Discovering task subsequences from a continuous video stream facilitates a robot imitation of sequential tasks. In this research, we develop unsupervised learning of the task subsequences which does not require a human teacher to give the supervised label of the subsequence. Task-discriminative feature, in the form of sparsely activated cells called task capsules, is proposed for self-training to preserve spatio-semantic information of a visual input. The task capsules are sparsely and exclusively activated with respect to the spatio-semantic context of the task subsequence: a type and location of the object. Therefore, the generalized purpose in multiple videos is unsupervisedly discovered according to the spatio-semantic context, and the demonstration is segmented into the task subsequences in an object-centric way. In comparison with the existing studies on unsupervised task segmentation, our work has the following distinct contribution: 1) the task provided as a video stream can be segmented without any pre-defined knowledge, 2) the trained features preserve spatio-semantic information so that the segmentation is object-centric. Our experiment shows that the recognition of the task subsequence can be applied to robot imitation for a sequential pick-and-place task by providing the semantic and location information of the object to be manipulated.

从连续视频流中发现任务子序列有助于机器人模仿序列任务。在这项研究中，我们开发了任务子序列的无监督学习，它不需要人类教师给出子序列的监督标签。任务区分特征，以称为任务胶囊的稀疏激活细胞的形式，被提议用于自我训练，以保留视觉输入的空间语义信息。相对于任务子序列的空间语义上下文：对象的类型和位置，任务胶囊被稀疏地和专门地激活。因此，根据空间语义上下文无监督地发现多个视频中的广义目的，并以对象为中心的方式将演示分割成任务子序列。与现有的无监督任务分割研究相比，我们的工作有以下明显贡献：1）作为视频流提供的任务可以在没有任何预定义知识的情况下进行分割，2）训练的特征保留空间语义信息，以便分割是以对象为中心的。我们的实验表明，通过提供要操纵的对象的语义和位置信息，任务子序列的识别可以应用于机器人模仿以进行顺序拾放任务。