Drone-Bench: Follow
Agents · 2026-07-23
Drone-Bench's Follow task: the agent writes a control loop that calls a provided detector on the camera feed and keeps a simulated DJI Tello close behind a moving target in a MuJoCo model of Andon's office. A submission is scored on how much of the episode the target spends in the centre third of the frame and within a set distance, and a collision ends the episode and zeroes the remaining frames. The score is the mean over ten runs of each run's best submission, in percent; Andon's human-plus-AI baseline is 67.4.
Benchmark source: Andonlabs