PILOC: A Pheromone Inverse Guidance Mechanism and Local-Communication Framework for Dynamic Target Search of Multi-Agent in Unknown Environments

Hengrui Liu; Qilong Zhang; Yi Feng

REVIEW 3 major objections 4 minor 29 references

PILOC: A Pheromone Inverse Guidance Mechanism and Local-Communication Framework for Dynamic Target Search of Multi-Agent in Unknown Environments

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read PILOC claims a pheromone anti-revisit signal plus local map sharing lets two agents find moving targets in unknown grids 95.6% of the time, beating standard MARL baselines.

desk verdict A useful incremental integration spoiled by an unreported test-time A* fallback and under-specified baselines, so the headline success margin is not established. read the letter →

arxiv 2507.07376 v1 pith:ZUCHUY66 submitted 2025-07-10 cs.RO cs.AI

Hengrui Liu , Yi Feng , Qilong Zhang This is my paper · ORCID

classification cs.ROcs.AI

keywords multi-agentsearchandrescuedeepreinforcementlearningpheromoneinverseguidancelocalcommunicationdynamictargetsunknownenvironmentsMAPPODec-POMDP

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

PILOC is a decentralized deep-reinforcement-learning framework for multi-agent search and rescue in unknown, partially observable environments. The paper's central claim is that a pheromone inverse guidance mechanism—an anti-revisit signal embedded directly into each agent's observation—combined with local map sharing between nearby agents, lets a small team of agents find randomly moving targets more often and faster than standard MARL algorithms. In the paper's dynamic-target grid-world task, PILOC reaches a 95.6% success rate against 87.2% for IPPO, 74.0% for MASAC, 62.8% for QMIX, and 18.4% for frontier-based exploration. The contribution matters because real search-and-rescue settings lack global communication and prior maps, so a method that works from local perception and local exchange could transfer to such deployments.

What carries the argument

The central mechanism is the Pheromone Inverse Guidance Mechanism: each agent deposits a pheromone value at its current cell, the value is capped and evaporates over time, and the agent's observation includes a local pheromone map so the policy learns to steer toward low-pheromone, less-visited areas. This is supported by a local-communication rule that unions the obstacle and exploration maps of agents within range, and a shaped reward with exploration, re-exploration, collision, and pheromone-change terms. The whole loop is trained with MAPPO under centralized training and decentralized execution, with a condition-triggered A* fallback during testing.

What would settle it

Retrain IPPO and MASAC using PILOC's exact three-channel observation (obstacle, exploration, pheromone), its reward function, and its curriculum, then rerun the 250-map test; if their success rates approach PILOC's, the advantage comes from the reward/observation design rather than the framework. Alternatively, count how often the A* fallback triggers during PILOC's successful test episodes; if it fires frequently, the learned policy is not the primary driver.

Watch

Extended reading notes

Core claim

The paper discovers that reversing the usual ant-colony pheromone logic—making pheromone a deterrent that marks where agents have already been, rather than an attractor—and feeding that signal into a MAPPO-trained policy observation space, removes redundant exploration and lets agents coordinate indirectly through the environment. When agents within a limited communication range merge their obstacle and exploration maps, the system keeps low communication overhead while still sharing coverage knowledge. The paper reports that the full PILOC framework outperforms all tested baselines on success rate, average completion steps, step variance, and average targets found, and that ablations removing either mechanism reduce performance. The framework's hybrid decision module switches to A* pathfinding when an agent revisits a cell more than three times in ten steps, which the paper credits with preventing agents from getting trapped.

Load-bearing premise

The reported success-rate gap assumes the comparison baselines were trained with the same observation space, reward function, and curriculum as PILOC, which the paper does not state explicitly.

Editorial extensions

If this is right

In the tested setting, PILOC finds all six dynamic targets in 95.6% of episodes, compared with 87.2% for IPPO, 74.0% for MASAC, 62.8% for QMIX, and 18.4% for frontier exploration.
Adding the pheromone inverse guidance alone raises success from 81.6% to 91.2% in the ablation, and adding local communication alone raises it to 88.8%; combining both yields the full 95.6%.
Scaling from two to five agents increases success rate monotonically to 100% and reduces average steps from 123.51 to 59.01, indicating the framework is scalable.
Because communication is local, the framework reduces reliance on global channels and is suited to communication-constrained scenarios.

