SyncSBC: Decentralized Swarm Behavior Prediction for Synchronized Autonomous Control

Connor Mattson; Daniel S. Brown; Varun Raveendra

REVIEW 3 major objections 5 minor 38 references

SyncSBC: Decentralized Swarm Behavior Prediction for Synchronized Autonomous Control

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A swarm of simple robots can infer its own collective behavior from each robot's local sensor history alone, and then synchronize its decisions tightly enough to switch behaviors and flag anomalies without any central controller.

desk verdict Genuinely new decentralized swarm behavior recognition plus synchronization, but the load-bearing transfer of the learned classifier to other swarm sizes and real robots is not validated by accuracy numbers. read the letter →

arxiv 2608.06587 v1 pith:VJWUAQX7 submitted 2026-08-06 cs.RO cs.AI

Varun Raveendra , Connor Mattson , Daniel S. Brown This is my paper · ORCID

classification cs.ROcs.AI

keywords swarmroboticsdecentralizedbehaviorclassificationdistributedconsensussynchronizeddecision-makinganomalydetectionevent-triggeredcommunicationtemporalconvolutionalnetworkswitching

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 a swarm of simple robots can infer the swarm's collective behavior from only their own local sensor history, without a central observer, and can then align their decisions tightly in time using a lightweight distributed consensus layer. It introduces SyncSBC, which pairs a learned Swarm Behavior Classifier (SBC) with two synchronization mechanisms, Boolean Gossip and Integrated Belief Synchronizer, plus event-triggered communication. The authors report that the SBC classifies three emergent behaviors with greater than 95% accuracy on held-out data, that synchronization delay drops below one second in simulation, and that real 8-robot experiments achieve sub-3-second synchronization, well under agreement-only baselines. If true, this would make fault detection, coordinated behavior switching, and self-healing possible on communication-constrained swarms without centralized control.

What carries the argument

The machinery is the two-stage SyncSBC pipeline. Stage one is a temporal convolutional network (TCN), trained with cross-entropy on windowed time-of-flight readings from individual robots, that outputs per-agent class probabilities; predictions are smoothed with an exponential moving average. Stage two is a consensus-and-synchronization layer: each robot maintains an internal binary belief over behavior classes using a Sample-and-Hold update, then exchanges beliefs with neighbors within communication radius using event-triggered communication. Boolean Gossip propagates internal beliefs through logical AND operations with Lamport-style sequence counters, while the Integrated Belief Synchronizer piggybacks a token-owner, epoch, and membership set on behavior fields to merge states and prevent split decisions. The stage-two layer is what turns independent classifications into temporally aligned swarm decisions.

What would settle it

Run the trained SBC on held-out test windows from 5- and 16-robot swarms and from real robots, and compare per-agent classification accuracy to the reported >95% figure; if accuracy drops substantially, for example below 80%, in any setting, the claim that SyncSBC scales without retraining is falsified. Alternatively, measure synchronization delay under a disconnected or intermittently partitioned communication graph; if delay grows linearly or consensus never forms, the synchronization claim is bounded by connectivity.

Watch

Extended reading notes

Core claim

The central claim is that collective swarm behavior is recoverable from egocentric time-of-flight sensor streams: a single shared neural network, trained once on labeled episodes of stable swarm behaviors, lets each robot predict the global behavior class from its own recent observations. The paper then claims that robots can convert these noisy local predictions into a synchronized swarm-wide decision by first forming an internal belief through local consensus, where Sample-and-Hold outperforms averaging, entropy fusion, and neighbor variance, and then aligning internal beliefs across the swarm with Boolean Gossip or the Integrated Belief Synchronizer, both driven by event-triggered communication. The result is a fully decentralized pipeline in which the swarm's first synchronized belief acts as a trigger for collective action and as a signal for detecting anomalous agents.

Load-bearing premise

The load-bearing premise is that a behavior classifier trained on eight simulated robots' single time-of-flight sensor streams produces reliable local predictions when transferred to 5- and 16-robot swarms and to real HeRo+ robots without retraining or recalibration; if local sensor signals are not informative under those conditions, the entire pipeline degrades.

