Social feedback in collective decision making

The rapid evolution of autonomous systems has catalysed a growing demand for decentralised robotic collectives capable of operating in complex, unpredictable, and physically constrained environments. While swarm intelligence offers highly scalable and resilient alternatives to traditional centralised control, the transition from idealised theoretical models to real-world physical deployment is hindered by the reality gap. In physical environments, issues such as obstacles restricting robots interactions, unreliable communication channels, and sensory noise compromise the reliability of the local peer-to-peer communication upon which macroscopic collective behaviour relies. This thesis investigates how low-level local communication algorithms shape and sustain macroscopic global consensus, decomposing swarm coordination into three sequential, interconnected layers: informational, cognitive, and topological. At the informational layer, we provide a comparative analysis of broadcasting protocols in mobile robot networks. We demonstrate that while identity-aware protocols show high estimation accuracy and robust performance even in spatially segregated environments, they incur a memory overhead. For minimalist swarms relying on little memory, we investigate lightweight anonymous communication and a stochastic filtering mechanism based on $k$-priority sampling. Our findings reveal that while this filtering successfully stabilises estimates reducing transient errors, it inherently trades operational responsiveness for steady-state stability, without matching the precision of identity-aware systems. At the cognitive layer, we address the challenge of spatial fragmentation and decision deadlocks under extreme sensory uncertainty. By transitioning Quorum Sensing from a static threshold to a continuous, adaptive control variable modulated by a non-linear, cubic polynomial response function, we enable agents to dynamically adjust their reliance on social influence. This continuous adaptation allows the swarm to autonomously balance environmental exploration and social exploitation, effectively breaking symmetric deadlocks and managing the speed-accuracy trade-off. Finally, at the topological layer, this thesis demonstrates the scalability of these decentralised principles by translating them to a macroscopic, safety-critical domain: real-time railway traffic management. By limiting the time window used to predict conflicts, we demonstrate that self-organising train fleets can efficiently negotiate track access and resolve local traffic conflicts. Our results show that short prediction horizons are more effective than looking far into the future; they keep local computation manageable while maintaining stable overall coordination, as continuous replanning corrects any short-sighted decisions. Ultimately, this work provides a coordination framework that scales from simple robot swarms to large physical infrastructures.