Supporting situated spoken Human-Robot Interaction through perceivable context

At a growing pace, the presence of robots in everyday environments is increasing day by day. In fact, their incredibly wide applicability, spanning over various environments and scenarios, is speeding up such spreading. Industrial and working environments, healthcare assistance in public or domestic areas are highly benefiting from robots’ services, that are able to accomplish manifold tasks that have become difficult and annoying for humans. As an example, in domestic environments robots can be deployed in a vast plethora of applications, supporting humans in everyday activities. However, robots are not yet comparable to humans in terms of reasoning and autonomy: a complete knowledge of the environment robots are deployed into is often required to both accomplish the desired task and effectively improve the interaction experience with the user. In this perspective, an active interaction with the end user is still a valuable solution for alleviating this lack of autonomy, as in the so-called Symbiotic Autonomy. My thesis analyzes the impact of such a contextual knowledge in several Human-Robot Interaction (HRI) sub-tasks, with a particular attention on when information, desiderata, and knowledge are shared through natural language. In fact, natural language can be considered one of the most natural way of communicating. To this end, three HRI-specific problems have been considered. First, the importance of the environmental context has been analyzed in a scenario where the robot is not able to achieve its tasks on its own and it needs to ask humans for help. The thesis will address how some perceivable characteristics of the environment might help in designing the robot’s behaviors. The second scenario refers to the ability of re-ranking the transcriptions hypothesized by an Automatic Speech Recognition (ASR) in a situated command interpretation task. This dissertation will show that context, encoded as domain-dependent information, can actually improve the accuracy of a free-form domain-independent ASR. The third scenario relates to the task of interpreting robotic commands, where the robot has to react to commands expressed through natural language. This thesis will provide evidence that a structured representation of the environmental knowledge is beneficial for coherently mapping the sentence to the correct interpretation in a situated scenario. The last scenario investigates to what extent, when acquiring the above mentioned structured representation of the environment through a dialogic guidance, an active perception of the environment improves the dialogic experience, by decreasing the tutoring cost. In this respect, different types of context have been considered and defined for each scenario, ranging from information that is actively perceivable and observable by the robot, to structured knowledge acquired through pre-processing stages.