All the secret settings are not mentioned, but the article does have a video of favorite toys by II. The robot does not know the rules of the game and makes his conclusions empirically, as he does with managing air conditioning, sensors and other devices. Ahh, watching such videos makes me want to get out my old video console and joystick.
By the way, this is the principle used by Google - accessible to all, yet used by everyone differently. Based on Deep learning, neural networks are developed, image recognition is constructed, specified objects are searched for and parameters are optimized, which was mentioned earlier. I'm sure many of you, colleagues, work with something like this (tell us about your experience in the comments, I'm very interested!).