NAVER's DroneCATS Finds MLLMs Fly, Fail to Declare Arrival

Qwen3.5-9B enters the 5-meter success zone in 90% of approaching episodes. It converts only 35% into wins. On average it declares arrival at 0.63× start distance, announcing the drone is there before it actually is.

That gap is the through-line of DroneCATS, a benchmark from a NAVER Cloud Drone AI Team led by Jaewoo Park and Geewook Kim. It slots multimodal LLMs into a swappable drone-agent role and tests them, without fine-tuning, on four capabilities: approaching a visible target, tracking a moving one, searching outside the initial view, and commanding a multi-drone fleet.

The paper's summation is blunt: "what separates a deployable edge model from a frontier model is not navigation, but the discipline to sustain a declared protocol and emit the correct terminating action." Smaller open models, the authors write, "navigate into the success radius more reliably than frontier models, yet lose the episode by declaring arrival prematurely or not at all."

Frontier models show the inverse shape. GPT-5 lands 60% of one-drone approaching runs. Claude Opus 5 lands 35%. Gemini 3.7 Flash lands 65% and holds under multi-drone commanding at 80% success flying four drones at once. GPT-5 collapses on that same commanding task to 20%. Small open models, handed multiple platforms, are "blindly copying a single coordinate across distinct views."

The write-up lands in a week when the compact-embodied-model theme is stacking up. LightNav-0, published the day before, argued a small VLM tops all ten navigation sims it tested. It sits inside a broader run of 61 multimodal stories we've logged in the last 90 days.