“OpenAI trained a control policy entirely in simulation to control a human-like robot hand for solving a Rubik's Cube.”