Looks like I own this section =)
Heres an old video of my first delving into random motor searches inside physics engines, and he walks along ok, besides the fact its skeleton isnt constrained properly, its getting one of its hip dimensions out of the dodgy rig, and it isnt hitting the ground properly. (It gumbies along at times.) The amount of motor searching is very small, its only 3 frames in front all directions, but even that small a time section can still get the robot to continually correct itself from where it is.
Heres a better one ive made a month ago, but I didnt finish the system to get it walking. Also! the cool thing (when i finally get it finished) is the new brain is using a more proper induction system that means I dont actually give it the physics code for it to sample from, it has to learn it before it can use it, and thats the main difference between a real robot and one thats stuck in the computer and cant get out. Two of the great practical benefits is the physics it learns is synched to its environment, and it possibly works without encoders if it at least has some clue to work out its body positions from.