Interactive Robotics in Medicine and Care

The goal of B05 is to achieve context-sensitive explainability in interactive task learning - an explainability that should be primarily implicit and does not involve pronounced explanations. The project considers the scenario of an AI system in the form of a humanoid robot equipped with a state-of-the-art learning mechanism. In addition, it learns physical tasks in interaction with human users without in-depth technical knowledge. This scenario is of great importance in various areas, as robots are entering our everyday lives as ubiquitous household helpers, medical assistants or assistants on factory floors. We therefore need to enable them to learn from everyday users.

For such interactive learning to be successful, the human user must be able to form an accurate mental model (i.e. internal representations that people build up about things) of how the system learns. Only in this case the user can understand the robot and provide good training input for learning. At the same time, an explanation of the learning (our explanandum) by the robot (our explainer) for the human user (our explaninee) should not dominate the interaction, but happen alongside it.

B05 goes beyond current human-in-the-loop systems and strives for what we call co-constructive training (CCT). The user's understanding of how the robot learns is co-constructed by combining two mechanisms: (a) monitoring the user's mental model against the training data and (b) adaptively supporting the user's understanding by visualizing internal (architectural) information. The long-term goal of B05 is the algorithmic modeling of the context of a training interaction in a system that is able to tailor co-constructive training to the individual human user. In the first phase, the project will focus on establishing the empirical and conceptual basis through a series of experimental human-robot interaction studies. B05 has two main research areas that will be pursued in collaboration with the Ö project. The first focus examines the interactions that take place during the training of a learning robot with regard to the co-constructive approach. As a result, explainability can be achieved within a co-constructive interaction, and B05 tests whether an accurate mental model that the user has of the AI system's learning process can be created in a CCT. In the second main research focus, for its long-term goal of developing a context-aware system, the project will investigate the social contextualization of this training to understand whether the required technology concepts differ for people with different roles in society. Specifically, the project will first consider factors such as gender and age as well as prior knowledge and try to understand their influence on the formation of users' mental models and their attitudes towards technology in the CCT environment.