Reading between the lines

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

If the pheromone inverse guidance transfers beyond grid worlds, it could be applied to continuous terrain or heterogeneous robots by replacing the grid pheromone field with a decayed visit-density function over a metric map.
The same anti-revisit signal could act as a cheap substitute for explicit team coordination in bandwidth-limited settings, since agents need only observe local pheromone, not send messages.
A testable extension: vary the evaporation rate and pheromone cap systematically; the paper fixes them at λ=0.02 and Pmax=10, leaving a sensitivity analysis that would show how robust the mechanism is.
The scalability result (100% success at four agents) suggests the marginal benefit of extra agents saturates; an inference is that the framework's value is largest in the two-to-three-agent regime where the anti-revisit mechanism matters most.

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.

Desk Editor's Note

read the letter

Quick take: this is a reasonable engineering paper that assembles known pieces (MAPPO, exploration bonuses, map merging, a rule-based fallback) into a system that beats its chosen baselines on a public grid-world benchmark. The result might even be true. But as reported, the main comparison doesn't isolate the mechanism the paper claims to sell, because the test-time A* fallback is never measured or ablated, and the baselines are not specified to match the reward and observation design.

What is actually new: the pheromone field is fed into the network as an observation channel and also used in the reward, with the sign inverted to discourage revisits. That's a modest, legitimate twist on count-based exploration. The local map merging with time-marking selection is standard but neat, and the hybrid decision mechanism is clearly described. The ablation study is the best part: removing both mechanisms drops success from 95.6% to 81.6%, and each component adds something. That passes the smell test.

Where it falls short: the §4.8 fallback is active at test time and fires whenever an agent revisits a cell more than 3 times in 10 steps. Its trigger rate is unreported, and no run disables it. Since the fallback sends the agent to the nearest unexplored cell or the cell with the highest time-marking value, it encodes exactly the coverage objective the pheromone reward is supposed to shape. If it fires often, the 95.6% might be mostly rule-based rescue, not learned policy. Independent of that, the baselines are not documented as receiving the same shaped reward, observation channels, or curriculum. Without those details, the margin over IPPO and MASAC could be reward shaping or observation engineering rather than the framework. There is also a numeric inconsistency: Table 1 says QMIX 62.8%, the text says 58.4%. No error bars or seeds are given, so we don't know how stable the 95.6% is.

None of this is fatal to the idea. The components are sensible, the ablations are directionally right, and the paper is honest about relying on a public map dataset. But the central claim as written is not yet supported. The authors need to report fallback trigger statistics, run a fallback-disabled variant, disclose baseline training details, and give seed variance.

Verdict: this is a potentially solid incremental contribution that deserves a serious referee, but with major revision required. Not something I'd cite for its performance claims until the confounds are addressed.

Referee Report

3 major / 4 minor

Summary. The manuscript presents PILOC, a decentralized multi-agent reinforcement learning framework for dynamic target search and rescue in unknown grid environments. Agents use three observation maps (obstacle, exploration with time-marking, and pheromone concentration), a pheromone inverse guidance mechanism that rewards movement toward less-frequented cells, and a local communication scheme that merges nearby agents' maps. Training uses MAPPO with a shaped reward (Eqs. 6-9) and a curriculum schedule. The paper reports that PILOC achieves a 95.6% success rate on 250 test scenarios, outperforming IPPO (87.2%), MASAC (74.0%), QMIX (62.8%), and Frontier (18.4%), with ablations showing that both the pheromone mechanism and local communication contribute. A separate hybrid decision mechanism (Section 4.8) switches to an A* rule-based planner when an agent revisits its position too often, but its activation statistics are not reported.

