Where it shines
A certified safety envelope
A controller does not need a sharp forecast of the next state, it needs a region the true state is guaranteed to stay in.
Robotics and control are a natural home for conformal prediction, because the question is almost always one of containment: will the true state lie inside a region I can plan around? Calibrate the prediction error of a state predictor and you get a radius \(q\) such that the realised state stays within \(\hat\mu(t)\pm q\) at least \(1-\alpha\) of the time, a distribution-free safety tube around the nominal trajectory. If that tube clears the obstacle, you have a certificate: the system stays out of the keep-out zone at the chosen confidence. Lower the risk budget \(\alpha\) and the tube widens; raise the obstacle and the margin returns. The certificate flips between “clear” and “not” exactly when the guaranteed tube touches the obstacle.
Green: the nominal trajectory and its conformal safety tube. Red zone: the obstacle. The realised states stay inside the tube at the target rate; the certificate is “clear” whenever the whole tube sits below the obstacle over the horizon.
Takeaway. Safety is a coverage property, and coverage is what conformal prediction guarantees with almost no assumptions. The tube is the deliverable; a better state predictor makes it tighter and buys back clearance, but the guarantee that the true state is contained at level \(1-\alpha\) holds regardless. This is the marginal guarantee doing exactly the job it is built for.