Teaching Robots Parametrized Executable Plans Through Spoken Interaction

While operating in domestic environments, robots will necessarily face difficulties not envisioned by their developers at programming time. Moreover, the tasks to be performed by a robot will often have to be specialized and/or adapted to the needs of specific users and specific environments. Hence, learning how to operate by in-teracting with the user seems a key enabling feature to support the introduction of robots in everyday environments. In this paper we contribute a novel approach for learning, through the interaction with the user, task descriptions that are defined as a combination of primitive actions. The proposed approach makes a significant step forward by making task descriptions parametric with respect to domain specific semantic categories. Moreover, by mapping the task representation into a task representation language, we are able to express complex execution paradigms and to revise the learned tasks in a high-level fashion. The approach is evaluated in multiple practical applications with a service robot