로봇 매니퓰레이터를 위한 공유 Q-러닝 기반 반복적 제어기 설계 방법
Shared Q-Learning Based Iterative Controller Design for Robot Manipulators
0
조회
0
다운로드
0
인용
0
참고문헌
키워드
초록
In this paper, a shared Q-learning based iterative controller design method is proposed for robot manipulators. In the proposed method, the data collected from the manipulators are used to obtain the optimal path and to gradually improve tracking performance. An optimal task-specific path is obtained through a shared Q-learning scheme. In this scheme, each agent explores the environment independently and periodically exchanges its Q-table with the others, accelerating exploration efficiency and convergence. The optimal path serves as the reference trajectory for a proportional-derivative (PD) tracking controller, while the PD gains are tuned via iterative gain optimization. Specifically, different gain sets are assigned to the manipulators, and the resulting root mean square errors (RMSE) are iteratively compared to find the gains that minimize the RMSE. To verify its effectiveness, a virtual environment is constructed with the kinematic and dynamic models of the 7-degree-of-freedom Panda robot manipulator. The proposed method is then applied to the robot manipulators, demonstrating improved control performance.
Abstract
In this paper, a shared Q-learning based iterative controller design method is proposed for robot manipulators. In the proposed method, the data collected from the manipulators are used to obtain the optimal path and to gradually improve tracking performance. An optimal task-specific path is obtained through a shared Q-learning scheme. In this scheme, each agent explores the environment independently and periodically exchanges its Q-table with the others, accelerating exploration efficiency and convergence. The optimal path serves as the reference trajectory for a proportional-derivative (PD) tracking controller, while the PD gains are tuned via iterative gain optimization. Specifically, different gain sets are assigned to the manipulators, and the resulting root mean square errors (RMSE) are iteratively compared to find the gains that minimize the RMSE. To verify its effectiveness, a virtual environment is constructed with the kinematic and dynamic models of the 7-degree-of-freedom Panda robot manipulator. The proposed method is then applied to the robot manipulators, demonstrating improved control performance.