Editorial extensions

If this is right

A swarm can switch between behavior controllers at nearly the same moment across all agents without any central coordinator, since each robot's synchronized belief is the switch trigger.
Anomalous agents can be detected from within the swarm: a robot whose internal belief repeatedly disagrees with the majority membership set can be flagged as misaligned, enabling future isolation or self-healing.
Event-triggered communication makes synchronization cheap; the Integrated Belief Synchronizer reaches consensus with far fewer messages than periodic gossip, which matters under fixed communication budgets.
Increasing swarm density and communication radius reduces synchronization delay, suggesting SyncSBC improves with connectivity rather than requiring a central node.

Reading between the lines

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

The paper reports synchronization delay and anomaly outcomes for the transferred model on 5-, 8-, and 16-robot swarms and on real hardware, but not the transferred classifier's per-agent accuracy in those settings; measuring that accuracy directly would separate classifier-transfer failures from consensus failures.
Because the classifier is trained on a single time-of-flight sensor, applying SyncSBC to other modalities such as LiDAR, cameras, or contact sensing would likely require retraining or domain adaptation; the consensus layer itself is sensor-agnostic.
The synchronization delay appears controlled primarily by graph connectivity and message rate, which suggests that communication radii or event-trigger thresholds could be chosen a priori to meet a required synchronization deadline rather than tuned per scenario.

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

The real contribution is the system integration: a learned local-to-global behavior classifier feeding into event-triggered consensus and two synchronization mechanisms that align agent decisions in time. The individual pieces are known, but the combination is new and the experiments are set up to show it works on real HeRo+ robots. I buy that the pipeline runs, and the code and videos are a plus. The TCN architecture choice is well motivated, and the comparison across six model families is fair.

The soft spot is exactly where the stress-test lands. The SBC is trained on 8 simulated robots with a single ToF sensor, then deployed on 5/16-robot swarms and on real robots without retraining or recalibration. The paper reports synchronization delay and message counts in those transferred settings, but never reports SBC classification accuracy, confusion matrices, or calibration. That matters because the synchronization stage only aligns whatever beliefs the SBC produces. Low delay alone does not tell you the swarm is synchronized on the correct behavior; it could be synchronized on a confident error. The real-robot behavior switching and anomaly detection figures show the mechanism fires, but they do not establish that the local ToF stream remains informative after the domain shift. This is a correctness risk, not a scope limitation, because the model is explicitly deployed without adaptation.

Two smaller issues. First, the ground-truth convergence time is a human-calibrated threshold on swarm-level metrics, and the consensus strategy parameters are tuned with Optuna to minimize delay relative to that proxy. So the headline delay numbers are tuned rather than predicted; that is fine for a system paper, but it should be stated more plainly. Second, the anomaly detection is defined as disagreement with the majority belief. That makes the anomaly experiment somewhat circular: the mechanism is designed to detect misalignment, so showing it detects misalignment is expected. It is still useful evidence for a self-healing loop, but the claim of inherent ability to detect faulty robots goes beyond what is demonstrated.

Who is this for? Swarm robotics people working on decentralized decision-making and communication-constrained coordination. They will get value from the consensus-and-synchronization design even if the perceptual transfer story is incomplete. It deserves a serious referee, not a desk reject, but the referee should demand transfer accuracy numbers and a less circular anomaly evaluation before publication. I would not reject the central idea; I would ask for the missing evidence.

Referee Report

3 major / 5 minor

Summary. The paper introduces SyncSBC, a two-stage decentralized pipeline in which each robot classifies the swarm's collective behavior from its own local time-of-flight sensor stream using a trained temporal convolutional network (the SBC model), and then a consensus layer (Sample-and-Hold internal beliefs combined with Boolean Gossip or an Integrated Belief Synchronizer, plus event-triggered communication) aligns agents' beliefs so that the swarm can execute synchronized actions. The authors evaluate the approach in Isaac Sim across 5, 8, and 16 robots and on a real 8-robot HeRo+ swarm, reporting >95% held-out classification accuracy on the training domain, sub-second average synchronization delay for the ESB/ESI variants in simulation, and sub-3-second delay on real robots, with demonstrations of autonomous behavior switching and anomaly detection.

