Real-Time Constrained Tracking Control of Redundant Manipulators Using a Koopman - Zeroing Neural Network Framework

This study proposes a combined Koopman-ZNN (Zeroing Neural Network) architecture for real-time control of redundant manipulators subject to input constraints. An autoencoder-based neural architecture is employed to discover the bilinear Koopman model for manipulator dynamics in joint space using input-output data, which is subsequently integrated with a feed-forward neural network that maps the joint coor- dinates to end-effector Cartesian coordinates. The proposed ar- chitecture allows for efficient learning of highly accurate models using a significantly lower number of observable states compared to the previous studies. This learning architecture is then coupled with a ZNN controller, which offers a computationally inexpen- sive alternative to state-of-the-art Nonlinear Model Predictive Control (NMPC) controllers whose computational burden might render real-time control infeasible. The low-dimensional nature of the learned model, combined with a computationally inexpensive ZNN controller, facilitates real-time control applications with improved tracking accuracy. Simulation and experimental studies of trajectory tracking, including performance comparisons with leading alternative designs, are used to verify the efficacy of the proposed scheme.