Generation of optimal configuration for a redundant manipulator with a trained neural network

Redundant manipulators have more degrees of freedom than what is absolutely necessary for performing a task. The extra degrees of freedom can be used for avoiding obstacles or to optimize certain performance indices like manipulability or task compatibility. Maximizing manipulability keeps the manipulator away from singularities and provides more velocity transmission ratios in all directions. Optimizing task compatibility improves the force/velocity transmission ratios in the specified directions. However, the real time implementation of various optimizing algorithms is difficult because of the need of large computing time. In the present work, robot configurations for an optimum performance index are computed throughout the workspace. These configurations are then used to train a layered feed forward neural network (FFNN). During operation of the robot, the trained neural net outputs optimal configurations in real-time. The neural net captures the gross behaviour of the training data rather than memorizing the individual data, as in a lookup table. Thus its output is smooth and ideally suited for control purposes. We have simulated this approach on a 3-DOF redundant planar manipulator and the results are discussed in this paper