Negotiating Spatial Goals with a Wheelchair

We present our iterative approach to enabling natural dialogic interaction between human users and a wheelchair, based on the alternation of empirical studies and dialogue modelling. Our approach incorporates empirically identified conceptual problem areas and a dialogue model designed to manage the available information and to ask clarification questions. In a Wizard-of-Oz experiment employing the first version of the model, we test how verbal robotic reactions can enable users to provide the information needed by the wheelchair to carry out the spatial task. Results show that the output must be extraordinarily coherent, temporally well-placed, and aligned with the user's descriptions, as even slightly deviating reactions systematically lead to confusion. The dialogue model is improved accordingly.