Interactive Teaching of Task-Oriented Visual Recognition

我们一直在研究一种个人服务机器人的互动教学，它可以在平常的家庭或办公室环境中帮助我们。本文的研究结果如下：1。任务模型的表示。任务模型描述了操作的步骤，每个操作需要哪些信息，以及每个信息的教学方法。当机器人执行一项操作时，它会检查任务模型并检查它是否具有该操作所需的信息。如果没有，机器人会要求人类操作员来教它。这样，机器人就主导了机器人与人之间的互动。对象的表示。我们研究了家庭或办公室中许多物体需要教给机器人什么样的信息，并将这些信息分为以下几类：形状、视图和机构作为不变部分，机构的姿态和状态作为可变部分。我们使用GUI开发具体的对象模型和教学方法几个对象,如上班 ... 更多的ators.3。对象信息教学。为了帮助机器人识别物体，我们开发了一种方便地教授场景中信息位置的方法。我们用一个基于led的设备教机器人。从这些位置信息以及任务模型和对象模型中，机器人决定关注哪个位置以及使用哪些图像特征进行识别。互动视觉。物体识别有时会因为光照条件的变化而失败。我们在分析当前识别结果和估计可能情况的基础上，提出了一种查询完成目标识别所需信息的方法。机器人要求提供对辨别这种情况有用的信息。实验系统的开发。我们开发了一种个人服务机器人，它有一个六自由度的机械手，一只手，三个摄像头，一个激光测距仪。我们对两项任务进行了实验：从冰箱里拿罐头和乘电梯。在实验中，我们成功地教会了机器人几种用于视觉识别的信息。教授信息的例子有：冰箱的形状和大小，门上把手的位置，电梯的门和按钮，到电梯大厅的路线。少

We have been investigating an interactive teaching of personal service robots, which helps us in usual home or office enviromnents. The results of this research are as follows :1.Representation for task models. A task model describes steps of operations, what information is necessary for each operation, and a teaching method for each piece of information. When a robot executes an operation, it examines the task model and check if it has necessary information for the operation. If not, the robot asks a human operator to teach it. In this way, the robot leads the interaction between the robot and the human.2.Representation for objects. We examined what kinds of information to be taught to the robot for many objects in home or offices, and classify the information into the following : shape, view, and mechanism as an invariable part, and pose and state of the mechanism as a variable part. We develop concrete object models and teaching methods using GUI for several objects such as refriger … More ators.3.Teaching of object information. We develop a method of easily teaching informative positions in the scene for assisting object recognition of the robot. We teach the robot them using an LED-based device. From those position information as well as a task model and object models, the robot determines which position to focus on and which image features to use for recognition.4.Interactive vision. Object recognition sometimes fails due to, for example, changes of illumination condition. We develop a method of asking necessary information for completing object recognition based on the analysis of the current recognition result and the estimation of possible situations. The robot asks for the information which is useful for discriminating such situations.5.Development of experimental system. We develop a personal service robot which has a 6-DOF manipulator, a hand, three cameras, a laser range finder.6.Experiments. We performed experiments for two tasks : fetching a can from a refrigerator and taking an elevator. In the experiments, we successfully taught several kinds of information, used for visual recognition, to the robot. Examples of taught information are : shape and size of refrigerators, position of the handles on their doors, doors and buttons of an elevator, the route to the elevator hall. Less