Adaptive NMPC-RBF with Application to Manipulator Robots

Abstract

Control algorithms must deal with model uncertainties and disturbances, making them perfect for real-world applications. In addition, increased computational power in the industrial field allows the implementation of advanced control algorithms such as nonlinear model predictive control (NMPC), which is an optimal control scheme that includes system and control constraints imposed by robot dynamics and the environment. Nevertheless, modeling the robot and its environment is a complicated task due to high nonlinearities, such as model uncertainties in the form of complex unmodeled dynamics, varying payloads, and parameter mismatch, leading to fast degradation of NMPC; therefore, online adaptation laws that improve the performance even in unknown environments are needed. Due to the facts presented before, this work combines the universal approximation of RBFNN and the optimality offered by NMPC in a unified adaptive framework that guarantees good performance even under uncertainties and unmodeled dynamics. The proposed framework is tested in simulation using a 2-link planar robotic arm (SCARA Bosch SR-800), where optimization techniques were used to identify the robot's dynamics. Finally, a comparison of the proposed architecture with a baseline nominal NMPC is made with particular attention to trajectory tracking performance, demonstrating the reduction of the tracking error over non-adaptive NMPC.