Significance. If the central claim holds, the paper is a useful contribution: it provides evidence that a single egocentric ToF sensor can support decentralized collective-behavior classification, and it shows that lightweight synchronization mechanisms can substantially reduce the temporal misalignment of decisions compared with agreement-only consensus. The paper's strengths include a systematic comparison of six neural architectures with calibration metrics, a modular treatment of consensus strategies, public code and videos, and deployment on real robots rather than simulation only. The main limitation is that the classifier's accuracy is not reported after transfer to different swarm sizes and to real hardware, so the evidence that synchronized beliefs correspond to the true collective behavior is incomplete.

major comments (3)

[Section VI-D and Section VII] The SBC model is trained exclusively on 8-robot simulated ToF streams (Section VI-A), but Sections VI-D and VII report synchronization delay and message counts for 5/8/16-robot swarms and for real robots without reporting SBC classification accuracy, confusion matrices, or calibration in those transferred settings. Because the synchronization stage aligns whatever beliefs the SBC produces, low synchronization delay is compatible with the swarm synchronizing on an incorrect behavior. To support the paper's central claim that agents map local observations into temporally aligned conclusions of the collective behavior, please report transfer accuracy (overall and per class) for the 5- and 16-robot simulations and for the real-robot runs, or otherwise demonstrate that the synchronized beliefs coincide with ground truth (e.g., correct-switch rates and anomaly precision/recall). Without this, the delay numbers alone do not establish correct behavior recognition after the domain shift.
[Section VI-C] The synchronization-delay results depend on ground-truth convergence times obtained from thresholds on scatter, radial variance, and average speed that are calibrated by a human observer on a subset of controller runs. The paper reports no information about the threshold values, the number of human-annotated runs, inter-annotator agreement, or sensitivity of the reported delays to threshold perturbations. Since the headline claims (<1s in simulation, <3s on real robots) are measured relative to this ground truth, a sensitivity analysis is needed to establish that the claimed delay advantage is not an artifact of the specific calibration.
[Section VII-B] The anomaly-detection mechanism defines anomaly likelihood as disagreement with the majority synchronized belief, so the deliberately modified robots in Scenarios 1 and 2 are anomalous essentially by construction. Figure 10 shows that the injected robots' likelihoods rise over time, but the paper does not report detection metrics such as precision, recall, false-positive rate, or AUC over runs that include normal, non-anomalous robots. To substantiate the claim that SyncSBC 'accurately identify[ies] anomalies,' please provide quantitative detection performance, including evidence that healthy robots are not flagged during normal operation.

minor comments (5)

[Section II] The citation for asymmetric Best-of-M problems, reference [2], appears to be a fuzzy-aggregation paper rather than a best-of-n study; please verify that this reference supports the statement or replace it with an appropriate citation.
[Section VI-B] Figure 4 reports the architecture comparison but provides no numerical values; including a table with accuracy, macro F1, Brier score, and ECE for all six models would improve reproducibility and make the claimed TCN advantage easier to assess.
[Section VII-A] In the table inside Figure 9, the entry 'Pure SH 18.0001 s 2 6.0000 s' appears to contain a typographical error (likely '26.0000 s'); please correct it.
[Section V-B] The condition for setting a synchronized belief in the Integrated Belief Synchronizer ('once all robots in the neighborhood agree and no active token holders remain') is informal; please specify the formal predicate on the fields F_k and the synchronized variable s_i.
[Section III] The definition of synchronization delay Delta assumes that every agent reaches a finite t*_i; please state how episodes are handled if some agents never synchronize within the evaluation horizon.

Circularity Check

1 steps flagged · score 4.0 of 10

Anomaly detection reduces to the update rule by construction; the SBC training and synchronization core remain independent.

self definitional [Section VII-B (Anomaly detection), anomaly-likelihood definition]
"Anomaly likelihood increases when robot i strongly rejects behavior k (b_i,k = 0 and arg max_k P_i,k != k) while others propagate s_i,k = 1 in ESB. Similarly, under ESI, the anomaly likelihood increases when the robot's internal belief b_i repeatedly disagrees with the largest membership set M_k."

