Inverted Kinematics of a Redundant Manipulator with a MLP Neural Network

The article describes the solution of the inverse kinematics of a serial redundant manipulator. Reachable endpoint positions are generated randomly based on forward kinematics, described by Denavit-Hartenberg notation. If the randomly generated position is part of the area in which the desired movement is to be solved, it is recorded in a special structure where each cell corresponds to a small range of the endpoint coordinates. Up to thousands of possible combinations can be recorded in each of the cells. Based on this data, inverse kinematics cannot be solved for a redundant manipulator because the same point can be reached by infinitely many combinations of arm settings. Therefore, the prepared angle settings for reaching an individual cell are first evaluated with a suitable additional fitness function. Additionally, solutions that do not represent continuous movement are filtered. After this process, described in this article, the few best solutions are then selected from each of the cells and used to train a simple MLP (multilayer perceptron) neural network. Based on data from forward kinematics, the network is trained to obtain an inverse kinematics solution. The result is a smooth motion whose accuracy is limited by the cell size used and the amount of samples generated.