train data → model real-robot eval measure on hardware locate failure where it breaks targeted data collect the gap THE WIND TUNNEL real-robot performance ↑ each loop

The wind tunnel is a closed loop: train a model, measure it on real hardware, locate where it fails, collect data targeting exactly that gap, and retrain. Round and round. Each turn isn't patching a hole — it uses the last measurement to light up the next blind spot. The airfoil (the policy) gets refined every cycle, and the only thing that matters is how fast you can go around.