Mauro Birattari

Reinforcement learning offers a powerful framework for automatically generating control policies for robot swarms, yet its application to physical multirobot systems has been limited by the simulation-to-reality gap. Despite substantial advances in multiagent reinforcement learning, systematic demonstrations of gradient-based sim-to-real transfer to physical robot swarms remain scarce. Most work remains confined to simulation, and the few real-robot demonstrations lack evaluation across multiple runs and tasks. We show that this barrier can be overcome by embedding structural constraints into policy learning. We introduce attention-based counterfactual policy gradients for behavior selection in robot swarms (SwarmACB), a multiagent actor-critic method in which robots arbitrate among predefined behavior modules—reactive sensorimotor routines such as phototaxis or aggregation—rather than learning low-level motor commands. This architectural choice reduces the variance of learned solutions, which we argue mitigates overfitting to simulation artifacts. Through systematic evaluation on five collective missions using 20 physical e-puck robots (250 real-robot runs), we demonstrate that SwarmACB transfers successfully from simulation to reality, substantially outperforming continuous-action baselines. By progressively constraining policy architecture, perception, and model capacity, we show that smaller, structured policies transfer more reliably than large, unconstrained ones. These findings challenge the assumption that scale is necessary for successful transfer, suggesting instead that principled structural design enables reinforcement learning to produce control software for robot swarms that is both learned automatically and directly deployable in the physical world.

From flocking birds to human crowds, collective motion patterns arise from the coordinated motion of individuals in a group. These patterns are notoriously hard to infer from the actions of each individual, yet are clearly distinguishable even by an untrained observer. Established formal approaches towards analysing collective motion patterns rely on the identification of relevant variables that are specific to the system under study. As a result, they fail to address the generality of collective motion as an emergent behaviour across systems. In this paper, we show how topological data analysis allows abstracting collective motion patterns and defining meaningful distances between them. The proposed method uses persistent homology to represent the topological features of velocity fields as sets of persistence diagrams. We show how a distance between these sets can discriminate the various behaviours exhibited by a system, and we demonstrate its generality by applying it to a fish school, a pedestrian crowd, and a robot swarm. We validate our results against those obtained through system-specific abstractions, and illustrate how to use topological representations to explain discrepancies. In short, we show how to quantify and interpret differences between collective motion patterns for any system for which a velocity field can be constructed. This will allow researchers in complex systems to adopt a common framework for collective motion analysis, enabling direct comparison of emergent behaviours across disciplines.

We introduce a method for decentralized person re-identification in robot swarms that leverages natural language as the primary representational modality. Unlike traditional approaches that rely on opaque visual embeddings -- high-dimensional feature vectors extracted from images -- the proposed method uses human-readable language to represent observations. Each robot locally detects and describes individuals using a vision-language model (VLM), producing textual descriptions of appearance instead of feature vectors. These descriptions are compared and clustered across the swarm without centralized coordination, allowing robots to collaboratively group observations of the same individual. Each cluster is distilled into a representative description by a language model, providing an interpretable, concise summary of the swarm's collective perception. This approach enables natural-language querying, enhances transparency, and supports explainable swarm behavior. Preliminary experiments demonstrate competitive performance in identity consistency and interpretability compared to embedding-based methods, despite current limitations in text similarity and computational load. Ongoing work explores refined similarity metrics, semantic navigation, and the extension of language-based perception to environmental elements. This work prioritizes decentralized perception and communication, while active navigation remains an open direction for future study.

Swarm perception enables a robot swarm to collectively sense and interpret the environment by integrating sensory inputs from individual robots. In this study, we explore its application to people re-identification, a critical task in multi-camera tracking scenarios. We propose a decentralized, feature-based perception method that allows robots to re-identify people across different viewpoints. Our approach combines detection, tracking, re-identification, and clustering algorithms, enhanced by a model trained to refine extracted features. Robots dynamically share and fuse data in a decentralized manner, ensuring that collected information remains up to date. Simulation results, measured by the cumulative matching characteristics (CMC) curve, mean average precision (mAP), and average cluster purity, show that decentralized communication significantly improves performance, enabling robots to outperform static cameras without communication and, in some cases, even centralized communication. Furthermore, the findings suggest a trade-off between the amount of data shared and the consistency of the Re-ID.

