Autonomous mobile robots (AMR) operating in the real world often need to make critical decisions that directly impact their own safety and the safety of their surroundings. Learning-based approaches for decision making have gained popularity in recent years, since decisions can be made very quickly and with reasonable levels of accuracy for many applications. These approaches, however, typically return only one decision, and if the learner is poorly trained or observations are noisy, the decision may be incorrect. This problem is further exacerbated when the robot is making decisions about its own failures, such as faulty actuators or sensors and external disturbances, when a wrong decision can immediately cause damage to the robot. In this paper, we consider this very case study: a robot dealing with such failures must quickly assess uncertainties and make safe decisions. We propose an uncertainty aware learning-based failure detection and recovery approach, in which we leverage Decision Tree theory along with Model Predictive Control to detect and explain which failure is compromising the system, assess uncertainties associated with the failure, and lastly, find and validate corrective controls to recover the system. Our approach is validated with simulations and real experiments on a faulty unmanned ground vehicle (UGV) navigation case study, demonstrating recovery to safety under uncertainties.