Zero-Fine-Tuning Safety-First Closed-Loop Decision Making for UAVs

Many UAV systems can already understand scenes and tasks, but they still fall short in two simple ways in real-world use. Task goals and safety rules are often written as free-form text, which makes it hard to verify what the system is actually allowed to do. When the model is unsure, it may still output actions that are not safe. In this work, we ask how a UAV can stay within a clear flight boundary and still make decisions that are reliable and traceable, without adding extra training cost. We design a simple decision loop: the task, current state, and safety rules are first written in a fixed template; the model then outputs an action together with a confidence score; if the confidence is low, the system chooses to hover or not execute the action; at the same time, an independent safety check blocks any action that may break the rules. On two public UAV test sequences, this loop keeps violations at zero and achieves higher decision accuracy than a standard baseline, while keeping the processing time in the same range, and a single threshold chosen on the first sequence can be reused on the second without any additional tuning.