Both stationary 'industrial' and autonomous mobile robots nowadays pervade many workplaces, but human-friendly interaction with them is still very much an experimental subject. One of the reasons for this is that computer and robotic systems are very bad at performing certain tasks well and robust. A prime example is classification of sensor readings: Which part of a 3D depth image is the cup, which the saucer, which the table? These are tasks that humans excel at. To alleviate this problem, we propose a team approach, wherein the robot records sensor data and uses an Augmented-Reality (AR) system to present the data to the user directly in the 3D environment. The user can then perform classification decisions directly on the data by pointing, gestures and speech commands. After the classification has been performed by the user, the robot takes the classified data and matches it to its environment model. As a demonstration of this approach, we present an initial system for creating objects on-the-fly in the environment model. A rotating laser scanner is used to capture a 3D snapshot of the environment. This snapshot is presented to the user as an overlay over his view of the scene. The user classifies unknown objects by pointing at them. The system segments the snapshot according to the user's indications and presents the results of segmentation back to the user, who can then inspect, correct and enhance them interactively. After a satisfying result has been reached, the laserscanner can take more snapshots from other angles and use the previous segmentation hints to construct a 3D model of the object.
展开▼
