To seamlessly transition from the simulation to the real world and thus overtake the reality gap, some generalization techniques that we presented earlier, such as domain adaptation and domain randomization, could be used. For example, in the paper Learning Dexterous In-Hand Manipulation, the authors trained a human-like robot to manipulate physical objects with incredible dexterity using domain randomization. The policy learned from many different parallel simulations that were designed to provide a variety of experiences with random physical and visual attributes. This mechanism that prefers generalization over realism overall has been key, considering that the system, when deployed, showed a rich set of in-hand dexterous manipulation strategies, many of which are used by humans as well.