The mobility limits of vehicle systems depend on internal and external parameters that are varying throughout the system’s life cycle due to aging or operational conditions changing. An intelligent vehicle system is supposed to have sufficient self-learning capability to continually adjust its decision and control algorithms to maintain high performance against such parameter changes. However, conventional learning methods (such as standard reinforcement learning) do not guarantee safety, in terms of strictly satisfying prescribed safety constraints, during the learning process. This makes them not suitable as online methods in safety-critical applications such as autonomous driving. This talk will introduce recently proposed governor-based approaches to safe learning. In these approaches, a governor is an add-on device that monitors and minimally adjusts the learning process to ensure all-time satisfaction of safety constraints. An overview of the governor theory is accompanied by simulation examples to demonstrate the application of these approaches in autonomous driving safety.