Learning predictive control based on extended fuzzy state observation for trajectory tracking of an uncertain...

Trajectory tracking is a fundamental aspect of robotics research and it is essential for robots to track tasks effectively. However, manipulators are multi-input, multi-output systems characterized by high nonlinearity and strong coupling, often functioning under uncertain conditions, such as external random disturbances, parameter fluctuations, and unmodeled dynamics. Therefore, this paper proposes a learning-based predictive control method with a fuzzy extended state observer (LPC-FESO), integrating nonlinear predictive control with reinforcement learning to address the challenge of slow reinforcement learning (RL) convergence in stochastic environments and achieve desired trajectory tracking. A nonlinear predictive control, utilizing a fuzzy backstepping approach to generate the initial control sequence, serves as the base controller for Deep Deterministic Policy Gradient (DDPG). This design minimizes dependency on precise system modeling, enhances computational efficiency, and constrains joint angles and velocities via the value function. A fuzzy extended state observer (FESO), balancing both position and velocity states, is also designed to improve the system's disturbance rejection capability, ensuring the required transient and steady-state tracking performance. The theoretical convergence properties of the LPC-FESO framework are provided firstly, considering disturbances and state constraints. The proposed framework targets a class of uncertain multi degree-of-freedom (DOF) manipulators that can be represented by the standard manipulator dynamics with bounded external disturbances and model uncertainties. In this paper, a 2-DOF manipulator is used as an example for demonstration and simulation. Simulation results demonstrate that the proposed approach effectively tracks desired trajectories in terms of both position and velocity, exhibits strong disturbance rejection capabilities, and meets the required performance criteria across various trajectory tracking tasks.