Organizations: IRIDIA, Université Libre de Bruxelles, Brussels, Belgium · Centre for the Advanced Study of Collective Behaviour, Universität Konstanz, Konstanz, Germany · Department of Collective Behaviour, Max Planck Institute of Animal Behavior, Konstanz, Germany
Robot swarms harness simple, local interactions among many individually limited robots to produce complex, coordinated collective behavior, making them valuable across domains from agriculture to disaster response. We envision the next frontier for such swarms as open swarms, in which robots join or leave the swarm dynamically at runtime. The ability to dynamically adjust swarm size has been shown to be beneficial, for instance by reducing inconsistencies and network partitioning, especially under sparse connectivity or in complex environments, ultimately improving task performance. For practical deployment, however, open swarms requires governance, particularly protection against Byzantine behavior (i.e., faulty or malicious robots), since allowing robots to join freely also makes the swarm easier to infiltrate. Blockchain technology has recently been incorporated into robot swarms, offering secure, decentralized coordination and improved resilience, though largely in closed settings with a fixed, predetermined membership. In this work, we extend these benefits to open swarms by introducing robot swarms governed by a blockchain-based decentralized autonomous organization (DAO), in which the swarm regulates itself using autonomous rules encoded in smart contracts. We outline several DAO-based governance mechanisms for robot swarms, depending on the needs of the application, and present a proof-of-concept implementation of a simple DAO that regulates swarm membership and prevents Byzantine infiltration. Our simulation results and real-robot validations show that DAO-governed access control and robot removal can maintain proper swarm functioning while keeping the number of Byzantine robots within acceptable limits.
Figures & tables
Figure 1: DAO implementation. We present a basic proof-of-concept DAO implemented in a robot swarm. (A) Every t timesteps, each robot checks for Byzantine behavior in the swarm. ‘When’ a robot initiates removal is determined by a threshold-based Byzantine detection mechanism. Once a robot detects a Byzantine, it signals the smart contract, which triggers a voting round. Robots then decide ‘who’ to remove using a non-tokenized, reputation-based mechanism where each robot submits a list of suspected Byzantines to the smart contract. The contract aggregates these reports, and ‘how’ a removal occurs is governed by the rule whereby a robot is removed if it appears on at least 50% of the submitted lists after a 5-block waiting period. (B) Our DAO operates effectively as long as the influx of robots does not exceed the system’s ability to remove Byzantine robots. For every robot removed, one or more new robots join the swarm. These new robots initially do not have the right to vote on the removal of other robots. The parameter α modulates how long newly added robots must provide valid estimates before they are allowed to participate in the removal decision.
Figure 2: Motivation for open swarms . (A–B) Average absolute error over time from the true 25 % value across 10 runs, with standard deviation (showing variability across runs). To demonstrate the need for open swarms, we consider a larger arena of 15.40 m 2 (instead of 3.55 m 2 ) for the results presented in this figure. (A) A static swarm of 12 robots (green) struggles, whereas adding honest robots every 120 s until the swarm reaches 30 robots (blue) drives the error toward zero. (B) Similarly, when the initial swarm has 33 % Byzantines, adding honest robots reduces Byzantine influence. The experiments in (A-B) are run without the outlier-based defense. (C-D) The collective estimation task becomes substantially harder in non-uniform environments. In a run with 12 honest robots, final estimates vary widely when tiles are clustered (D) compared to the uniform case (C), illustrating increased variance and higher risk of misclassifying honest robots.
Figure 3: Experimental setup and baseline analysis . (A) The arena, measuring 3.55 m 2 , used for collective sensing, where Pi-puck robots perform a random walk to estimate the proportion of white tiles (actual value: 25 %). Each robot acts as a blockchain node using Toychain [ 63 ] to collectively converge on a global estimate. Robots communicate locally within 13 cm. The tiles are clustered in two corners, increasing sensing variance and making the task more challenging compared to a uniformly random tile distribution [ 9 ] . Byzantine robots can further disrupt the task by sending incorrect sensor readings. All experiments are performed in simulation using the ARGoS swarm simulator [ 67 ] . (B) The physical arena, measuring 1 m 2 , used for the real-robot demonstrations, with white tiles covering 25% of the arena. (C–D) Average swarm estimates in an 8-robot swarm implementing the baseline method of Strobel et al. [ 77 ] , but in an open swarm rather than a closed one. (C) When a naive Byzantine robot is introduced into an initially healthy swarm every 120 s, the static outlier-based method still converges to the correct estimate. (D) When the base swarm begins with 50 % Byzantines and only honest robots are added, the mechanism fails: the on-chain outlier filtering in Strobel et al. [ 77 ] depends heavily on initial estimates of the swarm and accepts only values within 0.10 of the current shared mean estimate. Once Byzantine robots drive the estimate toward 0, newly added honest robots cannot recover the true value. (Inset of C) We also illustrate the failure of the outlier filtering mechanism using ‘smarter’ Byzantines, which send estimates equal to the system mean minus 0.09 to evade outlier detection and shift the estimate towards 0.
