GPT-6 Astra flew a drone through an office to follow a person

Given the instruction “find this person and follow them,” GPT-6 Astra flew a drone through an office on its own, located a specific man and tracked him. In the footage, a green box locks onto him.

What Andon Labs tested

AI lab Andon Labs posted the test on September 10, 2026. Astra navigated the drone through Andon Labs’ office autonomously, without a human pilot.

The lab’s benchmark, Drone-Bench, splits the mission into five tasks: building a 3D map from video, working out where the drone is, planning a collision-free route, detecting the person, and following them. Andon Labs says Astra is the first AI model to beat a human on each of the five tasks in at least one of its attempts.

The lab frames the demo as a way to discuss where society does and does not want AI, based on an accurate picture of what the technology can do.

How the benchmark scores it

The caveat is in the scoring. Each task is scored separately, so a strong attempt on mapping and a strong attempt on following do not have to come from the same run. Andon Labs itself notes that a real end-to-end run is harder because errors compound: a noisy map produces bad paths, and bad paths produce a crash.

That makes the result narrower than the video suggests. Beating a human on each task in at least one attempt shows that Astra can perform every step of autonomous tracking on a good try. It does not show that it can perform all of them reliably, one after another, in a single flight.

What to watch

Physical AI can now handle every step of finding and following a person when an attempt goes well. The missing piece is reliability across a full run, and that gap is the one worth tracking. If it closes, following a person indoors stops being a benchmark exercise and becomes a working capability, which is exactly the kind of use Andon Labs wants society to decide on while the picture is still accurate.