Original title: Evaluating Multimodal LLMs as Generalist Vision-Language-Action Agents for Drone Control: Commanding, Approaching, Tracking and Searching
Jaewoo Park, Minyoung Lee, Sukmin Seo, Moonbin Yim +10 · September 2026
Can smart artificial intelligence models that process pictures fly drones using only written instructions? Most systems limit what these models can choose. In this work, researchers gave models broad control. They introduced DroneCATS-Agent, which lets different models plug directly into the drone loop without extra training or preset tool structures. The drone relies on the model to turn, search, think when unsure, and state when it arrives. Testing frontier and open models, scaling down to 2B parameters, revealed a surprise. Flying was not the issue. Small open models often steered into the goal area more reliably than frontier models. Yet, they still failed missions because they announced arrival too early or never said it at all. In multi-drone tasks, small models failed by blindly copying one coordinate across different camera views. Overall, spatial vision works, but following the action rules breaks down.
Imagine guiding a friend through a maze by text messages. They follow the directions well but keep claiming they've reached the exit before actually getting there, or never say when they're out. In this paper, drones controlled by AI models navigate to targets but fail to correctly announce arrival, showing that the challenge isn't the navigation but the protocol for completion.
For industries like delivery or surveillance, autonomous drones need to operate reliably. If the AI controlling them can't properly signal task completion, they can't be trusted in real-world settings.
The study only evaluates models on the DroneCATS benchmark and doesn't provide a solution for improving action protocol adherence at low compute costs. It identifies the gap but doesn't close it.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Evaluating Multimodal LLMs as Generalist Vision-Language-Action Agents for Drone Control: Commanding, Approaching, Tracking and Searching — Jaewoo Park, Minyoung Lee, Sukmin Seo et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.