FlockGPT: Guiding UAV Flocking with Linguistic Orchestration
Artem Lykov, Sausar Karaf, Mikhail Martynov, Valerii Serpiva, Aleksey Fedoseev, Mikhail Konenkov, Dzmitry Tsetserukou
Abstract
The paper introduces the first rapid drone flocking control using natural language via generative AI. This approach enables the intuitive orchestration of a drone flock of any size to form desired geometries. The core innovation is a new interface based on Large Language Models (LLMs) that allows user interaction and target geometry description. Users can modify or comment on the flock geometry model interactively. By integrating flocking technology with target surface definition through a signed distance function, smooth and adaptive swarm movement is achieved. A user 1study on FlockGPT showed high intuitive control over drone flocking. Participants without prior experience constructed complex shapes in a few iterations and accurately recognized the figures. The study demonstrated a high recognition rate for six geometric patterns generated through the LLM-based interface, with an 80% mean and up to 93% for cube and tetrahedron patterns. Users reported low temporal demand (NASA-TLX score of 19.2), high performance (NASA-TLX score of 26), attractiveness (UEQ score of 1.94), and hedonic quality (UEQ score of 1.81). The FlockGPT demo code is available at: https://github.com/Taintedy/flock_gpt