The example is an aircraft flying from one point to another through uncertain crosswinds while avoiding a zone of poor connectivity, controlled by a model predictive policy. The paper is explicit that the aircraft is a vehicle for the method rather than the subject.

Calibration is the real topic. An MPC policy carries parameters governing how cautious it is, and choosing them well matters because the uncertainty is real: the wind is not known in advance and the avoidance zone is a hard constraint. Too conservative wastes fuel and time, too aggressive risks the constraint.

Pick-to-Learn addresses this by selecting which observations to learn from rather than treating all data equally, which suits a setting where a handful of difficult cases determine whether the policy is safe and the easy majority tells you almost nothing.