and horizon = 5 and 5
    Fixed in all LMAPF experiments; throughput results may depend on this hand-chosen protocol setting.
assumptions (4)
  • standard math Grid-world MAPF uses Manhattan distance as both the heuristic and the APF distance metric.
    Invoked in Equations (1) and (4); requires four-connected grid graphs and assumes Manhattan distance is a reasonable proxy for true path distance.
  • domain assumption Adding APF penalties to the g-value does not alter the search space, so completeness is preserved.
    Section 4.3 argues the search spaces with and without APFs are identical and only node expansion order changes; this relies on standard A* and SIPP correctness.
  • domain assumption The RHCR benchmark protocol with four 32x32 maps, 15 random instances, and a 100-step horizon is representative of lifelong MAPF workloads.
    Section 6 derives all conclusions from this protocol; if the protocol is unrepresentative, the general claim about LMAPF throughput weakens.
  • ad hoc to paper The hand-picked APF parameters are stable enough across maps and agent densities to support the reported gains.
    The paper selects w, d_max, gamma, and t_max based on performance on the same benchmark maps; if these choices are overfit, the 7-fold claim does not generalize.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Lifelong Multi-Agent Path-finding by Using Artificial Potential Fields." pith.science (2026). https://pith.science/paper/DRYQJL5D

@misc{pith2026250522753,
  author       = {Pith},
  title        = {Pith review of: Enhancing Lifelong Multi-Agent Path-finding by Using Artificial Potential Fields},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DRYQJL5D}},
  note         = {Machine review of arXiv:2505.22753}
}
read the original abstract

We explore the use of Artificial Potential Fields (APFs) to solve Multi-Agent Path Finding (MAPF) and Lifelong MAPF (LMAPF) problems. In MAPF, a team of agents must move to their goal locations without collisions, whereas in LMAPF, new goals are generated upon arrival. We propose methods for incorporating APFs in a range of MAPF algorithms, including Prioritized Planning, MAPF-LNS2, and Priority Inheritance with Backtracking (PIBT). Experimental results show that using APF is not beneficial for MAPF but yields up to a 7-fold increase in overall system throughput for LMAPF.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 58 canonical work pages

  1. [1]

    Agrawal, S

    A. Agrawal, S. Hariharan, A. S. Bedi, and D. Manocha. Dc-mrta: De- centralized multi-robot task allocation and navigation in complex envi- ronments. InIROS, pages 11711–11718. IEEE, 2022

    work page 2022

  2. [2]

    Barraquand, B

    J. Barraquand, B. Langlois, and J.-C. Latombe. Numerical potential field techniques for robot path planning.IEEE Transactions on Systems, Man, and Cybernetics, 22(2):224–241, 1992

    work page 1992

  3. [3]

    Barták, J

    R. Barták, J. Švancara, V . Škopková, D. Nohejl, and I. Krasi ˇcenko. Multi-agent path finding on real robots.AI Communications, 2019

    work page 2019

  4. [4]

    Bennewitz, W

    M. Bennewitz, W. Burgard, and S. Thrun. Optimizing schedules for prioritized path planning of multi-robot systems. InICRA, volume 1, pages 271–276, 2001

    work page 2001

  5. [5]

    Bettini, R

    M. Bettini, R. Kortvelesy, J. Blumenkamp, and A. Prorok. Vmas: a vectorized multi-agent simulator for collective robot learning.arXiv preprint arXiv:2207.03530, 2022

    arXiv 2022

  6. [6]

    S.-H. Chan, R. Stern, A. Felner, and S. Koenig. Greedy priority-based search for suboptimal multi-agent path finding. InSoCS, pages 11–19, 2023

    work page 2023

  7. [7]

    Z. Chen, D. Harabor, J. Li, and P. J. Stuckey. Traffic flow optimisation for lifelong multi-agent path finding. InAAAI, volume 38, pages 20674– 20682, 2024

    work page 2024

  8. [8]

    Daily and D

    R. Daily and D. M. Bevly. Harmonic potential field path planning for high speed vehicles. In2008 American Control Conference, pages 4609–4614. IEEE, 2008

    work page 2008

Show all 59 references
  1. [9]

    P. S. Daniel Kornhauser, Gary Miller. Coordinating pebble motion on graphs, the diameter of permutation groups, and applications. InFOCS, 1984

    1984

  2. [10]

    Dergachev and K

    S. Dergachev and K. Yakovlev. Distributed multi-agent navigation based on reciprocal collision avoidance and locally confined multi- agent path finding. InCASE, 2021

    2021

  3. [11]

    H. T. Dinh, R. van Lon, and T. Holvoet. Multi-agent route planning using delegate mas. InWorkshop on Distributed and Multi-Agent Plan- ning, pages 24–32. London, UK, 2016

    2016

  4. [12]

    T. Fan, P. Long, W. Liu, and J. Pan. Distributed multi-robot collision avoidance via deep reinforcement learning for navigation in complex scenarios.IJRR, 39(7):856–892, 2020

    2020

  5. [13]

    D. Fox, W. Burgard, and S. Thrun. The dynamic window approach to collision avoidance.IEEE Robotics & Automation Magazine, 4(1): 23–33, 1997

