Autonomous systems such as automated driving systems, autonomous ships or industrial collaborative robots are data-intensive software systems that embed self-adaptive and self-reasoning capabilities. Even though increased autonomy is highly desirable, validation challenges come with it as autonomous systems are considered non-testable. Indeed, their exact behaviour is hardly predictable as they depend on datasets, trained models and observable execution environments. Fortunately, Metamorphic Testing techniques have been developed for a quarter of a century to support test engineers in validating safety- and business-critical software systems. My talk will review the main principles of Metamorphic Testing and detail two specific approaches where reinforcement learning and automated AI planning are used to address the testability problem of selected non-testable autonomous systems. A few lessons learned will allow me to draw some research perspectives for the progress of Metamorphic Testing.