Research On Visual Detection And Batting Decision For 7-DOF Table-tennis Robot

Table tennis requires its participants to be agile in decision and reaction. Based on these features, Table-tennis robot is regarded as an ideal benchmark for high-speed intelligent robot systems. In this dissertation, the kinematic control of redundant manipulators, visual tracking of the high-speed ball and learning-based batting decision are addressed using table tennis robot as the benchmark.A modified analytical inverse kinematic solution based on the arm angle parametrization method is proposed for the 7-DOF anthropomorphic manipulators.The arm angle, which is defined as the dihedral angle between the arm and the reference plane, is adopted to describe the redundancy of the manipulators. Using this parametrization technique, the constraints of the redundant parameter imposed by joint limits can be represented as triangle inequalities that can be solved analytically. On each self-motion manifold, the feasible arm angle interval can be obtained based on simple and easy-to-implement trigonometric computation. Then valid joint configurations can be calculated for every arm angle in the feasible range. The validity of the proposed method is verified theoretically and experimentally in this dissertation. In addition, the forward kinematics is implemented based on MATLAB®symbolic computation. And the weighted least-norm method is chosen as the differential inverse kinematic algorithm for the 7-DOF manipulator after comparing state-of-the-art differential inverse kinematic solutions.To deal with the motion blur of the high-speed ball, a stereo-vision based framework is presented: firstly the corresponding region of the ball is extracted using background subtraction. Then the blur parameters are identified by minimizing the L2 norm of the directional derivatives of the blurred image and the motiondeblurred image is obtained based on Richardson-Lucy deconvolution. Lastly 3D location and velocity of the ball are reconstructed using RANSAC based circle-fitting.This method can mitigate the motion-blur effect greatly so that the visual-tracking precision can be increased notably.Lastly, a method based on support vector regression is proposed to learn the batting policy. Table tennis playing is formalized as the batting evaluation function,which maps the state of the incoming ball and the parameters of the batting trajectory to the reward. This function is then obtained by generalizing the training data using?-support vector regression. During online decision, the optimal batting trajectory is computed by maximizing the batting evaluation function with respect to the parameters of the batting trajectory. As this learning based method is model-free, it makes the success rate of the batting free from unmodeled dynamics and parameter errors.The proposed methods are implemented on a 7-DOF robotic table-tennis platform. And the experiments, namely the motion plan of the 7-DOF manipulator, the visual tracking and the table tennis rallying verify their effectiveness.