Significance. If properly validated, the idea of embedding a pheromone-inspired anti-revisit signal directly into the DRL observation space, combined with local map merging, is a plausible and useful step for decentralized search in communication-constrained settings. The paper also provides a concrete curriculum and reward design. However, the central empirical claim currently rests on a confounded comparison: the test-time A* fallback is not isolated, and the baseline training conditions are unspecified. With supplementary trigger statistics and fallback-free evaluations, the contribution could be assessed fairly.

major comments (3)

[Section 4.8 and Section 5.2] The condition-triggered A* fallback is enabled at test time, but the paper reports no trigger frequency, no fraction of episodes in which the fallback fired, and no evaluation of PILOC or the ablations with the fallback disabled. Because the fallback targets essentially the same coverage/anti-revisit objective as the pheromone and time-marking signals, the headline 95.6% success rate may conflate the learned mechanism with a rule-based escape. Please report trigger statistics and add fallback-free runs for PILOC and for the Section 5.3 ablations.
[Section 5.1 and Table 1] The training setup for the baselines IPPO, MASAC, and QMIX is not described. It is unclear whether these baselines received the same observation channels (obstacle map, exploration map with time markings, and pheromone map), the same shaped reward of Eqs. (6)-(9), and the same curriculum learning schedule. If they did not, the reported performance gap could reflect reward shaping and observation design rather than the proposed framework. Please specify the baseline observation and reward configurations and report means and standard deviations over multiple seeds.
[Section 4.6, Eq. (9)] The pheromone reward r_ph depends on I'_ph and I_ph, described only as "pheromone concentrations in the perceptual range," but the aggregation over the perceptual range is not defined, and the division by I'_ph raises a potential division-by-zero concern when the previous pheromone concentration is zero. Please define these quantities precisely and state how zero values are handled.

minor comments (4)

[Table 1 and Section 5.2] The QMIX success rate is reported as 62.8% in Table 1 but as 58.4% in the text; please reconcile the discrepancy.
[Title] The title contains a spacing artifact: "P HEROMONE" should be "PILOC: A Pheromone Inverse Guidance Mechanism..."
[Section 4.1] The Dec-POMDP tuple is written as <N, S, A, O, Omega, R, T, gamma>, but the text later uses per-agent rewards; please clarify how the reward function R(s) relates to the individual rewards used in Eqs. (6)-(9).
[Section 5.1] The evaluation uses 250 randomly selected scenarios, but the number of evaluation seeds or independent runs is not stated; please state the variance across repeated evaluations.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper is an empirical MARL comparison against external baselines on a public dataset; the DRL mechanism is not defined in terms of its reported success metric, and no fitted parameter is renamed as a prediction.

full rationale

PILOC is an empirical systems paper: its central claim is a performance comparison of a learned search policy with external baselines (IPPO, MASAC, QMIX, Frontier) measured on the public map dataset [26]. The pheromone map and time-marking map are generated from agent trajectories, and the reward terms in Eqs. 6-9 shape exploration behavior; they do not define the success metric, and no parameter is fitted to the reported test success rates. The cited prior work is external, not authored by the PILOC authors, and no uniqueness theorem or self-citation chain forces the architecture. A non-circularity confound should be flagged: Section 4.8 enables a condition-triggered A* fallback at test time ('it switches from neural network-based decision making to a rule-based SAR algorithm'), and the paper does not report how often the fallback triggers or provide a fallback-disabled ablation, so the incremental contribution of the learned mechanism is not fully isolated. This is an attribution/control-variable concern rather than circularity, because the fallback is an external rule-based component, not an input that is equivalent by construction to the measured success rate. Therefore the circularity score is 0.

Assumptions & free parameters 11 free parameters · 6 assumptions · 1 invented entities

The central claim rests on a large set of hand-chosen reward coefficients and thresholds, on domain assumptions about target motion and map representativeness, and on an algorithmic construct (virtual pheromone) with no external falsifiable handle. The paper does not report the values of several key parameters, notably the communication range and the pheromone observation window.

free parameters (11)

Pmax = 10
Upper bound on pheromone concentration, set by hand in Section 4.3; affects the dynamics of the guidance signal.
lambda (evaporation rate) = 0.02
Pheromone decay rate per time step, set by hand in Section 4.3.
alpha, beta (pheromone reward weights) = 0.1, 0.1
Weights in Eq. 9 balancing relative and absolute pheromone changes; hand-chosen in Section 4.6.
Exploration reward coefficient = 0.5
Coefficient in Eq. 7 that determines reward per newly discovered cell (Section 4.6).
Re-exploration rewards = 0.01 and 0.1
Reward terms in Eq. 8 for revisiting aged areas versus finding a target (Section 4.6).
Collision penalty = -3
Fixed penalty per obstacle collision (Section 4.6).
Perception radius v = 5
Agent visibility range in grid cells (Section 5.1); controls the observation maps.
Communication range c
Radius within which agents exchange maps; its value is never reported, though the method's central claim depends on it (Section 4.4).
Pheromone observation window l
Side length of the pheromone map given to the network; not reported (Section 4.2).
Curriculum threshold M
Number of episodes without improvement before increasing episode length; not reported (Section 5.1).
Revisit fallback threshold = 3 visits in 10 steps
Trigger for switching to A* rule-based search in testing; set in Section 4.8 and not ablated.

