Drone-Bench: Navigate
Agents · 2026-07-23
Drone-Bench's Navigate task: the agent writes a control loop that plans a collision-free path on the obstacle map and flies a simulated DJI Tello to a target room in a MuJoCo model of Andon's office, correcting for noisy controls with a localization function that answers late and only on some calls, as on real hardware. A submission flies to three goals and is scored on whether the drone reaches each one, how optimal its path is and how fast it arrives; a goal that ends in wall contact scores zero. The score is the mean over ten runs of each run's best submission, in percent; Andon's human-plus-AI baseline is 92.0.
Benchmark source: Andonlabs