Human’s imagination capability provides recognition of unseen environment which should be improved in robots in order to have better mapping, planning, navigation and exploration capabilities in the fields where the robots are utilized such as military, disasters, and industry. The task of completion of a partial scene via estimating the unobserved parts relied on the known information is called scene extrapolation. It increases performance and satisfies a valid approximation of unseen content even if it is impossible or hard to obtain it due to the issues related with security, environment, etc. In this survey paper, the studies related to learning-based scene extrapolation in robotics are presented and evaluated taking the efficiencies and limitations of the methods into account to provide researchers in this field a general overview on this task and encourage them to improve the current studies for higher success. In addition, the methods which use common datasets and metrics are compared. To the best of our knowledge, there isn’t any survey on this essential topic and we hope this survey will compensate this.