论文部分内容阅读
针对7自由度仿人型乒乓球机器人的定点回球问题,提出了一种基于支持向量回归的击球策略学习方法.首先,把机器人的击球过程形式化为击球评价函数,该函数以来球状态和击球轨迹参数为输入,以回报值为输出.然后,提出一种基于物理模型置信域的随机搜索方法以提高训练数据的采集效率,并基于ε支持向量回归(ε-SVR)对经验数据集进行泛化从而得到击球评价函数.最后,在决策过程中,采用多初值拟牛顿法最大化击球评价函数以求解出最优击球轨迹.将该方法应用于7自由度乒乓球机器人系统中,实验结果验证了其有效性.
Aiming at the problem of fixed-point return ball of a 7-DOF humanoid pingpong robot, a batting strategy learning method based on support vector regression is proposed. Firstly, the robot’s batting process is formalized as batting evaluation function Ball state and ball trajectory parameters as input, the output value is output.Then, a random search method based on confidence region of physical model is proposed to improve the collection efficiency of training data, and based on ε-Support Vector Regression (ε-SVR) Experience data set to obtain the batting evaluation function.Finally, in the decision-making process, the multi-initial quasi-Newton method is used to maximize the batting evaluation function to find out the optimal batting trajectory.This method is applied to 7 degrees of freedom Table tennis robot system, the experimental results verify its effectiveness.