On - line Inference of O - line Learned

Learning in robotics has received more and more attention in recent years. It eases bridging the gap between low-level sensor data and high-level concepts. A high-level representation language is necessary in order to support the communication between robot and user in both directions when the robot navigates in unknown environments. Controlling robots in terms of a high-level representation formalism like rst-order logic is often said to be too slow. In this paper, we present a performance system capable of inferring previously learned high-level concepts from the sensory input of a mobile robot on-line in real-time. Furthermore, we describe an inference engine tailor-made to the requirements of our representation. 1 Learning in robotics In recent years, machine learning becomes more and more an important topic in robotics. Learning eases adaptation of robots to diierent tasks, to diierent environments, and diierent robot settings. Learning can be applied in very diierent ways. Biologically inspired work in robotics has developed artiicial beings which adapt to their environment (Brooks, 1991; Steels, 1993). This type of learning is restricted to reeex-like behavior. Learning skills by human demonstrations (Dillmann et al., 1995) eases programming of the robot's basic behavior, but does not enable the robot to report on its actions. In order to enhance a robot's high-level processing, planning can be tailored to particular robot tasks by learning techniques (Segre, 1988; Bennett, 1989). Learning can also be used to link the environment model with perceptions when executing a plan (DeJong and Bennett, 1993). These approaches have in common, that learning takes place at one horizontal level of control. Our approach, in contrast, concerns vertical learning to close the gap between numerical data and high-level concepts. The concepts, learned at higher levels of the hierarchy will be used as items for communication between user and robot in both directions, from the user to the robot to control the system, and from the robot to the user to report the performed actions. The report will explain whether the task was performed correctly, or which circumstances, e.g. a wrong representation of some concept, have forced an error. The learning tasks and the symbol grounding aspects are described in more detail in (Klingspor and Morik, 1995). In contrast to many robot learning approaches, learning phase and performance phase are two separate steps. During the learning phase, we show the robot in a known environment, how objects look like. From this information, …