These videos demonstrate the gaits produced by neural
network controllers designed using the ENSO neuroevolution
method for a physical quadruped robot. These controllers
were evaluated in a physical simulation of the robot for
walking on flat ground (1) when all four legs of the robot
are functional and (2) when one leg is disabled to
simulate a real-world motor failure. The controllers
produced in the first experiment were evaluated further
for generalization by reducing the maximum speed of the
motors and by initializing one of the legs with a large
error. In each experiment, generalization was also tested
by placing the robot on different surfaces. These
experiments show that the evolved controllers generalize
well and are more robust against faults than a
hand-designed PID controller, demonstrating the potential
of the ENSO approach for real-world applications.