Observing the gait behavior of the robots trained using neural networks, they are much more stable and robust than the gait behavior when controlled directly by the half-ellipse, because the neural network uses sensor information (bumpers) and have the capacity to generalize to non expected situations. The Figure 7 shows the