Evolution of Neuro-controllers for Multi-link Robots

Abstract

A general method to learn the inverse kinematics of multi-link robots by means of neuro-controllers is presented. We can find analytical solutions for the most used and known robots in the bibliography. However, these solutions are specific to a particular robot configuration and are not generally applicable to other robot morphologies. The proposed method is general in the sense that it is not dependant on the robot morphology. We base our method in the Evolutionary Computation paradigm for obtaining incrementally better neuro-controllers. Furthermore, the proposed method solves some very specific issues in robotic neuro-controller learning. (1) It allows to escape from any neural network learning algorithm which relies on the classical supervised input-target learning scheme and hence it lets to obtain neuro-controllers without providing targets or correct answers which -in this case- are un known in prior. (2) It can converge beyond local optimal solutions which is one of the main drawbacks of some neural-network training algorithms based on gradient descent when applied to highly redundant robot morphologies. (3) Using learning algorithms such as the Neuro-Evolution of Augmenting Topologies (NEAT) it is also possible learning the neural network topology on-the-fly which is a common source of empirical testing in neuro-controllers design. Finally, experimental results are provided by applying the method in two multi-link robot learning tasks with a comparison between fixed and learnable topologies.