Publications

Collective perception, the process whereby multiple agents integrate partial local observations into a shared understanding of their environment, remains a core challenge in swarm robotics. Many existing approaches rely on numeric or categorical message passing, which effectively supports distributed consensus but limits the formation of rich, semantically grounded representations of the world. This paper addresses this limitation by introducing a framework for language-based collective perception using a swarm of agents powered by Large Language Models (LLMs). Each agent has limited local visual perception, and perform conversion of image input into linguistic representation. In addition, agents use local communication to share knowledge, resulting in improved individual understanding of the environment. The approach is evaluated by comparing each agent's knowledge over time to the ground truth built using global information. Using both individual and social learning achieves higher coverage of the environment on average than a swarm using only individual learning, demonstrating that natural language can serve as an effective coordination medium for decentralized perception. This work establishes a foundation for emergent, language-mediated world modeling in multi-agent systems.

This paper introduces ARIEL, an open-source Python framework for the development of robots through evolution and learning. ARIEL combines body-brain co-design, persistent database-backed execution, and a global genotype-to-phenotype interface (blueprint). Initially built for a modular mobile robot system, it is easily extensible to other types, including aerial robots and robot manipulators. The framework supports evolutionary development of simulated robots in a design space that allows the construction of ‘physical twins’, enabling a direct connection between the simulated and real world. ARIEL offers standardised tools for experiment setup, analysis, and visualisation. It natively supports asynchronous evolution and learning, allowing selection, reproduction, learning, and evaluation to proceed without global synchronisation. This makes the framework particularly suitable for experiments in which learning and evolution are integrated, evaluation costs vary across individuals, and flexible orchestration of execution is required. A key concept introduced by ARIEL is that of a blueprint, which serves as an intermediate layer between genotype and phenotype. Traditionally, the phenotype is directly derived from the genotype; in our system, the phenotype is derived from the blueprint. This decoupling enables different genotypes to encode morphologies independently of the particular robot system employed.

Evolutionary computation has produced many successful algorithms and tools. The main challenge in flexible evolutionary computation lies not only in varying and selecting individuals, but also in how they are represented, stored, scheduled, and retrieved over time. This paper introduces ARIEL, a framework that shifts evolutionary computation from a generation focus to persistent, stateful individuals. We present three configurations: (1) synchronous, (2) archive-assisted, and (3) asynchronous. The experiments show that population management can support different evolutionary workflows without changes to the underlying engine or operators. These workflows can all be achieved within the same infrastructure by adjusting eligibility conditions and orchestration logic.

Following an environmentally constrained path in a timely fashion can be crucial for swarms in scenarios such as disaster response or emergency evacuation. In such situations, swarms must rapidly follow potentially challenging paths while not losing cohesion. We benchmark a robust, decentralized gradient-following behavior against varying path sinuosity, and swarm size. The swarm races to reach a minimum distance in a fixed time budget. We measure the success rate and the mean completion time. Our findings show that the algorithm allows the swarm to successfully reach the target line in the allotted time as long as the swarm size is relatively small. Furthermore, we find that the algorithm handles low and medium degrees of sinuosity well but struggles with high sinuosity, where the turns become too sharp. This work is the first to study swarm racing in constrained environments, revealing that “more is not always better”: larger swarms hinder rapid path traversal. This paves the way for future research on scale-invariant racing collective behaviors.