RL With Human Feedback: The Most Prevalent Application Today

The first RL with human feedback paper from Oakland AI was actually even earlier, I think maybe even 2017 or 2016. The reason was because normally when you want to do reinforced learning with robots, the typical paradigm would be you write some piece of code that is the reward function. And then it tries to optimize that reward that's equivalent to the score in the game. But for a backflip, how do you write a pieces of code that says this was the quality of the backflip? It's not clear how you how you do that part.