The Reality Gap refers to the disparity between a  system's simulation and its behavior in reality. For a simulation based optimization of a robot controller it poses the risk of finding results that transfer poorly to the real robot.

In my talk, I will present the Transferability Approach, recently published by Koos et al. [2013], which optimizes both the quality of the result in the simulation and the transferability, i.e., how close the simulation resembles the reality for the performed behavior. The transferability of individual controllers is estimated by a surrogate model minimizing the required number of evaluations on the real robot.

In parallel, the goals of my master thesis are presented: (1) the implementation of the \emph{Transferability Approach} with Multi-Objective Evolutionary Algorithms for a robotic application and (2) the expansion of the surrogate model by probabilistic regression approaches. The latter aims at a more generic representation of the behavioral features and at a faster convergence due to a ­­­ ­­­- possibly - more efficient testing policy.

S. Koos, J.-B. Mouret, and S. Doncieux. The transferability approach: Crossing the reality gap in evolutionary robotics. IEEE Transactions on Evolutionary Computation, 17(1):122-145, February 2013.