Robust Data-driven Model Predictive Control via On-policy Reinforcement Learning for Robot Manipulators

摘要

In this paper, a robust data-driven model predictive control (MPC) via on-policy reinforcement learning (RL) is presented for the regulation of constrained robot manipulators subject to both model mismatch and bounded additive disturbances. Based on the Euler-Lagrangian model of the robot manipulator, the model mismatch is characterized by the difference between the system matrix describing the Coriolis and centrifugal torques and its initial approximation. The additive disturbances considered in this work affect the torques applied to the joints. To reduce the complexity of the prediction model adopted by MPC and thus the computational load of solving the online control problem in MPC, the inverse dynamics controller is incorporated into the control framework. The rigid tube-based MPC with adaptive design of the terminal weighting matrix and terminal set is presented to ensure robust constraint satisfaction. The use of the inverse dynamics control policy leads to the selection of the on-policy RL algorithm: Sarsa for designing the policy to update key parameters of the MPC optimization problem. The efficacy of the proposed control framework is validated by a case study using a robot manipulator with two revolute joints in simulation.