Generalized-rank action selection in multi-robot belief space planning with limited connectivity

Abstract

Effective coordination among autonomous robots in dynamic, communication-constrained environ- ments remains a significant challenge in multi-robot systems. This thesis addresses such coordination problems in real-world scenarios, such as search-and-rescue missions, where communication failures and inconsistent beliefs hinder effective collaboration. We build upon the decentralized algorithm VerifyAC[1], which verifies consistency in multi-robot coordination and triggers communication only when required. However, VerifyAC is restricted to rank-1 action preferences, which incurs both high communication and computational costs. To address this, we previously introduced VerifyAC-Gen[2], a decentralized variant that generalizes rank selection via backward reasoning, pruning, and heuristic-based ambiguity resolution. While VerifyAC-Gen effectively reduces unnecessary communication, its extension to multiple agents introduces scalability and complexity issues. In this thesis, we present two extended strategies to scale our framework to environments involving N > 2 robots: a decentralized Min-Heap Tree approach and a centralized cluster registry protocol. The decentralized Min-Heap Tree method reduces communication complexity by assigning non- leader robots to clusters based on KL divergence from entropy-minimized leaders. Communication is hierarchically structured, significantly reducing overhead while preserving coordination integrity. Complementarily, the centralized registry approach maintains a global cluster-to-robot mapping, enabling dynamic reconfiguration upon leader failure or agent arrival. It utilizes entropy-based leader selection, KL-divergence-based clustering, and min-heap structures to ensure optimal com- munication and reallocation of robots across clusters. Together, these strategies extend the scalability, robustness, and efficiency of our coordination framework under dynamic and uncertain environments. Ongoing experimental validation focuses on communication reduction, computational efficiency, and adaptability in real-time applications.