Lightweight AI for UAV-Mounted RIS: An Overview

A taxonomy and comparative analysis of RL, meta-learning, federated learning, and multi-armed bandits for energy-constrained airborne intelligent surfaces in 6G

Sherief Hashima; Kohei Hatano; Eiji Takimoto; Mohamed Rihan; Basem. M. Elhalawany; Hamada Rizk· 2026· DOI 10.48550/arXiv.2608.25402

The core problem

Unmanned Aerial Vehicles (UAV)-mounted Reconfigurable Intelligent Surfaces (RIS) have emerged as a promising architecture for enhancing wireless coverage, spectral efficiency, and energy performance in 6G networks. By combining programmable electromagnetic wave manipulation with aerial mobility, UAV-RIS systems enable dynamic blockage mitigation, adaptive beamforming, and flexible deployment across terrestrial, maritime, and satellite-integrated environments. The core appeal lies in the ability to reposition the intelligent surface on demand, thereby creating favorable propagation conditions where static infrastructure cannot reach. However, joint optimization of UAV trajectory, RIS phase configuration, and resource allocation incurs high computational complexity, which is incompatible with the strict energy and onboard processing constraints of UAV platforms. This tension between optimization quality and resource feasibility motivates the central research question of the paper: how can lightweight AI techniques deliver practical solutions to this challenge? The authors argue that a comprehensive overview of lightweight AI for UAV-mounted RIS is needed, covering Reinforcement Lear

Innovation

The survey reports a structured taxonomy and comparative analysis of existing work. Key findings include the observation that RL-based approaches can achieve near-optimal trajectory and phase control but often require extensive training episodes and substantial onboard computation, which may exceed UAV energy budgets. Meta-learning reduces adaptation time to new deployment scenarios by learning initialization parameters that generalize across tasks, yet it still demands episodic memory and meta-training overhead. Federated Learning distributes model training across UAVs and ground nodes, lowering backhaul usage and preserving data privacy, but introduces communication rounds that can dominate energy consumption in bandwidth-limited aerial links. Multi-Armed Bandits emerge as particularly attractive for UAV-mounted RIS because they offer low-complexity online decision-making with provable regret bounds. The case study demonstrates that MAB schemes improve throughput and energy efficiency relative to baseline selection strategies, confirming their suitability for resource-constrained airborne platforms. Energy-aware optimization techniques further refine performance by explicitly mo

Unmanned Aerial Vehicles (UAV)-mounted Reconfigurable Intelligent Surfaces (RIS) have emerged as a promising architecture for enhancing wireless coverage, spectral efficiency, and energy performance in 6G networks. By combining programmable electromagnetic wave manipulation with aerial mobility, UAV-RIS systems enable dynamic blockage mitigation, adaptive beamforming, and flexible deployment across terrestrial, maritime, and satellite-integrated environments. The core appeal lies in the ability to reposition the intelligent surface on demand, thereby creating favorable propagation conditions where static infrastructure cannot reach.

However, joint optimization of UAV trajectory, RIS phase configuration, and resource allocation incurs high computational complexity, which is incompatible with the strict energy and onboard processing constraints of UAV platforms. This tension between optimization quality and resource feasibility motivates the central research question of the paper: how can lightweight AI techniques deliver practical solutions to this challenge? The authors argue that a comprehensive overview of lightweight AI for UAV-mounted RIS is needed, covering Reinforcement Learning (RL), meta-learning, Federated Learning (FL), Multi-Armed Bandits (MAB), and energy-aware optimization. The paper contributes a detailed taxonomy, a comparative analysis of existing work, an examination of computational-energy trade-offs, and a case study demonstrating the effect of MAB schemes on throughput and energy efficiency in UAV-mounted RIS.

Why it matters

The paper's central analytical contribution is the articulation of computational-energy trade-offs as a first-class design constraint for UAV-mounted RIS. The authors argue that the strict energy and onboard processing constraints of UAV platforms render many centralized, high-complexity optimization methods impractical, and that lightweight AI is not merely an efficiency improvement but a feasibility requirement. The taxonomy reveals that technique selection should be driven by the dominant bottleneck: RL when sequential control is critical, meta-learning when rapid adaptation across missions is needed, FL when privacy and backhaul are limiting, MAB when online low-complexity decisions suffice, and energy-aware optimization when mission endurance is paramount. Open research challenges identified include scalable learning architectures for large RIS arrays, robustness to channel aging and mobility, energy-efficient federated orchestration across heterogeneous aerial nodes, and theoretical guarantees for hybrid lightweight AI schemes. The authors also call for standardized benchmarks that jointly report throughput, energy efficiency, and computational overhead, since isolated metrics can mislead deployment decisions. The case study on MAB schemes serves as a template for such benchmarking, demonstrating that throughput and energy efficiency can be evaluated together under realistic UAV-mounted RIS conditions. Overall, the overview positions lightweight AI as the enabling layer for scalable, energy-efficient airborne intelligent surfaces in 6G, while cautioning that significant work remains to translate promising techniques into field-deployable systems.

Who should read this

CS practitioners and researchers

Opening member content…