The anomaly score is defined, not learned or independently validated, as disagreement between a robot's internal belief and the synchronized majority belief (s_i,k=1 in ESB, largest M_k in ESI). The two reported scenarios are constructed by halting robot #7 or altering the controllers of robots #7 and #8, so those robots are by construction the ones whose beliefs differ from the rest of the swarm. Plotting their rising anomaly likelihood therefore re-derives the update rule: if a robot's belief disagrees with the majority, the rule labels that robot anomalous. The claim that SyncSBC can 'accurately identify anomalies' is thus a restatement of the definition of the anomaly likelihood, with no independent anomaly ground truth, false-positive analysis, or normal-robot baseline provided.

full rationale

The core learning and synchronization chain is not circular. The SBC is trained by supervised cross-entropy on per-robot ToF windows with episode-level labels, and evaluation uses disjoint 70/15/15 train/validation/test segments before window extraction, so the reported >95% held-out accuracy is a standard held-out evaluation. The synchronization delay is measured against a separately constructed convergence ground truth, with synchronized belief defined as s_i(t)=y*, so delay is only credited when the swarm belief is correct. The Stage-1 and Stage-2 consensus comparisons tune each method fairly and do not assume the claimed outcome. Scaling to 5/8/16 robots and real HeRo+ deployment is an empirical transfer test; the absence of reported SBC accuracy in those settings is a correctness risk, not circularity. The one genuine reduction is in the anomaly-detection application: the anomaly likelihood is defined as disagreement with the synchronized majority, and the experiments use exactly the robots whose controllers were stopped or modified, so high anomaly likelihood for those robots follows by construction. The central claims of decentralized classification and synchronized consensus still stand on independent supervised learning and consensus evaluation.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The main things the paper pulls in are standard ML and consensus machinery. The free parameters are the trained classifier weights plus several threshold and tuning hyperparameters. The most fragile input is the human-calibrated ground-truth convergence time, which all delay numbers depend on, and the implicit assumption that local ToF readings reveal global behavior across swarm sizes and real hardware.

free parameters (7)

SBC neural network weights (TCN) = unknown, trained on simulated ToF windows
Learned via cross-entropy on 8-robot Isaac Sim data; the central behavior classifier in the pipeline.
EWMA smoothing coefficient alpha = not reported
Smooths per-agent class probabilities in Section IV; chosen by the authors and affects prediction stability and delay.
Event-triggered communication threshold sigma = not reported
Hyperparameter in the ETC rule in Section V-C; controls when messages are transmitted.
Sample-and-Hold disagreement threshold = tuned with Optuna
In Stage-1, determines when an agent locks its internal belief; tuned to minimize delay while penalizing early convergence.
Consensus strategy hyperparameters (Averaging, Entropy, Neighbor Variance) = tuned with Optuna
Each Stage-1 strategy is tuned with Optuna on the delay metric used in the evaluation.
Ground-truth convergence thresholds (scatter, radial variance, average speed) = calibrated by human annotations on a subset
Used in Section VI-C to label behavior convergence time; all delay measurements are relative to this proxy.
Observation window length T and stride = not reported
Chosen without an ablation; affects the accuracy versus latency trade-off of the classifier.

assumptions (5)

domain assumption The time-varying communication graph G(t) remains connected.
Stated in Section III; needed for beliefs to propagate across the entire swarm, but dispersal behavior may push robots apart and threaten this.
domain assumption Single time-of-flight sensor streams contain enough information to distinguish Cyclic Pursuit, Aggregation, and Dispersal across robot counts.
Assumed in Sections IV and VI; validated only indirectly through classifier accuracy in the 8-robot simulation.
domain assumption Each controller rollout has a well-defined stable portion with a consistent episode label.
The data collection in Section VI-A labels windows only within stable portions and assumes these labels describe collective behavior.
domain assumption Human annotations of visual stabilization are accurate enough to calibrate convergence thresholds.
Section VI-C uses these thresholds as the ground truth baseline for all synchronization delay measurements.
standard math Standard supervised learning assumptions hold for the classifier.
Cross-entropy training and disjoint train-validation-test splits before window extraction are assumed to prevent leakage; the paper states this split procedure.

how reviews work

0 comments

Cite this review

Pith. "Pith review of SyncSBC: Decentralized Swarm Behavior Prediction for Synchronized Autonomous Control." pith.science (2026). https://pith.science/paper/VJWUAQX7

