1. Train a neural network A to reproduce a movie screenplay by watching a movie video and audio as input.
2. Train a neural network B to act out character X in a screenplay like a movie using a robotic body R. The robotic body R should have a video camera and mic. X should be specified as an input to neural network B. The screenplay should also be input to network B.
3. Train a neural network D to generate screenplays of movies that have the archetype of the “Hero’s Journey“.
4. Train a neural network E to detect the protagonist in a screenplay of movies.
5. Train a neural network F to measure the well being of the protagonist in the screenplay of movies.
6. Train a neural network G to generate one future screenplay of an incomplete screenplay (generated by D) such that it increases the well being of the protagonist as calculated by neural network E and F operating on the aforementioned incomplete screenplay as input.
7. Take the video camera and mic input from robotic body R and feed it to neural network A. Take the screenplay generated by neural network A and feed it to neural network G (which uses networks E and F to calculate who the protagonist of the screenplay is and also maximize the well being of the protagonist), and take the future screenplay generated by neural network G and feed it to neural network B which uses robotic body R. Also feed output of neural network E as the input X used by B.

There is no protagonist. But the above ensemble will act out the part of a protagonist in an incomplete screenplay such that it increases its "well being".

Networks A and D could be trained simultaneously. A and D could even be the same network operating in different modes or hyper-parameters. But conceptually they should generate different screenplays. Network A generates screenplays of how things are. Network D generates screenplays of how things should be for an incomplete screenplay using Hero’s Journey stories as a training set.