Route instruction dialogues with a robotic wheelchair

Our aim is to enable natural and intuitive spoken dialogue between users and the Bremen robotic wheelchair "Rolland" (Röfer & Lankenau 1998), in order to instruct the robot to move to other places autonomously. The purpose of our interdisciplinary work is to develop specific functionalities matching potential users’ intuitive expectations. Our current focus is on a detailed qualitative analysis of the discourse flow between human and robot, using a realistic interaction scenario with uninformed users that is tailored to the actual technological requirements. This approach is useful to establish and improve the relationship between implemented functionalities and humans’ intuitive reactions at being confronted with an autonomous transportation device. In earlier work (Shi & Tenbrink 2005) we identified a range of potential problems and devised a dialogue model to address them. Our dialogue modelling approach (see Ross et al. 2005) is based primarily on the COnversational Roles model (Sitter & Stein 1992) combined with the information state based approach (Traum & Larsson 2003). Figure 1 shows a depiction of a clarification subdialogue initiated by the robot, a part of the dialogue model that we start from in the present study. Following an utterance by the user, the robot can request something or inform the user; or it can make a suggestion, which can be rejected or accepted by the user. Following such a rejection or a request by the robot, the user instructs the robot, reaching the final state of this specific subdialogue. Based on this model, we carried out a second study, this time in a Wizard-of-Oz scenario, to test the communicative success of a number of systematic robot reactions. Here we present the results of this study, outlining the range of problems that could successfully be handled by the robotic reactions, and pointing to a number of novel problems that arise precisely because of the robotic output. Our results show that highlevel or generic robotic reactions will lead to increased confusion, while specific and aligned clarification questions enable smooth and efficient dialogue between humans and robots. This leads to an improved dialogue model. Figure 1: Clarification subdialogue