In this thesis we deal with the problem of using deep reinforcement learning to generate robust policies for real robots. We identify three key issues that need to be tackled in order to make progress along these lines. ...