A Reinforcement Learning Approach to Timing Varying Image Jacobian Matrix

• Main Authors: Chi-yuan Tai, 戴啓原
• Other Authors: Kao-Shing Hwang
• Format: Others
• Language: zh-TW
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/uj9h2d

碩士 === 國立中山大學 === 電機工程學系研究所 === 107 === The behavior of traditional image-based visual servoing method is unsatisfactory when current camera pose is very different from the desired, especially the rotational error along or around the optical axis. During the movement, it might encounter the singular point of image Jacobian matrix, causing the robot arm to lose control, or to leave the feature points out of FOV, resulting servoing failure. Therefore, this thesis propose a reinforcement learning approach to time varying image Jacobian matrix. It is implemented by Q-learning, because Q-learning is easy to realize and model-free. This thesis discretize image plane into state space according to the location of feature points, and the action space is composed of the linear combination of image Jacobian matrix. Then, according to current state, learning agent choose an action by ε-greedy policy. The agent finally get a reward by environment, and use it to update the policy. The agent will learn a policy approximate to the best solution through fully interaction with the environment. In order to verify the method proposed in this paper, a six-axis robot arm and a single-lens camera are used to form a visual servoing system. The method proposed in this thesis is verified by comparing the result with visual servo system using fixed image Jacobian matrix in the Webots simulation software and real world environment, respectively.