    1997

  6. [14]

    Hagelback and S

    J. Hagelback and S. Johansson. A multi-agent potential field-based bot for a full rts game scenario. InAAAI, volume 5, pages 28–33, 2009

    2009

  7. [15]

    Hagelbäck and S

    J. Hagelbäck and S. J. Johansson. Using multi-agent potential fields in real-time strategy games. InAAMAS, pages 631–638, 2008

    2008

  8. [16]

    S. D. Han and J. Yu. Ddm: Fast near-optimal multi-robot path plan- ning using diversified-path and optimal sub-problem solution database heuristics.IEEE RA-L, 5(2):1350–1357, 2020

    2020

  9. [17]

    S. D. Han and J. Yu. Optimizing space utilization for more effective multi-robot path planning. InICRA, pages 10709–10715. IEEE, 2022

    2022

  10. [18]

    O. Khatib. The potential field approach and operational space formu- lation in robot control. InAdaptive and Learning Systems: Theory and Applications, pages 367–377. Springer, 1986

    1986

  11. [19]

    Koren and J

    Y . Koren and J. Borenstein. Potential field methods and their inherent limitations for mobile robot navigation. InICRA, pages 1398–1404, 1991

    1991

  12. [20]

    E. Lam, P. Le Bodic, D. Harabor, and P. J. Stuckey. Branch-and- cut-and-price for multi-agent path finding.Computers & Operations Research, 144:105809, 2022. URL https://www.sciencedirect.com/ science/article/pii/S0305054822000946

    2022

  13. [21]

    C. Leet, J. Li, and S. Koenig. Shard systems: Scalable, robust and per- sistent multi-agent path finding with performance guarantees. InAAAI, pages 9386–9395, 2022

    2022

  14. [22]

    J. Li, Z. Chen, D. Harabor, P. Stuckey, and S. Koenig. Anytime multi- agent path finding via large neighborhood search. InIJCAI, 2021

    2021

  15. [23]

    J. Li, A. Tinka, S. Kiesel, J. W. Durham, T. K. S. Kumar, and S. Koenig. Lifelong multi-agent path finding in large-scale warehouses.AAAI, May 2021

    2021

  16. [24]

    J. Li, Z. Chen, D. Harabor, P. J. Stuckey, and S. Koenig. Mapf-lns2: Fast repairing for multi-agent path finding via large neighborhood search. In AAAI, 2022

    2022

  17. [25]

    X. Liu, S. S. Ge, and C.-H. Goh. Formation potential field for trajec- tory tracking control of multi-agents in constrained space.International Journal of Control, 90(10):2137–2151, 2017

    2017

  18. [26]

    R. J. Luna and K. E. Bekris. Push and swap: Fast cooperative path- finding with completeness guarantees. InIJCAI, 2011

    2011

  19. [27]

    H. Ma, J. Yang, L. Cohen, T. K. S. Kumar, and S. Koenig. Feasibil- ity study: Moving non-homogeneous teams in congested video game environments. InAIIDE, 2017

    2017

  20. [28]

    H. Ma, D. Harabor, P. J. Stuckey, J. Li, and S. Koenig. Searching with consistent prioritization for multi-agent path finding. InAAAI, pages 7643–7650, 2019

    2019

  21. [29]

    T. T. Mac, C. Copot, D. T. Tran, and R. De Keyser. Heuristic approaches in robot path planning: A survey.Robotics and Autonomous Systems, 86:13–28, 2016

    2016

  22. [30]

    Morag, R

    J. Morag, R. Stern, and A. Felner. Adapting to planning failures in lifelong multi-agent path finding. InSoCS, 2023

    2023

  23. [31]

    Morris, C

    R. Morris, C. S. Pasareanu, K. S. Luckow, W. Malik, H. Ma, T. S. Ku- mar, and S. Koenig. Planning, scheduling and monitoring for airport surface operations. InAAAI Workshop: Planning for Hybrid Systems, 2016

    2016

  24. [32]

    K. Okumura. Lacam: Search-based algorithm for quick multi-agent pathfinding. InAAAI, volume 37, pages 11655–11662, 2023

    2023

  25. [33]

    Okumura, M

    K. Okumura, M. Machida, X. Défago, and Y . Tamura. Priority inheri- tance with backtracking for iterative multi-agent path finding.AI, 310: 103752, 2022

    2022

  26. [34]

    Phillips and M

    M. Phillips and M. Likhachev. Sipp: Safe interval path planning for dynamic environments. InICRA, pages 5628–5635. IEEE, 2011

    2011

  27. [35]

    Rezaee and F

    H. Rezaee and F. Abdollahi. Adaptive artificial potential field approach for obstacle avoidance of unmanned aircrafts. InAIM, pages 1–6. IEEE, 2012

    2012

  28. [36]

    Salzman and R

    O. Salzman and R. Z. Stern. Research challenges and opportunities in multi-agent path finding and multi-agent pickup and delivery problems blue sky ideas track. InAAMAS, 2020

    2020

  29. [37]

