[qwert7890]

Simplest RL algorithm (Q-learning) achieves 100m in QWOP: https://www.youtube.com/watch?v=e27TUmMkOA0

Although it found and exploited a local maximum of "knee scraping" technique (which humans can replicate) :)