In this talk, it is shown how input-output data can be leveraged to improve the performance of robots in an industrial setting. A common task for industrial robots is trajectory tracking, where a robot needs to follow a predefined path in space and time. These trajectories are often periodic, e.g., if the same path is followed repeatedly through a large warehouse by mobile robots or if a robotic manipulator does the same motion repeatedly at an assembly line. The model mismatch to a nominal model, for instance based on first principles, is learned using Gaussian process regression and then leveraged using a robust model predictive controller. A special focus is put on the trade-off between exploiting the existing model for immediate performance gains and exploring for further model refinement. As more input-output data becomes available, the reference is adapted in an iterative fashion lowering the tracking error (exploitation) while at the same time exploring previously unseen reference solutions (exploration). This learning procedure is done in a safe and controlled manner by leveraging the posterior variances of the Gaussian processes, indicating the underlying model uncertainty. Guaranteeing the safe operation of the robots, especially during these learning phases, is an essential requirement for the application of any data-based control method in practice. The effectiveness of this approach is validated in meaningful real-world experiments on custom-built omnidirectional mobile robots as well as robotic manipulators.