| Abstract | Multi-joint manipulators hold significant potential across various applications; however, achieving optimized performance while ensuring constraint satisfaction remains challenging. To address this, a reinforcement learning-based model predictive control (RLMPC) framework is employed to optimize the manipulator's motion while simultaneously tuning the terminal weighting. To reduce the computational burden and meet real-time requirements, the terminal constraint is removed from the optimization problem. However, the absence of a terminal constraint in conventional RLMPC frameworks necessitates a sufficiently large prediction horizon for convergence, since a longer horizon helps approximate the long-term cost and guides the system toward stability. Meanwhile, efficiently obtaining feasible samples in the state space remains challenging for manipulators. To overcome these limitations, a stability governor is introduced to generate a reference target at each time step, which enhances sampling efficiency and guides the RLMPC optimization toward a feasible solution that balances path efficiency and control performance. The proposed framework is validated through comparison simulations using a numerical model of the UR10e robot manipulator, demonstrating improved tracking performance, reduced computational complexity, and enhanced constraint satisfaction, showing its potential for real-world applications. |
|---|
