Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 68 of the 68 references it cites were found in a registry or an index; 11 of the 43 sources its claims cite were checked for support.

Automatic design of stigmergy-based behaviours for robot swarms

Open access
Salman, Muhammad, Garzón Ramos, David, Birattari, Mauro · 2024 · Communications Engineering

Abstract

Stigmergy is a form of indirect communication and coordination in which individuals influence their peers by modifying the environment in various ways, including rearranging objects in space and releasing chemicals. For example, some ant species lay pheromone trails to efficiently navigate between food sources and nests. Besides being used by social animals, stigmergy has also inspired the development of algorithms for combinatorial optimisation and multi-robot systems. In swarm robotics, collective behaviours based on stigmergy have always been designed manually, which is time consuming, costly, hardly repeatable, and depends on the expertise of the designer. Here, we show that stigmergy-based behaviours can be produced via automatic design: an optimisation process based on simulations generates collective behaviours for a group of robots that can lay and sense artificial pheromones. The results of our experiments indicate that the collective behaviours designed automatically are as good as—and in some cases better than—those produced manually. By taking advantage of pheromone-based stigmergy, the automatic design process generated collective behaviours that exhibit spatial organisation, memory, and communication.

Overview

CoChat Summary

TL;DR

Habanero is an automatic off-line AutoMoDe method that designs stigmergy-based collective behaviours for e-puck swarms able to lay and sense artificial pheromone trails. In four missions, its real-robot performance was comparable to or better than manually designed and neuroevolutionary control software, with Habanero outranking all other design methods in aggregate.

Key findings

  • Habanero produced effective pheromone-based behaviours across AGGREGATION, DECISION MAKING, RENDEZVOUS POINT, and STOP. A Friedman rank sum test on real-robot performance across the four missions indicated that Habanero outranked all other design methods with at least 95% confidence; Human-Designers performed significantly better than both EvoPheromone and Random-Walk.
  • Per-mission real-robot results: in AGGREGATION, Habanero performed significantly better than all other design methods; in RENDEZVOUS POINT, Habanero performed significantly better than all other methods while EvoPheromone performed significantly worse than all others; in STOP, Habanero and Human-Designers performed similarly and significantly better than EvoPheromone; in DECISION MAKING, Habanero correctly selected the blue region in all runs, though real-robot performance was lower than simulation because some robots spilled out of the region.
  • The automatically designed behaviours exhibited spatial organisation, external memory, and communication. Pheromone semantics were mission-specific: for example, in STOP the trail meant “stop where you are,” whereas in AGGREGATION it meant “come here.” These properties were not hand-coded in the modules but emerged from how Habanero combined them per mission.
  • Habanero and Human-Designers used the same modules but combined them differently: Habanero used Exploration less, relied more on pheromone-reacting modules, used wall-colour modules less, and used the waggle module more than Human-Designers.

Methods

  • Habanero is an AutoMoDe instance that assembles seven low-level behaviours and six transition conditions into probabilistic finite-state machines with at most 4 states and at most 4 outgoing transitions per state. It tunes 105 parameters using Iterated F-race in ARGoS3 with the argos3-epuck-phormica library, with a budget of 100,000 simulation runs per design execution.
  • The robot platform was an extended e-puck with an Overo Gumstix Linux board, a UV-light module to lay artificial pheromone trails, an omnidirectional camera to detect them, and a ground sensor. The arena floor was coated with photochromic material—acrylic binder with 20% (w/w) photochromic pigments—which turns magenta under UV light and fades back to white in about 50 s; walls were modular RGB blocks, and an overhead tracking camera measured performance only.
  • The empirical study used four missions (AGGREGATION, DECISION MAKING, RENDEZVOUS POINT, STOP) with N = 8 robots and T = 180 s. For each mission, 10 instances of control software were produced per method, giving 160 instances total; each was evaluated once in simulation and once with physical robots. Comparisons were Habanero vs. EvoPheromone (neuroevolution: 61 input nodes, 7 output nodes, no hidden layer, 427 real-valued weights), Human-Designers (10 human designers, 4 h per mission), and Random-Walk. Analysis used Wilcoxon paired rank sum tests and a Friedman rank sum test.

Limitations

  • The paper states that the study leaves open whether automatic design can leverage pheromone intensity and decay time, since in these experiments a robot either did or did not sense pheromone in a binary fashion. It also leaves open whether automatic methods can select or combine direct and indirect communication.
  • The paper states that the photochromic artificial pheromone system is only suitable for indoor applications where the environment can be prepared beforehand, and that no technology currently provides robots with a universally applicable ability to mark their environment. It also states that the experimental setup cannot adequately explain the rationale behind module selection, tuning, and combination, or its relationship to performance, and that the missions are not suitable for drawing conclusions about more complex missions or more complex stigmergy-based interactions.
  • Evident from the text: the study is limited to four missions, a swarm of 8 robots, one simulation and one real-robot evaluation per control-software instance, a fixed design budget of 100,000 simulation runs, and comparisons against only EvoPheromone, Human-Designers, and Random-Walk. Real-robot performance was sometimes lower than simulation—for example, in DECISION MAKING—indicating a reality gap.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s44172-024-00175-7

