“Initial attempts to transfer a policy trained in simulation to a real robot for the Rubik's Cube task failed due to unmodeled differences between the simulation and the real world.”