Swarm robotics addresses the design, deployment, and analysis of large groups of robots that collaborate to perform tasks in a decentralized manner. Research in this field has predominantly relied on simulations or small-scale robots with limited sensing, actuation, and computational capabilities. Consequently, despite significant advancements, swarm robotics has yet to see widespread commercial or industrial application. A major barrier to practical deployment is the lack of affordable, modern, and robust platforms suitable for real-world scenarios. Moreover, a narrow definition of what swarm robotics should be has restricted the scope of potential applications. In this paper, we argue that the development of more advanced robotic platforms—incorporating state-of-the-art technologies such as SLAM, computer vision, and reliable communication systems—and the adoption of a broader interpretation of swarm robotics could significantly expand its range of applicability. This would enable robot swarms to tackle a wider variety of real-world tasks and integrate more effectively with existing systems, ultimately paving the way for successful deployment.

Robot Operating System (ROS) is a set of development tools that has become a de facto standard both in the academia and the industry. However, little effort has been devoted to enable ROS in small robots widely used in swarm robotics for education and research, such as the e-puck. We developed DeimOS: a Linux-based ROS-ready operating system for the e-puck robot extended with the Gumstix Overo COM. DeimOS integrates support for ARGoS3—a swarm-dedicated simulator that facilitates the realization of control software for robot swarms. Thanks to DeimOS, e-puck robots can be modernized to conduct educational and research experiments in modern engaging robotics tasks such as robot mapping. DeimOS is a free and open-source software-only solution to extend the lifetime and the capabilities of e-puck robots—a cost effective alternative to existing hardware-based solutions.

Automatic design is an appealing approach to realizing robot swarms. In this approach, a designer specifies a mission that the swarm must perform, and an optimization algorithm searches for the control software that enables the robots to perform the given mission. Traditionally, research in automatic design has focused on missions specified by a single design criterion, adopting methods based on single‐objective optimization algorithms. In this study, we investigate whether existing methods can be adapted to address missions specified by concurrent design criteria. We focus on the bi‐criteria case. We conduct experiments with a swarm of e‐puck robots that must perform sequences of two missions: each mission in the sequence is an independent design criterion that the automatic method must handle during the optimization process. We consider modular and neuroevolutionary methods that aggregate concurrent criteria via the weighted sum, hypervolume, or ‐norm. We compare their performance with that of Mandarina, an original automatic modular design method. Mandarina integrates Iterated F‐race as an optimization algorithm to conduct the design process without aggregating the design criteria. Results from realistic simulations and demonstrations with physical robots show that the best results are obtained with modular methods and when the design criteria are not aggregated.

Swarm perception refers to the ability of a robot swarm to utilize the perception capabilities of each individual robot, forming a collective understanding of the environment. Their distributed nature enables robot swarms to continuously monitor dynamic environments by maintaining a constant presence throughout the space.In this study, we present a preliminary experiment on the collective tracking of people using a robot swarm. The experiment was conducted in simulation across four different office environments, with swarms of varying sizes. The robots were provided with images sampled from a dataset of real-world office environment pictures.We measured the time distribution required for a robot to detect a person changing location and to propagate this information to increasing fractions of the swarm. The results indicate that robot swarms show significant promise in monitoring dynamic environments.

Most studies in swarm robotics treat the swarm as an isolated system of interest. We argue that the prevailing view of swarms as self-sufficient, independent systems limits the scope of potential applications for swarm robotics. A robot swarm could act as a support in an heterogeneous system comprising other robots and/or human operators, in particular by quickly providing access to a large amount of data acquired in large unknown environments. Tasks such as target identification & tracking, scouting, or monitoring/surveillance could benefit from this approach.