Drone-Bench
Agents · 2026-07-23
Andon Labs' Drone-Bench measures how well AI models write the code for a simple surveillance demonstration on a low-cost DJI Tello drone, split into five tasks with hidden test sets: Reconstruct (a 3D model and 2D obstacle map from office videos), Localize (the drone's pose from one camera frame), Navigate (flying to a room), Detect (a target person's bounding box in drone video) and Follow (keeping that person in frame); flight runs in a MuJoCo simulation of Andon's office and perception on real recordings. In each run a ReAct agent with code-editing, image-reading and submission tools submits up to ten implementations in one context and keeps its best score, and each model has ten runs per task. Progress is the mean over the five tasks of the average run score, capped at Andon's human-plus-AI baseline and divided by it, in percent; 100 means the average run matches the baseline on every task.
Benchmark source: Andonlabs