Robots, compute servers, and workstations for semantic robot vision

视觉引导移动机器人技术和视觉物体识别技术的最新进展有望使我们能够制造出能够与人合作的家庭机器人助手，了解家庭、家中的物品及其用途。在UBC，我们开发了一个移动机器人，好奇的乔治，它可以搜索一个小房间，并从目标列表中拍摄物体。机器人通过搜索网络，收集显示物体的图像来学习如何识别物体，然后学习如何在房间里找到物体。然后，机器人绘制房间地图，并找到一条路线，让它搜索物体，然后返回一组显示它找到的物体的图像。好奇的乔治赢得了最近举行的语义机器人视觉挑战赛。更一般地说，这样的机器人可以学习物体类别的外观，然后识别它们，有时会使用它们出现的上下文。但这些机器人需要梳理图像，找到物体的独特特征，并从图像中构建复杂的分类器。机器人本身需要灵活和紧凑。我们现在正在申请新的机器人和计算服务器，这将使我们能够在现实环境中进行实验，比如在家里。服务器需要大量的内存，因为我们必须理解的特征学习数据包括成千上万的图像，例如，机器人在家里的经验记录。测绘还扩展了计算系统的能力，因为机器人在穿越房屋时获得了数以万计的地标。最后，为了在视觉搜索中利用上下文，我们需要学习许多不同对象及其属性之间的空间关系。

Recent advances in visually guided mobile robotics and visual object recognition  promise to enable us to build home robot assistants that can work with people, knowing about the home, the objects in the home, and their use.  At UBC we have developed a mobile robot, Curious George, that can search a small room and photograph objects from a target list.  The robot learns how to identify objects by searching the web, gathering images showing the objects, and then learns how to find the objects in the room.  The robot then maps the room and finds a route that lets it search for the objects, and returns with a set of images showing the objects it has found. Curious George won the recently held Semantic Robot Vision Challenge.More generally, such robots can learn the appearance of classes of objects and then recognize them, sometimes using the context in which they appear.  But these robots need to comb through the images, finding distinctive features of the objects, and build complex classifiers from the images.  The robots themselves need to be agile and compact.We are now applying for new robots and computation servers that will enable us to experiment in realistic environments such as a home.  The servers need substantial memory since the data we must understand for feature learning comprises thousands of images, for example, the record of a robot's experience in the home.  Mapping also stretches the capabilities of computing systems since the robot acquires tens of thousands of landmarks while traversing a home.  Finally, to utilize context in visual search we need to learn the spatial relations between many different objects and their properties.