Research questionHow can multimodal models control drones reliably under prompt-defined action protocols and terminate at the right time?A drone agent may navigate into a target area yet fail because it loses the required action protocol or declares completion too early—or never does. Reliable control therefore depends on sustained action selection and correct termination, not navigation alone. Latest papersRecent research connected to this question, newest first.Evaluating Multimodal LLMs as Generalist Vision-Language-Action Agents for Drone Control: Commanding, Approaching, Tracking and SearchingThe source evaluates a swappable multimodal language model in a drone control loop whose action space is specified in the prompt, without fine-tuning or function-calling schemas. It tests frontier and open models, including 2B-parameter models, on approaching, tracking, searching, and multi-drone command; reported failures include premature or missing arrival declarations and copying one coordinate across distinct views.research paper · Sep 1, 2026