@misc{pith2026260806587,
  author       = {Pith},
  title        = {Pith review of: SyncSBC: Decentralized Swarm Behavior Prediction for Synchronized Autonomous Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VJWUAQX7}},
  note         = {Machine review of arXiv:2608.06587}
}

read the original abstract

Robot swarms utilize many independent limited-sensing agents to produce complex emergent behaviors without requiring centralized control. However, little research explores how agents can infer swarm-level behavior from purely local perception, a capability critical for detecting faults and behavior changes. In this paper, we introduce Synchronized Swarm Behavior Classification (SyncSBC), which combines improvements in machine learning and distributed consensus to classify collective swarm behavior and synchronize swarm decision-making in an entirely decentralized manner. We show that SyncSBC achieves high classification accuracy and low synchronization delay, making it suitable for real-world deployment. Finally, we use SyncSBC to demonstrate two promising swarm applications on real robots where we show that swarms utilizing SyncSBC can accurately identify anomalies in robot behavior and autonomously coordinate collective changes in swarm behavior. Videos, code and supplemental experiments are available at https://sites.google.com/view/sync-sbc/home.

Figures

Figures reproduced from arXiv: 2608.06587 by the authors.

**Figure 1.** Synchronous Swarm Behavior Classification (SyncSBC) is a dual-stage fully-decentralized framework for transforming agent-level behavioral predictions into synchronized action execution using limited-range communication and consensus. (Left) A trained deep neural network predicts the behavioral properties of the swarm using only egocentric local observations. (Right) A consensus algorithm measures neighborhood agreem… view at source ↗

**Figure 2.** Synchronization and Communication in SyncSBC. We study two consensus methods for use in our framework based on system design preferences: (Left) when minimal synchronization delay is desired, we propose using event-triggered communication and boolean gossip and (Right) when reducing communication messages is preferred, then a combination of event-triggered communication and integrated belief synchronizer is optimal.… view at source ↗

**Figure 3.** HeRo+ Robots: (a) A single HeRo+ robot uses unicycle commands to locomote and time-of-flight sensing to detect other robots. (b) 3 emergent behaviors deployed on 8 robots. From Left to Right: Aggregation, where agents aggregate at the swarm’s centroid, Cyclic Pursuit, where agents travel about the centroid in a circular motion, and Dispersal, where agents maximize the inter-agent distance to span (explore) the envir… view at source ↗

Figures from the paper (4 more)

**Figure 6.** Figure 6: Synchronization Error Results (Stage-2): We compare 4 synchronization methods and measure the delay between the first robot forming its synchronized belief and the last robot. Across all methods, we find that combining Event-triggered, Sample-and-Hold with Boolean Goss…

**Figure 7.** Figure 7: Message Count Results: For constrained deployments with a fixed communication budget, we examine the number of total messages transmitted under each of the different synchronization methods. Across all methods, we find that combining Sample-and-Hold with Integrated Bel…

**Figure 8.** Figure 8: Robots vs Radius: We examined SyncSBC on swarms with varying number of robots and communication radii. We find that ESB and ESI achieve the lowest synchronization delay, especially in domains with many robots. While ESB has a higher communication rate, ESI performs wel…

**Figure 10.** Figure 10: Application: Anomaly Detection. (left) SyncSBC provides a mechanism where robots can self-inform themselves of incorrect behavior, i.e., anomaly detection. We studied two anomaly detection scenarios where 1-2 robots’ behaviors have been deliberately halted or modified…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 37 canonical work pages

[1]

Arnold, Hiroyuki Yamaguchi, and Toshiyuki Tanaka

Ross D. Arnold, Hiroyuki Yamaguchi, and Toshiyuki Tanaka. Search and rescue with autonomous flying robots through behavior-based cooperative intelligence.Journal of International Humanitarian Action 2018

work page 2018
[2]

Fuzzy-based aggregation with a mobile robot swarm

Farshad Arvin, Ali Turgut, and Shigang Yue. Fuzzy-based aggregation with a mobile robot swarm. pages 346–347, 09 2012

work page 2012
[3]

Classifying swarm behavior via compressive subspace learning

Matthew Berger, Lee M Seversky, and Daniel S Brown. Classifying swarm behavior via compressive subspace learning. In2016 IEEE International Conference on Robotics and Automation (ICRA)

work page
[4]