Cites

68 of 68 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 AutoMoDe-Habanero. a Habanero automatically produces control software for e-puck robots by assembling predefined and mission-independent software modules into a probabilistic finite-state machine. A set of seven low-level behaviours and six transition conditions function as states and edges of the finite-state machine, respectively. Using the Iterated F-race algorithm, the design process determines the topology of the finite-state machine by maximising the performance of the robot swarm. The performance of an instance of control software is assessed in simulation, before the swarm is deployed. b The set of low-level behaviours: operations a robot can execute. c The set of transition conditions: criteria to switch from a low-level behaviour to another one.
Fig. 1 AutoMoDe-Habanero. a Habanero automatically produces control software for e-puck robots by assembling predefined and mission-independent software modules into a probabilistic finite-state machine. A set of seven low-level behaviours and six transition conditions function as states and edges of the finite-state machine, respectively. Using the Iterated F-race algorithm, the design process determines the topology of the finite-state machine by maximising the performance of the robot swarm. The performance of an instance of control software is assessed in simulation, before the swarm is deployed. b The set of low-level behaviours: operations a robot can execute. c The set of transition conditions: criteria to switch from a low-level behaviour to another one.CC BY · PubMed Central
Fig. 2 The e-puck robot, its reference model, and the experimental setup. a An e-puck robot equipped with a Linux board, a hardware module to focus UV light onto the ground, and an omni-directional camera. b The experimental arena. The floor is coated with photochromic material. It changes in colour from white to magenta when exposed to UV light, and gradually returns to its normal white colour when the UV light is removed. The walls of the arena are constructed using modular RGB (Red, Green, Blue) blocks, which have the ability to display various colours using the RGB colour code. A tracking system is used to automatically measure performance indicators. c The reference model RM 4.1, which formally describes the interface between the robot and the control software.
Fig. 2 The e-puck robot, its reference model, and the experimental setup. a An e-puck robot equipped with a Linux board, a hardware module to focus UV light onto the ground, and an omni-directional camera. b The experimental arena. The floor is coated with photochromic material. It changes in colour from white to magenta when exposed to UV light, and gradually returns to its normal white colour when the UV light is removed. The walls of the arena are constructed using modular RGB (Red, Green, Blue) blocks, which have the ability to display various colours using the RGB colour code. A tracking system is used to automatically measure performance indicators. c The reference model RM 4.1, which formally describes the interface between the robot and the control software.CC BY · PubMed Central
Fig. 3 Construction of the arenas for the four missions. Technical drawings of the arena with dimensions and positions of different regions, along with photos of the real arena, in the four mission configurations: a AGGREGATION, b DECISION MAKING, c RENDEZVOUS POINT, and d STOP. All measurements are expressed in meters. The missions are described in the Methods section.
Fig. 3 Construction of the arenas for the four missions. Technical drawings of the arena with dimensions and positions of different regions, along with photos of the real arena, in the four mission configurations: a AGGREGATION, b DECISION MAKING, c RENDEZVOUS POINT, and d STOP. All measurements are expressed in meters. The missions are described in the Methods section.CC BY · PubMed Central
Fig. 4 Pictorial representation of the design methods under analysis. a Habanero, b EvoPheromone, c Human-Designers, and d Random-Walk. Habanero is an automatic modular design method. EvoPheromone is an implementation of the neuroevolutionary approach. Human-Designers is a manual design method. Although Random-Walk is not a design method, we include it to serve as a lower bound of performance. See the Methods section for the details.
Fig. 4 Pictorial representation of the design methods under analysis. a Habanero, b EvoPheromone, c Human-Designers, and d Random-Walk. Habanero is an automatic modular design method. EvoPheromone is an implementation of the neuroevolutionary approach. Human-Designers is a manual design method. Although Random-Walk is not a design method, we include it to serve as a lower bound of performance. See the Methods section for the details.CC BY · PubMed Central
Fig. 5 Results of the empirical analysis. We report results of the evaluation of 160 instances of control software, 10 per method and per mission. All instances of control software were evaluated once in simulation and once with physical robots—more details on the protocol are provided in Methods. The results are presented using boxplots on a per-mission basis: a AGGREGATION, b DECISION MAKING, c RENDEZVOUS POINT, and d STOP. In all missions, for each method, we report the performance obtained in simulation and with physical robots using thin and thick boxes, respectively. e Friedman rank sum test on real-robot performance to aggregate the overall performance of each method across the four missions—the lower the rank, the better. An explanation of the graphical convention adopted in the boxplots and in the Friedman test are provided in the Methods section under the heading Statistics.
Fig. 5 Results of the empirical analysis. We report results of the evaluation of 160 instances of control software, 10 per method and per mission. All instances of control software were evaluated once in simulation and once with physical robots—more details on the protocol are provided in Methods. The results are presented using boxplots on a per-mission basis: a AGGREGATION, b DECISION MAKING, c RENDEZVOUS POINT, and d STOP. In all missions, for each method, we report the performance obtained in simulation and with physical robots using thin and thick boxes, respectively. e Friedman rank sum test on real-robot performance to aggregate the overall performance of each method across the four missions—the lower the rank, the better. An explanation of the graphical convention adopted in the boxplots and in the Friedman test are provided in the Methods section under the heading Statistics.CC BY · PubMed Central
Fig. 6 Behaviours produced by Habanero and Human-Designers. For each mission, we show: a snapshot of robots executing an instance of Habanero control software in (a–d) simulation and (e–h) real-robot experiments, as well as a plot of the aggregate execution time of each software module in the control software produced by (i–l) Habanero and (m–p) Human-Designers. We use the aggregate execution time of the modules to qualify the behaviour we observe in the robot swarms. In the aggregate plots, the colour gradient shows the percentage of time one behaviour was executed throughout all instances of control software produced for a mission. We identify the behaviour modules using the labels defined in Fig. 1.
Fig. 6 Behaviours produced by Habanero and Human-Designers. For each mission, we show: a snapshot of robots executing an instance of Habanero control software in (a–d) simulation and (e–h) real-robot experiments, as well as a plot of the aggregate execution time of each software module in the control software produced by (i–l) Habanero and (m–p) Human-Designers. We use the aggregate execution time of the modules to qualify the behaviour we observe in the robot swarms. In the aggregate plots, the colour gradient shows the percentage of time one behaviour was executed throughout all instances of control software produced for a mission. We identify the behaviour modules using the labels defined in Fig. 1.CC BY · PubMed Central

