From Video Game to Real Robot

Can you train a deep learning agent in video game and then transfer this learned information to a real world robot? How can we handle the differences in visuals and action spaces between the game and reality?