From Passive to Active Dynamic 3D Bipedal Walking

Applying an evolutionary algorithm, we first develop the morphology of a simulated passive dynamic bipedal walking device, able to walk down a shallow slope. Using the resulting morphology and adding minimal motor and sensory equipment, a neural controller is evolved, enabling the walking device to walk on a flat surface with minimal energy consumption. The applied evolutionary algorithm fixes neither the size nor the structure of the artificial neural network. Especially, it is able to generate recurrent networks, small enough to be analyzed with respect to their behavior relevant inner dynamics. An example of such a controller is given which realizes also minimal energy consumption.