Figure 4: Experiment 1 – Neutralizing Byzantine robots already present in the swarm. A swarm of 12 robots is evaluated under two types of adversarial behavior: naive Byzantines, which always report an estimate of 0 % (A,C,E,G), and smart Byzantines, which report a value approximately 0.09 below the swarm mean (B,D,F, H). For each Byzantine type, we test two DAO settings: α=10 and α=1 . The plots show: (A–B) the classification performance in terms of accuracy (the fraction of Byzantine robots removed from the swarm) and error (the fraction of non-Byzantine robots removed from the swarm); (C–D) the average swarm size over time, with the evolution of the number of Byzantine robots shown in the insets; (E–F) the average absolute error in the swarm’s estimates relative to the true value of 25%, computed at the time when all robots in each run have obtained a mean estimate; and (G–H) the total number of robot removals across the experimental conditions (the black diamond indicates the mean, the horizontal line indicates the median, and the orange or grey dots represent removals in each run). For each value of α and experimental setting (naive and smart), we conduct 10 simulation runs, with the resulting data used to generate all the plots shown in the figure.
Figure 5: Experiment 2 – Preventing Byzantine takeover. A swarm of 12 robots is evaluated under two types of adversarial behavior: naive Byzantines, which always report an estimate of 0 % (A,C,E,G), and smart Byzantines, which report a value approximately 0.09 below the swarm mean (B,D,F,H). For each Byzantine type, we test two DAO settings: α=10 and α=1 . The plots show: (A–B) the classification performance in terms of accuracy (the fraction of Byzantine robots removed from the swarm) and error (the fraction of non-Byzantine robots removed from the swarm); (C–D) the average swarm size over time, with the evolution of the number of Byzantine robots shown in the insets; (E–F) the average absolute error in the swarm’s estimates relative to the true value of 25%, computed at the time when all robots in each run have obtained a mean estimate; and (G–H) the total number of robot removals across the experimental conditions (the black diamond indicates the mean, the horizontal line indicates the median, and the orange or grey dots represent removals in each run). For each value of α and experimental setting (naive and smart), we conduct 10 simulation runs, with the resulting data used to generate all the plots shown in the figure.
Figure 6: Experiment 3 – Swarm growth under continuous robot influx. A swarm of 12 robots is evaluated under two types of adversarial behavior: naive Byzantines, which always report an estimate of 0 % (A,C,E,G), and smart Byzantines, which report a value approximately 0.09 below the swarm mean (B,D,F,H). For each Byzantine type, we test two DAO settings: α=10 and α=1 . The plots show: (A–B) the classification performance in terms of accuracy (the fraction of Byzantine robots removed from the swarm) and error (the fraction of non-Byzantine robots removed from the swarm); (C–D) the average swarm size over time, with the evolution of the number of Byzantine robots shown in the insets; (E–F) the average absolute error in the swarm’s estimates relative to the true value of 25%, computed at the time when all robots in each run have obtained a mean estimate; and (G–H) the total number of robot removals across the experimental conditions (the black diamond indicates the mean, the horizontal line indicates the median, and the orange or grey dots represent removals in each run). For each value of α and experimental setting (naive and smart), we conduct 10 simulation runs, with the resulting data used to generate all the plots shown in the figure.