Reliable communication in the presence of failures.ACM Trans

Kenneth P Birman and Thomas A Joseph. Reliable communication in the presence of failures.ACM Trans. on Computer Systems, 1987

work page 1987
[5]

Managing a Mobile Agricultural Robot Swarm for a seeding task

Timo Blender, Thiemo Buchner, Benjamin Fernandez, Benno Pichlmaier, and Christian Schlegel. Managing a Mobile Agricultural Robot Swarm for a seeding task. InIECON 2016

work page 2016
[6]

Limited bandwidth recognition of collective behaviors in bio-inspired swarms

Daniel S Brown and Michael A Goodrich. Limited bandwidth recognition of collective behaviors in bio-inspired swarms. InAAMAS, 2014

work page 2014
[7]

Discovery and exploration of novel swarm behaviors given limited robot capabilities

Daniel S Brown, Ryan Turner, Oliver Hennigh, and Steven Loscalzo. Discovery and exploration of novel swarm behaviors given limited robot capabilities. InDistributed Autonomous Robotic Systems: The 13th International Symposium, pages 447–460. Springer, 2018

work page 2018
[8]

Giulia De Masi, Judhi Prasetyo, Elio Tuci, and Eliseo Ferrante.Zealots Attack and the Revenge of the Commons: Quality vs Quantity in the Best-of-n. 10 2020

work page 2020

Show all 38 references

[9]

Dimarogonas, Emilio Frazzoli, and Karl H

Dimos V . Dimarogonas, Emilio Frazzoli, and Karl H. Johansson. Distributed event-triggered control for multi-agent systems.IEEE Transactions on Automatic Control, 57(5):1291–1297, 2012

2012
[10]

Evo- lution of Collective Behaviors for a Real Swarm of Aquatic Surface Robots.PLOS ONE, 11(3):e0151834, March 2016

Miguel Duarte, Vasco Costa, Jorge Gomes, Tiago Rodrigues, Fernando Silva, Sancho Moura Oliveira, and Anders Lyhne Christensen. Evo- lution of Collective Behaviors for a Real Swarm of Aquatic Surface Robots.PLOS ONE, 11(3):e0151834, March 2016

2016
[11]

Weinberger

Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q. Weinberger. On calibration of modern neural networks. InICML, 2017

2017
[12]

A multi-dimensional failure factor affecting the on- site adoption of construction robots.Developments in the Built Environment, 25:100841, 2026

Hakpyeong Kim, Juwon Hong, Eunseong Song, Taehoon Hong, and Jun-Ki Choi. A multi-dimensional failure factor affecting the on- site adoption of construction robots.Developments in the Built Environment, 25:100841, 2026

2026
[13]

Time, clocks, and the ordering of events in a distributed system.Communications of the ACM, 1978

Leslie Lamport. Time, clocks, and the ordering of events in a distributed system.Communications of the ACM, 1978

1978
[14]

Distributed consensus with limited communication data rate.IEEE Transactions on Automatic Control, 56(2):279–292, 2011

Tao Li, Minyue Fu, Lihua Xie, and Ji-Feng Zhang. Distributed consensus with limited communication data rate.IEEE Transactions on Automatic Control, 56(2):279–292, 2011

2011
[15]

Probabilistic consensus decision making algorithm for artificial swarm of primitive robots.SN Applied Sciences

Yang Liu and Kiju Lee. Probabilistic consensus decision making algorithm for artificial swarm of primitive robots.SN Applied Sciences
[16]

A survey of deep learning for time series forecasting: Theories, datasets, and state-of-the-art techniques.Computers, Materials and Continua, 2025

Gaoyong Lu, Yang Ou, Zhihong Wang, Yingnan Qu, Yingsheng Xia, Dibin Tang, Igor Kotenko, and Wei Li. A survey of deep learning for time series forecasting: Theories, datasets, and state-of-the-art techniques.Computers, Materials and Continua, 2025

2025
[17]

Preserving region consistency for wide-area distributed programs.IEEE Trans

R Mahy, R Verghese, and C Torgerson. Preserving region consistency for wide-area distributed programs.IEEE Trans. on Parallel and Distributed Systems, 2004

