MAVLingo: ‘A Natural Language Command Classification’ Dataset for MAVLink-Based Drones

Unmanned vehicles (UVs) hold significant potential across various applications, but their operation execution with User Interface(UI) can be complex and demanding for human operators. High quality datasets are required to enhance the command classification for unmanned systems. To address this gap and enhance the ability of complex commands execution with natural text, we introduce MAVLingo, a novel dataset for classifying drone commands. MAVLingo comprises 2,629 data points, synthesized from gold data and augmented with synthetic data generated from ChatGPT using few shot prompting. We benchmark the dataset on leading open source large language models (LLMs) LLaMa2 and LLaMa3 by fine-tuning and Retrieval-Augmented Generation (RAG) based approach. The MAVLingo dataset is utilized in a search and rescue drone system via a chat window, enhancing unmanned systems' operational efficiency and responsiveness.