The Curious Robot learns grasping in multi-modal interaction
Ingo Lütkebohle, Julia Peltason, Robert Haschke, Britta Wrede, Sven Wachsmuth · PUB – Publications at Bielefeld University (Bielefeld University) · 2010
Practical robots have come a long way, from being confined to strong cages in industrial manufacturing halls to open environments shared with humans. One consequence if robots are to share spaces with humans is that they must be able to learn from them – that much is well accepted. The reverse – that the robot becomes the “teacher” and the human the “student” – is less commonly seen, however. This is despite the fact that many applications tacitly assume that humans learns about the robot, e.g. from a manual or through instruction by an expert. We surmise that there is a great deal of potential in an explicit reversal of the roles. Therefore, we have investigated how this reversal of the traditional roles can improve HRI. Concretely: How could a robot structure the dialog such that a naive human partner is aware of her/his possible dialog actions? The goal is to make humans able to act with confidence despite having absolutely no prior knowledge of either the robot’s goals or its capabilities. To achieve this ambitious goal, several hard problems must be adressed. One important issue is the vocabulary problem [1], that describes the fact that humans do not know what the system understands, in particular at the beginning of an interaction [2]. Another well known problem is that user’s expectations about a system are strongly shaped by appearance [3], [4], which may lead to erroneous assumptions [5]. Last, but not least, it is not clear how to provide guidance in an easy to understand way and this requires an iterative, study-based approach towards system development [6]. To investigate how robot guidance can improve upon this, we have introduced the “Curious Robot” interactive scenario for learning about real-world objects [5]. In it, we have used a mixed-initiative [7] approach, that has the robot query the human for information at appropriate points during the interaction. For example, the robot queries a human about object labels and how to grasp an object. Initiative is guided by visual saliency information [8]. In this scenario, our results indicate that closed questions provide excellent guidance to the human, resulting in con-