assumptions (6)

domain assumption The random-walk target model with uniform action selection is representative of dynamic targets in search-and-rescue scenarios.
Section 3.2 defines target movement; the generality of the results depends on this model matching real dynamic targets.
domain assumption The training and test grid maps from dataset [26] are representative of unknown environments beyond the dataset.
Section 5.1 uses a public dataset for training/testing; generalization to other terrain types is assumed in the conclusion.
domain assumption Agents have perfect, noiseless local perception and exact map merging within communication range.
Sections 3.3 and 4.4 assume observations and communicated maps are accurate; sensor noise and localization error are absent.
domain assumption MAPPO with shared parameters and CTDE converges to a policy that transfers to unseen test maps.
Section 4.7 relies on standard MARL convergence expectations without proofs or convergence diagnostics.
ad hoc to paper The environment grid's passable area is fully connected.
Section 3.1 restricts maps to fully connected passable regions, simplifying navigation and the A* fallback.
domain assumption Reward shaping (Eqs. 6-9) is aligned with the true objective of locating all targets.
Section 4.6 hand-designs rewards that encourage exploration; if this shaping conflicts with the task, the learned policy's success is not a fair measure.

invented entities (1)

Virtual pheromone field
purpose: A scalar concentration per grid cell that accumulates when agents visit and evaporates over time, used to discourage re-exploration and enable indirect coordination.
It is an algorithmic construct added to the environment state with no physical counterpart; its contribution is only validated through the end-to-end system performance, not independently.

how reviews work

0 comments

Cite this review

Pith. "Pith review of PILOC: A Pheromone Inverse Guidance Mechanism and Local-Communication Framework for Dynamic Target Search of Multi-Agent in Unknown Environments." pith.science (2026). https://pith.science/paper/ZUCHUY66

@misc{pith2026250707376,
  author       = {Pith},
  title        = {Pith review of: PILOC: A Pheromone Inverse Guidance Mechanism and Local-Communication Framework for Dynamic Target Search of Multi-Agent in Unknown Environments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZUCHUY66}},
  note         = {Machine review of arXiv:2507.07376}
}

read the original abstract

Multi-Agent Search and Rescue (MASAR) plays a vital role in disaster response, exploration, and reconnaissance. However, dynamic and unknown environments pose significant challenges due to target unpredictability and environmental uncertainty. To tackle these issues, we propose PILOC, a framework that operates without global prior knowledge, leveraging local perception and communication. It introduces a pheromone inverse guidance mechanism to enable efficient coordination and dynamic target localization. PILOC promotes decentralized cooperation through local communication, significantly reducing reliance on global channels. Unlike conventional heuristics, the pheromone mechanism is embedded into the observation space of Deep Reinforcement Learning (DRL), supporting indirect agent coordination based on environmental cues. We further integrate this strategy into a DRL-based multi-agent architecture and conduct extensive experiments. Results show that combining local communication with pheromone-based guidance significantly boosts search efficiency, adaptability, and system robustness. Compared to existing methods, PILOC performs better under dynamic and communication-constrained scenarios, offering promising directions for future MASAR applications.

Figures

Figures reproduced from arXiv: 2507.07376 by the authors.

