Survivor-Aware UAV Positioning for RIS-Assisted 6G Emergency Communications: A Deep Reinforcement Learning Approach

Abstract

Sustaining service for survivor IoT devices in disaster zones where terrestrial infrastructure has failed is an open challenge in sixth-generation (6G) emergency communications. This paper studies a single UAV flying base station that serves survivor devices over power-domain NOMA, with a passive reconfigurable intelligent surface (RIS) providing a reflected bypass for far survivors blocked by deep rubble obstruction, and a Deep Q-Learning (DQL) agent placing the UAV in 3D. We establish three results. First, the conventional sum-rate positioning objective is structurally blind to blocked survivors: because the far users contribute only a negligible fraction of the optimised aggregate throughput, the positioning reward is dominated by the near users, so a throughput-optimal controller abandons the very survivors the network exists to reach. Second, we expose the governing geometry: for a blocked survivor the achievable rate is dictated by the UAV-to-RIS separation rather than the UAV-to-survivor separation (because only the UAV-to-RIS hop of the two-hop cascade is controllable), yielding a counterintuitive design rule: to serve a blocked survivor, move the UAV toward the reflector, not the survivor. The throughput-optimal and survivor-optimal UAV positions are consequently distinct, creating an irreducible positioning tension. Third, because a single passive RIS cannot phase-align to all blocked survivors simultaneously, we introduce an on-demand, per-beacon RIS scheduling model and characterise its delivery capacity with a queuing analysis, resolving a physical contradiction common to multiuser RIS studies. We cast the throughput-reliability tension as a Pareto frontier and resolve it with a survivor-aware reward; re-trained under this reward, the DQL agent autonomously reaches a position delivering near-100% far-user telemetry-beacon reliability at an aggregate-throughput cost of about 7%. All findings are reproduced by openly available simulation code. The study is positioned as a sub-6 GHz emergency-connectivity framework for future 6G architectures: the rates reported here are specific to a 3.5 GHz, 20 MHz disaster-coverage configuration and are not intended to represent general 6G performance.