Large Language Model-based Multi-Task UAVs – Towards Distilled Real-Time Interactive Control

In recent years, unmanned aerial vehicles (UAVs) have gained widespread application due to their advantages such as high durability, low cost, ease of implementation, and flexibility. However, traditional UAV applications often focus on a singular task, requiring redesign and retraining when there are changes in scenarios and objectives. To address this issue, we propose a construction approach for a general UAVs task framework based on Large language models (LLMs), enabling adaptive execution of various UAV tasks in a zero-shot manner. In our methodology, we initiate the extraction of capabilities from black-box LLMs serving as victims, employing a mimicry attack to transfer these abilities into the mimic model CodeT5. Subsequently, fine-tuning is performed using UAV domain data, ensuring that the imitation model is specialized in generating UAV control code. To further compress the model for deployment on UAVs, we employ knowledge distillation techniques, utilizing the fine-tuned imitation model as a teacher to train a high-precision lightweight student model. This generative approach facilitates on-the-edge task execution and auxiliary decision-making. During the execution phase, users can achieve real-time online interaction with UAVs through natural language commands.