**Figure 1.** Search and rescue scenario map 3.2 Target model We consider a dynamic target model with randomly moving search and rescue targets. The SAR scenario consists of multiple targets, denoted as G, which are distributed across the passable area. The initial position of each target gi ∈ G is randomly generated, and at each time step t, target gi randomly selects an action agi from the action space Ag. The probability of se… view at source ↗

**Figure 2.** Agent observation inputs: (a) The obstacle map [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Schematic diagram of agent communication: the left side shows the exploration map update method, and the [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: (a) Overall architecture of the network, consisting of four main components: the convolutional module, the [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: The PILOC framework proposed in this paper, which combines local communication and pheromone mechanism, [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 6.** Figure 6: Results of the ablation experiment: shows the comparison results of task success rates (red bars) and average [PITH_FULL_IMAGE:figures/full_fig_p011_6.png]

**Figure 7.** Figure 7: Effect of the number of agents on Success Rate and Average Steps in scalability experiments [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 26 canonical work pages

[1]

A review of the applications of multi-agent reinforcement learning in smart factories

Fouad Bahrpeyma and Dirk Reichelt. A review of the applications of multi-agent reinforcement learning in smart factories. Frontiers in Robotics and AI, 9:1027340, 2022

work page 2022
[2]

A multi-agent reinforcement learning framework for intelligent manufacturing with autonomous mobile robots

Akash Agrawal, Sung Jun Won, Tushar Sharma, Mayuri Deshpande, and Christopher McComb. A multi-agent reinforcement learning framework for intelligent manufacturing with autonomous mobile robots. Proceedings of the Design Society, 1:161–170, 2021

work page 2021
[3]

A multi-agent reinforcement learning method for swarm robots in space collaborative exploration

Yixin Huang, Shufan Wu, Zhongcheng Mu, Xiangyu Long, Sunhao Chu, and Guohong Zhao. A multi-agent reinforcement learning method for swarm robots in space collaborative exploration. In 2020 6th international conference on control, automation and robotics (ICCAR) , pages 139–144. IEEE, 2020

work page 2020
[4]

Formation control for multi-unmanned vehicles via deep reinforcement learning

Zeng Yuling, Hao Yuqing, Yu Ying, and Wang Qingyun. Formation control for multi-unmanned vehicles via deep reinforcement learning. Chinese Journal of Theoretical and Applied Mechanics , 56(2):460–471, 2023

work page 2023
[5]

Reinforcement learning for multi-agent formation navigation with scalability

Yalei Gong, Hongyun Xiong, MengMeng Li, Haibo Wang, and Xiaohong Nian. Reinforcement learning for multi-agent formation navigation with scalability. Applied Intelligence, 53(23):28207–28225, 2023

work page 2023
[6]

Heterogeneous Multi-Agent Reinforcement Learning for Unknown Environment Mapping

Ceyer Wakilpoor, Patrick J Martin, Carrie Rebhuhn, and Amanda Vu. Heterogeneous multi-agent reinforcement learning for unknown environment mapping. arXiv preprint arXiv:2010.02663, 2020

work page Pith review arXiv 2010
[7]

H2gnn: Hierarchical-hops graph neural networks for multi-robot exploration in unknown environments

Hao Zhang, Jiyu Cheng, Lin Zhang, Yibin Li, and Wei Zhang. H2gnn: Hierarchical-hops graph neural networks for multi-robot exploration in unknown environments. IEEE Robotics and Automation Letters , 7(2):3435–3442, 2022

work page 2022
[8]

Deep reinforcement learning for decentralized multi-robot exploration with macro actions

Aaron Hao Tan, Federico Pizarro Bejarano, Yuhan Zhu, Richard Ren, and Goldie Nejat. Deep reinforcement learning for decentralized multi-robot exploration with macro actions. IEEE Robotics and Automation Letters , 8(1):272–279, 2022

work page 2022

Show all 29 references

[9]

Deep reinforcement learning robot for search and rescue applications: Exploration in unknown cluttered environments

Farzad Niroui, Kaicheng Zhang, Zendai Kashino, and Goldie Nejat. Deep reinforcement learning robot for search and rescue applications: Exploration in unknown cluttered environments. IEEE Robotics and Automation Letters , 4(2):610–617, 2019

2019
[10]

Hierarchical integration of model predictive and fuzzy logic control for combined coverage and target-oriented search-and-rescue via robots with imperfect sensors

Christopher de Koning and Anahita Jamshidnejad. Hierarchical integration of model predictive and fuzzy logic control for combined coverage and target-oriented search-and-rescue via robots with imperfect sensors. Journal of Intelligent & Robotic Systems , 107(3):40, 2023

2023
[11]

Adaptive optimal receding-horizon robot navigation via short-term policy development