2004
[18]

Leveraging human feedback to evolve and discover novel emergent behaviors in robot swarms

Connor Mattson and Daniel S Brown. Leveraging human feedback to evolve and discover novel emergent behaviors in robot swarms. In Proceedings of the Genetic and Evolutionary Computation Confer- ence, pages 56–64, 2023

2023
[19]

Discovery and deployment of emergent robot swarm behaviors via representation learning and real2sim2real transfer

Connor Mattson, Varun Raveendra, Ricardo Vega, Cameron Nowzari, Daniel Drew, and Daniel S Brown. Discovery and deployment of emergent robot swarm behaviors via representation learning and real2sim2real transfer. InAAMAS. ACM, 2025

2025
[20]

Event-triggered communication and control of networked systems for multi-agent consensus.Automatica, 105:1–27, 2019

Cameron Nowzari, Eloy Garcia, and Jorge Cort ´es. Event-triggered communication and control of networked systems for multi-agent consensus.Automatica, 105:1–27, 2019

2019
[21]

Distributed kalman filtering for sensor networks

Reza Olfati-Saber. Distributed kalman filtering for sensor networks. InProceedings of the 46th IEEE Conference on Decision and Control
[22]

Alex Fax, and Richard M

Reza Olfati-Saber, J. Alex Fax, and Richard M. Murray. Consensus and cooperation in networked multi-agent systems.Proceedings of the IEEE, 95(1):215–233, 2007

2007
[23]

Op- tunaHub:a platform for black-box optimization.arXiv:2510.02798, 2025

Yoshihiko Ozaki, Shuhei Watanabe, and Toshihiko Yanase. Op- tunaHub:a platform for black-box optimization.arXiv:2510.02798, 2025

2025 arXiv
[24]

Shepherding with robots that do not compute

Anil ¨Ozdemir, Melvin Gauci, and Roderich Groß. Shepherding with robots that do not compute. InArtificial Life Conference Proceedings. MIT Press, 2017

2017
[25]

Performance analysis of fault tolerance algorithm for pattern formation of swarm agents

Shanu Rakesh and Manish Shrivastava. Performance analysis of fault tolerance algorithm for pattern formation of swarm agents. Knowledge-Based Systems, 240:108020, 12 2021

2021
[26]

Paulo Rezeck, H ´ector Azp ´urua, Maur ´ıcio F. S. Corr ˆea, and Luiz Chaimowicz. HeRo 2.0: a low-cost robot for swarm robotics research. Autonomous Robots, 47(7):879–903, October 2023

2023
[27]

Swarm robotic behaviors and current applications.Frontiers in Robotics and AI, V olume 7 - 2020, 2020

Melanie Schranz, Martina Umlauft, Micha Sende, and Wilfried Elmen- reich. Swarm robotic behaviors and current applications.Frontiers in Robotics and AI, V olume 7 - 2020, 2020

2020
[28]

Gossip algorithms.Foundations and Trends in Networking, 2009

Devavrat Shah. Gossip algorithms.Foundations and Trends in Networking, 2009

2009
[29]

Discrete collective estimation in swarm robotics with distributed bayesian belief sharing.Swarm Intelligence, 15, 12 2021

Qihao Shan and Sanaz Mostaghim. Discrete collective estimation in swarm robotics with distributed bayesian belief sharing.Swarm Intelligence, 15, 12 2021

2021
[30]

A distributed mutual exclusion algorithm

I Suzuki and T Kasami. A distributed mutual exclusion algorithm. ACM Trans. on Computer Systems, 1985

1985
[31]

An immune-inspired swarm aggregation algorithm for self-healing swarm robotic systems.Biosystems, 146:60–76, 2016

Jonathan Timmis, Amelia Ritahani Ismail, Jan Dyre Bjerknes, and Alan FT Winfield. An immune-inspired swarm aggregation algorithm for self-healing swarm robotic systems.Biosystems, 146:60–76, 2016

2016
[32]

The best-of-n problem in robot swarms: Formalization, state of the art, and novel perspectives.Frontiers in Robotics and AI, 2017

Gabriele Valentini, Eliseo Ferrante, and Marco Dorigo. The best-of-n problem in robot swarms: Formalization, state of the art, and novel perspectives.Frontiers in Robotics and AI, 2017

