Controle preditivo de trajetórias com restrição no sinal de controle via redes neurais

Abstract

This dissertation investigates the application of non-linear predictive control with constraints on the control signal for robotic systems characterized by a cascade control architecture. The proposed control strategy uses a feedforward neural network model to assist the controller when applying restrictions on the secondary loop control signal. This secondary loop has a proportional-derivative (PD) controller. The neural network model is trained based on the initial and final conditions of position and velocity for the system to be controlled. The trained model infers, as output, the peak derivative of torques or future forces that the PD controller must respect based on the commands from the outer loop. As a case study, model predictive control (MPC) is applied as the outer loop, and proportional-derivative (PD) controllers are used as the inner loop for a 2R planar manipulator robot and a quadrotor drone. The control application with the proposed architecture was tested on both systems through computational simulations in MATLAB. The approach resulted in smoother trajectories for the manipulator robot during obstacle avoidance. However, during some experiments with obstacles, the controller allowed the drone to realize significant deviations from the planned route. In conclusion, the developed control strategy presents good potential for applications in robotic systems with “kinematically decoupled” state variables.