Machine learning - balancing a stack of balls

A genetic algorithm is used to train a neural network controller to balance a stack of balls.

A genetic algorithm is used to train a neural network controller to balance a stack of balls by applying a lateral force and a torque to the bottom ball. The capsim physics engine is used to produce a realistic simulation for each attempt. It can learn up to 4 balls.
http://www.youtube.com/watch?v=c4vADjwOFNc