RSF-MPC: A Manipulator Dynamic Modeling and Control Method Based on Deep Koopman Operator

This article proposes a deep extended dynamic mode decomposition structure based on the deep Koopman operators, residual and multihead attention modules (RS-DEDMD). The RS-DEDMD structure is designed for manipulator systems, using a 13-D Koopman operator to build the high-precision approximate linear model of the manipulator. Experiments show that the Koopman operator trained by RS-DEDMD have lower dimensions, higher prediction accuracy, fewer training iterations and faster convergence than other structures. A model predictive control system based on RS-DEMD (RSF-MPC) is constructed to verify the feasibility of using linearized model-based control. Experiments and simulations show the RSF-MPC is stable and outperforms other control methods in prediction and control capabilities. This demonstrates the accuracy of the linear model constructed by RS-DEDMD.