he operational design domain (ODD) of a system defines the conditions under which its components can be trusted to maintain safety. A runtime monitor of an ODD predicts, based on a sequence of observable data, whether the system is about to exit its ODD. For black-box systems, one key challenge in learning an ODD monitor is achieving high accuracy. Earlier approaches to learning ODD monitors have relied on passive learning techniques that generate monitors based on simple binary definitions of ODDs. These traditional monitors typically decide whether to switch between a safe and optimized controller based on a prediction of whether the system is about to exit its ODD. An ODD monitor should therefore decide based on the current valuation of its feature space, to switch to an appropriate controller. Our goal is to introduce a contextual formalization of the ODD monitoring problem and present a framework for learning ODD monitors using an approach based on contextual bandits. This approach can be applied to improve safety in autonomous vehicles. For the experimentation, we will (1) train several Convolutional Neural Networks (Deep Learning) to act as controllers of the autonomous vehicle, and (2) apply Machine Learning and Multi-armed bandit techniques to monitor and find the best controller in every situation. Multiple simulations will be required to get information about the stability of the algorithm.