Energy-Aware Multilingual Vision Language Models for Drone Smart Sensing

[EN] Drone-based smart sensing increasingly relies on Vision¿Language Models (VLMs) for real-time scene interpretation, obstacle detection, and autonomous navigation reasoning. Deploying such systems at scale demands not only high perceptual accuracy but also energy efficiency, a critical constraint...

Descripción completa

Detalles Bibliográficos
Autores: de Curtò, J., Liz, Mauro, de Zarzà, I., Tavares De Araujo Cesariny Calafate, Carlos Miguel|||0000-0001-5729-3041
Tipo de recurso: artículo
Fecha de publicación:2026
País:España
Institución:Universitat Politècnica de València (UPV)
Repositorio:RiuNet. Repositorio Institucional de la Universitat Politécnica de Valéncia
Idioma:inglés
OAI Identifier:oai:dnet:riunet______::812e66af200f3c91792a0510e8e2c1b8
Acceso en línea:https://riunet.upv.es/handle/10251/236065
Access Level:acceso abierto
Palabra clave:Large language models
Vision language models
Drone smart sensing
AI energy score
Autonomous navigation
UAV perception
Scene understanding
Descripción
Sumario:[EN] Drone-based smart sensing increasingly relies on Vision¿Language Models (VLMs) for real-time scene interpretation, obstacle detection, and autonomous navigation reasoning. Deploying such systems at scale demands not only high perceptual accuracy but also energy efficiency, a critical constraint on battery-powered Unmanned Aerial Vehicle (UAV) platforms, and linguistic flexibility for multinational operational contexts. We present a systematic benchmarking framework that jointly evaluates perception performance and inference energy for five open-source VLMs across thirteen languages spanning six language families, including three low-resource varieties (Arabic, Basque, and Luxembourgish). Using imagery sampled from the Berkeley DeepDrive 10K (BDD10K), each model is evaluated on four sensing tasks of increasing difficulty scored via a sentence-transformer backbone, with energy measured following the AI Energy Score methodology (Wh per 1000 queries) through continuous NVML-based GPU power sampling. Across 65 language¿model observations, LLaVA-1.6 achieves the highest perception score (¿¿¿¿¿¿=0.160 ) while Phi-3-Vision attains the best energy efficiency (66.3 Wh/1000 queries); energy consumption and task accuracy are statistically uncorrelated (Spearman ¿=0.001 ; ¿=0.995 ). A formal UAV inference energy model instantiated for four commercial platforms confirms LLaVA-1.6 as Pareto-optimal on heavy-lift platforms (DJI Matrice 300/350 RTK) and LLaVA-1.5 on the energy-constrained Matrice 30; compact UAVs such as the Mavic 3 Enterprise exceed the budget of all evaluated models at standard query rates. Friedman tests reveal significant cross-language variability in energy demands (¿2=40.43 ; ¿=3.5×10¿8 ) and navigation reasoning performance (¿2=13.35 ; ¿=0.010 ). Critically, we document a double penalty for low-resource languages, which simultaneously incur higher inference energy costs and lower task accuracy, with direct implications for equitable multilingual UAV deployments.