one of the bets i'm making is that "learning the environment" is going to give my team the ability to learn policies that exploit the physics of our world

basically, replacing the physics with a neural network

unfortunately that is going to cost too much time, so our own physics simulator will have to suffice. i'm going to spend this weekend thinking about how to fit things, at which level to cut it. e.g., what part of my simulator gets replaced by the neural network?

I think what I really want to do is create a physical room with pillow walls and have a bunch of robots lined up for deployment that can easily recover themselves, and let them move around all night collecting data and finding the situations they can't predict