engleski

Learning control for positionally controlled manipulators

The majority of the widely available robotic arms employ the joint position control paradigm. Additionally, these kind of arms are usually closed architecture, meaning that the user has little knowledge or control over the inner workings of the controller. The user can only specify a position trajectory that the arm needs to follow. In the unstructured environment this can be a serious drawback. By exploiting the knowledge of the system, the performance of the closed architecture robotic arm system can be improved. Recently, nonparametric regression methods have been shown to improve performance of torque controlled arms. In this paper, we investigate the effectiveness of those methods in the case of closed architecture robotics arms. We apply Gaussian Process Regression (GPR) to learn the dynamic model between the input and the output signal, including the dynamics of the robot plant and controller. We also consider a sparse variant of GPR, called Sparse Spectrum Gaussian Process Regression, which enables faster training and prediction times. It is demonstrated by simulation that the proposed approach significantly enhances the trajectory following performance of closed architecture robotic arms.

Robot Manipulators; Learning Control; Gaussian Process Regression

nije evidentirano