基于改进动态系统稳定估计器的机器人技能学习方法

提出一种基于改进动态系统稳定估计器的机器人技能学习方法. 现有的动态系统稳定估计器方法可以通过非线性优化来确保学习系统的全局稳定性, 但是存在确定高斯混合分量个数困难以及稳定性和精度无法兼顾的问题. 因此, 根据贝叶斯非参数模型可以自动确定合适分量个数的特性, 采用狄利克雷过程高斯混合模型对演示进行初始拟合. 随后利用参数化二次李雅普诺夫函数重新推导新的稳定性约束, 有效地解决了动态系统稳定估计器方法中稳定性和精度难以兼顾的问题. 最后, 在LASA数据库和Franka-panda机器人上的实验验证了新方法的有效性和优越性.

A Robot Skill Learning Method Based on Improved Stable Estimator of Dynamical Systems

This paper presents a novel robot skill learning method based on improved stable estimator of dynamical systems (SEDS). The original SEDS method can ensure the global stability of the learning system through nonlinear optimization. However, it cannot automatically determine the optimal number of Gaussian components and is difficult to make a trade-off between reconcile the stability and accuracy. Therefore, note that the Bayesian non-parametric model can be used to determine the appropriate number of components, the Dirichlet process Gaussian mixture model is applied to perform the initial fitting of the demonstrations in this paper. Then, the stability constraints are reformulated by using the parameterized Lyapunov function. The problems of stability and accuracy in the SEDS method are solved effectively. Finally, experiments on a LASA dataset and a Franka-panda cooperative robot validate the effectiveness and superiority of the proposed method.