Figure 7: Swarm size and Byzantine evolution in simulation and real-robot demonstration A swarm of eight robots is evaluated under two types of adversarial behavior: naive Byzantines, which always report an estimate of 0 % (A–B), and smart Byzantines, which report values approximately 0.09 below the swarm mean (C–D). For each Byzantine type, two DAO settings are tested for α=10 and α=1 . The experiments are conducted in a scenario in which a healthy swarm is prevented from being overtaken by newly introduced Byzantines (Experiment 2), and correspond to the real-robot demonstration setup described in Section Real-robot demonstrations . Each of the four conditions is repeated 10 times in simulation (thick solid lines represent the average; thin solid lines represent individual runs) and the red or green fills represent the 99% confidence interval of the simulation runs, with results from a real-robot run overlaid for each α value on each plot (dashed lines). The plots show the average swarm size over time (A, B), along with the evolution of the number of Byzantine robots (C, D).
Figure 8: DAO architectures. (A) Swarm-hosted architecture where the robots themselves maintain the blockchain and the swarm is simultaneously the subject of governance and the infrastructure that enforces it. (B) Externally-hosted architecture in which the robots that are part of DAO access external network to participate in governance decisions on a public blockchain, maintained by a distributed set of nodes rather than the swarm. (C) Multi-stakeholder architecture with DAO membership extending beyond robots to include human participants, who take part in the same governance process alongside the swarm.
Figure 9: DAO-based governance of participation rights in open robot swarms. Illustration detailing how the DAO regulates swarm membership through voting by defining when removal is triggered, which robots participate in the decision process, and how removal is executed, with a particular focus on Byzantine robots (the e-puck robot illustration taken from Webots, Cyberbotics). As underperforming or malicious robots are removed, new robots can be added to maintain or adjust the swarm size. The voting logic used to identify removal candidates can remain private to each robot. Detailed explanation of these mechanisms are provided in section Methods .
Robots are improving their autonomy with minimal human supervision. However, auditable actions, transparent decision processes, and new human-robot interaction models are still missing requirements to achieve extended robot autonomy. To tackle these challenges, we propose RODEO (RObotic DEcentralized Organization), a blockchain-based framework that integrates trust and accountability mechanisms for robots. This paper formalizes Decentralized Autonomous Organizations (DAOs) for service robots. First, it provides a ROS-ETH bridge between the DAO and the robots. Second, it offers templates that enable organizations (e.g., companies, universities) to integrate service robots into their operations. Third, it provides proof-verification mechanisms that allow robot actions to be auditable. In our experimental setup, a mobile robot was deployed as a trash collector in a lab scenario. The robot collects trash and uses a smart bin to sort and dispose of it correctly. Then, the robot submits a proof of the successful operation and is compensated in DAO tokens. Finally, the robot re-invests the acquired funds to purchase battery charging services. Data collected in a three day experiment show that the robot doubled its income and reinvested funds to extend its operating time. The proof validation times of approximately one minute ensured verifiable task execution, while the accumulated robot income successfully funded up to 88 hours of future autonomous operation. The results of this research give insights about how robots and organizations can coordinate tasks and payments with auditable execution proofs and on-chain settlement.
Milan Groshev, Eduardo Castelló Ferrer
CyPhy Life, Robotics & AI Lab, School of Science & Technology, IE University, Spain
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.
Varun Raveendra, Connor Mattson, Daniel S. Brown
Kahlert School of Computing, University of Utah, Salt Lake City, USA
Robot swarms coordinate through directed, time-varying local communication links, some of which may be compromised to allow the transmission of arbitrary malicious data. We provide an algorithm that acts as a wrapper around robot communication---when a robot seeks to update its state using data from some incoming links, our wrapper provides a subset of trusted links whose data can be safely incorporated. It selects these links by testing a stream of trust values---data collected from an anomaly detector on each robot---using Safe Anytime-Valid Inference. We require only that the detector be better than chance. This wrapper ensures safe and performant autonomy by providing two powerful statistical guarantees: it uniformly excludes all compromised links for all time with high user-chosen probability, but nevertheless quickly learns to trust legitimate links. Both guarantees are vital to secure generic tasks, but no previous work provides them simultaneously without restrictive assumptions on connectivity structure or the number of compromised links that are not guaranteed in practice. We study our Safe Wrapper's performance in hardware for two contrasting tasks---neural implicit mapping and heading consensus---deploying on 3 LoCoBots and 8 GoPiGos, respectively. We provide simulations using trust values sampled from hardware to validate that our approach scales well to swarms of hundreds of robots.
Thomas Kaminsky, Áron Vékássy, Hammad Izhar +2
School of Engineering and Applied Sciences, Harvard University, Cambridge, USA · School of Electrical, Computer, and Energy Engineering, Arizona State University, Tempe, USA