Decentralized collectives such as robot swarms, animal groups, programmable particles, and populations of AI agents are typically studied in isolation. Yet in natural settings, collectives rarely operate alone: they coexist, cooperate, compete, and adapt in shared environments, and each collective forms part of the environment of the others. Similarly, engineered and digital collectives are now entering the same regime as multiple robot swarms share physical spaces and populations of AI agents increasingly learn and act alongside other adaptive agents. This shift from single collectives to interacting collectives introduces qualitatively new challenges for theory and practice. Existing results in distributed computing, control, robotics, biology, and machine learning largely address single homogeneous populations and do not extend naturally to settings involving multiple collectives.

Interactions between collectives constitute a new level of organization. We understand a collective as a population of autonomous agents whose system-level behavior arises primarily from decentralized interactions among its members, rather than from a single centralized controller. This behavior is further shaped by environmental signals and by interactions with other collectives. A central question throughout is how individual agents make decisions under uncertainty, from perception and internal state to action, including bounded rationality and threshold/response rules, and how these individual decision rules aggregate into collective behavior, the micro-to-macro link that also governs how one collective’s choices reshape another’s.

We will bring together researchers from theoretical computer science, control theory and robotics, animal behavior, ecology and active-matter physics, and multi-agent learning. This Dagstuhl Seminar will focus on three challenges, mirroring its focus on theory, control, and learning. First, how can a collective perceive, identify, and anticipate another collective from partial and noisy local information? This includes questions about what agents can accomplish given limited sensing, communication, memory, and prior knowledge. Second, how can interacting collectives be modeled, analyzed, and controlled across scales? Agent-level algorithms, dynamical-systems models, and continuum descriptions must be connected through tools such as mean-field and networked control to establish computability, stability, convergence, safety, and control under cooperative, competitive, asymmetric, or nonreciprocal interactions, with game-theoretic formulations for the competitive and nonreciprocal cases. Shepherding, where one collective steers another through indirect influence, provides a particularly rich example. Third, how do learning and adaptation change inter-collective dynamics? Populations of AI agents may co-learn and continually alter one another’s learning environments, while biological systems show how short-term interaction, ecological change, and longer-term adaptation become coupled.

Interdisciplinary working groups will examine scenarios using the concepts and tools of their respective fields. Our goals are to develop a conceptual framework for different forms of collective-collective interaction; formulate a small number of precise, cross-disciplinary benchmark problems; and establish a research agenda identifying where existing methods fail and what new foundations are required. By bringing these communities together at a formative moment, the seminar aims to lay the groundwork for understanding and designing the interacting collectives that will increasingly shape natural, technological, and digital environments.

Creative Commons BY 4.0

Mario di Bernardo, Paola Flocchini, Heiko Hamann, and Guy Theraulaz

This seminar qualifies for Dagstuhl's a LZI Junior Researchers program. Schloss Dagstuhl wishes to enable the participation of junior scientists with a specialization fitting for this Dagstuhl Seminar, even if they are not on the radar of the organizers. Applications by outstanding junior scientists are possible until November 13, 2026.