Learning the Context in Programming by Demonstration of Manipulation Tasks

最先进的服务机器人能够基于人类教师的观察，学习由多个定义的动作组成的新操作任务，称为演示编程(PBD)。以分支或可选动作为特征的任务的逻辑结构可以使用符号PBD方法来学习。子符号PBD-方法允许学习原子动作，该动作生成可执行的机器人运动。但是，不会生成关于可以在其中执行原子操作的上下文的信息。为了自主地执行操作任务，需要对上下文进行操作描述。需要解决的问题是，什么是相关的、可测量的对象属性，以及机器人如何能够高可靠性地自主观察它们。由于人类环境中对象的多样性导致大量的歧义和误报，仅凭视觉感知是不够的。解决这些问题的基础是选择合适的对象分类器，减少待分类对象的数量，以及搜索区域的定义。此外，必须定义传感器动作来确定无法用视觉测量的对象属性。在大多数最先进的系统中，这种知识是人工定义的。因此，本项目的目标是扩展PBD方法，以基于人类观察来学习操纵任务的上下文。为了实现这一目标，我们将把场景分析的方法整合和扩展到PBD方法中。第二个目标是使用学习的上下文来解决视觉感知算法的歧义和检测误报。我们将把交互式对象检测的方法集成并扩展到PBD方法中，以有效地学习传感器动作来测量非视觉对象属性，例如权重，从而解决歧义和错误检测。基于所学习的上下文，我们可以例如基于与已知对象的空间关系来推断未知对象在环境中的角色。用一个未知的对象执行学习的操作任务，如果没有泛化过程是不可能的。因此，通过交互地将学习到的操作任务的约束和目标适应于新对象来增加机器人的泛化能力。变形方法的应用是为了在三维物体模型的基础上变换约束和目标。在仿真技术的帮助下，交互地进行了变换和调整的验证。所开发的算法将在真实的机器人拟人系统上实现，并使用真实世界的例子进行评估。

State-of-the-art service robots are able to learn new manipulation tasks consisting of multiple, defined actions, based on the observation of a human teacher, known as Programming by Demonstration (PbD). The logic structure of the task characterized by branches or alternative actions, can be learned using symbolic PbD-approaches. Subsymbolic PbD-approaches allow to learn atomic actions, which generate executable robot motions.. However, no information about the context, in which the atomic action can be executed, is generated. In order to execute a manipulation task autonomously, an operational description of the context is necessary. The problem to solve is, what are relevant, measurable object properties and how can they be autonomously observed with high reliability by the robot.Visual perception alone is insufficient since the variety of objects in the human environment leads to a large number of ambiguities and false positives. The basis to resolve these issues is a selection of suitable object classifiers, a reduced number of objects to classify and the definition of search regions. Additionally, sensor actions have to be defined to determine object properties, which can't be measured visually. In most state-of-the-art systems, this knowledge is defined manually.Thus, the goal of this project is to extend the PbD-approach to learn the context of a manipulation task based on human observation. In order to achieve this goal, we will integrate and extend methods from scene analysis into the PbD-approach. The second goal is to resolve ambiguities and detect false positives of visual perception algorithms using the learned contexts. We will integrate and extend methods from interactive object detection into the PbD-approach to learn sensor actions efficiently to measure non-visual object properties, e.g. weight, and thereby resolve ambiguities and false detections. Based on the learned context, we can infer the role of unknown objects in the environment, e.g. based on the spatial relation to known objects. The execution of learned manipulation tasks with an object, which was unknown , is without generalization processes not possible. Thus, increasing the generalization capabilities of the robot by interactively adapting learned constraints and goals of a manipulation task to a novel object. The application of morphing methods is planned to transform constraints and goals on the basis of 3D-object-models. With the help of simulation techniques the verification of the transformation and its adjustment is done interactively. The developed algorithms will be implemented on real robot anthropomorphic systems and evaluated using real world examples.