The _rst step in the neural gait control was to create the examples data set to be employed in the ANN learning. The half-ellipse controller, described above, was adapted in order to generate a log _le, containing records of: the current joint angles at time t, the sensors state (bumpers below the paws indicating the leg phase . 0 = swing phase; 1 = sustained phase), and the desired joint angles for the next time t+1. In the near future we plan to use motion capture devices to provide the examples for the training database. The learning database was created with 6000 examples, (3000 for the learning and 3000 for the generalization test), each one with 16 inputs and 12 outputs. The_rst 12 inputs are the current joint angles, and the 4 last inputs are obtained through bumpers. The 12 outputs are the expected joint angles at time t + 1. With the neural network controller, there is no need to calculate the inverse kinematics. The neural network used in the experiments was an Elman network, as shows the Figure 5. This _gure is just illustrative, because we used in our experiments more neurons and connections per layer.