Anahita Jamshidnejad and Emilio Frazzoli. Adaptive optimal receding-horizon robot navigation via short-term policy development. In 2018 15th International Conference on Control, Automation, Robotics and Vision (ICARCV), pages 21–28. IEEE, 2018

2018
[12]

A complete methodology for generating multi-robot task solutions using asymtre-d and market-based task allocation

Fang Tang and Lynne E Parker. A complete methodology for generating multi-robot task solutions using asymtre-d and market-based task allocation. In Proceedings 2007 IEEE international conference on robotics and automation , pages 3351–3358. IEEE, 2007

2007
[13]

Frontier-based exploration using multiple robots

Brian Yamauchi. Frontier-based exploration using multiple robots. In Proceedings of the second international conference on Autonomous agents, pages 47–53, 1998

1998
[14]

A survey on coverage path planning for robotics

Enric Galceran and Marc Carreras. A survey on coverage path planning for robotics. Robotics and Autonomous systems, 61(12):1258–1276, 2013

2013
[15]

Multi-agent coordination across diverse applications: A survey

Lijun Sun, Yijun Yang, Qiqi Duan, Yuhui Shi, Chao Lyu, Yu-Cheng Chang, Chin-Teng Lin, and Yang Shen. Multi-agent coordination across diverse applications: A survey. arXiv preprint arXiv:2502.14743, 2025

2025 arXiv
[16]

A Modern Approach

Norvig Peter and Russell Stuart Artificial Intelligence. A Modern Approach. Pearson Education, USA, 2021

2021
[17]

Scrimp: Scalable communication for reinforcement-and imitation-learning-based multi-agent pathfinding

Yutong Wang, Bairan Xiang, Shinan Huang, and Guillaume Sartoretti. Scrimp: Scalable communication for reinforcement-and imitation-learning-based multi-agent pathfinding. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , pages 9301–9308. IEEE, 2023

2023
[18]

Opportunistic multi-robot environmental sampling via decentralized markov decision processes

Ayan Dutta, O Patrick Kreidl, and Jason M O’Kane. Opportunistic multi-robot environmental sampling via decentralized markov decision processes. In International Symposium Distributed Autonomous Robotic Systems , pages 163–175. Springer, 2021

2021
[19]

Multirobot systems: a classification focused on coordination

Alessandro Farinelli, Luca Iocchi, and Daniele Nardi. Multirobot systems: a classification focused on coordination. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics) , 34(5):2015–2028, 2004. 13 A PREPRINT - SEPTEMBER 8, 2025

2015
[20]

A critical review of communications in multi-robot systems

Jennifer Gielis, Ajay Shankar, and Amanda Prorok. A critical review of communications in multi-robot systems. Current robotics reports, 3(4):213–225, 2022

2022
[21]

Challenge: Unlicensed lpwans are not yet the path to ubiquitous connectivity

Branden Ghena, Joshua Adkins, Longfei Shangguan, Kyle Jamieson, Philip Levis, and Prabal Dutta. Challenge: Unlicensed lpwans are not yet the path to ubiquitous connectivity. In The 25th Annual International Conference on Mobile Computing and Networking , pages 1–12, 2019

2019
[22]

Multi-agent systems for search and rescue applications

Daniel S Drew. Multi-agent systems for search and rescue applications. Current Robotics Reports, 2:189–200, 2021

2021
[23]

Victor Lesser

Chongjie Zhang. Victor Lesser. Coordinating multi-agent reinforcement learning with limited communication. In Proceedings of the 2013 International Conference on Autonomous Agents and Multi-Agent Systems , pages 1101–1108, 2013

2013
[24]

Low-bandwidth communication emerges naturally in multi-agent learning systems

Niko A Grupen, Daniel D Lee, and Bart Selman. Low-bandwidth communication emerges naturally in multi-agent learning systems. arXiv preprint arXiv:2011.14890, 2020

2011 arXiv
[25]

Hma-sar: Multi-agent search and rescue for unknown located dynamic targets in completely unknown environments

Xiao Cao, Mingyang Li, Yuting Tao, and Peng Lu. Hma-sar: Multi-agent search and rescue for unknown located dynamic targets in completely unknown environments. IEEE Robotics and Automation Letters , 2024

2024
[26]

Self-learning exploration and mapping for mobile robots via deep reinforcement learning

