A V2V-Inspired Adaptive Clustering Protocol for Resilient Coordination in Disaster-Response Robotic Swarms

Robotic swarms offer scalable, autonomous support for disaster response, yet coordination stays fragile once communication degrades and agents fail. This work sits within a layered federated multi-agent coordination framework in which agents share model updates rather than raw data, aggregated at elected cluster heads. That clustering layer is the focus here. Most clustering schemes assume stable links and reliable nodes, assumptions that fail in a disaster zone. The paper presents an adaptive clustering protocol inspired by vehicle-to-vehicle (V2V) network formation that forms, repairs, and reconfigures clusters as conditions deteriorate, electing cluster heads from a weighted combination of connectivity, link-quality, mobility, and energy indicators. Evaluation at this stage is conducted entirely in simulation; no physical robots were used. The model is not idealised: radio links follow distance-based path loss (exponent 3.5, 3 dB Gaussian variation, receive threshold), and the severest regime adds an obstacle field, stochastic link drops, and scheduled agent and cluster-head losses. Against lowest-identifier and highest-degree election (three regimes, thirty seeds each), with churn measured over an identical window for every protocol, the protocol held clusters about five times more stably than highest-degree election under the severest regime (p < 0.001), but was indistinguishable from lowest-identifier election and less stable than both in milder regimes. Once settled, its clusters changed head two orders of magnitude less often than either baseline, yet it was settled for only 40 percent of a severe-regime run and never under sustained coordinated drift: the protocol trades availability for stability. The discussion addresses disturbances the simulator omits: asymmetric and multipath radio behaviour, localisation error, timing jitter, and terrain effects on motion. Because decisions are local and confirmed over a dwell period, noisy inputs should delay decisions rather than corrupt them. A physical testbed will quantify the residual sim-to-real gap.