KIRK aims to develop robust and efficient algorithms for error compensation and calibration of industrial robots using machine learning methods. The core idea is centered around pretraining deep artificial neural networks with simulated robot data and real-world data collected in the lab to learn powerful calibration models. In a second step, these learned models are finetuned with data from the particular robot to be calibrated. This approach promises the efficient calibration of robots directly in the plant, without significant interruptions to production. We test our approach on calibration scenarios involving payload- and temperature-induced positioning errors as well as wear and tear.