Fanfei Chen, Shi Bai, Tixiao Shan, and Brendan Englot. Self-learning exploration and mapping for mobile robots via deep reinforcement learning. In Aiaa scitech 2019 forum , page 0396, 2019

2019
[27]

Multi-agent actor-critic for mixed cooperative-competitive environments

Ryan Lowe, Yi I Wu, Aviv Tamar, Jean Harb, OpenAI Pieter Abbeel, and Igor Mordatch. Multi-agent actor-critic for mixed cooperative-competitive environments. Advances in neural information processing systems , 30, 2017

2017
[28]

Is independent learning all you need in the starcraft multi-agent challenge? arXiv preprint arXiv:2011.09533, 2020

Christian Schroeder De Witt, Tarun Gupta, Denys Makoviichuk, Viktor Makoviychuk, Philip HS Torr, Mingfei Sun, and Shimon Whiteson. Is independent learning all you need in the starcraft multi-agent challenge? arXiv preprint arXiv:2011.09533, 2020

2011 arXiv
[29]

Monotonic value function factorisation for deep multi-agent reinforcement learning

Tabish Rashid, Mikayel Samvelyan, Christian Schroeder De Witt, Gregory Farquhar, Jakob Foerster, and Shimon Whiteson. Monotonic value function factorisation for deep multi-agent reinforcement learning. Journal of Machine Learning Research, 21(178):1–51, 2020. 14

2020

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

A review of the applications of multi-agent reinforcement learning in smart factories

Fouad Bahrpeyma and Dirk Reichelt. A review of the applications of multi-agent reinforcement learning in smart factories. Frontiers in Robotics and AI, 9:1027340, 2022

work page 2022

[2] [2]

A multi-agent reinforcement learning framework for intelligent manufacturing with autonomous mobile robots

Akash Agrawal, Sung Jun Won, Tushar Sharma, Mayuri Deshpande, and Christopher McComb. A multi-agent reinforcement learning framework for intelligent manufacturing with autonomous mobile robots. Proceedings of the Design Society, 1:161–170, 2021

work page 2021

[3] [3]

A multi-agent reinforcement learning method for swarm robots in space collaborative exploration

Yixin Huang, Shufan Wu, Zhongcheng Mu, Xiangyu Long, Sunhao Chu, and Guohong Zhao. A multi-agent reinforcement learning method for swarm robots in space collaborative exploration. In 2020 6th international conference on control, automation and robotics (ICCAR) , pages 139–144. IEEE, 2020

work page 2020

[4] [4]

Formation control for multi-unmanned vehicles via deep reinforcement learning

Zeng Yuling, Hao Yuqing, Yu Ying, and Wang Qingyun. Formation control for multi-unmanned vehicles via deep reinforcement learning. Chinese Journal of Theoretical and Applied Mechanics , 56(2):460–471, 2023

work page 2023

[5] [5]

Reinforcement learning for multi-agent formation navigation with scalability

Yalei Gong, Hongyun Xiong, MengMeng Li, Haibo Wang, and Xiaohong Nian. Reinforcement learning for multi-agent formation navigation with scalability. Applied Intelligence, 53(23):28207–28225, 2023

work page 2023

[6] [6]

Heterogeneous Multi-Agent Reinforcement Learning for Unknown Environment Mapping

Ceyer Wakilpoor, Patrick J Martin, Carrie Rebhuhn, and Amanda Vu. Heterogeneous multi-agent reinforcement learning for unknown environment mapping. arXiv preprint arXiv:2010.02663, 2020

work page Pith review arXiv 2010

[7] [7]

H2gnn: Hierarchical-hops graph neural networks for multi-robot exploration in unknown environments

Hao Zhang, Jiyu Cheng, Lin Zhang, Yibin Li, and Wei Zhang. H2gnn: Hierarchical-hops graph neural networks for multi-robot exploration in unknown environments. IEEE Robotics and Automation Letters , 7(2):3435–3442, 2022

work page 2022

[8] [8]

Deep reinforcement learning for decentralized multi-robot exploration with macro actions

Aaron Hao Tan, Federico Pizarro Bejarano, Yuhan Zhu, Richard Ren, and Goldie Nejat. Deep reinforcement learning for decentralized multi-robot exploration with macro actions. IEEE Robotics and Automation Letters , 8(1):272–279, 2022

work page 2022