Low-Altitude Hunter: A System for Vision-Language Anti-UAV Task With MLLM
Abstract
With the rapid expansion of the low-altitude economy (LAE), safety governance in urban airspace has become increasingly important, and physical capture is widely regarded as a promising, safe, and efficient approach to countering non-cooperative uncrewed aerial vehicles (UAVs). However, physical capture in complex and rapidly changing urban environments remains difficult to achieve: traditional end-to-end control methods suffer from limited generalization, while aerial vision-language action (VLA) paradigms, despite their stronger generalization potential, are still primarily developed for static target-reaching tasks. To bridge this gap, we present the Low-altitude Hunter system for the vision-language anti-UAV (VLAU) task with a multimodal large language model (MLLM). Specifically, we build a high-fidelity simulation platform that integrates diverse urban scenes, multiple UAV models with realistic flight dynamics, and flexible onboard sensor configurations. Building upon the proposed platform, we further construct a large-scale VLAU that covers diverse spatial layouts, motion trajectories, and task difficulty levels. Then, we develop an MLLM-based framework that unifies offline imitation learning (IL), closed-loop dataset aggregation (DAgger), and online reinforcement learning (RL) to jointly improve multimodal reasoning, continuous control, and online adaptability under distribution shifts and long-horizon interactions. Extensive experiments show that the proposed framework achieves higher success rates, stronger robustness, and better policy adaptability than conventional baselines in VLAU scenarios.