How do you propagate and handle uncertainty throughout the autonomous driving stack?

Uncertainty handling: 1) Perception uncertainty - object detection confidence scores, probabilistic tracking with state covariance, ensemble methods for multiple estimates; 2) Localization uncertainty - covariance from SLAM/localization filter, map uncertainty representation; 3) Prediction uncertainty - multimodal trajectory distributions, scenario probabilities; 4) Propagation through planning - robust planning considering uncertainty bounds, contingency planning for multiple scenarios, probabilistic safety verification; 5) Control - stochastic MPC, tube-based approaches guaranteeing constraint satisfaction under uncertainty. Decision making incorporates uncertainty: Conservative behavior when uncertainty is high, active perception (moving to reduce uncertainty), and transition to safer states when confidence is low. Uncertainty calibration ensures confidence scores are meaningful.