7 Conclusions and perspectives
The main goal of this paper was to describe the use of a gradient descent learning algorithm (RPROP) to adjust weights of a recurrent neural network used in a legged robot gait control. For this, it was described the properties of the neural network used and the methodology for the generation of the examples database used in the neural learning. The neural networks results are more robust than the results obtained by the controller based in the locus based functions, and the ANN has the advantage that it does not require the inverse kinematics calculation.
The perspectives of this work includes to use motion capture devices to generate the learning database files, and to use more complex robots, like bipeds. We want too use more sensor information (gyroscope, accelerometers, sonars) to improve the quality of the obtained gaits. The perspectives also includes to adapt gait control in order to make possible control robots moving over irregular surfaces and to climb or to descend stairs.