2017
[33]

Collective decision with 100 kilobots: speed versus accuracy in binary discrimination problems.AAMAS, (3), May 2016

Gabriele Valentini, Eliseo Ferrante, Heiko Hamann, and Marco Dorigo. Collective decision with 100 kilobots: speed versus accuracy in binary discrimination problems.AAMAS, (3), May 2016

2016
[34]

Pratt, and Spring Berman

Sean Wilson, Aur ´elie Buffin, Stephen C. Pratt, and Spring Berman. Multi-robot replication of ant collective towing behaviours.Royal Society Open Science, 5(10):180409, October 2018

2018
[35]

Swarm behavior tracking based on a deep vision algorithm.arXiv:2204.03319, 2022

Meihong Wu, Xiaoyan Cao, and Shihui Guo. Swarm behavior tracking based on a deep vision algorithm.arXiv:2204.03319, 2022

2022 arXiv
[36]

Conformal prediction enhanced svms for swarm behavior classification

Zepu Xi, Hongbo Chen, Xiaoqian Chen, and Wen Yao. Conformal prediction enhanced svms for swarm behavior classification. In2023 International Joint Conference on Neural Networks (IJCNN)
[37]

Autonomous environment-adaptive microrobot swarm navi- gation enabled by deep learning-based real-time distribution planning

Lidong Yang, Jialin Jiang, Xiaojie Gao, Qinglong Wang, Qi Dou, and Li Zhang. Autonomous environment-adaptive microrobot swarm navi- gation enabled by deep learning-based real-time distribution planning. Nature Machine Intelligence, 4(5):480–493, 2022

2022
[38]

Consensus decision-making in artificial swarms via entropy-based local negotiation and preference updating.Swarm Intelligence, 2023

Cheng Zheng and Kyungjoon Lee. Consensus decision-making in artificial swarms via entropy-based local negotiation and preference updating.Swarm Intelligence, 2023

2023

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

Arnold, Hiroyuki Yamaguchi, and Toshiyuki Tanaka

Ross D. Arnold, Hiroyuki Yamaguchi, and Toshiyuki Tanaka. Search and rescue with autonomous flying robots through behavior-based cooperative intelligence.Journal of International Humanitarian Action 2018

work page 2018

[2] [2]

Fuzzy-based aggregation with a mobile robot swarm

Farshad Arvin, Ali Turgut, and Shigang Yue. Fuzzy-based aggregation with a mobile robot swarm. pages 346–347, 09 2012

work page 2012

[3] [3]

Classifying swarm behavior via compressive subspace learning

Matthew Berger, Lee M Seversky, and Daniel S Brown. Classifying swarm behavior via compressive subspace learning. In2016 IEEE International Conference on Robotics and Automation (ICRA)

work page

[4] [4]

Reliable communication in the presence of failures.ACM Trans

Kenneth P Birman and Thomas A Joseph. Reliable communication in the presence of failures.ACM Trans. on Computer Systems, 1987

work page 1987

[5] [5]

Managing a Mobile Agricultural Robot Swarm for a seeding task

Timo Blender, Thiemo Buchner, Benjamin Fernandez, Benno Pichlmaier, and Christian Schlegel. Managing a Mobile Agricultural Robot Swarm for a seeding task. InIECON 2016

work page 2016

[6] [6]

Limited bandwidth recognition of collective behaviors in bio-inspired swarms

Daniel S Brown and Michael A Goodrich. Limited bandwidth recognition of collective behaviors in bio-inspired swarms. InAAMAS, 2014

work page 2014

[7] [7]

Discovery and exploration of novel swarm behaviors given limited robot capabilities

Daniel S Brown, Ryan Turner, Oliver Hennigh, and Steven Loscalzo. Discovery and exploration of novel swarm behaviors given limited robot capabilities. InDistributed Autonomous Robotic Systems: The 13th International Symposium, pages 447–460. Springer, 2018

work page 2018

[8] [8]

Giulia De Masi, Judhi Prasetyo, Elio Tuci, and Eliseo Ferrante.Zealots Attack and the Revenge of the Commons: Quality vs Quantity in the Best-of-n. 10 2020

work page 2020