PUB - Publikationen an der Universität Bielefeld

Autonomous robots are going to become an increasingly important part of humans' everyday life. Currently, not many tasks could be automated completely, so it is still necessary, and sometimes even preferable, to take the human in the loop. There is a wide range of unsolved challenges in such a Human Robot Cooperation. One of them is the modeling of competence, which concerns the estimation of the capabilities of both interaction partners towards a common goal. From a robot's point of view, this does not only mean awareness of its own but also of human capabilities.

The thesis focuses on competence modeling regarding active and incremental classification. Within a generic scenario where the human uses his world knowledge about the environment for teaching the robot, we have identified three different areas where competence modeling is elementary:

First, the robot should be able to assess its own capabilities. We show that it is possible to estimate classification accuracy in a semi-supervised manner using a static regression model trained on histograms of confidences computed on unlabeled samples in active learning.

Second, the robot should assess its teacher's capabilities because sometimes the human teacher does not have perfect domain knowledge. We propose an additional model that is, by excluding ambiguous areas in feature space, capable of increasing training speed by showing uncertain but valuable samples to the user for labeling.

Finally, bringing the two complementary partners together is the last cornerstone of this thesis. We introduce an active learning querying technique utilizing dimension-reduction approaches for realizing a visualization of deep image feature spaces for the human. Non-expert humans can label robot object recordings faster and more qualitatively using this new teaching interface. That was revealed by a user study we conducted. However, the proposed method adds specific demands to the Machine Learning model for handling non-i.i.d. training data. Therefore, we investigate several approaches enabling instance-based classifier for compensating these conditions.