Their machine-learning model uses images of the robot's environment to generate a 2D action space, which includes the robot and the area around it. They simulate robot motion using what is known as ...