    S. H. Semnani, H. Liu, M. Everett, A. De Ruiter, and J. P. How. Multi- agent motion planning for dense and dynamic environments via deep reinforcement learning.IEEE RA-L, 5(2):3221–3226, 2020

    2020

  30. [38]

    Sharon, R

    G. Sharon, R. Stern, M. Goldenberg, and A. Felner. The increasing cost tree search for optimal multi-agent pathfinding.AI, 195:470–495, 2013

    2013

  31. [39]

    Sharon, R

    G. Sharon, R. Stern, A. Felner, and N. R. Sturtevant. Conflict-based search for optimal multi-agent pathfinding.AI, 2015

    2015

  32. [40]

    Shin and E

    Y . Shin and E. Kim. Hybrid path planning using positioning risk and artificial potential fields.Aerospace Science and Technology, 112, 2021

    2021

  33. [41]

    D. Silver. Cooperative pathfinding. InAIIDE, 2005

    2005

  34. [42]

    Skrynnik, A

    A. Skrynnik, A. Andreychuk, M. Nesterova, K. Yakovlev, and A. Panov. Learn to follow: Decentralized lifelong multi-agent pathfinding via planning and learning. InAAAI, volume 38, pages 17541–17549, 2024

    2024

  35. [43]

    Song, K.-I

    S. Song, K.-I. Na, and W. Yu. Anytime lifelong multi-agent pathfinding in topological maps.IEEE Access, 11:20365–20380, 2023

    2023

  36. [44]

    T. S. Standley. Finding optimal solutions to cooperative pathfinding problems. InAAAI, 2010

    2010

  37. [45]

    Stern, N

    R. Stern, N. R. Sturtevant, A. Felner, S. Koenig, H. Ma, T. T. Walker, J. Li, D. Atzmon, L. Cohen, T. K. S. Kumar, E. Boyarski, and R. Bar- tak. Multi-agent pathfinding: Definitions, variants, and benchmarks. In SoCS, pages 151–158, 2019

    2019

  38. [46]

    P. Surynek. An optimization variant of multi-robot path planning is intractable. InAAAI, 2010

    2010

  39. [47]

    Surynek, A

    P. Surynek, A. Felner, R. Stern, and E. Boyarski. Efficient sat approach to multi-agent path finding under the sum of costs objective. InECAI, pages 810–818, 2016

    2016

  40. [48]

    Švancara, M

    J. Švancara, M. Vlk, R. Stern, D. Atzmon, and R. Barták. Online multi- agent pathfinding. InAAAI, 2019

    2019

  41. [49]

    Vadakkepat, K

    P. Vadakkepat, K. C. Tan, and W. Ming-Liang. Evolutionary artificial potential fields and their application in real time robot path planning. In Congress on Evolutionary Computation, 2000

    2000

  42. [50]

    Van den Berg, S

    J. Van den Berg, S. J. Guy, M. Lin, and D. Manocha. Reciprocal n-body collision avoidance. InRobotics Research, pages 3–19, 2011

    2011

  43. [51]

    Varambally, J

    S. Varambally, J. Li, and S. Koenig. Which mapf model works best for automated warehousing? InSoCS, 2022. URL https://www. autostoresystem.com/

    2022

  44. [52]

    Wagner and H

    G. Wagner and H. Choset. M*: A complete multirobot path planning algorithm with performance bounds. In2011 IEEE/RSJ international conference on intelligent robots and systems, pages 3260–3267. IEEE, 2011

    2011

  45. [53]

    Wahid, H

    N. Wahid, H. Zamzuri, M. A. A. Rahman, S. Kuroda, and P. Raksin- charoensak. Study on potential field based motion planning and control for automated vehicle collision avoidance systems. InICM, pages 208– 213, 2017

    2017

  46. [54]

    Wang and G

    Y . Wang and G. Chirikjian. A new potential field method for robot path planning. InICRA, volume 2, pages 977–982 vol.2, 2000. doi: 10.1109/ROBOT.2000.844727

  47. [55]

    P. R. Wurman, R. D’Andrea, and M. Mountz. Coordinating hundreds of cooperative, autonomous vehicles in warehouses.AI magazine, 29(1): 9–9, 2008

    2008

  48. [56]

    Yu and S

    J. Yu and S. M. LaValle. Structure and intractability of optimal multi- robot path planning on graphs. InAAAI, 2013

    2013

  49. [57]

    Zhang, W.-m

    H.-y. Zhang, W.-m. Lin, and A.-x. Chen. Path planning for the mobile robot: A review.Symmetry, 10(10):450, 2018

    2018

  50. [58]

    Zhang, J

    S. Zhang, J. Li, T. Huang, S. Koenig, and B. Dilkina. Learning a priority ordering for prioritized planning in multi-agent path finding. InSoCS,

  51. [2022]

    A Justification for Parameters To demonstrate the role of each parameter, consider Figure 5

    URL www.aaai.org. A Justification for Parameters To demonstrate the role of each parameter, consider Figure 5. The x-axis is the distance from the agent (in the middle), they-axis is the value of APFs, and the height of the red bars represents the specific APFs cost for the lo...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.