Self-Organized Collective Decision-Making in a 100-Robot Swarm

Gabriele Valentini, Heiko Hamann, Marco DorigoAAAI 2015 · DOI

Introduction

Collective decision-making is a cornerstone of swarm robotics, enabling a group of simple agents to reach a consensus on global tasks without centralized control. A classic benchmark is the "site-selection problem," where a swarm must evaluate multiple locations and agree on the one with the highest quality. This study explores a robust, self-organized method for a 100-robot swarm to solve this problem using local interactions and quality-proportional feedback.

Methods and Model

The researchers utilized a swarm of 100 Kilobots—low-cost, autonomous robots that communicate via infrared pulses. The decision-making process is governed by a probabilistic finite-state machine consisting of two main phases: exploration and dissemination.

The collective decision-making model and experimental results. (a) State transition diagram. (b) Experimental arena with 100 Kilobots. (c) Consensus progress over time for different neighborhood sizes.

In the Exploration state ($E_A, E_B$), robots visit a site to estimate its quality. Upon returning to the "nest," they enter the Dissemination state ($D_A, D_B$), where they broadcast their preferred site to nearby peers. The core innovation is Quality-Proportional Feedback: the time a robot spends disseminating is directly proportional to the quality of the site it visited.

Before transitioning back to exploration, robots apply a Majority Rule based on their local neighborhood. By listening to its neighbors, a robot adopts the opinion held by the majority. Because robots favoring higher-quality sites broadcast for longer durations, they exert a stronger influence on the swarm, initiating a positive feedback loop toward the optimal choice.

Experimental Results

The experimental setup involved two sites, with Site A being twice as good as Site B ($\rho_A = 1, \rho_B = 0.5$). The study investigated the impact of the maximum neighborhood size ($G_{max}$) on the swarm's performance.

The results, shown in Figure 1(c), demonstrate that the swarm consistently reached a consensus for Site A. Increasing the neighborhood size from $G_{max} = 5$ to $G_{max} = 25$ significantly accelerated the decision-making process, reducing the time to reach a 90% majority from 70 minutes to approximately 60 minutes.

The system proved highly resilient to individual hardware failures. Even when some robots exhibited poor motion or failed to update their opinions due to IR communication issues, the collective was still able to identify the superior site. This robustness is a key advantage of the decentralized, majority-rule approach.

Discussion and Conclusion

This research highlights how simple agents can achieve sophisticated collective behavior by using time as a proxy for information. By weighting the duration of communication rather than its complexity, the swarm bypasses the need for advanced individual processing.

While larger neighborhoods improved speed in this study, future research is required to investigate potential speed-accuracy trade-offs in more complex environments. Additionally, moving beyond emulated sensing to real-world quality estimation will be a critical step for deploying such swarms in hazardous or unknown environments.

Summary updated

BibTeX
@inproceedings{valentini2015self,
  author = {Gabriele Valentini and Heiko Hamann and Marco Dorigo},
  title = {{Self-Organized Collective Decision-Making in a 100-Robot Swarm}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2015},
  doi = {10.1609/aaai.v29i1.9720},
}
Figure