As per the tradition [0]:

In the days when Sussman was a novice, Minsky once came to him as he sat hacking at the PDP-6.

“What are you doing?”, asked Minsky.

“I am training a randomly wired neural net to play Tic-Tac-Toe” Sussman replied.

“Why is the net wired randomly?”, asked Minsky.

“I do not want it to have any preconceptions of how to play”, Sussman said.

Minsky then shut his eyes.

“Why do you close your eyes?”, Sussman asked his teacher.

“So that the room will be empty.”

At that moment, Sussman was enlightened.

[0] http://catb.org/jargon/html/koans.html

The more times I read this, the deeper it gets, and the more I appreciate modern machine learning.

Well the very first attempt this guys made (https://arxiv.org/pdf/1804.10332.pdf#subsection.6.1) was let the model learn from scratch (open loop, large bounds feedback). An agile galloping gait emerged automatically also in my simulation (https://github.com/nicrusso7/rex-gym#galloping-gait---from-s...), even if the gait was 'noisy'

I've seen a crawling robot trained with a string which would pull it back to the starting position after each episode so it could automatically try over and over again. I'd love to see someone doing that with a biped and just lift it up on a hoist to reset its orientation and position whenever it fell over. But really, it might not make a lot of economic sense if you can simulate the physics like these guys are doing.

I'm not sure about doing that with physical robots, but I've seen a couple interesting iterations using simulated robots

https://www.youtube.com/watch?v=gn4nRCC9TwQ