摘要
In this chapter we have discussed various imitation learning techniques based on kinesthetic teaching where the expert (teacher) provides demonstrations of certain task while holding the robotic manipulator and guiding it through the task trajectory. The dynamic movement primitives (DMPs) presented here, has been shown to learn task in joint space and the other methods which use Gaussian mixture model, learn the task in the Cartesian space. The task trajectory is asymptotically stable in DMP models as it use an inbuilt PD controller. Parameters of DMP based model are learnt from single demonstration in an unconstrained optimization process which make them computationally efficient. SED motion encoders which are great at generating asymptotically stable trajectories have been presented in this chapter.The parameters of the SED models are learnt in a constrained optimization process. The motion learning architecture C-FuzzStaMP is also presented in this chapter. The multitask learning framework exploits additional information to enhance multitasking capability of the proposed technique. Three algorithms are presented to describe learning using different regression techniques namely GMR, LWPR and SVR.
摘要译文
在本章中，我们讨论了基于动觉教学的各种模仿学习技术，其中专家（老师）在握住机器人操纵器并指导其通过任务轨迹的同时提供特定任务的演示。此处展示的动态运动原语（DMP）已被证明可以学习关节空间中的任务，而其他使用高斯混合模型的方法则可以学习笛卡尔空间中的任务。由于使用内置的PD控制器，因此在DMP模型中任务轨迹是渐近稳定的。基于DMP的模型的参数是通过无限制的优化过程中的单个演示来学习的，从而使它们的计算效率更高。本章介绍了非常适合生成渐近稳定轨迹的SED运动编码器。在受限的优化过程中学习了SED模型的参数。本章还介绍了运动学习体系结构C-FuzzStaMP。多任务学习框架利用附加信息来增强所提出技术的多任务能力。提出了三种算法来描述使用不同回归技术的学习，即GMR，LWPR和SVR。
Laxmidhar Behera, Swagat Kumar, Prem Kumar Patchaikani, Ranjith Ravindranathan Nair, Samrat Dutta. Imitation Learning. Intelligent Control of Robotic Systems[M].UK: Taylorfrancis, 2020: 319-384