A lightweight AI controller called CTS kept a simulated UAV-mounted communications system above 90% of an optimal data-rate benchmark and outperformed the comparison policies tested alongside it, according to a new preprint. The finding comes from a modeled case study, not a test of an embedded controller or a field deployment.
The paper examines whether lightweight AI can make UAV-mounted RIS optimization feasible when onboard energy and processing capacity are tightly constrained. In its case study, a UAV-mounted RIS assists a millimeter-wave base station serving multiple distributed hotspots with uncertain traffic and changing channel conditions.
Making the decision small enough to run onboard
The overview compares lightweight AI methods by their performance and complexity characteristics, then adds a case study. Its conclusion groups the approaches into model-efficiency, adaptive decision-making, topology-aware distributed-learning and hybrid model-driven families.
At the center of the case study is a multi-armed bandit, or MAB, formulation. Each hotspot is treated as an arm, or choice available to the learner, and the UAV is the learner. The reward is the achieved data rate, while the context includes location, historical payoffs and residual energy.
The comparison pits CTS against classical upper-confidence-bound, or UCB, and Thompson-sampling, or TS, policies, plus nearest-hotspot selection and a random policy. The simulation varies the number of hotspots and uses 128 RIS elements with the UAV at a height of 50 metres.
High throughput, sharply better energy efficiency
CTS was reported to achieve data-rate performance exceeding 90% of the optimal benchmark while outperforming UCB, TS, nearest-hotspot selection and random policies. Because the supplied results give no absolute data-rate values, the benchmark comparison does not show the throughput a deployed system would deliver.
Energy efficiency was defined as data rate divided by energy consumed per hotspot. On that measure, CTS was reported to show orders-of-magnitude improvement over non-contextual or greedy strategies, especially as hotspot numbers or the covered region increased. The paper does not provide exact energy-efficiency values.
The qualitative assessment of MAB methods rates them low in computational complexity, high in energy efficiency, medium in adaptability and very high in onboard feasibility. Those are descriptive ratings, not numerical scores, and no formal scoring or validation procedure is reported for them.
The authors interpret the case-study results as showing that MAB-based decision-making can extend UAV operational lifetime while preserving high communication throughput. The supplied analysis treats that as an interpretation of a modeled comparison, not a causal experimental estimate.
Evidence that stops at the model
That distinction matters because the evidence consists of a narrative overview, qualitative table comparisons and a simulation case study. No inferential statistical tests or uncertainty estimates are reported. The supplied text also omits exact data-rate, energy-efficiency and energy-use results, along with variance, confidence intervals and hypothesis-test results.
The study does not report the number of reviewed studies, simulated hotspots, decision steps or repetitions. The paper identifies embedded-controller prototypes, real-time field trials, standardized benchmark datasets and interoperability standards as necessary to bridge simulation and deployment.
The authors conclude that hardware-software co-design, embedded implementation and experimental validation should be future priorities for this area.
Publication note
The document is an arXiv preprint, version 1, dated 26 August 2026. It reports support from JSPS KAKENHI grant numbers JP25K07743 and JP23K24905 in Japan.
Paper data and sources
Original title: Lightweight AI for UAV-Mounted RIS: An Overview
Authors: Sherief Hashima, Kohei Hatano, Eiji Takimoto et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text