The fundamental challenge in autonomous systems is how to learn to perform useful tasks autonomously across fields, such as robotics or customer service. The key bottleneck in learning the operation logic in such tasks involves exploration, which in the real world is costly and unsafe. Simulators of the world can be built but they are imperfect models of the real world. We will bridge this reality gap between real and simulated worlds by learning accurate simulators, matching them with real world while ensuring safety. We combine currently disjoint modelling methods into a complete next-generation autonomous system with a team of complementary expertise in robotics, dynamical modelling and machine learning from Aalto University and Finnish Center for Artificial Intelligence FCAI. The developed results will be taken into practise by a global cargo company Cargotec.