Imagine you can do every athletic move perfectly, but you don't know when to use which one. This is a problem when teaching parkour to a virtual human. A new way of teaching is trying to solve this problem. It wants to remember how humans move and also be able to do new things when needed. The training data is very small, only 30 seconds of internet parkour videos. This new technique is trying to do what previous ones couldn't.
The new technique is tested on different levels, and it can solve them like a ninja. It can adapt and come up with new skills to get through the levels. The question is, can it solve longer levels? The answer is yes, it can. It can even solve levels with unseen obstacle arrangements.
The secret to this new technique is that it learns in two classrooms at the same time. One classroom teaches it to copy human movement, and the other teaches it to solve new obstacle courses. The AI controller learns from both classrooms and gets better at moving like a human. The judge and athlete keep improving together, and the results are stunning. The research paper is available for free, and this knowledge can be used by everyone.
