An open benchmark for AI approach control at a real, busy airport, starting with London Gatwick. A controller takes over each real arrival as it comes within 40 NM of the airport and works it onto final while you watch on a radar screen. The benchmark scores it on the same measures as the real controllers who handled the same aircraft.
The simulator is built and runs on real traffic. That’s 21,160 Gatwick arrivals from 60 days of open ADS-B data, each hour flown in its real winds. A fixed safety check blocks any unsafe instruction. There’s no learned AI yet. For now a hand-written rules controller flies the traffic as the first yardstick. It keeps up in a quieter hour and falls well behind the real controllers in a busy one. The first learned controller comes next. It’ll be a small neural network trained by reinforcement learning in the simulator, on a laptop. It starts with speed control only, then builds up stage by stage to full vectoring, holds, go-arounds, runway changes and storms.
I’m looking for collaborators. The biggest gap is controller knowledge, so I’d most like to hear from anyone who controls Gatwick, on VATSIM or for real. I’d also welcome help with the reinforcement learning or the aviation data. If that sounds like you, get in touch.