Verification

Claims the paper makes

10 Valid · 3 Not valid · 20 Not checked · 81 Own results and interpretations

58 more claims

Create a free account to see all 114 claims and every source verdict.

See all claims — free account
  • Probably not a substantial claim. 21
  • It may state two separate claims as one. 35

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

2 of 2

Citations

Cites

Show all 66 references

Research works printed but not linked: 5

  1. Wilson, E. O. Sociobiology: The New Synthesis. (Harvard University Press, Cambridge, MA, USA, 1975).
  2. Goss, S., Deneubourg, J.-L., Bourgine, P. & Varela, E. Harvesting by a group of robots. In Toward a Practice of Autonomous Systems, Proceedings of the First European Conference on Artificial Life, Complex adaptive systems, 195–2054 (MIT Press, Cambridge, MA, USA, 1992).
  3. Brooks, R. A. Artificial life and real robots. In Varela, F. J. & Bourgine, P. (eds.) Towards a Practice of Autonomous Systems: Proceedings of the First European Conference on Artificial Life, 3–10 (MIT Press, Cambridge, MA, USA, 1992).
  4. Kuckling, J. Optimization in the automatic modular design ofcontrol software for robot swarms. (Ph.D. thesis, Université libre de Bruxelles, Brussels, Belgium 2023).
  5. Chambers, J. M., Cleveland, W. S., Kleiner, B. & Tukey, P. A. Graphical Methods For Data Analysis. (CRC Press, Belmont, CA, USA, 1983).

Cited by

See all 26

Metadata

Authors
Salman, M., Garzón Ramos, D., & Birattari, M.
Affiliations
Université Libre de Bruxelles (BE)
Corresponding author
Birattari, Mauro
Venue
Communications Engineering
Published
2024-02-14
Volume
3
Issue
1
ISSN
2731-3395 (ISSN-L)
OpenAlex
W4391808535
Cited by
26
References
66
Citations by year
2026: 8 · 2025: 9 · 2024: 9
FWCI
3.48
Type
journal article
Language
en
Open access
Licence
Locations
  • Communications Engineering (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • Dépôt institutionnel de l'Université libre de Bruxelles (Université Libre de Bruxelles) (submittedVersion, other-oa)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
EC | Horizon 2020 Framework Programme (681872)
Topics
  • Physical Sciences › Engineering › Mechanical Engineering › Modular Robots and Swarm Intelligence
  • Physical Sciences › Computer Science › Artificial Intelligence › Metaheuristic Optimization Algorithms Research
  • Physical Sciences › Engineering › Control and Systems Engineering › Distributed Control Multi-Agent Systems
Keywords
stigmergy, swarm robotics, artificial pheromone, collective behavior, multi-robot systems, combinatorial optimization, spatial organization

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-01, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-01 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Automatic design of stigmergy-based behaviours for robot swarms. https://app.cochat.ai/papers/10.1038/s44172-024-00175-7
The paper
Salman, M., Garzón Ramos, D., & Birattari, M. (2024). Automatic design of stigmergy-based behaviours for robot swarms. Communications Engineering. https://doi.org/10.1038/s44172-024-00175-7

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.