Kinesthetic teaching and predictive control of interaction tasks in robotics

作为工业制造任务的一部分，精确的交互通常非常复杂，难以表征和实现。其中一个原因是运动和控制行为的任务特定要求的异质性。因此，将任务直接执行到机器人程序中需要高素质的专家，并且仅在大批量时才有利可图。为了使机器人系统具有灵活的适用性和易于重新配置，本项目开发了一种通过动觉演示进行编程的方法。机器人由用户引导完成整个操作任务，同时记录机器人运动以及相互作用力。通常，为了补偿人类演示的次优性和不精确性，需要多次重复演示。 这对于复杂的运动序列或交互情况尤其重要，例如周期性运动或组件组装，这些运动序列或交互情况难以演示，但同时对于成功执行任务至关重要。该项目的基础是以前开发的模型预测交互控制（MPIC）的一般框架。操纵任务被分成一系列的基本任务，所谓的操纵原语（MP）与个人的运动和控制特性，这是在一个整体的方式处理的模型预测控制方法。MPIC方法在该项目中阐述了关于操纵任务的动觉演示，例如，通过考虑MPC预测范围内MP之间的切换。进一步的重点在于自动生成的MP序列的重复演示的操作任务，而不需要额外的专业知识。在演示的基础上，通过学习MP的设定点和过渡条件，并最终优化整体操作任务，将迭代地细化操作任务。

Precise interactions as part of industrial manufacturing tasks are typically very complex to characterize and implement. One reason for this is the heterogeneity of the task-specific requirements for the motion and control behavior. A direct implementation of the task into a robot program therefore requires highly qualified specialists and is only profitable for large lot sizes. For a flexible applicability and easy (re-)configuration of the robot system, an approach to programming by kinesthetic demonstration is developed in this project. The robot is guided by the user through the entire manipulation task, while the robot motion as well as the interaction forces are simultaneously recorded. Typically, several repetitions of the demonstration are necessary in order to compensate for the suboptimality and imprecision of the human demonstration. This is particularly important for complex motion sequences or interaction situations, such as periodic movements or the assembly of components, that are difficult to demonstrate but at the same time are crucial for a successful task execution. The basis for this project is a previously developed general framework for model predictive interaction control (MPIC). The manipulation task is split into a sequence of elementary tasks, so-called manipulation primitives (MPs) with individual motion and control characteristics, which are treated in a holistic manner by a model predictive control approach. The MPIC approach is elaborated in this project regarding the kinesthetic demonstration of manipulation tasks, e.g. by considering the switching between MPs over the prediction horizon of the MPC. A further focus lies on the automatic generation of the MP sequence from the repeated demonstration of the manipulation task without requiring additional expert knowledge. Based on the demonstration, the manipulation task will be iteratively refined by learning the setpoints and the transition conditions of the MPs and finally by optimizing the overall manipulation task.