Drone-Bench is a new benchmark designed to evaluate the capabilities of AI models in writing code for autonomous drone surveillance in real-world environments. The benchmark simulates a low-cost drone navigating an office to find and follow a person, encompassing five key tasks: reconstruction, localization, navigation, detection, and following.

Each task is scored against a human-programmed baseline, allowing researchers to assess how closely AI models can replicate or surpass human-level performance in these surveillance functions. The benchmark aims to highlight the growing physical autonomy of AI models, which are increasingly moving beyond chat interfaces into practical applications. This development raises important questions about the potential for misuse of increasingly capable AI systems.

The benchmark's tasks are sequential, with the output of one feeding into the next. For example, reconstructing the office environment (Reconstruct) provides a map used for drone navigation (Navigate) and localization (Localize). Detection identifies targets, and Following uses this information to keep the drone on track. The evaluation process involves iterative refinement, with models attempting to improve their code over multiple submissions within a single run. Frontier models like gpt-4o-may and gemini-3.1 were tested, showing varying degrees of progress towards the human baseline across the tasks.