[karpathy]

I've worked on similar projects during my Master's degree. An important trick I learned was to make the controller a function of some kind of discrete feedback in the world (e.g. when the foot hits the ground). In your equation above for example, a simple way to incorporate this would be to reset simulation_steps to zero every time a foot hits the ground.

This is also an important feature in the SIMBICON controller, which is arguably the simplest and most robust walker system. (http://www.cs.ubc.ca/~van/papers/Simbicon.htm)

EDIT: and yes, Reinforcement Learning is much more effective way of attacking this type of problem (e.g. see some recent work from Sergey Levine http://www.eecs.berkeley.edu/~svlevine/), but I also agree with the author that GA are fun :)