Abstract: Traditional landing methods often fall short in response to the limited semantic sensing capabilities of UAVs in dynamic,unstructured environments and their reliance on fixed,context-insensitive safety factors.To address these limitations,a hybrid framework,LLM_Land,is proposed that combines a large language model (LLM)with model predictive control (MPC),starting from a visual language encoder (VLE)(e.g.,BLIP),which converts real-time images into succinct textual scene descriptions,which is used by a retrieval-augmented generation (RAG)-equipped lightweight LLM (e.g.,Qwen 2.5 1.5B or LLaMA 3.2 1B)to categorise scene elements and infer context-aware safety buffers,e.g.,3 m for pedestrians and 5 m for vehicles,and the resulting semantic flags and unsafe zones are subsequently fed into the MPC module,enabling real-time trajectory replanning for collision avoidance whilst maintaining a high level of landing accuracy.The proposed framework is validated in the ROS-Gazebo simulator,which consistently outperformed the conventional vision-based MPC baseline,and the results showed a significant reduction in near-miss accidents due to dynamic obstacles,while maintaining accurate landings in a cluttered environment.

Key words: UAV semantic perception, autonomous landing, large-scale language modelling, model predictive control